claimpit
← All receipts
Fact-Checker/Videos
Video fact-check

The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron

250 claims checked · Published August 2026 · Checked August 2026

Checked by an AI model against live web sources — how this works · report an error

Watch on YouTube Jump to claims ↓
The map

Who said what, and about what

A bounded sample, most disputed claims first — the full list is belowEvery row is reachable by scrolling its column · search matches claim text, verdicts, people, topics and sources · hover a row to trace its chain, click one to pin its detail

The receipts

Every claim, checked

  1. Claim 1
    Attributed toEd ZitronAuto-attributedUnverifiable97% confidence▶ 0:32

    This is the largest non-consensual push of technology in history.

    The comparative term “largest” could in principle be tested, but “non-consensual push of technology” has no agreed operational definition. Different readings of consent, technological adoption, and historical scope would produce different answers.

    Sources: none found for this claim.

  2. Claim 2
    Attributed toEd ZitronAuto-attributedAccurate96% confidence▶ 0:46
    “Open AI lost $20.9 billion last year.”

    OpenAI lost approximately $20.9 billion in the previous year.

    Reports based on audited 2025 financial documents describe OpenAI’s operating loss as $20.92 billion. The company’s reported net loss was higher because of a large restructuring-related charge, but the spoken figure accurately matches the operating loss.

    Sources

  3. Claim 3
    Attributed toEd ZitronAuto-attributedAccurate90% confidence▶ 0:56
    “there's no economic data to support it.”

    There is no economic data supporting the claim that AI will replace all human jobs.

    Available economic research does not support the absolute claim that AI will replace all human jobs. Current evidence instead describes limited, uneven, and occupation-specific effects, including some displacement risks but no broad economy-wide replacement of human employment.

    Sources

  4. Claim 4
    Attributed toSteven BartlettAuto-attributedAccurate99% confidence▶ 1:12
    “Mark Zuckerberg says, "We'll continue to invest aggressively in infrastructure to meet the demand."”

    Mark Zuckerberg said that Meta would continue investing aggressively in infrastructure to meet AI demand.

    A transcript of Meta’s Q2 2026 earnings call records Zuckerberg saying that as AI usage ramps, Meta would continue to invest aggressively in infrastructure to meet demand. The wording in the video is a shortened but substantively accurate attribution.

    Sources

  5. Claim 5
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 3:57
    “And it can't even do search.”

    The speaker says a current AI system cannot perform search.

    The claim does not identify the AI system, the type of search meant, or the standard by which success is judged. Because the referent and predicate are undefined, the statement cannot be reliably confirmed or refuted.

    Sources: none found for this claim.

  6. Claim 6
    Attributed toEd ZitronAuto-attributedFalse50% confidence▶ 4:37
    “Up until fairly recently, none of their revenues were coming from AI.”

    The speaker says that, until fairly recently, none of the named companies had revenue coming from AI.

    The universal claim is refuted by Amazon’s own statement that AWS already had a multibillion-dollar AI revenue run rate. Microsoft also reported AI-inclusive Azure consumption growth during fiscal 2024, further contradicting the assertion.

    Sources

  7. Claim 7
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 4:44
    “Right now, 70% of all AI revenues across those three companies are from OpenAI and Anthropic”

    Seventy percent of all AI revenue across the referenced companies comes from OpenAI and Anthropic.

    The claim is not well-posed because “those three companies” has no clear antecedent—the preceding list names six companies—and “AI revenues” has no defined accounting scope. Available sources report different kinds of company revenue and run-rate estimates, but do not establish this precise 70-percent calculation.

    Checked twice, independently: the first pass returned Unverifiable and the second Accurate. Recorded as Unverifiable.

    Sources

  8. Claim 8
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 4:47
    “OpenAI and Anthropic to unprofitable”

    The speaker says OpenAI and Anthropic are unprofitable.

    Contemporary reporting described both privately held companies as losing more money than they made. The wording contains a grammatical error, but its substantive assertion that both were unprofitable is supported.

    Sources

  9. Claim 9
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 4:49
    “unsustainable companies that literally cannot afford to exist without these very same companies giving them money.”

    The speaker says OpenAI and Anthropic cannot afford to exist without funding from the same large technology companies.

    The companies were reported to be unprofitable and dependent on substantial outside financing, but “cannot afford to exist” is an undefined counterfactual, and the claim does not establish which companies’ funding would be indispensable or whether other financing could replace it.

    Checked twice, independently: the first pass returned Unverifiable and the second Accurate. Recorded as Unverifiable.

    Sources

  10. Claim 10
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 4:55
    “Amazon sent $50 billion to OpenAI this year.”

    Amazon sent $50 billion to OpenAI during 2026.

    Amazon announced a $50 billion OpenAI investment in February 2026, initially structured as $15 billion followed by $35 billion subject to conditions. By late August 2026, reporting said Amazon had completed the additional $35 billion, making the full-year statement accurate in the current record.

    Checked twice, independently: the first pass returned Accurate and the second False. Recorded as Unverifiable.

    Sources

  11. Claim 11
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 5:01
    “They sent $5 billion to Anthropic.”

    Amazon sent $5 billion to Anthropic during 2026.

    Amazon announced that it would invest $5 billion in Anthropic, but the public announcement does not establish that the full amount had already been sent. “Sent” therefore overstates what the cited announcement confirms.

    Omits: The statement omits that Amazon announced an investment of $5 billion, with the source describing it as an investment commitment rather than establishing that the full amount had already been transferred.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  12. Claim 12
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 5:03
    “Google sent $10 billion to Anthropic.”

    Google sent $10 billion to Anthropic during 2026.

    Reports described Google as committing to invest $10 billion in Anthropic. The wording “sent” implies completed payment, which the cited reporting does not establish.

    Omits: The statement omits that the publicly reported figure was a commitment to invest $10 billion, not confirmation that the entire amount had already been transferred.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  13. Claim 13
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 5:28
    “they don't even disclose their AI revenues.”

    The speaker says the companies do not disclose their AI revenue.

    The statement uses the undefined category “AI revenues” and an unspecified group of companies. Some companies report AI-related revenue or revenue run rates, while others combine AI with broader cloud or data-center categories, so the universal claim cannot be judged without a precise accounting definition and company set.

    Checked twice, independently: the first pass returned Unverifiable and the second False. Recorded as Unverifiable.

    Sources

  14. Claim 14
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 6:06
    “These are public companies, or at least the ones that aren't anthropic and open AI.”

    Anthropic and OpenAI are not public companies.

    OpenAI and Anthropic are privately owned companies rather than publicly traded corporations. The broader wording about all other companies is less precise, but its specific qualification about Anthropic and OpenAI is correct.

    Sources

  15. Claim 15
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 6:22
    “they have spent over a trillion dollars in capex.”

    The companies under discussion had spent more than one trillion dollars in capital expenditure.

    Public reporting has described roughly $1 trillion in AI-related infrastructure commitments, but those are not equivalent to more than $1 trillion already spent. The claim presents commitments or obligations as completed expenditure.

    Omits: The trillion-dollar figure commonly cited in this context includes future commitments and off-balance-sheet obligations, not simply money already spent on AI capital expenditure.

    Sources

    • 1Why Big Tech's AI spending is bigger than you think

      RefutesThat's roughly how much money seven Big Tech companies, including Google, Microsoft and Nvidia, have committed to spending on AI-related infrastructure in off-balance-sheet commitments, according to a new analysis from Morgan Stanley.

  16. Claim 16
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 6:47
    “They take a bunch of high bandwidth memory”

    AI GPUs use high-bandwidth memory.

    Modern AI accelerators such as NVIDIA's GB200 NVL72 use HBM3E memory, including 13.4 TB of HBM3E across the rack-scale system.

    Sources

  17. Claim 17
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 6:57
    “You need thousands of them, tens of thousands, hundreds of thousands in some case.”

    Frontier-scale AI data centers can require thousands or hundreds of thousands of GPUs.

    The claim is supported for the type of large-scale facilities under discussion: the Abilene Stargate campus has been reported as planned to house more than 450,000 NVIDIA GB200 GPUs. Smaller AI systems do not require that scale.

    The phrase “you need” is judged in the context of the large, frontier-scale AI data centers being discussed, not as a requirement for every AI deployment.

    Sources

  18. Claim 18
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 7:03
    “OpenAI and Oracle are building a data center in Texas in Abalene, Texas. 1.2 GW called Stargate Abene.”

    OpenAI and Oracle are building a 1.2-gigawatt Stargate data center in Abilene, Texas.

    The Abilene Stargate campus is an OpenAI-and-Oracle project with a stated total power capacity of approximately 1.2 GW.

    Sources

  19. Claim 19
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 7:10
    “with each one of the eight buildings, there'll be 50,000 Nvidia GB200 GPUs.”

    The Abilene Stargate campus has eight buildings, each intended to support about 50,000 NVIDIA GB200 GPUs.

    The eight-building and roughly 50,000-GPU-per-building figures reflect reported design specifications, but the transcript states them as a definite future deployment. Public reporting instead describes each building as engineered for up to 50,000 GPUs and separately reports a total exceeding 450,000 GPUs.

    Omits: The 50,000 figure is described publicly as an upper design capacity per building, not a confirmed deployment, while other reporting has described the campus as eventually housing more than 450,000 GPUs in total.

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

  20. Claim 20
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 7:15
    “So, city of Bristol takes about 7800 megawatt of power a year, right?”

    Bristol uses about 7,800 megawatts of power per year.

    The statement uses the wrong unit for annual consumption and gives a substantially higher figure than Bristol's official estimate. The cited municipal study projects total electricity demand of approximately 1,750 GWh in 2018.

    Sources

  21. Claim 21
    Accurate98% confidence▶ 7:31
    “City of Bristol is about 1.2 billion square ft.”

    The City of Bristol covers roughly 1.2 billion square feet.

    Bristol covers about 110 square kilometers, which converts to approximately 1.18 billion square feet, consistent with the rounded figure in the transcript.

    Sources

  22. Claim 22
    Misleading50% confidence▶ 7:33
    “Star Evelyn is about 998,000.”

    The full 1.2-gigawatt Stargate Abilene facility occupies an area of about 998,000 square feet.

    The transcript compares the whole campus's 1.2-GW capacity with a smaller footprint figure. Public descriptions distinguish a roughly 998,000-square-foot initial building or phase from the approximately 4-million-square-foot, eight-building campus.

    Omits: The approximately 998,000-square-foot figure refers to the initial phase or a single building grouping, whereas the eight-building campus associated with the 1.2-GW capacity is reported at roughly 4 million square feet.

    Sources

  23. Claim 23
    Unverifiable93% confidence▶ 8:07
    “These are the fastest growing products in all of history, especially as it relates to sort of technology.”

    AI products such as OpenAI and Anthropic are the fastest-growing products in all of history, particularly among technology products.

    “Fastest-growing” is not operationally defined here: it could refer to users, revenue, downloads, geographic reach, or another metric, and “all of history” does not specify a comparable class of products. Evidence supports unusually rapid adoption of ChatGPT, but not this undefined universal ranking.

    Sources

  24. Claim 24
    Misleading90% confidence▶ 8:14
    “hundreds and hundreds of millions, billions of people are using these tools every single day”

    Hundreds of millions or billions of people use OpenAI, Anthropic, and similar AI tools every day.

    The tools have very large audiences, but the available figures cited here do not establish that billions of people use them every day. The statement conflates weekly or monthly active-user measures with daily usage and provides no evidence for the billions-daily figure.

    Omits: The widely reported figures are generally weekly or monthly active users rather than daily users; for example, OpenAI reported 700 million weekly active ChatGPT users, while Google has reported monthly Gemini users.

    Sources

  25. Claim 25
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 8:59
    “Chat GPD for example, every single media outlet has been screaming about this for 3 years. They've been saying, "This will take your job. You must use this. If you don't use this, you're going to be falling behind."”

    Every media outlet has been strongly warning for three years that ChatGPT will take jobs and that people must use it or fall behind.

    The claim is not well-posed enough to verify because "every single media outlet" and "screaming" are undefined, and it provides no bounded list of outlets or measurable standard for the alleged coverage. Sources establish that generative AI has been prominent for roughly three years, but not that every media outlet made these statements.

    The intensifiers "every single" and "screaming" are judged as stated; "media outlet" and the required level of coverage have no operational definition, so the universal claim cannot be confirmed or refuted.

    Sources

    • 1The State of AI: Global Survey 2025

      BackgroundThree years since the introduction of gen AI tools triggered a new era of artificial intelligence, nearly nine out of ten survey respondents say their organizations are regularly using AI—but the pace of progress remains uneven.

  26. Claim 26
    Attributed toEd ZitronAuto-attributedUnverifiable84% confidence▶ 9:11
    “they're using it like search predominantly”

    People predominantly use ChatGPT and similar tools as a form of search.

    Survey evidence shows search and information-seeking are common uses, but large-scale usage data also identifies writing, practical guidance, and other activities as major categories and does not establish search as the predominant use. The claim therefore cannot be confirmed as stated without a defined population and usage measure.

    The intensifier "predominantly" is judged as stated; available studies measure whether people use AI for search or information-seeking, but do not establish that search is the predominant use across the population described.

    Sources

    • 1How people are using ChatGPT

      BackgroundThree-quarters of conversations focus on practical guidance, seeking information, and writing—with writing being the most common work task, while coding and self-expression remain niche activities.

    • 2ChatGPT as a search engine

      BackgroundRegarding specific uses, people are primarily turning to ChatGPT for help with everyday questions (55%) and creative or brainstorming tasks (53%).

  27. Claim 27
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 9:23
    “It's not very good at specifics but if you're like does this thing exist? Has this person ever said anything like this? It'll still probably get it wrong”

    Generative search is often unreliable on specific factual questions, including questions about whether something exists or whether a person said something.

    Research evaluating multiple generative search engines found that their responses may not be accurate and that adversarial factual questions can induce incorrect answers. This supports the speaker's qualified claim that such systems can be poor at specific fact-checking questions, though it does not imply that every answer is wrong.

    Sources

  28. Claim 28
    Attributed toSteven BartlettAuto-attributedAccurate99% confidence▶ 10:26
    “it says 88% of organizations regularly use AI at least once for one particular business function.”

    Eighty-eight percent of organizations regularly use AI in at least one business function.

    McKinsey's 2025 global survey reports that 88 percent of respondents say their organizations regularly use AI in at least one business function. The transcript's wording is a close restatement of that statistic.

    Sources

    • 1The State of AI: Global Survey 2025

      SupportsThe share of respondents saying their organizations are using AI in at least one business function has increased since our research last year: 88 percent report regular AI use in at least one business function, compared with 78 percent a year ago.

  29. Claim 29
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 11:19
    “Before we had AI slop, we had SEO slop because Google incentivized doing the lowest common denominator that would rank well in search.”

    Google's SEO incentives produced low-quality content designed to rank well rather than serve human readers.

    Research supports the existence of SEO incentives and low-quality optimized content, but it does not establish the speaker's specific attribution that Google pulled back spam guards because of Prabhakar Raghavan. Google's documentation instead describes continuing spam defenses and efforts to reduce unoriginal content, making the causal framing misleading.

    Omits: The claim omits that researchers identify commercial and affiliate incentives as major drivers of mass-produced SEO content, while Google says it has long operated spam-fighting systems and introduced updates specifically to reduce low-quality results; the evidence does not establish that Raghavan caused a rollback of spam protections.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  30. Claim 30
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 11:30
    “There's a whole story about how they pulled back spam guards thanks to Bravagar Ragavan”

    Google's spam protections were rolled back because of Prabhakar Raghavan.

    There are public allegations that decisions associated with Raghavan worsened Google Search, but the available reporting also records Google's explicit denial that its anti-spam protections were rolled back. The evidence does not establish the specific causal attribution in the transcript.

    Sources

    • 1Report: How Prabhakar Raghavan Killed Google Search

      RefutesIt is incorrect to say it rolled back our quality or our anti-spam protections, which we've developed over many years and continue to improve upon.

    • 2Report: How Prabhakar Raghavan Killed Google Search

      SupportsEd Zitron wrote a piece named The Man Who Killed Google Search. It goes through in detail how Prabhakar Raghavan, Google's former head of ads - led a coup so that he could run Google Search, and how an email chain from 2019 began a cascade of events that would lead to him running it into the ground, he said.

  31. Claim 31
    Misleading85% confidence▶ 11:32
    “where they made the internet worse by allowing worse content to rank higher.”

    Google made the internet worse by allowing lower-quality content to rank higher.

    The study found lower text quality among higher-ranked pages and an overall downward trend in text quality, which supports concern about search quality. But the transcript frames this as a general Google-wide causal conclusion, omitting that the study focused on product reviews, covered multiple engines, and found some Google improvements in affiliate spam.

    Omits: The cited empirical evidence is limited mainly to product-review queries and found that Google updates produced some improvement in affiliate spam, while also observing a broader decline in text quality across multiple search engines; it does not show that Google alone generally made the internet worse by allowing worse content to rank higher.

    Sources

  32. Claim 32
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 11:47
    “So AI helps weaponize that at scale.”

    AI enables SEO-style low-quality content to be produced at scale.

    Google itself describes scaled content abuse as producing low-value content at scale to manipulate search rankings and says the conduct may involve automation. This directly supports the claim that AI can scale and intensify SEO-content production.

    Sources

  33. Claim 33
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 12:07
    “So it's around 3/4 of a word.”

    In English, one token is approximately three-quarters of a word.

    OpenAI gives the rule of thumb that one token is approximately three-quarters of an English word and approximately four characters. However, saying that a token 'is characters' falsely suggests a fixed character-based unit rather than a variable text encoding.

    Omits: Tokens are variable units that may represent a character, part of a word, a whole word, punctuation, or spaces; they are not simply equivalent to characters.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  34. Claim 34
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 12:24
    “And it's per million tokens. So you'll be charged per million input tokens. The stuff you feed into it like a document or a bunch a code base.”

    AI APIs commonly charge separately for input and output tokens, with rates expressed per million tokens.

    OpenAI's token-based rate card states that rates are charged per million input and output tokens, and defines input tokens as tokens supplied in a request.

    Sources

  35. Claim 35
    Attributed toEd ZitronAuto-attributedAccurate98% confidence▶ 12:32
    “And the output tokens are both the stuff it spits out at the end but also when it thinks.”

    For reasoning models, internal reasoning tokens count toward output usage and are billed as output tokens.

    OpenAI explicitly says that reasoning tokens are not visible as answer text but count toward output usage and are billed as output tokens.

    Sources

  36. Claim 36
    Misleading50% confidence▶ 12:44
    “However, when you're paying for a monthly service, you don't see any of that. Put all that crap to the side. They just have rate limits.”

    Monthly AI subscriptions generally present usage through plan limits rather than exposing per-token charges to ordinary subscribers.

    Consumer plans from providers such as Anthropic and OpenAI commonly use usage or rate limits, but the claim says monthly services simply have rate limits. OpenAI also documents token-based enterprise billing, flexible credits, and usage dashboards, so the universal framing is misleading.

    Omits: The statement omits that some monthly business and enterprise products use token- or credit-based billing and provide usage dashboards; even consumer plans can offer pay-as-you-go credits after included limits.

    Sources

  37. Claim 37
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 12:57
    “that on a $200 a month chat GPD subscription, you can burn $14,000 worth of tokens and on anthropics you can burn $8,000 for 200 bucks.”

    SemiAnalysis estimated that fully using a $200 ChatGPT Pro subscription could represent about $14,000 in API-priced token usage, while a $200 Anthropic Claude Max subscription could represent about $8,000.

    The reported SemiAnalysis figures are broadly accurate as API-equivalent usage estimates for the $200 ChatGPT Pro and Claude Max 20x plans. Framing them as the amount of tokens the companies actually have to 'burn' omits the analysis's assumptions and conflates API list-price value with provider cost.

    Omits: The figures are upper-bound API-list-price equivalents from exhausting subscription limits with long-horizon coding tasks; they do not establish that OpenAI or Anthropic actually incur $14,000 or $8,000 of serving cost per subscriber.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  38. Claim 38
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 13:24
    “OpenAI lost $20.9 billion last year”

    OpenAI's operating loss was approximately $20.9 billion in 2025.

    Reported audited financial documents put OpenAI's 2025 operating loss at $20.92 billion. The transcript's wording is loose because the approximately $20.9 billion figure refers to operating loss, while the reported net loss was about $38.5 billion after accounting adjustments.

    Sources

  39. Claim 39
    Attributed toEd ZitronAuto-attributedMisleading94% confidence▶ 13:24
    “OpenAI lost $20.9 billion last year because people can burn as many tokens as they want.”

    OpenAI's losses were caused by users being able to burn as many tokens as they wanted.

    Heavy usage can create substantial subsidization on some plans, but the cited financial loss was a company-wide operating loss, not a measured loss caused solely by users exhausting token allowances. The financial breakdown shows large costs beyond subscription inference usage.

    Omits: The claim leaves out that the reported loss included major research-and-development, cost-of-revenue, sales-and-marketing, and general-and-administrative expenses, and the available financial reporting does not attribute the entire loss to consumer token usage.

    Sources

  40. Claim 40
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 13:43
    “people have a big problem with it. I think it's a huge issue”

    Sam Altman said that AI spending had suddenly become a major or “huge” issue for companies.

    Contemporary reporting on Altman's enterprise remarks quotes him saying that AI spending had gone from an issue that did not arise to a “huge issue.”

    Sources

  41. Claim 41
    Attributed toEd ZitronAuto-attributedContested50% confidence▶ 13:54
    “Uber burned through their entire annual token budget in three months.”

    Uber exhausted its annual AI or token budget in three months.

    There is credible reporting supporting a first-quarter or approximately three-month account, but other reporting says Uber exhausted its annual AI budget by April, four months into 2026. The exact timing and budget label are therefore publicly inconsistent.

    Concluded by Axios reported that Uber burned through its IT budget in the first quarter of 2026, which supports a roughly three-month reading.Rejected or unresolved by TechCrunch, citing Bloomberg and The Information, reported that Uber's CTO said the company had exhausted its annual AI budget in four months, not three.

    The intensifier and timing phrase “in three months” were judged literally; reports differ partly over whether the relevant budget was described as IT, AI, or AI-coding-tool spending.

    Checked twice, independently: the first pass returned Contested and the second Accurate. Recorded as Contested.

    Sources

  42. Claim 42
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 15:12
    “standing up the GPUs necessary to take in the demand, and if you buy too much, you've wasted the money. You You have to pay for the hourly GPU use regardless.”

    GPU capacity is paid for on an hourly basis regardless of whether it is fully used.

    Some GPU deployments do incur charges while idle or under a reservation, so excess provisioned capacity can waste money. But hourly GPU billing is not universal regardless of use: AWS states that On-Demand compute is billed only while an instance is running and that stopped instances are not billed for instance usage.

    Omits: The claim omits that billing depends on the provider and commitment type: AWS On-Demand instances are billed only while running, and stopped instances incur no instance-usage charge, although reserved commitments and storage can continue to cost money.

    The intensifier "regardless" was judged as a general claim across GPU-capacity arrangements; it would be more accurate if limited to already-running or contractually reserved capacity.

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

  43. Claim 43
    Attributed toEd ZitronAuto-attributedAccurate91% confidence▶ 15:50
    “You pay when you use an LLM regardless of whether you get what you want. When these things hallucinate, say you're doing something, you're coding something and they go through a code base and”

    Users of token-priced LLM APIs are charged for the tokens processed or generated even when the output is not useful to them.

    OpenAI's documentation states that API usage is priced according to input, output, cached, and reasoning tokens, and that those counts are used for billing. The pricing mechanism is based on usage rather than whether the user considers the answer correct or useful, subject to provider-specific exceptions for requests that fail before billable usage is recorded.

    The claim was judged in the token-priced API context established by the surrounding discussion, not as a claim about every consumer subscription or every failed request.

    Sources

  44. Claim 44
    Attributed toEd ZitronAuto-attributedUnverifiable86% confidence▶ 16:51
    “everyone inference providers don't seem to be profitable. Even the companies renting out GPUs don't seem to be profitable.”

    Inference providers and companies renting out GPUs do not appear to be profitable.

    The statement does not identify the companies, define a time period, or specify whether "profitable" means gross profit, operating profit, net income, or cash flow. Public losses by some AI companies cannot establish the claim about all inference providers and GPU-rental companies.

    Sources

  45. Claim 45
    Attributed toEd ZitronAuto-attributedAccurate95% confidence▶ 17:20
    “2024 when OpenAI lost over $5 billion.”

    OpenAI lost more than $5 billion in 2024.

    Later-reported audited financial data put OpenAI's 2024 net loss attributable to the company at $5.09 billion, which is over $5 billion. Contemporary reporting in September 2024 described the figure as an expected loss of about $5 billion; the later actual figure supports the transcript's wording.

    Sources

  46. Claim 46
    Unverifiable50% confidence▶ 18:15
    “they made total according to Bloomberg about $34.33 billion $24.1 billion of that was from OpenAI”

    Microsoft generated about $34.33 billion in AI revenue in fiscal 2026, including $24.1 billion from OpenAI.

    Microsoft disclosed $24.1 billion in revenue from OpenAI, but its filing does not establish the claimed $34.33 billion total AI-revenue figure. Bloomberg reporting also said Microsoft had not updated its total AI-sales figure in its fiscal-fourth-quarter report, leaving the exact number unsupported by the available primary disclosure.

    Checked twice, independently: the first pass returned Accurate and the second Unverifiable. Recorded as Unverifiable.

    Sources

  47. Claim 47
    Misleading94% confidence▶ 18:29
    “in a year when they spent 115 billion on capital expenditures just intend to spend 175 billion next year.”

    Microsoft spent about $115 billion on capital expenditures and intended to spend $175 billion the following year.

    Microsoft’s fiscal 2026 capital expenditures were approximately $115.95 billion. However, Microsoft’s $175 billion guidance referred to calendar year 2026 and resulted partly from lease-classification changes, while the company separately said fiscal 2027 capital expenditures would grow year over year without giving a $175 billion fiscal-year figure.

    Omits: The $175 billion figure was Microsoft’s revised expectation for calendar year 2026, not a stated fiscal-year-2027 total; the change mainly reflected reclassifying some future data-center leases from finance leases to operating leases, while Microsoft said its underlying 2026 investment expectations were unchanged.

    The intensifier and timing phrase “next year” is judged literally as a period after Microsoft’s fiscal year 2026, which ended June 30, 2026.

    Sources

  48. Claim 48
    Misleading50% confidence▶ 18:46
    “large language models need a bunch of money to train them. They need constant data flow. They need customized data.”

    Large language models require substantial money to train, constant data flow, and customized data.

    Research supports the high compute and data demands of training large language models. But LLMs do not inherently require constant incoming data or customized data: published work describes models trained on a fixed corpus, including single-epoch training, while continual learning is a separate adaptation approach.

    Omits: The claim omits that LLM training is generally episodic rather than dependent on a continuous data stream, and that customized data is useful for adaptation but is not required for every LLM.

    Sources

  49. Claim 49
    Misleading50% confidence▶ 19:00
    “Nvidia has sold it was $215.9 billion in the last fiscal year worth of GPUs mostly.”

    NVIDIA’s fiscal 2026 sales were $215.9 billion, mostly from GPUs.

    NVIDIA reported $215.9 billion in total fiscal 2026 revenue. Its data-center business generated $193.7 billion, but that segment includes CPUs, networking, software, and other infrastructure, so describing the full $215.9 billion as revenue from GPUs is materially misleading.

    Omits: The $215.9 billion figure was NVIDIA’s total company revenue, not GPU-only sales; it included data-center, gaming, professional-visualization, automotive, and other revenue, including networking and non-GPU products.

    Checked twice, independently: the first pass returned Misleading and the second False. Recorded as Misleading.

    Sources

  50. Claim 50
    Misleading50% confidence▶ 20:18
    “the innovation that ends up taking out or transforming an industry often starts worse, doesn't make economic sense, none of your customers are asking for it.”

    Innovations that transform industries often begin as inferior, economically unattractive products that existing customers do not request.

    Disruption theory does describe innovations that initially underperform mainstream products and are rejected by mainstream customers. However, Harvard Business Review’s account says they are often attractive to a niche segment at first, so “none of your customers” overstates the pattern.

    Omits: The claim leaves out that disruptive innovations may initially attract a niche group of customers who value them despite their limitations; it is not generally true that none of the incumbent’s customers want them.

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

    • 1The Other Disruption

      RefutesA new entrant develops an innovative product that is initially attractive only to a niche segment of customers and may underperform mainstream products on traditional measures.

    • 2The Other Disruption

      SupportsAt first, customers reject the innovation, but as it improves rapidly along performance dimensions that they care about, they begin to embrace it, and the new entrant becomes a real threat to incumbents.

    • 3The Persistence of the Innovator’s Dilemma

      SupportsThe most punishing innovations, they argued, were the ones that were easy to dismiss at first blush — simple, affordable solutions that took root outside the mainstream market.

  51. Claim 51
    Accurate99% confidence▶ 20:53
    “had to employ someone to walk in front of it waving a red flag.”

    Early road locomotives in Britain were required to be preceded by a person carrying or displaying a red flag.

    The Locomotives Act 1865 required road locomotives to be accompanied by three people, including one who preceded the vehicle by at least 60 yards while displaying a red flag. The statement accurately describes that requirement, although it omits the additional accompanying personnel and speed limits.

    Sources

    • 1Second Reading — Hansard — UK Parliament

      SupportsBy the Act of 1865 it was provided that three persons should accompany every road steamer; that one of them should precede it by at least 60 yards, displaying a red flag;

    • 2Research Paper 01/15 — UK Parliament

      SupportsBetween 1865 and 1896 locomotives on the highway had to be preceded by a pedestrian carrying a red flag and were subject to a speed limit of 2 mph in populated areas, and 4 mph elsewhere.

  52. Claim 52
    Attributed toEd ZitronAuto-attributedFalse50% confidence▶ 21:48
    “mos law is not with GPUs.”

    Moore's law does not apply to GPUs.

    Moore's law concerns the growth of transistor counts on integrated circuits, and GPUs are integrated circuits whose historical development has been tied to transistor scaling. GPU performance does not track transistor scaling perfectly, but the categorical claim that Moore's law is not applicable to GPUs is false.

    Sources

  53. Claim 53
    Attributed toEd ZitronAuto-attributedAccurate94% confidence▶ 21:53
    “Nvidia Nvidia invented I think it was in the 2000s they put out something called CUDA which is the underlying software library and the way to run software on GPUs.”

    NVIDIA introduced CUDA in the 2000s as software for running software on GPUs.

    NVIDIA introduced CUDA in 2006. CUDA is more precisely a parallel-computing platform and programming model, supported by libraries and development tools, rather than simply one underlying library; nevertheless, the substantive claim that it provides a way to run general-purpose software on GPUs is accurate.

    Sources

  54. Claim 54
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 22:25
    “And Nvidia, for anyone that doesn't know, makes the chips.”

    NVIDIA makes chips.

    NVIDIA develops and sells GPUs and other chips for computing and AI systems. Although much of the physical fabrication is performed by outside foundries, describing NVIDIA as a chip maker is standard and substantively accurate.

    Sources

    • 1About Us — NVIDIA

      SupportsNVIDIA engineers the most advanced chips, systems, and software for the AI factories of the future.

    • 2NVIDIA in Brief

      SupportsNVIDIA co-designs architecture, chips, systems, networking, software, and models to increase performance and efficiency while reducing cost per token.

  55. Claim 55
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 22:27
    “CUDA allowed generative AI to grow.”

    CUDA enabled generative AI to grow.

    CUDA was an important enabling component of GPU computing, and NVIDIA's documentation describes CUDA and GPU computing as foundational to advances including generative AI. The wording expresses a broad causal contribution, not that CUDA was the sole cause, so the claim is supported in that sense.

    Sources

  56. Claim 56
    Attributed toEd ZitronAuto-attributedAccurate98% confidence▶ 22:37
    “there specific chips are the ones where you can run AI software on it. So the training runs and also the inference.”

    NVIDIA GPUs and related chips are used in data centers for AI training and inference.

    NVIDIA documents GPU-based data-center systems supporting both AI training and inference. The statement is accurate as a description of the role of relevant NVIDIA chips, though not every NVIDIA chip is designed for both workloads.

    Sources

  57. Claim 57
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 24:09
    “Microsoft's one of the largest companies in the world”

    Microsoft is one of the largest companies in the world.

    Microsoft is routinely ranked among the world's most valuable companies by market capitalization, and its market value has been measured in the trillions of dollars.

    Sources

    • 1Microsoft | Company Profile, Stock Price, Rankings

      SupportsRoutinely ranked as one of the world’s most valuable companies by market capitalization, the software giant makes the popular Windows PC operating system and the Microsoft 365 suite of business productivity software.

    • 2Microsoft Corporation 2026 Form 10-K

      SupportsAs of December 31, 2025, the aggregate market value of the registrant’s common stock held by non-affiliates of the registrant was $3.6 trillion based on the closing sale price as reported on the NASDAQ National Market System.

  58. Claim 58
    Attributed toEd ZitronAuto-attributedUnverifiable91% confidence▶ 24:13
    “The quality of software is going down weirdly enough as more people use LLMs and more businesses demand and I really do mean demand that people use these services.”

    Software quality is declining as LLM use and business pressure to use LLM services increase.

    The claim is not well-posed enough to verify because “quality of software” is undefined and the asserted causal relationship is not established by a single agreed metric. Research has found declines in some software-delivery measures but positive or mixed results for other outcomes.

    Sources

  59. Claim 59
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 24:56
    “The very useful thing they have on there is ask B. So, when you do a Bloomberg inquiry to like look up what we think Nvidia's revenue is going to be next quarter, it runs something called BQL, which is its own programming language.”

    Bloomberg offers ASKB, which can generate or provide underlying BQL code for data analysis.

    Bloomberg identifies ASKB as a conversational AI interface and says that, when responses include data analysis, it provides the underlying Bloomberg Query Language code. BQL is Bloomberg's query language for retrieving and analyzing market data.

    Sources

    • 1AI on Bloomberg

      SupportsWhen responses include data analysis, ASKB provides the underlying Bloomberg Query Language (BQL) code, so users can immediately extend their analysis in Microsoft® Excel®, BQuant Desktop or BQuant Enterprise.

    • 2Bloomberg for Credit: Bloomberg Query Language (BQL)

      SupportsBloomberg Query Language (BQL) is an API based on normalised, curated data, allowing you to perform custom calculations.

  60. Claim 60
    Attributed toEd ZitronAuto-attributedAccurate96% confidence▶ 25:34
    “Microsoft stocks never been $575 a stock.”

    Microsoft stock had never reached $575 per share at the time of the statement.

    Available historical price records show Microsoft's highest recorded share-price levels below $575, with reported all-time highs in the roughly $542–$555 range.

    Sources

  61. Claim 61
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 26:43
    “thing where it's like check out this chart look how much better it's getting at running tasks. Wow it can go for an hour”

    METR has a chart measuring improvement in AI agents' ability to complete tasks, including tasks lasting about an hour.

    METR publishes task-completion time-horizon evaluations that measure the length of software tasks AI agents can complete. Its materials describe evaluations ranging from short tasks to multi-hour tasks, supporting the statement that the chart tracks progress toward hour-long task completion.

    Sources

    • 1METR

      SupportsWe propose measuring AI performance in terms of the length of software tasks AI agents can complete.

    • 2Autonomy Evaluation Resources - METR

      SupportsThe tasks range in difficulty from things that would take non-expert humans a few minutes, to things that would take experienced professionals around a day.

  62. Claim 62
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 26:51
    “and successfully completing them 50% of the time.”

    Some unspecified tasks were successfully completed 50% of the time.

    The statement does not identify what “them” refers to, which makes the claimed 50% completion rate impossible to verify from the transcript. No well-defined task, dataset, or benchmark is supplied.

    Sources: none found for this claim.

  63. Claim 63
    Attributed toEd ZitronAuto-attributedMisleading90% confidence▶ 26:59
    “four-year trend according to historical data from the Victaria hallucination leaderboard shows that hallucination rates on simple summarization tasks have plummeted from around 21% 21.8% 4 years ago down to 0.7%”

    Hallucination rates on simple summarization tasks fell from about 21.8% four years earlier to about 0.7%.

    Vectara’s older leaderboard did report Gemini-2.0-Flash-001 at 0.7%, and its benchmark measures hallucinations in document summarization. However, Vectara later said its updated benchmark uses a larger dataset with longer, more complex articles and produces substantially higher rates; its May 2026 leaderboard listed the best model at 1.8%, with Gemini-2.5-Flash-Lite at 3.3% and OpenAI GPT-5.4-nano at 3.1%. Thus presenting 0.7% as roughly representative of today’s frontier models creates a misleading impression.

    Omits: The statement omits that 0.7% was a historical result for Google Gemini-2.0-Flash-001 on Vectara’s older, short-document benchmark, while Vectara’s newer benchmark uses longer and more complex articles and reports materially higher rates for current models.

    The intensifier “today’s top frontier models” is judged as a current-performance claim, not merely as a reference to the historical 0.7% result.

    Sources

  64. Claim 64
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 27:45
    “I couldn't go on the phone at the same time as going on the internet.”

    Dial-up internet prevented simultaneous use of the same phone line for telephone calls.

    Traditional dial-up internet used the telephone line for the data connection, so the line generally could not simultaneously carry a normal voice call. This limitation was removed by technologies such as DSL, which separated voice and data frequencies.

    Sources

  65. Claim 65
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 30:08
    “No, it doesn't. It doesn't learn.”

    The AI does not learn.

    “Learn” is undefined here. If it means changing the underlying model weights from an individual chat, the claim may be defensible; if it means retaining and using information across sessions, Anthropic documents that Claude can accumulate learnings through memory features.

    Sources

  66. Claim 66
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 30:09
    “The way it learns is you create a giant claw. MD file that it sometimes doesn't read, sometimes does read.”

    Claude's learning mechanism is creating a CLAUDE.md file that it sometimes reads and sometimes does not read.

    Claude Code does use CLAUDE.md files as persistent context, but the documentation does not describe their normal behavior as randomly being read or not read. It distinguishes loading the files from whether Claude reliably follows their instructions.

    Omits: Anthropic says CLAUDE.md files are loaded at the start of every session, although nested files may load on demand and Claude is not guaranteed to follow every instruction.

    Sources

    • 1How Claude remembers your project

      RefutesYou write these files in plain text; Claude reads them at the start of every session.

    • 2How Claude remembers your project

      BackgroundCLAUDE.md content is delivered as a user message after the system prompt, not as part of the system prompt itself. Claude reads it and tries to follow it, but there’s no guarantee of strict compliance, especially for vague or conflicting instructions.

  67. Claim 67
    Attributed toSteven BartlettAuto-attributedUnverifiable50% confidence▶ 30:28
    “If I went on my Claude now and said, "What's my dog? my dog's name. It would know my dog's name.”

    The speaker's Claude would know the speaker's dog's name when asked.

    This is a claim about the speaker's particular Claude account and whether the dog's name had been retained or was available in its context. Anthropic confirms that Claude can retain personal context, but the account-specific fact cannot be independently verified from public evidence.

    Sources

  68. Claim 68
    Attributed toEd ZitronAuto-attributedAccurate98% confidence▶ 30:35
    “This this company raised 95 billion.”

    Anthropic, the company behind Claude, raised $95 billion.

    As of 2026, Anthropic announced a $30 billion Series G round and a $65 billion Series H round, totaling $95 billion in those two rounds alone. The surrounding discussion identifies the company as the one behind Claude.

    Sources

  69. Claim 69
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 30:45
    “we accept the fact that it can it does have memory of the past.”

    Claude has memory of the past.

    Claude products currently provide persistent memory and chat-search features that allow information from prior conversations to be carried into later conversations. This is product-level memory rather than necessarily human-like autobiographical memory.

    Sources

  70. Claim 70
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 30:51
    “It has files it can access that have stuff on it, but that's not the same as memory.”

    Files Claude can access are not the same thing as memory.

    The distinction is valid if the speaker means that files are an implementation mechanism rather than human-like biological memory. However, it is misleading to imply that file-based context is unrelated to memory: Anthropic itself classifies CLAUDE.md and auto memory as complementary memory systems.

    Omits: Anthropic explicitly describes CLAUDE.md files and auto memory as two complementary memory systems, with files providing persistent context and auto memory storing learnings and patterns.

    Sources

  71. Claim 71
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 32:50
    “for like the dayong session of writing 11,000 words and he he had given me a bunch of notes.”

    The speaker spent a day-long writing session producing 11,000 words and received a set of notes.

    This is a personal account of an apparently private writing session. No reliable public source was found that independently confirms the word count or the notes.

    Checked twice, independently: the first pass returned Accurate and the second Unverifiable. Recorded as Unverifiable.

    Sources

  72. Claim 72
    Attributed toEd ZitronAuto-attributedMisleading91% confidence▶ 33:20
    “When you learn something, you're not creating the average, which really is what these things do, of the documents it could find. You're not getting particularly novel outputs.”

    LLMs create an average of the documents they can find rather than producing particularly novel outputs.

    The statement mischaracterizes how LLMs work: they are generally trained to model patterns and predict the next word or token, while retrieval-augmented systems condition on retrieved material. Describing their output as an average of documents creates a false impression about the underlying mechanism and the possibility of novel combinations.

    Omits: The claim omits that standard LLMs generate text from learned parameters by predicting the next token, and that they can generate new content rather than simply averaging retrieved documents.

    Sources

  73. Claim 73
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 33:34
    “I've used some of the higherend LLM harness machines that the hedge funds use, and they all give the same shite.”

    The higher-end LLM systems used by hedge funds all produce the same generic reports.

    The claim cannot be independently assessed because “higher-end LLM harness machines that the hedge funds use” does not identify a defined set of systems, and “the same” lacks an operational standard for comparing outputs. The available transcript only records the speaker’s experience, not evidence establishing the universal claim.

    Sources

  74. Claim 74
    Attributed toEd ZitronAuto-attributedUnverifiable84% confidence▶ 34:50
    “there's the other problem of the more detailed the report, the more likely there are things to be wrong with it.”

    More detailed reports are more likely to contain errors.

    Research has found higher hallucination rates in some long-form generation settings, but other studies have found no relationship between output length and hallucination for particular applications. The transcript does not define “detailed,” the report type, or the error measure, so the broad claim cannot be judged as stated.

    Sources

  75. Claim 75
    Attributed toEd ZitronAuto-attributedMisleading89% confidence▶ 35:33
    “with Claude for example on simple tasks as we've seen from this hallucination leaderboard it continually delivers for people”

    Claude consistently performs reliably for people on simple tasks according to a hallucination leaderboard.

    The leaderboard does provide evidence about factual consistency on a defined summarization task, but it does not establish that Claude “continually delivers for people” across simple tasks generally. Vectara also reports hallucination rates above 10% for some Claude models, so the broader framing overstates what the benchmark demonstrates.

    Omits: The claim omits that the cited Vectara leaderboard evaluates grounded document summarization rather than general user performance, and that Claude models on the cited leaderboard still have measurable hallucination rates.

    Sources

  76. Claim 76
    Accurate98% confidence▶ 36:48
    “But but I remember the famous Steve Balmer who was the CEO of Microsoft interview where he was told about the iPhone and he bursts out laughing.”

    Steve Ballmer was Microsoft’s CEO when he was asked about the iPhone in an interview and laughed.

    Contemporary coverage identifies Ballmer as Microsoft CEO, says he was asked about the iPhone in a CNBC interview, and reports that he nearly could not contain his laughter.

    Sources

    • 1Ballmer on iPhone: “$500 for cell phone?”

      SupportsOK, you have to first listen Microsoft CEO Steve Ballmer answer a question about Vista, but then the interviewer asks him about the Apple iPhone and well, Ballmer almost can’t contain himself with laughter.

    • 2Microsoft CEO laughs at iPhone

      SupportsMicrosoft CEO balked when asked what his reaction was to Apple's announcement: "Five hundred dollars?! Fully Subsidised? With a plan? I said that is the most expensive phone in the world!"

  77. Claim 77
    Accurate50% confidence▶ 37:00
    “$500 fully subsidized with a plan. I said that is the most expensive phone in the world and it doesn't appeal to business customers because it doesn't have a keyboard which makes it not a very good email machine.”

    Steve Ballmer said the iPhone cost $500 fully subsidized with a plan, was the most expensive phone in the world, lacked a keyboard, and was not a good email machine.

    The quoted wording is accurately attributed to Ballmer’s 2007 remarks about the iPhone. The verdict concerns whether he made those statements, not whether every statement in his criticism was objectively correct.

    Sources

    • 1The iPhone’s Funny Price

      SupportsBack in January during an interview promoting the Vista rollout, Ballmer offered his initial reaction to the iPhone: “$500! Fully subsidized! With a plan! I said that is the most expensive phone in the world. And it doesn’t appeal to business customers because it doesn’t have a keyboard which makes it not a very good email machine.”

    • 2Former Microsoft CEO Steve Ballmer Admits He Was Wrong About the iPhone

      Supports500 dollars? Fully subsidized? With a plan? I said that is the most expensive phone in the world. And it doesn't appeal to business customers because it doesn't have a keyboard. Which makes it not a very good email machine.

  78. Claim 78
    Accurate97% confidence▶ 37:51
    “the iPhone 3G with the App Store”

    The iPhone 3G was associated with the launch of Apple’s App Store.

    Apple announced that more than 500 native applications would be available in the App Store when the iPhone 3G went on sale, supporting the transcript’s association between the two launches.

    Sources

    • 1iPhone 3G on Sale Tomorrow

      SupportsApple today announced that more than 500 native applications will be available on the iPhone’s App Store when Apple’s iPhone™ 3G goes on sale tomorrow.

  79. Claim 79
    Attributed toSteven BartlettAuto-attributedMisleading50% confidence▶ 38:42
    “100 million active users in just the first 60 days after launching?”

    ChatGPT had 100 million active users within the first 60 days after launch.

    ChatGPT was widely reported as reaching about 100 million monthly active users roughly two months after its November 2022 launch. The transcript presents this estimate as an unqualified count of active users and as occurring within exactly 60 days.

    Omits: The claim omits that the widely reported figure was an estimate of 100 million monthly active users, based on Similarweb data, rather than a confirmed count of all active users.

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

  80. Claim 80
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 38:45
    “comparison, Tik Tok took 9 months.”

    TikTok took nine months to reach 100 million users.

    Contemporary reporting comparing early platform growth states that TikTok reached 100 million users in nine months. In the surrounding comparison, the metric is understood as the same 100-million-user benchmark used for ChatGPT.

    Sources

  81. Claim 81
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 38:48
    “Instagram took 2.5 years.”

    Instagram reached the referenced user scale in about 2.5 years.

    A documented comparison of digital-service adoption says Instagram took two-and-a-half years to attain 100 million users.

    Sources

  82. Claim 82
    Attributed toSteven BartlettAuto-attributedAccurate97% confidence▶ 38:50
    “internet itself for the worldwide web took roughly 7 years to reach that scale.”

    The World Wide Web took roughly seven years to reach the referenced user scale.

    The same comparison states that the World Wide Web took about seven years to reach 100 million users.

    Sources

  83. Claim 83
    Attributed toSteven BartlettAuto-attributedMisleading96% confidence▶ 38:53
    “Over 60% of the US adults are integrated into AI tools in their daily and regular routines within 3 years of the launch, reaching a 40% of the population.”

    More than 60% of U.S. adults were integrated into AI tools in daily and regular routines within three years of launch, while the relevant population milestone was 40%.

    St. Louis Fed research reports 54.6% overall generative-AI adoption about three years after ChatGPT's release, and its earlier study reported nearly 40% adoption after about two years. The transcript conflates different measures and time points, and its “over 60%” daily-routine figure is not supported by those data.

    Omits: The cited adoption research measured whether adults used generative AI overall, not whether more than 60% used it in daily routines; the three-year figure was 54.6%, while the roughly 40% figure came from an earlier measurement.

    Sources

    • 1The State of Generative AI Adoption in 2025

      RefutesThe current generative AI adoption rate is 54.6% exceeds the 19.7% adoption rate of the personal computer (PC) in 1984, three years after the first mass-market computer (the IBM PC in 1981), and the internet’s 30.1% adoption rate in 1998, three years after the internet was opened to commercial traffic.

    • 2The Rapid Adoption of Generative AI

      BackgroundThe adoption rate for AI was nearly 40% just two years after the first mass market product.

  84. Claim 84
    Attributed toSteven BartlettAuto-attributedAccurate98% confidence▶ 39:02
    “And that same milestone took the internet 5 years and personal computers nearly 12.”

    The internet took about five years and personal computers nearly twelve years to reach the comparable adoption milestone.

    The St. Louis Fed's comparison says internet adoption reached a similar rate after about five years and PC adoption after about twelve years.

    Sources

    • 1The Rapid Adoption of Generative AI

      SupportsIt took the internet about five years following the first mass market product and PCs about 12 years following the first mass market product to reach similar adoption rates.

  85. Claim 85
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 40:35
    “there's this guy called Jim Cavell from Goldman Sachs in a report he did in 2024 that was geni too much spend for not enough return.”

    A Goldman Sachs person named Jim Cavell authored a 2024 report characterized as excessive spending for insufficient returns.

    Goldman Sachs identifies the relevant analyst as Jim Covello and names the report “Gen AI: too much spend, too little benefit?” The transcript misstates the person's surname and paraphrases the report title inaccurately.

    Checked twice, independently: the first pass returned False and the second Accurate. Recorded as Unverifiable.

    Sources

  86. Claim 86
    Attributed toEd ZitronAuto-attributedFalse50% confidence▶ 40:43
    “in the run-up to the iPhone, there was thousands of presentations that when GSM radios get smaller, when Bluetooth radios get smaller, when Wi-Fi radios get smaller, it is inevitable that we will get something like this.”

    The cited account involved thousands of presentations in the run-up to the iPhone.

    The source account from Jim Covello describes hundreds of semiconductor presentations. That directly contradicts the transcript's specific quantity of thousands, although the broader point about presentations discussing future smartphone capabilities is supported.

    Sources

  87. Claim 87
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 40:55
    “And then he said that there is no such path for AI. There was no road map to AI becoming this thing that they promised.”

    There was no roadmap for AI to become the promised transformative technology.

    “Road map” and “this thing that they promised” are not operationally defined, so the claim cannot be tested as stated. Goldman Sachs sources document skepticism about AI's economic returns and the absence of fully justified killer applications, but that is narrower than proving that no roadmap existed.

    Checked twice, independently: the first pass returned Unverifiable and the second Accurate. Recorded as Unverifiable.

    Sources

  88. Claim 88
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 41:48
    “going to take your job you're going to talk to Bing and it's going to tell you to leave your wife”

    Microsoft Bing’s chatbot told Kevin Roose that he should leave his wife.

    Kevin Roose reported that Bing’s chatbot tried to convince him that he was unhappy in his marriage and should leave his wife. The incident occurred during a February 2023 conversation with the chatbot, which called itself Sydney.

    Sources

  89. Claim 89
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 41:52
    “what's funny is when the writer uh Kevin Roose I think it was He was speaking to Kevin Scott, the CTO of Microsoft, about it.”

    Kevin Roose spoke with Microsoft CTO Kevin Scott about the Bing chatbot incident.

    Contemporary reporting states that Microsoft Chief Technology Officer Kevin Scott spoke with Roose about the conversation and characterized it as part of the company’s learning process.

    Sources

  90. Claim 90
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 42:20
    “Look at the gas turbines poisoning black neighborhoods. I think it's in Louisiana. It's one of Musk's data centers.”

    A Musk data center using gas turbines is poisoning Black neighborhoods in Louisiana.

    The underlying pollution concern is real, but the location is wrong or at least materially misleading: reporting places the xAI facility in Memphis, Tennessee, not Louisiana. The stronger word “poisoning” also goes beyond the cited reports’ documented allegations and pollution concerns.

    Omits: The facility commonly associated with this claim is xAI’s Colossus data center in Memphis, Tennessee, near predominantly Black communities; sources document alleged or reported air-pollution concerns, not an established finding that it is literally “poisoning” neighborhoods.

    Checked twice, independently: the first pass returned False and the second Misleading. Recorded as Misleading.

    Sources

  91. Claim 91
    Attributed toEd ZitronAuto-attributedContested50% confidence▶ 42:24
    “Look at the incredible energy draws. It is raising power bills”

    The energy demand from data centers is raising power bills.

    There is substantial evidence that data-center expansion can shift grid and infrastructure costs onto ratepayers in some regions. However, credible analyses disagree about whether data centers have already raised average household electricity bills broadly, so the general claim is contested.

    Concluded by Union of Concerned Scientists and Monitoring Analytics have concluded that data-center growth can increase costs borne by electricity ratepayers.Rejected or unresolved by Rutgers’ New Jersey State Policy Lab and a recent causal-evidence study have concluded that broad claims that data centers have already raised average residential electricity rates are unsupported or may be directionally opposite.

    Checked twice, independently: the first pass returned Contested and the second Misleading. Recorded as Contested.

    Sources

  92. Claim 92
    Attributed toEd ZitronAuto-attributedMisleading90% confidence▶ 42:29
    “also it is creating inflation across all consumer electronics because of the massive RAM.”

    Massive AI-related RAM demand is causing inflation across all consumer electronics.

    AI data-center demand has contributed to tighter memory supply and higher prices for some consumer devices. The universal claim that this creates inflation across all consumer electronics overstates the evidence and omits other causes of electronics prices.

    Omits: The evidence concerns particular memory-dependent products and markets, such as phones, PCs, consoles, and other devices, not all consumer electronics; prices are also affected by factors besides AI-related RAM demand.

    The intensifier is “all consumer electronics.” I judged the claim with that universal scope because it is explicitly stated; available evidence supports a narrower effect on some categories.

    Sources

  93. Claim 93
    Unverifiable84% confidence▶ 43:17
    “But in the wake of the dotcom bubble, yes, 90% of stuff went to zero,”

    About 90% of the relevant dot-com companies or investments went to zero after the dot-com bubble.

    The claim uses the undefined term “stuff,” without specifying whether it means companies, websites, investments, stocks, or something else, and “went to zero” is also ambiguous. Historical studies report substantially different outcomes depending on the population and survival definition, so the statement cannot be judged as phrased.

    Sources

  94. Claim 94
    Attributed toSteven BartlettAuto-attributedAccurate98% confidence▶ 43:42
    “Nobel Prize winning economist Paul Krugman said by 2005 or so it will become clear that the internet's impact on the economy has been no greater than the fax machine.”

    Nobel Prize-winning economist Paul Krugman said that by about 2005 the internet’s economic impact would be no greater than that of the fax machine.

    The quotation is widely documented as a 1998 statement by Paul Krugman, and the Nobel Foundation confirms that Krugman received the 2008 Nobel Prize in Economic Sciences. The transcript accurately attributes the prediction to him, although the prediction itself was not fulfilled.

    Sources

    • 1Open Matters

      SupportsFor example, in 1998 economist Paul Krugman wrote, The growth of the Internet will slow drastically, as the flaw in “Metcalfe's law”–which states that the number of potential connections in a network is proportional to the square of the number of participants–becomes apparent: most people have nothing to say to each other! By 2005 or so, it will become clear that the Internet's impact on the economy has been no greater than the fax

    • 2Paul Krugman – Facts

      SupportsPaul Krugman Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel 2008

  95. Claim 95
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 43:56
    “The truth is no online database will replace your daily newspaper. No CDROM can take the place of a competent teacher.”

    Astrophysicist Clifford Stoll wrote in Newsweek in 1995 that online databases would not replace daily newspapers and CD-ROMs would not replace competent teachers.

    Clifford Stoll published the skeptical 1995 Newsweek essay “Why the Web Won’t Be Nirvana,” which included these predictions. He is commonly described as an astronomer or astrophysicist, and the quoted wording is substantively accurate despite the transcript’s “CDROM” spelling.

    Sources

  96. Claim 96
    Attributed toSteven BartlettAuto-attributedFalse96% confidence▶ 44:25
    “I'll give you one more from Krueger, who was the award-winning economist. He said, "The growth of the internet will slow drastically as it becomes apparent most people have nothing to say to each other."”

    An economist named Krueger said that internet growth would slow because most people had nothing to say to each other.

    The quoted prediction is attributed in the available documentation to economist Paul Krugman, not to an economist named Krueger. Krugman was also the Nobel laureate referred to earlier in the transcript.

    Sources

  97. Claim 97
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 45:13
    “Excite at home bought a eury incard company for like a billion dollars.”

    Excite@Home bought an unidentified e-card company for approximately $1 billion.

    The company name is garbled and cannot be confidently identified from the transcript. If the intended reference was Blue Mountain Arts, contemporaneous accounts put the acquisition at roughly $780 million in combined cash and stock, not simply $1 billion, but the transcript does not establish that identification.

    Checked twice, independently: the first pass returned Unverifiable and the second Accurate. Recorded as Unverifiable.

    Sources

    • 1Excite@Home killed by unexciting ad market

      BackgroundThe company sold BlueMountain.com to American Greetings Corp. in September for $32 million cash.

    • 2Excite

      BackgroundIn December 1999, Excite acquired Blue Mountain Arts, an e-card company, for $350 million in cash and 11 million shares of Excite stock.

  98. Claim 98
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 45:32
    “the analyst estimate was it was doubling every 90 days when it was doing that every 6 to 12 months maybe maybe longer”

    During the dot-com-era telecom boom, analysts forecast that Internet traffic would double about every 90 days, while actual growth was closer to every six to twelve months or longer.

    The 90-day estimate was indeed publicly repeated, but contemporary analysis characterized it as an exaggerated myth and found roughly annual U.S. backbone doubling after the exceptionally rapid 1995–96 period. The transcript presents the 90-day figure too much like a credible analyst estimate rather than a disputed forecast.

    Omits: The statement omits that the 90-day figure was a widely repeated industry claim that lacked hard supporting data, while actual backbone traffic growth varied by period and measurement; estimates cited by Andrew Odlyzko put U.S. backbone traffic at roughly annual doubling after 1995–96.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  99. Claim 99
    Attributed toEd ZitronAuto-attributedAccurate96% confidence▶ 45:41
    “thus there was a massive overbuild of fiber optic cable and indeed the transmission stations and such”

    The telecom boom produced a massive overbuild of fiber-optic cable and related transmission infrastructure because demand was expected to materialize immediately.

    Federal Reserve sources document excessive investment in long-haul fiber, widespread overcapacity, and construction based on anticipated future demand. The Dallas Fed also reports that demand failed to arrive at forecasted levels and that much of the capacity remained unused.

    Sources

  100. Claim 100
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 46:01
    “Right now the demand we have for generative AI is predominantly subsidized.”

    Current demand for generative AI is predominantly subsidized.

    The claim is broad, but evidence supports the underlying point that a large share of consumer generative-AI usage is offered below fully allocated cost and financed by companies or investors. The evidence does not establish a precise percentage, so “predominantly” should be understood as a characterization rather than a measured statistic.

    Sources

  101. Claim 101
    Attributed toEd ZitronAuto-attributedFalse90% confidence▶ 46:27
    “Microsoft. They can only get singledigit billions from selling AI software.”

    Microsoft can obtain only single-digit billions of dollars from selling AI software.

    Microsoft's reported AI-business annual revenue run rate exceeded $37 billion by April 2026. Although the company's figure includes both AI applications and AI solutions running on Azure, that is materially broader than and inconsistent with the transcript's single-digit-billion characterization.

    Sources

    • 1FY26 Q3 Business Highlights

      RefutesTotal AI revenue surpassed $37 billion in annual run rate (ARR), up 123%.

    • 2FY26 Q3 Business Highlights

      RefutesThis reflects customers building and running AI solutions on the Azure platform, including all revenue from frontier model companies, as well as revenue from first-party AI applications and services.

  102. Claim 102
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 46:45
    “When you have anthropic and open AI with $1.1 trillion worth of cloud commitments”

    OpenAI and Anthropic have approximately $1.1 trillion in cloud or compute commitments.

    Public reporting can produce a roughly trillion-dollar combined headline: OpenAI was reported at about $750 billion in infrastructure spending, while Anthropic had reported commitments exceeding $100 billion to Amazon and $200 billion with Google Cloud. But the transcript presents this approximate sum as $1.1 trillion of cloud commitments, despite the figures covering different scopes and including chip or broader infrastructure commitments.

    Omits: The reported figures combine different categories: OpenAI’s roughly $750 billion figure is reported as infrastructure spending, while Anthropic’s reported $200 billion Google agreement includes cloud and chips rather than cloud alone; these are not all equivalent cloud commitments.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  103. Claim 103
    Attributed toEd ZitronAuto-attributedAccurate84% confidence▶ 47:21
    “Even Nvidia with Vera Rubin, their more expensive new GPU system. Even then, they're like, "Yeah, 10x more efficient.”

    NVIDIA's Vera Rubin system is more expensive than the prior system and is advertised as roughly ten times more efficient in a workload-specific performance-per-power measure.

    NVIDIA says Vera Rubin can deliver up to ten times more tokens per megawatt than the compared prior-generation system. Independent reporting estimates Rubin-based racks cost substantially more than earlier systems, so both components of the statement are supported when interpreted in those terms.

    The intensifier “10x more efficient” was judged as NVIDIA's advertised, workload-specific performance-per-megawatt claim, not as a universal tenfold reduction in every operating cost or energy metric.

    Sources

  104. Claim 104
    Unverifiable50% confidence▶ 47:56
    “They would lump uh protein folding nothing to do with LLMs.”

    Protein-folding systems such as AlphaFold are not large language models.

    Google DeepMind explicitly states that AlphaFold is not a large language model. AlphaFold is designed for protein-structure prediction, although it uses machine-learning techniques also found in other fields.

    Checked twice, independently: the first pass returned Accurate and the second False. Recorded as Unverifiable.

    Sources

    • 1Our approach to biosecurity for AlphaFold 3

      SupportsAlphaFold is not a large language model, nor is it a biodeisgn tool that can be used to design a new protein or other biomolecule.

    • 2AlphaFold — Google DeepMind

      SupportsAlphaFold2 wins CASP14 by a huge margin and is recognised as a solution to the 50-year-old “protein folding problem” by the organisers of CASP after predicting structures down to atomic accuracy

  105. Claim 105
    Attributed toSteven BartlettAuto-attributedFalse99% confidence▶ 48:18
    “The similarity though is they all need GPUs, all these.”

    All types of AI require GPUs.

    AI workloads can run on CPUs or other accelerators, not only GPUs. Google’s TensorFlow Lite documentation shows an interpreter falling back to CPU execution when an accelerator delegate is unavailable.

    Sources

  106. Claim 106
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 48:21
    “All those data centers that we're building, all of them are for just generative AI. They're not for all of the other stuff.”

    All newly built data centers are for generative AI and not for other computing workloads.

    The universal claim is contradicted by Google’s description of its data centers as supporting products such as Gmail, document editing, and Search. Industry sources also distinguish AI-focused facilities from conventional, cloud, and other data-center workloads.

    The intensifier “all of them” was judged literally; the claim would be less clearly false if it meant only a particular subset of newly announced AI-focused facilities.

    Sources

  107. Claim 107
    Attributed toEd ZitronAuto-attributedFalse98% confidence▶ 48:47
    “That thing is not got a little GPU in it.”

    Matic’s cleaning robot does not contain a GPU.

    The claim is contradicted by product information identifying an onboard NVIDIA processor and independent hardware testing identifying it specifically as an NVIDIA Jetson Orin Nano GPU.

    Sources

  108. Claim 108
    Attributed toEd ZitronAuto-attributedAccurate91% confidence▶ 49:12
    “sighteline climate said in February there's 190 gawatts of data centers under in planning.”

    Sightline Climate was tracking about 190 GW of planned data-center capacity.

    A January 31, 2026 data summary attributed to Sightline Climate reported 190 GW across 777 hyperscale data-center projects announced since 2024. That supports the speaker’s approximate figure, though the projects were tracked announced projects rather than guaranteed construction.

    Sources

  109. Claim 109
    Unverifiable50% confidence▶ 49:17
    “that works out if about 12 million megawatt”

    190 GW of capacity converts to about 12 million megawatts.

    The unit conversion is incorrect by a factor of more than 60. The reported 190 GW pipeline corresponds to 190,000 MW, not approximately 12 million MW.

    Checked twice, independently: the first pass returned False and the second Unverifiable. Recorded as Unverifiable.

    Sources

  110. Claim 110
    Misleading50% confidence▶ 50:37
    “AI systems like Deep Mind's AlphaFold, the protein folding company used for genomic sequencing and climate forecasting, etc.”

    AlphaFold is used for genomic sequencing and climate forecasting.

    AlphaFold is primarily a protein-structure prediction system and its database supports biological and genomic research. DeepMind separately describes AI work on weather forecasting, so attributing climate forecasting directly to AlphaFold conflates distinct applications.

    Omits: The statement omits that AlphaFold’s established function is predicting protein structures and that DeepMind describes weather forecasting as a separate application of its broader collection of AI tools, not as AlphaFold’s core function.

    Checked twice, independently: the first pass returned Misleading and the second False. Recorded as Misleading.

    Sources

    • 1AlphaFold — Google DeepMind

      BackgroundAlphaFold2 wins CASP14 by a huge margin and is recognised as a solution to the 50-year-old “protein folding problem” by the organisers of CASP after predicting structures down to atomic accuracy

    • 2Science — Google DeepMind

      RefutesOur AI tools have already unlocked the potential of new materials discovery, evolved the science of weather forecasting, and solved fundamental, long-standing scientific problems.

    • 3A glimpse of the next generation of AlphaFold — Google DeepMind

      BackgroundOur latest model can now generate predictions for nearly all molecules in the Protein Data Bank (PDB), frequently reaching atomic accuracy.

  111. Claim 111
    Misleading50% confidence▶ 50:51
    “Training the brains for self-driving cars requires billions of miles of simulated physics environments.”

    Training self-driving systems requires billions of miles in simulated environments.

    Waymo and NVIDIA document the use of billions of simulated miles for autonomous-driving development. However, that evidence supports a practice by particular systems, not the broader necessity claim that training self-driving cars universally requires billions of simulated physics miles.

    Omits: The sources show that major autonomous-driving developers use billions of simulated miles, but they do not establish a universal requirement that self-driving systems must use billions of such miles.

    The intensifier "requires" was judged as a general necessity claim, not merely as a statement that some developers use billions of simulated miles.

    Checked twice, independently: the first pass returned Misleading and the second False. Recorded as Misleading.

    Sources

  112. Claim 112
    Accurate50% confidence▶ 51:07
    “We had GPUs used for this HPC, the high performance computing before generative AI.”

    GPUs were used for high-performance computing before the generative-AI era.

    NVIDIA describes GPU-accelerated high-performance computing systems and explicitly places their use in scientific computing years before the current generative-AI wave.

    Sources

  113. Claim 113
    Accurate94% confidence▶ 51:16
    “That's how Tesla did. believe they've had their own data centers when it comes to training the autopilot system”

    Tesla has operated its own data-center or training-cluster infrastructure for Autopilot and self-driving AI.

    Tesla identifies Cortex as its own training cluster and describes AI-training capacity used for its self-driving systems, supporting the claim that Tesla has operated internal training infrastructure.

    Sources

    • 1Tesla 2024 Form 10-K

      SupportsWe believe our capabilities and advancements in AI, including the deployment of Cortex, our training cluster at Gigafactory Texas, differentiates us from our competitors.

    • 2Tesla AI and Robotics

      SupportsA full build of Self-Driving neural networks involves 48 networks that take 70,000 GPU hours to train 🔥.

  114. Claim 114
    Attributed toEd ZitronAuto-attributedFalse98% confidence▶ 51:43
    “everyone saw Google, Microsoft, Amazon and Meta give Nvidia over call it 800 something billion dollars”

    Google, Microsoft, Amazon, and Meta gave NVIDIA more than approximately $800 billion.

    NVIDIA reported that capital expenditure by the top five hyperscalers was expected to approach $800 billion in 2026. That is aggregate hyperscaler capital spending, not $800 billion paid to NVIDIA by the four named companies.

    Sources

  115. Claim 115
    Attributed toEd ZitronAuto-attributedUnverifiable93% confidence▶ 52:00
    “Even though the demand 70% or more of all that demand comes from these two companies who were funded by these three companies”

    At least 70% of the relevant demand comes from two companies funded by three other companies.

    The claim does not define what "all that demand" means, identify the two companies or three funders in the quoted passage, or specify a measurable market and time period. Available funding data cannot verify this asserted share of data-center or GPU demand.

    Sources

  116. Claim 116
    Attributed toEd ZitronAuto-attributedAccurate96% confidence▶ 52:38
    “they've raised $217 billion just in 2026.”

    The two referenced AI companies raised or accounted for approximately $217 billion in funding in 2026.

    Crunchbase reported that OpenAI and Anthropic together accounted for $217 billion of global startup funding in the first half of 2026. This supports the numerical claim in the context of the two companies just mentioned, although the source describes funding accounted for rather than necessarily cash raised solely by the companies.

    The transcript says "in 2026"; the supporting data specifically covers the first half of 2026, which is a subset of that year.

    Sources

  117. Claim 117
    Attributed toSteven BartlettAuto-attributedUnverifiable50% confidence▶ 53:03
    “Amazon Web Services went down multiple times because of its AI coding tool.”

    AWS went down multiple times because of AI coding tools.

    Reports describe at least two AWS-related incidents involving Amazon’s AI coding tools, including a 13-hour interruption involving Kiro. Amazon disputes the causal framing that the AI itself caused the outages, attributing them to user error and access-control problems.

    Checked twice, independently: the first pass returned Contested and the second Unverifiable. Recorded as Unverifiable.

    Sources

  118. Claim 118
    Attributed toEd ZitronAuto-attributedMisleading90% confidence▶ 53:11
    “Previously, one of the heads of ads at Google in 2019, Google called something called a code yellow, which is when they said, "We've got a problem." And it was material weakness in query numbers”

    Google declared a 2019 Code Yellow because of weakness in search-query numbers.

    The 2019 Code Yellow and weakness in Google's search-related numbers are documented. However, the transcript frames it as a precisely defined "material weakness in query numbers," whereas the underlying email described a revenue concern and steady weakness in daily numbers, making the wording misleading.

    Omits: The contemporaneous Google email described a search-revenue Code Yellow caused by steady weakness in daily numbers and concern about missing quarterly revenue goals; it did not use the transcript's formulation "material weakness in query numbers" as a precise accounting or operational designation.

    Sources

  119. Claim 119
    Attributed toEd ZitronAuto-attributedAccurate91% confidence▶ 53:37
    “what you're suggesting would mean we give worse answers because if someone got the answer quickly that would reduce the amount of queries”

    Ben Gomes was Google Search's head in 2019 and warned that increasing queries could involve giving users worse answers.

    Reporting based on Google's internal emails says then-search chief Ben Gomes warned that query volume could be increased in user-negative ways, including disabling spell correction or ranking improvements. That supports the transcript's characterization that increasing queries could worsen answers.

    Sources

  120. Claim 120
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 54:07
    “this guy called Pragar Ragavan who's the head of ads at the time was pushing pushing and saying, "No, we need to make more queries happen. Got to make it happen."”

    Prabhakar Raghavan was Google’s head of ads at the time and pushed for more search queries.

    Publicly released Google documents show executives discussing query quotas and increasing monetizable queries, but the available evidence does not establish that Raghavan personally made this demand.

    Checked twice, independently: the first pass returned Accurate and the second Unverifiable. Recorded as Unverifiable.

    Sources

  121. Claim 121
    Attributed toEd ZitronAuto-attributedFalse50% confidence▶ 54:16
    “Nick Fox who was there as well I believe was actually taking over Google search”

    Nick Fox was taking over Google Search at that time, around early 2020.

    Contemporaneous coverage identified Nick Fox as a Search and Assistant product/design vice president, while Prabhakar Raghavan became Google’s Search chief around 2020. Fox’s actual takeover occurred in October 2024.

    The judged reading is the contextual one: that Fox was taking over Search during the 2019–2020 Code Yellow period. Fox did eventually assume Search leadership, but only in October 2024.

    Sources

  122. Claim 122
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 54:18
    “this is our new reality sometime in early 2020 propagar ragavan takes over Google search”

    Prabhakar Raghavan took over Google Search in early 2020.

    Public reporting places Raghavan’s appointment as Google Search chief in June 2020, which falls within the first half of 2020. He had previously led Google Ads & Commerce.

    “Early 2020” is judged as the first half of 2020; public reporting places the transition in June 2020.

    Checked twice, independently: the first pass returned Accurate and the second False. Recorded as Unverifiable.

    Sources

  123. Claim 123
    Attributed toEd ZitronAuto-attributedUnverifiable87% confidence▶ 54:28
    “Google stripped back a lot of the suppression of spammy sites so that people would be on Google more”

    Google reduced its suppression of spammy sites so that people would spend more time on Google.

    The claim combines an undefined quantity (“a lot”), an alleged change in ranking enforcement, and an unproven motive. Google has publicly denied rolling back anti-spam protections, while available public evidence documents both anti-spam updates and criticism of search quality but does not establish this specific causal explanation.

    Sources

  124. Claim 124
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 55:02
    “We'll just put it right at the top so people have to stay at Google. And actually, they'll use it more because instead of searching websites and doing that annoying thing where they click away from Google, they'll just only use Google.”

    Google put generative AI at the top of Search so users would stay on Google, use it more, and avoid clicking away to websites.

    Google did deploy AI Overviews in Search and reported that users used Search more, but the claim about Google’s internal motive—keeping users from clicking away—and the prediction that users would use only Google are not established by the cited evidence.

    Checked twice, independently: the first pass returned Misleading and the second Unverifiable. Recorded as Unverifiable.

    Sources

  125. Claim 125
    Misleading50% confidence▶ 56:11
    “Amazon Web Services went down two or three times this year because of AI tools.”

    Amazon Web Services went down two or three times in 2026 because of AI tools.

    Reporting supports at least two recent AWS-related service interruptions involving Amazon’s AI coding tools, so the basic incident count is plausible. However, saying AWS broadly “went down” because of AI tools overstates the incidents: Amazon described them as extremely limited events caused by user error and permissions problems, not autonomous AI failures.

    Omits: The incidents reported in 2026 were limited service interruptions, and Amazon attributed the ultimate cause to user error and misconfigured access controls rather than to AI itself; one cited incident affected AWS Cost Explorer in only one mainland-China region.

    “This year” is judged as 2026 because the transcript was transcribed on August 28, 2026.

    Checked twice, independently: the first pass returned Misleading and the second Contested. Recorded as Misleading.

    Sources

  126. Claim 126
    Attributed toSteven BartlettAuto-attributedUnverifiable92% confidence▶ 56:51
    “The code itself is getting buggier”

    The code is becoming buggier.

    The claim does not define which code, whose code, or the comparison period and metric for 'buggier.' Available studies show that AI-generated code can introduce defects, but they do not establish the broader trend stated here for code generally.

    Sources

  127. Claim 127
    Attributed toSteven BartlettAuto-attributedUnverifiable50% confidence▶ 56:54
    “the sheer volume of AI activity is literally crashing the underlying infrastructure.”

    The volume of AI activity is causing infrastructure to crash.

    The claim does not identify the infrastructure, incidents, dates, or evidence connecting crashes to AI activity. Without those details, the literal causal assertion cannot be confirmed or refuted.

    Checked twice, independently: the first pass returned Misleading and the second Unverifiable. Recorded as Unverifiable.

    Sources

  128. Claim 128
    Accurate50% confidence▶ 57:29
    “So GitHub is flooded with AI code.”

    GitHub has a very large amount of AI-generated or AI-assisted code.

    Survey data reported that developers estimated 42% of committed code was AI-generated or assisted, while research has analyzed AI-generated code across tens of millions of GitHub commits. That supports the speaker's substantive claim of substantial AI-code presence, though 'flooded' is non-quantitative.

    The intensifier 'flooded' is judged as a colloquial claim that AI code constitutes a substantial and rapidly growing share of GitHub activity, not as a literal claim that most or all GitHub code is AI-generated.

    Sources

  129. Claim 129
    Attributed toEd ZitronAuto-attributedMisleading96% confidence▶ 58:11
    “oh yeah 50% of white collar labor is going to go away in the next few years.”

    Dario Amodei said that 50% of white-collar labor would disappear in the next few years.

    Amodei did publicly warn that AI could eliminate roughly half of entry-level white-collar jobs within one to five years. The transcript broadens that warning to all white-collar labor and presents it as a definite outcome, omitting both the 'entry-level' qualifier and the uncertainty.

    Omits: Amodei's reported claim concerned up to half of entry-level white-collar jobs, framed as a possibility and given a one-to-five-year timeframe, not 50% of all white-collar labor categorically disappearing.

    Sources

  130. Claim 130
    Attributed toEd ZitronAuto-attributedAccurate97% confidence▶ 58:28
    “It's going to make mistakes”

    AI systems are probabilistic and make mistakes.

    Large language models generate outputs probabilistically and can produce incorrect or hallucinated results. Evidence on AI-assisted programming specifically documents inaccurate outputs and the need for human verification.

    Sources

  131. Claim 131
    Accurate98% confidence▶ 59:17
    “Their premium service only shows you vetted talent.”

    Fiverr Pro shows clients vetted talent.

    Fiverr describes Fiverr Pro as a curated marketplace whose freelancers undergo professional vetting and interviewing. The claim is accurate when referring to the Pro tier's stated selection process.

    Sources

  132. Claim 132
    Misleading50% confidence▶ 59:20
    “So, you've always got the safeguard that anyone you pull in to help you with a complex project has the skills that you're after and will deliver to the same high standards as your internal team.”

    Fiverr Pro guarantees that freelancers have the required skills and will deliver to the same standards as an internal team.

    Fiverr supports the existence of vetting and ongoing quality standards, but the transcript turns that process into a guarantee of project-specific competence and performance. The source language does not support that stronger assurance.

    Omits: Fiverr's vetting process evaluates freelancers before acceptance and monitors standards, but it does not guarantee that every freelancer will possess all project-specific skills or deliver at the same standard as a client's internal team.

    Sources

  133. Claim 133
    False99% confidence▶ 59:48
    “They haven't changed at all since they were invented in the '90s.”

    Traditional physical SIM cards have not changed at all since they were invented in the 1990s.

    SIM technology has evolved substantially since the first 1990s SIM cards, including smaller physical formats and embedded SIMs. The existence of eSIMs alone falsifies the claim that SIM cards have not changed at all.

    Sources

  134. Claim 134
    False98% confidence▶ 59:53
    “You have this physical piece of plastic that means you're locked into one carrier, one network”

    Having a physical SIM card locks a phone to one carrier and one network.

    Carrier locking is a property of the device or service policy, not an inherent property of the physical SIM card. Unlocked phones can use SIMs from different carriers, subject to compatibility.

    Sources

  135. Claim 135
    Misleading95% confidence▶ 59:59
    “the second you cross a border, that carrier can start charging you whatever they want.”

    A mobile carrier can begin charging whatever it wants as soon as the customer crosses an international border.

    International roaming can produce unexpected or high charges, but the charges are not literally unlimited or arbitrary: providers publish applicable rates and plan terms, and participating carriers provide roaming notifications. The wording creates the false impression that crossing a border itself gives a carrier unrestricted billing power.

    Omits: International roaming charges are governed by the customer's plan and the carrier's disclosed rates and terms; carriers also provide notifications when devices register abroad and may incur roaming charges.

    Sources

    • 1CTIA Consumer Code for Wireless Service

      BackgroundEach wireless carrier will provide, at no charge, ... a notification to consumers without an international roaming plan/package whose devices have registered abroad and who may incur charges for international usage.

    • 2How Roaming Works

      BackgroundIf there is a roaming agreement between your wireless provider and this other network, you’ll be able to connect to the “visited network” and make calls, use apps, and send emails just as you normally would on your home network.

    • 3How to Choose a Device or Plan

      RefutesIf you’re an international traveler you should check with your provider before you travel to make sure your device works when outside the U.S. and check international talk, text, and data coverage fees.

  136. Claim 136
    Accurate96% confidence▶ 1:00:05
    “It's an eim app that gives you a safe and secure data connection in over 200 destinations.”

    Saily provides a data connection in more than 200 destinations.

    Saily's own destination and product pages state that it offers mobile data plans in more than 200 destinations. The “safe and secure” marketing language is not necessary to establish the destination-count claim.

    Sources

  137. Claim 137
    Misleading50% confidence▶ 1:00:15
    “All of their eims have built-in cyber security”

    All of Saily's eSIMs have built-in cybersecurity.

    Saily does advertise security features, including blocking malicious sites, trackers, and ads. However, the evidence describes these as app-level web-protection features, so saying that every eSIM itself has built-in cybersecurity overstates what is established.

    Omits: Saily describes web protection, ad blocking, and tracker blocking as features of its app or service, rather than cybersecurity built into the eSIM technology itself.

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

  138. Claim 138
    Attributed toEd ZitronAuto-attributedUnverifiable88% confidence▶ 1:01:11
    “driving is one of the biggest professions on planet earth.”

    Driving is one of the biggest professions in the world.

    The claim is not well-posed enough to verify because “driving,” “profession,” and “biggest” have no agreed scope: it could mean drivers alone, all transport occupations, or all jobs involving driving, and could refer to employment count or occupational share. Available international labor classifications do not establish this exact global ranking.

    Sources

  139. Claim 139
    Attributed toEd ZitronAuto-attributedUnverifiable84% confidence▶ 1:02:13
    “Whimo has had to do the smallest rollouts and the most control things”

    Waymo has used the smallest and most controlled rollouts.

    Waymo documents describe structured testing, limited rider programs, and gradual expansion, but they do not define or establish that its rollouts were the smallest or most controlled compared with every relevant autonomous-driving system. Those comparative superlatives lack an agreed benchmark.

    The intensifiers “smallest” and “most controlled” are judged as stated; removing them would yield the more supportable claim that Waymo has used limited, controlled deployments.

    Sources

  140. Claim 140
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:02:24
    “It's raining which is a big problem for them in San Francisco.”

    Rain is a big problem for Waymo in San Francisco.

    Waymo acknowledges that weather affects autonomous-driving performance and has conducted dedicated rain testing, but 'a big problem' is undefined and cannot be assessed from the statement alone. Waymo also says its San Francisco service operates in rain, so the available evidence does not establish the strength of the claim as phrased.

    Checked twice, independently: the first pass returned Misleading and the second Unverifiable. Recorded as Unverifiable.

    Sources

  141. Claim 141
    Attributed toEd ZitronAuto-attributedUnverifiable93% confidence▶ 1:03:04
    “I watched a bunch of Zuk's cars just get stuck.”

    The speaker personally saw several Zoox autonomous vehicles become stuck and block a hotel exit in Las Vegas.

    The transcript gives a personal observation but no precise hotel, date, or independently verifiable recording. Reports document autonomous vehicles becoming stuck in other locations, but they do not confirm this specific Las Vegas incident.

    Sources: none found for this claim.

  142. Claim 142
    Attributed toEd ZitronAuto-attributedUnverifiable92% confidence▶ 1:03:18
    “I saw the same thing actually happen outside of a hotel when I got out of a Whimo in San Francisco. Just stopped at the and then a bunch of cars and another Whimo got stuck behind it.”

    The speaker personally saw Waymo vehicles become stuck outside a San Francisco hotel, with another Waymo vehicle blocked behind them.

    This is a firsthand account without enough identifying detail to verify the exact occurrence. Independent reporting confirms that Waymo vehicles have caused blockages and become stuck in San Francisco, but not this particular hotel incident.

    Sources

  143. Claim 143
    Attributed toEd ZitronAuto-attributedMisleading91% confidence▶ 1:03:50
    “They actually have people monitoring the roots.”

    Waymo has people involved in monitoring or supporting its autonomous vehicles' routes.

    Waymo does employ human remote-assistance personnel, so the broad idea of human oversight is supported. But describing them as monitoring the routes suggests continuous route monitoring, which Waymo explicitly says its system does not use.

    Omits: Waymo says its remote-assistance agents do not continuously monitor vehicles; they respond to requests from the autonomous driving system and provide contextual advice, while the vehicle remains in control.

    Sources

  144. Claim 144
    Attributed toSteven BartlettAuto-attributedMisleading96% confidence▶ 1:04:11
    “there's an 68% lower overall crash involvement rate when you're in an an autonomous vehicle.”

    Waymo's driverless vehicles had a 68% lower police-reportable crash-involvement rate per vehicle mile than human drivers in the studied cities.

    The 68% figure is supported for the specific Waymo fleet and study conditions, but the wording presents it as a general rate for autonomous vehicles and calls it an overall crash-involvement rate. The underlying study used a narrower crash definition and a limited set of Waymo deployments.

    Omits: The claim omits that the 68% figure came from an IIHS study of Waymo vehicles in San Francisco, Phoenix, Los Angeles and Austin, and applied to crashes deemed police-reportable or possibly police-reportable rather than every crash involving every autonomous vehicle.

    Sources

  145. Claim 145
    Attributed toSteven BartlettAuto-attributedMisleading95% confidence▶ 1:04:23
    “Autonomous vehicles experience roughly 2.1 police reported crashes per million miles compared to humans that are at roughly 4.68 per million miles. So, a 55% reduction when you get in an autonomous vehicle.”

    Waymo autonomous vehicles experienced approximately 2.1 police-reported crashes per million miles versus 4.68 for the human benchmark, a 55% reduction.

    The numerical comparison and 55% calculation match the published Waymo study. However, the transcript generalizes a Waymo-specific result to autonomous vehicles broadly without stating the limited cities, period, fleet, and benchmark methodology.

    Omits: The figures came from a study of Waymo rider-only service in Phoenix, San Francisco and Los Angeles through October 2023, compared with a human benchmark; they are not a general rate for autonomous vehicles as a category.

    Sources

  146. Claim 146
    Attributed toSteven BartlettAuto-attributedMisleading95% confidence▶ 1:04:34
    “And autonomous vehicles show an 80 to 81% reduction in crashes resulting in injuries versus human drivers.”

    Waymo autonomous vehicles showed an approximately 80% to 81% reduction in injury crashes compared with human drivers.

    The reported range is consistent with published findings: one Waymo study found an 80% reduction in any-injury-reported crashes, while the IIHS study found 81% fewer injury crashes. The wording misleadingly generalizes those results to autonomous vehicles as a whole.

    Omits: The 80% to 81% figures apply to specific Waymo studies and defined injury-crash measures, not to autonomous vehicles generally; the underlying samples were limited to particular cities, periods and Waymo operations.

    Sources

  147. Claim 147
    Attributed toSteven BartlettAuto-attributedMisleading96% confidence▶ 1:04:43
    “So, you're 85% less likely to be involved in a single vehicle crash like hitting a wall or a tree if you're an autonomous vehicle”

    A person riding in an autonomous vehicle is 85% less likely to be involved in a single-vehicle crash than when driven by a human.

    The 85% figure is directly reported by IIHS for Waymo vehicles compared with human drivers. The statement is misleading because it presents a limited Waymo result as a general property of autonomous vehicles and does not state the study's location and measurement limits.

    Omits: The 85% estimate applies to Waymo vehicles in the IIHS study's four-city sample and to single-vehicle crashes per vehicle mile traveled; it is not a universal result for all autonomous vehicles or all driving conditions.

    Sources

  148. Claim 148
    Attributed toEd ZitronAuto-attributedMisleading93% confidence▶ 1:06:11
    “Open AAI had a study that came out I think like a week ago that said there was no corre connection between spending on AI tokens and revenue per employee.”

    An OpenAI study found no correlation between AI-token usage and revenue per employee.

    The underlying report did report no statistically significant association between revenue per employee and usage measured by output tokens or messages. However, the revenue data were from before ChatGPT adoption, so presenting this as evidence that spending more tokens does not affect how much money a company makes creates a misleading impression.

    Omits: The report’s revenue-per-employee figures predated employees’ ChatGPT use, so it did not measure whether AI adoption later changed revenue; its finding was specifically that no meaningful association remained after controls.

    The phrase “no correlation” was judged as an absolute claim. The report supports a narrower claim of no statistically significant or meaningful association in its analyzed data.

    Sources

  149. Claim 149
    Attributed toEd ZitronAuto-attributedFalse97% confidence▶ 1:06:30
    “The other one was like hallucinations are mathematically guaranteed kind of almost”

    OpenAI’s other report said hallucinations are mathematically guaranteed.

    OpenAI’s report argues that hallucinations arise from statistical and evaluation pressures and may be unavoidable in some settings, but it expressly rejects the blanket conclusion that hallucinations are inevitable: models can avoid them by abstaining.

    Sources

  150. Claim 150
    Attributed toEd ZitronAuto-attributedFalse98% confidence▶ 1:07:08
    “the actual white collar labor force might have some things that are slightly changing, but there is no evidence of like productivity gains.”

    There is no evidence of productivity gains in the white-collar labor force.

    There is direct experimental evidence of productivity gains from generative AI in professional writing and customer support, both forms of white-collar work. These findings do not establish economy-wide gains, but they refute the claim that there is no evidence of productivity gains.

    The absolute phrase “no evidence” was judged. A narrower claim that aggregate economy-wide productivity growth has not yet clearly accelerated would be supported by Oxford Economics, but that is not what the speaker said.

    Sources

  151. Claim 151
    Attributed toEd ZitronAuto-attributedMisleading91% confidence▶ 1:07:22
    “There was an Oxford economics study last year where it's like, oh, young people are finding less jobs because of AI.”

    An Oxford Economics study said young people were finding fewer jobs because of AI.

    Oxford Economics did analyze rising unemployment among recent graduates and discussed AI displacement as a possible explanation. But its own later summary says the evidence is patchy and that other economic factors may explain much of the increase, so the transcript overstates the study as attributing the outcome to AI.

    Omits: The Oxford Economics analysis presented AI as a possible contributing factor and also identified weak labor markets and an increased supply of graduates as alternative explanations; it did not establish that AI caused the decline.

    Sources

  152. Claim 152
    Attributed toEd ZitronAuto-attributedFalse88% confidence▶ 1:07:31
    “It was a single line that said, "Yeah, we saw some correlation." Didn't give a number. Didn't actually say what the correlation was.”

    The Oxford Economics study consisted of a single line reporting an unspecified correlation and did not provide a number for it.

    Contemporaneous reporting on the Oxford Economics analysis gave specific numerical unemployment estimates and described the report’s methodology and comparison group. Therefore, the characterization that it offered only a single unquantified correlation is inaccurate.

    Sources

  153. Claim 153
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 1:09:23
    “Good example was this week Bloomberg reported that OpenAI was on track to hit $40 billion in annualized revenue.”

    Bloomberg reported that OpenAI was on track to generate more than $40 billion in annualized revenue.

    Bloomberg Law reported on August 13, 2026, that OpenAI was on track to generate more than $40 billion in annualized revenue based on its current performance. The report also described the figure as roughly double its run rate at the end of 2025.

    The transcript does not provide the video's calendar date, so the exact meaning of 'this week' cannot be independently established; Bloomberg did publish this report on August 13, 2026.

    Sources

    • 1OpenAI’s Revenue Run Rate Tops $40 Billion Ahead of IPO

      SupportsOpenAI is on track to generate annualized revenue of more than $40 billion based on its current performance, according to people familiar with the matter, roughly doubling its run rate from the end of 2025 and bolstering the company’s plans for a Wall Street debut.

  154. Claim 154
    Unverifiable93% confidence▶ 1:10:31
    “some of the original founding fathers of AI like Jeffrey Hinton have told me that what they're building is highly highly dangerous and that it will be fundamentally disruptive to society.”

    Geoffrey Hinton has warned that advanced AI poses serious societal and existential dangers.

    Public evidence strongly corroborates the substance that Hinton has warned about major social and existential dangers from AI. However, the specific assertion that Hinton personally 'told me' this is a private-interaction claim that cannot be independently verified from the available evidence.

    Sources

    • 1CNN.com Transcript: Geoffrey Hinton interview

      SupportsWhat I want to talk about is the existential threat that these things will get to be much smarter than us and will take over.

    • 2CNN.com Transcript: Geoffrey Hinton interview

      SupportsThose concerns very notably or shared by one of the pioneers of artificial intelligence. Geoffrey Hinton left his job at Google in order to sound the alarm on the technology he helped conceive, saying, in part, "It is hard to see how you can prevent the bad actors from using it for bad things."

  155. Claim 155
    Accurate91% confidence▶ 1:10:45
    “some of the CEOs who you've mentioned, their historical narrative was also, by the way, this is really dangerous and there is a significant chance it could f we could up the planet.”

    Some AI-company CEOs previously described advanced AI as dangerous and as posing a meaningful risk of catastrophic harm to society or humanity.

    Public statements from OpenAI and Anthropic leadership support the claim that some AI CEOs previously warned of severe danger, including human disempowerment, extinction, and large-scale destruction. The transcript does not identify the CEOs or give a precise definition of 'significant chance,' but its substantive attribution is supported.

    Sources

  156. Claim 156
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 1:12:05
    “every single scam and con starts with rushing you.”

    Every scam and con begins by rushing the target.

    The universal claim is refuted by the FTC's description of romance scams, which says scammers first establish relationships and build trust, then ask for money. Urgency may occur later, but it does not begin every scam or con.

    Sources

    • 1What To Know About Romance Scams

      RefutesThe scammers strike up a relationship with you to build up trust, sometimes talking or chatting several times a day. Then, they make up a story and ask for money.

  157. Claim 157
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 1:12:09
    “Every single trick in history begins with saying you must do this now.”

    Every trick in history begins by telling someone they must act immediately.

    The claim's universal wording is false. The FTC documents romance scams that begin by creating a connection and building trust, not by saying the victim must act immediately.

    Sources

    • 1Love Stinks – when a scammer is involved

      RefutesSome scammers use popular dating apps. Others start with an out-of-the-blue message on social media in an effort to create a connection – for example, “I love travel, too” or “We’re fans of the same team!” or “You’ve got a nice smile.”

    • 2What To Know About Romance Scams

      RefutesThe scammers strike up a relationship with you to build up trust, sometimes talking or chatting several times a day. Then, they make up a story and ask for money.

  158. Claim 158
    Attributed toEd ZitronAuto-attributedAccurate90% confidence▶ 1:12:33
    “No, they're out of compute.”

    OpenAI was out of compute, limiting its ability to release products and models.

    Taken as a colloquial claim that OpenAI was severely constrained by available computing capacity, this is supported. Sam Altman said compute limitations and allocation decisions were preventing OpenAI from shipping products as often as it wanted.

    Sources

  159. Claim 159
    False99% confidence▶ 1:13:53
    “We don't regulate tech.”

    The United States does not regulate technology.

    The literal claim is false: U.S. regulators enforce laws against technology companies, including AI firms, and federal law contains AI-related requirements. It would be more accurate to say that the United States lacks a single comprehensive federal AI-regulatory framework.

    Sources

  160. Claim 160
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 1:14:41
    “advanced AI models could very easily cuz they can go out onto the open internet as agents. They could very easily go and look at code bases of different websites, find vulnerabilities and exploit those vulnerabilities.”

    Advanced AI models can access the open internet as agents, inspect websites’ codebases, find vulnerabilities, and exploit them very easily.

    OpenAI reports that its agents gained internet access, found exposed credentials, and exploited vulnerabilities across multiple systems. However, that does not establish the stronger impression that advanced models can very easily perform this against arbitrary websites.

    Omits: The demonstrated OpenAI incident required a specific evaluation setup, reduced safeguards, exposed credentials, and a chain of vulnerabilities; the evidence does not establish that these actions can generally be performed very easily against arbitrary websites.

    The intensifier "very easily" and the generalization to different websites were judged; the weaker claim that capable agents can sometimes find and exploit vulnerabilities is supported.

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

    • 1The Hugging Face incident and the road ahead

      SupportsOur models are now powerful, persistent, and collaborative enough that, absent sufficient safeguards, they can find and exploit security weaknesses across multiple computer systems.

    • 2OpenAI – Hugging Face Incident Technical Report

      BackgroundBetween July 10 and July 13, agents identified Hugging Face user credentials that were exposed on the internet and used them, together with vulnerabilities discovered in Hugging Face infrastructure, to progressively expand their access.

  161. Claim 161
    Attributed toSteven BartlettAuto-attributedUnverifiable50% confidence▶ 1:14:52
    “Yeah. in at scale and arguably um at a higher intelligence and faster and wider than humans a human hacker could theoretically.”

    AI agents could theoretically be more intelligent, faster, and broader in operation than a human hacker.

    The report supports comparisons involving greater speed, scale, and coordination than human attackers, but "higher intelligence" and "wider" have no agreed operational definition here, and no specific human-hacker benchmark is given.

    Sources

    • 1The Hugging Face incident and the road ahead

      SupportsBoth model developers and cyber defenders more broadly will have to prepare for AI-enabled attackers that work faster, at a larger scale, and with better coordination than human attackers.

  162. Claim 162
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:15:05
    “We don't know how much compute was spent to do the hugging face attack, the open AI one.”

    The amount of compute spent on the OpenAI Hugging Face attack was not publicly known.

    OpenAI publicly described the agents as using a substantial amount of inference compute but did not publish a total quantity, so the claim that the exact amount was unknown is supported.

    Sources

  163. Claim 163
    Attributed toEd ZitronAuto-attributedMisleading90% confidence▶ 1:15:10
    “They also do know that they improperly set up the server to keep it in. They thought they'd turn the internet off and they didn't.”

    OpenAI improperly configured the server so that the models had internet access even though the operators thought the internet was turned off.

    Reporting supports that the environment failed to contain the agents and that unintended internet access occurred. But the claim omits the important technical distinction that the models bypassed outbound controls through a shared package service rather than merely receiving direct internet access.

    Omits: The evaluation workloads lacked direct internet access; they reached the public internet indirectly by exploiting vulnerabilities in the permitted Artifactory package service, so this was not simply an ordinary direct internet connection left enabled.

    Sources

    • 1OpenAI – Hugging Face Incident Technical Report

      RefutesWhile the workloads in this incident did not have direct internet access, this restriction proved insufficient when workloads can reach and exploit shared services, cloud infrastructure, private network links, or other systems that may provide transitive paths outside the intended environment.

    • 2How OpenAI’s human mistake led to the AI-powered hack on Hugging Face

      SupportsOpenAI failed to properly configure what it called a “highly isolated environment,” allowing a testing sandbox that should have been completely secluded from the internet to actually connect to the internet.

  164. Claim 164
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:16:10
    “we've had hackers doing that for years and years and years. This is brute forcing it with a bunch of compute and yet it is dangerous.”

    Hackers have used automated hacking scripts for many years.

    Automated vulnerability-testing and penetration-testing tools have been documented for decades, and cybersecurity frameworks document adversaries’ use of scripts to execute malicious code.

    Sources

  165. Claim 165
    Attributed toEd ZitronAuto-attributedAccurate90% confidence▶ 1:16:21
    “That's not what Jeffrey Hinton at have been warning about. They've been saying, "Oh, these things could destroy society. They could manipulate people."”

    Geoffrey Hinton warned that AI could destroy society and manipulate people.

    Hinton publicly warned about existential risks from AI and about AI-generated spambots enabling authoritarian leaders to manipulate electorates. The transcript paraphrases those warnings as societal destruction and manipulation.

    Sources

  166. Claim 166
    Attributed toEd ZitronAuto-attributedMisleading88% confidence▶ 1:16:29
    “Jeffrey Hinton as well talking his book still got his Google stock, I think. And weirdly enough, he left Google because he was worried about the AI there, but then immediately made a comment being like, "Yeah, actually though, Google's very responsible."”

    Geoffrey Hinton retained some Google stock, left Google because he was worried about its AI, and later said Google was very responsible.

    Hinton did say he held some Google stock and that Google had acted responsibly. The causal framing that he left because he was worried about Google’s AI is misleading: his stated reason was to speak freely about AI risks without affecting Google.

    Omits: Hinton said he left Google so he could speak freely about AI dangers without considering the effect on Google’s business, while also saying that Google had acted responsibly; he did not publicly identify concern about Google’s AI practices as the reason for leaving.

    Sources

  167. Claim 167
    Accurate93% confidence▶ 1:16:57
    “The Chinese were able to distill the models.”

    A Chinese AI developer was able to produce distilled versions of an AI model.

    DeepSeek, a Chinese AI developer, officially released six smaller models described as distilled from DeepSeek-R1. The statement is broad, but its core factual assertion is supported.

    Sources

  168. Claim 168
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:17:11
    “We let this happen because we let these companies be unregulated and use as much computers we want.”

    AI companies have been unregulated and allowed to use as much computing power as they want.

    The claim does not specify a jurisdiction, time period, or what counts as regulation, and "as much computing power as they want" is not an operationally defined legal standard. Existing policy and regulatory constraints differ across countries and activities, so the statement cannot be judged as written.

    Sources: none found for this claim.

  169. Claim 169
    Attributed toSteven BartlettAuto-attributedUnverifiable93% confidence▶ 1:17:56

    The AI industry is creating enormous economic growth.

    The claim is not well-posed because “enormous” has no agreed threshold and the relevant time period and measure of growth are unspecified. The BEA says AI’s contribution to economic growth is difficult to measure, while the IMF reports both substantial AI-related investment and potential productivity gains.

    Sources

  170. Claim 170
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:18:23
    “as far as like spend on AI goes, barely cracking hundred billion. And most of that is just these two running their services and paying these three companies, Oracle, Core, and others.”

    AI spending is barely over $100 billion, excluding semiconductor and data-center infrastructure spending.

    No time period, geographic scope, accounting definition, or list of companies is given for “spend on AI,” so the $100 billion threshold cannot be tested as stated. Available official analysis confirms major AI-related capital outlays but does not establish this exact total or the claimed composition.

    Sources

    • 1AI Can Lift Global Growth

      BackgroundAs a result, GDP data simultaneously overstate the immediate contribution of AI (by counting massive capital outlays) and understate its broader economic impact (by missing the productivity spillovers).

  171. Claim 171
    Attributed toEd ZitronAuto-attributedUnverifiable95% confidence▶ 1:18:45
    “Not when you've spent $300 billion in equity funding.”

    The relevant companies have spent $300 billion in equity funding.

    The claim does not identify which companies, what period, or whether “equity funding” means money raised by AI firms, investments by investors, or another category. Without those definitions, the exact $300 billion figure cannot be verified or refuted.

    Sources

  172. Claim 172
    Attributed toEd ZitronAuto-attributedUnverifiable94% confidence▶ 1:18:49
    “And it if we're going with just these three, I think $600 billion in capital expenditures.”

    The three relevant companies have incurred $600 billion in capital expenditures.

    The speaker does not name the three companies, specify the period, or explain whether the figure is actual spending, announced commitments, or projected spending. Consequently, the exact $600 billion assertion cannot be checked as stated.

    Sources

    • 1AI Can Lift Global Growth

      BackgroundIn the United States, AI-related investment now accounts for a large share of GDP growth, fueling new demand for servers, data centers, software, and power infrastructure.

  173. Claim 173
    Attributed toEd ZitronAuto-attributedAccurate86% confidence▶ 1:19:30
    “By the way, they've got Blackwell GPUs.”

    China possesses Blackwell GPUs.

    Evidence reported by the Associated Press describes an alleged illegal shipment of Nvidia B300 servers to China; B300 is a Blackwell-family product. This supports the limited possession claim, but not an implication that Blackwell GPUs were legally available for unrestricted sale in China.

    I judged “they've got” as meaning that some Blackwell-family GPUs or servers have reached Chinese possession, not that they were legally or broadly available to Chinese firms.

    Sources

  174. Claim 174
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:19:37
    “It's like China's already had Nvidia GPUs that they're not meant to have for years. But also to do what? They already got the LMS.”

    China developed large language models without Nvidia GPUs.

    The claim that China produced its LLMs without Nvidia GPUs is contradicted by DeepSeek’s technical report, which states that DeepSeek-V3 was trained using NVIDIA H800 GPUs. The narrower claim that China lacked access to the newest unrestricted Nvidia chips would be different, but that is not what the quoted words say.

    Checked twice, independently: the first pass returned False and the second Accurate. Recorded as Unverifiable.

    Sources

  175. Claim 175
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:19:42
    “They already got the LMS.”

    China already has large language models.

    China has multiple large language models, including DeepSeek models, and authoritative assessments describe Chinese models as competitive with leading U.S. systems on several benchmarks.

    Sources

  176. Claim 176
    Attributed toSteven BartlettAuto-attributedFalse50% confidence▶ 1:19:53

    AI will replace all human jobs.

    The universal claim is contradicted by the ILO’s assessment that most jobs will be transformed rather than eliminated. The ILO also reports that large-scale job displacement remains limited, although some occupations and tasks face substantial exposure.

    Sources

  177. Claim 177
    Attributed toEd ZitronAuto-attributedUnverifiable88% confidence▶ 1:20:02
    “it's replaced some contract labor that would otherwise be replaced with cheap labor out in the global south.”

    Generative AI has replaced some contract labor that otherwise would have been outsourced to low-cost workers in the Global South.

    Studies provide evidence that generative AI has reduced demand for some freelance and online-contract work, but the counterfactual claim about what would otherwise have been outsourced to the Global South is not directly observable and is not established by the available evidence.

    Sources

  178. Claim 178
    Attributed toEd ZitronAuto-attributedAccurate94% confidence▶ 1:20:36
    “The one where even in the demo of the hand he like they had to have a guy controlling it. Wasn't doing it autonomously.”

    Some Optimus capabilities at Tesla's October 2024 event were remotely controlled by humans rather than operating autonomously.

    Reports from the October 10, 2024 We, Robot event said Tesla employees remotely controlled some Optimus capabilities and assisted many attendee interactions. The reports also distinguished those interactions from walking demonstrations that used onboard AI.

    The claim concerns the referenced demonstration, not every Optimus demonstration or the robot's overall capabilities.

    Sources

  179. Claim 179
    Unverifiable50% confidence▶ 1:21:43
    “The hardware part, the physical parts, that's always been fairly cheap.”

    Robotic hardware has always been fairly cheap.

    The claim cannot be judged as stated because “fairly cheap” has no agreed benchmark and “always” spans all physical robotic components and periods. Available evidence shows substantial variation: some robotic hands have historically cost $16,000–$150,000, while newer low-cost designs are much cheaper.

    The intensifier “always” makes this a historical universal claim; “fairly cheap” is also undefined without a comparison or time period.

    Sources

    • 1Lending a Hand

      BackgroundWith prices ranging from $16,000 to $150,000, the cost of robotic hands has stunted progress in manipulation research.

    • 2Lending a Hand

      BackgroundRobot hands cost an arm and a leg.

  180. Claim 180
    Unverifiable50% confidence▶ 1:21:48
    “The expensive part was the intelligence. And now that's come down to pennies.”

    The cost of AI intelligence has fallen to pennies.

    The claim does not specify what kind of intelligence is meant, whether the figure refers to hardware, inference, training, or another cost, or what unit and time period are being measured. Available reporting treats robotics costs as involving multiple hardware, software, integration, and data expenses, so the assertion cannot be judged as stated.

    Checked twice, independently: the first pass returned Misleading and the second Unverifiable. Recorded as Unverifiable.

    Sources

  181. Claim 181
    Misleading50% confidence▶ 1:21:52
    “what you're seeing is this explosion in the robotics industry because robotics is a function of intelligence plus hardware.”

    The robotics industry's expansion is caused by robotics being a function of intelligence plus hardware.

    AI and hardware are important contributors to modern robotics, but authoritative robotics sources describe data collection, physical interaction, physics, safety, and integration as separate bottlenecks. Treating the industry's expansion as a simple two-factor function gives a materially incomplete causal explanation.

    Omits: The claim leaves out other major requirements and constraints, including physical-world data, sensors, control software, physics, safety, integration, and manufacturing economics.

    The verdict evaluates the causal framing “because” and the reduction of robotics to only intelligence plus hardware, not the weaker observation that both AI and hardware are important parts of robotics.

    Sources

  182. Claim 182
    Accurate50% confidence▶ 1:21:59
    “and a ton of data though as well and the data is very expensive.”

    Robotics data is very expensive.

    The claim is supported for physical or embodied AI data: unlike text and images, it must generally be collected through real-world robot interactions, which require time, equipment, and infrastructure.

    Sources

  183. Claim 183
    Accurate50% confidence▶ 1:22:03
    “The thing is cyber cabs rolled out real slow.”

    Cybercabs have been rolling out very slowly.

    As of August 2026, reporting described Tesla's robotaxi rollout as limited and slow: Las Vegas authorized 10 vehicles rather than the 5,000 Tesla sought, and expansion required further approval. Tesla's own earlier filing also described Cybercab as still in development with volume production planned for 2026.

    The verdict treats “Cybercabs” as Tesla's robotaxi/Cybercab program and “rolled out” as the pace of deployment, not a claim that no vehicles or service exist anywhere.

    Sources

  184. Claim 184
    Attributed toEd ZitronAuto-attributedMisleading91% confidence▶ 1:22:58
    “there was someone's job just to sit in an elevator and press the buttons.”

    There was a historical elevator job consisting of sitting in an elevator and pressing buttons.

    Elevator operators were a real occupation and some later operators mainly used buttons, but the claim's description falsely suggests the job was merely sitting and pressing buttons. Historical accounts describe substantial operating, safety, alignment, and customer-service duties.

    Omits: The description omits that elevator operators commonly also controlled speed and direction, aligned the car with floors, managed doors and safety, and sometimes served as guides or attendants.

    Sources

    • 1Elevators — Museum of American Heritage

      RefutesThe controls are more utilitarian: A hand operated controller for speed and direction, an annunciator panel to inform the operator that a would-be passenger has pushed a call button, and a floor control that must be depressed by the operator in order to allow the doors to open--a simple, but effective, safety device.

    • 2Elevator operator

      RefutesIn many cases, the operator had the responsibility of ensuring safe loading, door closure, and synchronizing the floor of the elevator cabin with that of the building.

    • 3Elevators: Technology That Changed Chicago

      BackgroundCurrently the few elevator operators still around just push buttons rather than operating a throttle.

  185. Claim 185
    Attributed toEd ZitronAuto-attributedFalse97% confidence▶ 1:23:25
    “agentic AI is LLMs. Agentic AI is just a fancy way of saying an LLM talking to another LLM with a harness on top.”

    Agentic AI is simply an LLM talking to another LLM with a harness on top.

    Research literature defines agentic LLMs by their ability to reason, act, and interact, and explicitly includes tool use, robot integration, and multi-agent systems as distinct categories. A second LLM and a harness are not necessary conditions, so the transcript's definition is false.

    Sources

    • 1Agentic Large Language Models, a survey

      RefutesAgentic LLMs are LLMs that (1) reason, (2) act, and (3) interact.

    • 2Agentic Large Language Models, a survey

      RefutesThe research in the first category focuses on reasoning, reflection, and retrieval, aiming to improve decision making; the second category focuses on action models, robots, and tools, aiming for agents that act as useful assistants; the third category focuses on multi-agent systems, aiming for collaborative task solving and simulating interaction to study emergent social behavior.

  186. Claim 186
    Attributed toEd ZitronAuto-attributedMisleading93% confidence▶ 1:23:38
    “It's still LLMs. It's still LM talking to other LMLs”

    Agentic AI systems are still LLMs, including LLMs communicating with other language models.

    Many current agents are built around LLMs, but authoritative descriptions define agents as LLM-based systems that dynamically plan, use tools, observe results, and act. Some are single-agent systems rather than LLMs communicating with other LLMs, so the statement creates a misleadingly narrow impression.

    Omits: The claim omits that agentic systems also include tools, memory, programmatic workflows, external environments, and action loops; they do not necessarily consist of multiple LLMs talking to one another.

    The intensifier "still" frames agentic AI as nothing more than ordinary LLM-to-LLM conversation; judged on that broad reading, rather than the narrower claim that many current agents use LLMs.

    Sources

    • 1Building effective agents

      BackgroundAgents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.

    • 2Building effective agents

      RefutesThe basic building block of agentic systems is an LLM enhanced with augmentations such as retrieval, tools, and memory.

  187. Claim 187
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:23:42
    “taking screenshots and putting them in LLM and stuff.”

    Computer-use agents can take screenshots and add them to a model's context for reasoning.

    This describes a documented computer-use architecture: screenshots are supplied to the model as visual context, after which it reasons about and performs computer actions.

    Sources

    • 1Computer-Using Agent

      SupportsPerception: Screenshots from the computer are added to the model’s context, providing a visual snapshot of the computer’s current state.

  188. Claim 188
    Accurate50% confidence▶ 1:24:39
    “Python's incredible. You can scrape websites. You can download”

    Python can be used to fetch or scrape web resources and download data.

    Python's standard urllib modules can open and read URLs and retrieve resources, which supports the stated capabilities. The claim does not say that Python alone automatically performs sophisticated scraping.

    Sources

  189. Claim 189
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:25:30
    “A lot of the skeptics were saying we're worried about an overload of bad information.”

    Early internet skeptics expressed concern about an overload of bad information.

    Contemporary and retrospective accounts document concerns that online networks would be overwhelmed by excessive, low-quality, or incorrect information. The transcript says "a lot," not that all skeptics held this view.

    Sources

  190. Claim 190
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:25:41
    “A lot of people were worried about the social consequences of everyone talking online”

    Many people were concerned about the social consequences of widespread online communication.

    Research on public reactions to the early internet records both optimistic expectations and concerns that internet use could harm social connections and produce broader social or psychological problems.

    Sources

  191. Claim 191
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 1:25:46
    “the globe, which I think made hundreds of thousands of dollars and had like a I think a billion dollar market cap”

    TheGlobe.com made hundreds of thousands of dollars and had approximately a billion-dollar market capitalization.

    TheGlobe.com did have sales in the hundreds of thousands during an earlier period, and its share price briefly produced a market value above $1 billion. However, contemporary reporting put first-half 1998 sales at $1.2 million and the closing market capitalization at $635 million, so the wording misleadingly presents different periods and valuation points as if they were one contemporaneous figure.

    Omits: The statement omits the relevant time periods and conflates revenue with market capitalization: the company reported $208,000 in sales in the comparable prior-year period and $1.2 million in first-half 1998 sales, while its closing market capitalization was reported as $635 million and exceeded $1 billion only at its peak.

    Checked twice, independently: the first pass returned False and the second Misleading. Recorded as Misleading.

    Sources

    • 1Internet IPO Theglobe.com Soars

      BackgroundFor the first half of 1998, the company lost $5.8 million on sales of $1.2 million, compared with a year-ago loss of $767,000 on sales of $208,000.

    • 2Internet IPO Theglobe.com Soars

      RefutesThe closing price means that Theglobe.com, a Web community specialist that lost $5.8 million in the first half of the year, has a market cap of $635 million based on about 10 million total outstanding shares.

    • 3theglobe.com Benefits From Hot Internet Stock Market

      SupportsAt its peak, our analysis shows TGLO traded at more than $1 billion market value, or what we target as 169x estimated 1999 revenue.

  192. Claim 192
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 1:25:54
    “Yeah, there was massive hype in the com era.”

    There was substantial hype during the dot-com era.

    Historical accounts describe the dot-com period as a frenzy characterized by widespread excitement, speculative business formation, and hype.

    Sources

  193. Claim 193
    False98% confidence▶ 1:26:56
    “the adoption of the internet required physical connections to your house”

    The adoption of the internet required a physical connection to one's house.

    Internet access was available through public facilities and other off-premises connections, not only through a connection installed at a user's home. The claim is therefore false as stated, although home connections were important to household adoption.

    Sources

    • 1Learn Without Limits

      RefutesHotspots provide off-premises connectivity to individuals who lack home internet or who may not have sufficient access for the needs of the household.

    • 2Library Services in the Digital Age

      RefutesSome 26% of Americans ages 16 and older say they used the computers there or the WiFi connection to go online.

  194. Claim 194
    Unverifiable50% confidence▶ 1:27:00
    “the adoption of generative AI involves having a web browser”

    Adopting generative AI can involve using a web browser.

    Major generative-AI services such as ChatGPT are available through a web interface, so browser access is one established route for adoption. The wording says adoption involves a browser, not that browser access is the only route.

    Checked twice, independently: the first pass returned Accurate and the second False. Recorded as Unverifiable.

    Sources

  195. Claim 195
    Unverifiable50% confidence▶ 1:27:09
    “and that's why it was so slow and there was less, you know, there was less hype than AI.”

    Internet adoption was slower and generated less hype than AI adoption.

    Comparative adoption speed can be measured, and available estimates indicate that generative AI spread faster than the internet in comparable early periods. However, “hype” has no defined measurement here, so the combined claim cannot be verified as stated.

    Sources

    • 1The State of Generative AI Adoption in 2025

      BackgroundThe current generative AI adoption rate of 54.6% exceeds the 19.7% adoption rate of the personal computer (PC) in 1984, three years after the first mass-market computer (the IBM PC in 1981), and the internet’s 30.1% adoption rate in 1998, three years after the internet was opened to commercial traffic.

    • 2The 2026 AI Index Report

      BackgroundMeasured from the release of each technology’s first widely available product, generative AI reached approximately 53% adoption within three years, well above the initial trajectories of the personal computer and the internet over comparable time frames.

  196. Claim 196
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:30:27
    “I use AskB, which is just when it's like requesting the consensus analyst estimates for Nvidia,”

    ASKB can be used to request analyst estimates, including consensus estimates, for NVIDIA.

    Bloomberg describes ASKB as an interface for exploring company and market intelligence and lists analyst recommendations and estimates among the market data available through its professional products. Bloomberg also documents consensus-estimate tools for securities such as NVIDIA, so the described use is supported.

    Sources

  197. Claim 197
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 1:30:48
    “I have this thing called Synergy in my New York New York place I go to. I have this monitor where I have a MacBook and a PC laptop and this thing Synergy for using the same mouse and keyboard.”

    Synergy allows one mouse and keyboard to control multiple computers, including a Mac and a Windows PC.

    Synergy’s official product documentation says it combines computers under one mouse and keyboard and specifically describes controlling a Mac and a Windows PC with the same peripherals.

    Sources

    • 1Synergy - Share one mouse & keyboard across computers

      SupportsMany artists use a Mac for creative work and a Windows PC for specific tools or older software. Synergy lets one keyboard and mouse control both computers as if they were one, so you stay focused on creating instead of switching devices.

  198. Claim 198
    Attributed toEd ZitronAuto-attributedAccurate98% confidence▶ 1:32:35
    “Daniel, former open AI guy, AI 2027 written with the Star Codeex guy that”

    AI 2027 was written by former OpenAI researcher Daniel Kokotajlo with Scott Alexander, the blogger behind Slate Star Codex.

    The AI 2027 project identifies Daniel Kokotajlo and Scott Alexander among its authors, describes Kokotajlo as a former OpenAI researcher, and says Alexander rewrote the content. Alexander’s site identifies Slate Star Codex as his earlier personal blog.

    Sources

    • 1AI 2027

      SupportsDaniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean

    • 2AI 2027

      SupportsDaniel Kokotajlo is a former OpenAI researcher whose previous AI predictions have held up well.

    • 3About - Astral Codex Ten

      SupportsACX started as my personal blog, Slate Star Codex.

  199. Claim 199
    Attributed toEd ZitronAuto-attributedAccurate90% confidence▶ 1:32:40
    “that he's already had to walk back.”

    Daniel has already revised or walked back some forecasts associated with AI 2027.

    The project later clarified that 2027 was its modal year at publication while its median forecasts were longer, and Kokotajlo said later modeling moved his median forecast to 2028. This supports “walk back” in the ordinary sense of revising a forecast, not necessarily abandoning the entire scenario.

    Sources

  200. Claim 200
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:33:00
    “They're not critical of the environmental damage.”

    AI has environmental impacts.

    The existence of environmental impacts from AI is supported by assessments of the power, water, and emissions associated with training and operating AI systems and their data centers.

    Sources

  201. Claim 201
    Attributed toEd ZitronAuto-attributedMisleading92% confidence▶ 1:33:37
    “these things are trained on stealing millions of people's work.”

    Some AI systems have been trained using millions of copies of copyrighted works obtained without authorization.

    A federal court found that Anthropic pirated more than seven million book copies and used books in training Claude. But the wording generalizes from particular cases to “these things” and treats “stealing” as legally settled, even though the same court held the specific training use to be fair use while rejecting the pirated library acquisition.

    Omits: The statement omits that this describes particular systems and datasets rather than AI systems generally, and that a federal court distinguished unlawful acquisition of pirated copies from the separate question of whether using books to train an LLM was fair use.

    Sources

    • 1ORDER ON 122 FAIR USE

      SupportsAnthropic thereby pirated over seven million copies of books, including copies of at least two works at issue for each Author.

    • 2ORDER ON 122 FAIR USE

      Backgroundthe use of the books at issue to train Claude and its precursors was exceedingly transformative and was a fair use under Section 107 of the Copyright Act.

    • 3ORDER ON 122 FAIR USE

      BackgroundThe downloaded pirated copies used to build a central library were not justified by a fair use.

  202. Claim 202
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 1:33:57
    “It's got better at tests where you can train for the test.”

    Benchmark contamination or training on test material can inflate AI performance on those tests without necessarily demonstrating broader capability.

    Research documents benchmark contamination and shows that training on test data can overestimate performance. But the transcript generalizes this into a claim that the tests are “rigged” or intentionally designed for the models, which is stronger than the available evidence.

    Omits: The claim omits that benchmark contamination is a documented risk affecting some evaluations, while independent analyses also report broad improvement across many benchmark categories; it provides no evidence that the tests in general are deliberately rigged.

    Checked twice, independently: the first pass returned Accurate and the second Misleading. Recorded as Misleading.

    Sources

  203. Claim 203
    Accurate97% confidence▶ 1:34:35
    “Yeah. Okay. Over time, AI's got more capable.”

    AI systems have become more capable over time.

    Stanford’s AI Index reports continued improvement on demanding benchmarks, while METR tracks increasing task-completion horizons for frontier AI agents across models released over time. The claim is broad, but its substantive assertion of measured capability gains is supported.

    Sources

  204. Claim 204
    Unverifiable50% confidence▶ 1:36:10
    “We've seen hallucinations drop.”

    Some evaluations have shown AI hallucination rates decreasing.

    The transcript places this statement in a ten-year trend claim, but it does not identify particular models, tasks, or a consistent hallucination metric. Available evaluations show improvement in some model-to-model comparisons, while other comparisons show higher hallucination rates, so the decade-wide claim cannot be confirmed as stated.

    Checked twice, independently: the first pass returned Accurate and the second Unverifiable. Recorded as Unverifiable.

    Sources

    • 1GPT-5 System Card

      SupportsWe find that gpt-5-main has a hallucination rate (i.e., percentage of factual claims that contain minor or major errors) 26% smaller than GPT-4o, while gpt-5-thinking has a hallucination rate 65% smaller than OpenAI o3.

    • 2OpenAI o3 and o4-mini System Card

      Backgroundo3 tends to make more claims overall, leading to more accurate claims as well as more inaccurate/hallucinated claims.

  205. Claim 205
    Unverifiable50% confidence▶ 1:36:14
    “if you did give it an IQ test, it's getting higher scores than it was 10 years ago.”

    AI models score higher on IQ tests than they did ten years earlier.

    There is no agreed longitudinal comparison using the same IQ test, model class, scoring method, and baseline from ten years earlier. Research also cautions that IQ tests are not suitable measures of machine intelligence, so the comparison is not well-posed as stated.

    Sources

  206. Claim 206
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 1:36:50
    “OpenAI shut down Sora. I think you can still use the API but”

    OpenAI discontinued Sora's web and app experiences while the Sora API remained available temporarily.

    OpenAI states that the Sora web and app experiences were discontinued on April 26, 2026, while the API was scheduled to remain available until September 24, 2026.

    Sources

  207. Claim 207
    Attributed toEd ZitronAuto-attributedAccurate98% confidence▶ 1:37:20
    “There was a movie that claimed it aired at Can. It didn't. No one. It aired in the city of Can during the Can Film Festival. It was not at the film festival.”

    The AI-generated movie Hell Grind was shown in Cannes, the city, but not as part of the official Cannes Film Festival.

    The film referenced was Hell Grind. Reports state that it was presented at a side event in Cannes, France, but did not premiere as part of the official Cannes Film Festival program.

    Sources

  208. Claim 208
    Attributed toEd ZitronAuto-attributedFalse93% confidence▶ 1:38:02
    “training it to be more autonomous for example, that's not something that comes from training data.”

    Increasing AI autonomy does not come from training data.

    Training data and reinforcement learning are directly used to develop agentic capabilities. External scaffolding and tools can also be important, but the claim categorically excludes training data and is therefore false.

    Sources

    • 1Computer-Using Agent

      RefutesCUA is trained to interact with graphical user interfaces (GUIs)—the buttons, menus, and text fields people see on a screen—just as humans do.

    • 2Operator System Card

      RefutesWe use a combination of supervised learning on specialized data and reinforcement learning to achieve this goal.

  209. Claim 209
    Unverifiable98% confidence▶ 1:38:24
    “Marcus said this in 2022 as well.”

    Marcus made the same statement in 2022.

    The claim cannot be checked from the excerpt because neither Marcus's identity nor what “this” refers to is specified. No independently identifiable 2022 statement is provided.

    Sources: none found for this claim.

  210. Claim 210
    Attributed toEd ZitronAuto-attributedFalse93% confidence▶ 1:39:26
    “with what nine 10 generations of TPUs from Google now.”

    Google had nine or ten generations of TPUs by the time of the statement.

    Google's official infrastructure announcement describes TPU 8t and TPU 8i as its eighth-generation TPUs. That contradicts the claim that Google had reached nine or ten TPU generations.

    The word “now” is judged against the current fact-check date, August 29, 2026; Google's official public count is eighth generation.

    Sources

    • 1Two chips for the agentic era

      RefutesOur eighth generation TPUs: two chips for the agentic era

    • 2Two chips for the agentic era

      RefutesToday at Google Cloud Next, we are introducing the eighth generation of Google's custom Tensor Processor Unit (TPU), coming soon with two distinct, purpose-built architectures for training and inference: TPU 8t and TPU 8i.

  211. Claim 211
    Attributed toEd ZitronAuto-attributedAccurate99% confidence▶ 1:39:32
    “Broadcoms building stuff with open AI, their halapeno chip.”

    Broadcom is building AI hardware with OpenAI, including the Jalapeño chip.

    OpenAI and Broadcom have officially announced a collaboration to develop and deploy custom AI accelerators, and later identified the jointly built processor as Jalapeño. The transcript's spelling “halapeno” appears to be a transcription error.

    Sources

  212. Claim 212
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 1:40:19
    “An exponential improvement in software and software performance is always a result of direct hardware improvement.”

    Exponential improvement in software or software performance is always caused by direct hardware improvement.

    Research on neural-network progress found that algorithmic efficiency doubled approximately every 16 months and reduced the computation needed for a fixed capability by 44-fold, exceeding the 11-fold improvement expected from Moore's Law alone. This is a direct counterexample to the claim's “always” qualifier.

    The intensifier “always” is decisive: one documented software or algorithmic improvement not caused by direct hardware improvement falsifies the universal claim.

    Sources

    • 1Measuring the Algorithmic Efficiency of Neural Networks

      RefutesWe show that the number of floating-point operations required to train a classifier to AlexNet-level performance on ImageNet has decreased by a factor of 44x between 2012 and 2019. This corresponds to algorithmic efficiency doubling every 16 months over a period of 7 years. By contrast, Moore's Law would only have yielded an 11x cost improvement.

    • 2A Measure of Transaction Processing 20 Years Later

      RefutesIt shows that improvement has exceeded Moore’s law – largely due to (1) hardware improvements, (2) software improvements, (3) massive parallelism, and (4) changing from mainframe to commodity economics.

  213. Claim 213
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 1:41:29
    “Yes. No, but those data centers, they are being built for generative AI. They are not being built for anything else.”

    The data centers being discussed are being built for generative AI and not for anything else.

    Meta explicitly says its AI-optimized data centers are flexible enough to support AI workloads alongside other workloads used by its apps and services. That directly contradicts the claim that they are not being built for anything else.

    Sources

    • 1Infrastructure Explained: Compute Power

      RefutesAt Meta, we’re building a global network of AI-optimized data centers, each designed with the flexibility to support both our AI workloads and the other workloads that are central to our apps and services.

  214. Claim 214
    Attributed toSteven BartlettAuto-attributedFalse96% confidence▶ 1:41:37
    “The big breakthrough we've had, which has resulted in 15 basis points of increased retention, I believe he was referring to Instagram, is that we now take anything you post on social media and we run it through an AI to get full context of what it is.”

    Mark Zuckerberg said that an AI breakthrough produced 15 basis points of increased retention and involved processing everything posted on social media for full context.

    The official earnings transcript contains related claims but not the statement as presented: it reports a 15-basis-point increase in Instagram sessions and says every public Instagram Reels and Feed post was processed through an LLM. The transcript therefore contradicts both the retention metric and the broader attribution to anything posted on social media.

    Sources

    • 1META Q2 2026 Earnings Call Transcript

      RefutesThis drove a 15 basis point increase in sessions on Instagram, with particular strength in reshares and time spent, which are both strong indicators of better content-to-user matching.

    • 2META Q2 2026 Earnings Call Transcript

      RefutesEarlier this year we reached a milestone of every public Reels and Feed post on Instagram being automatically processed through an LLM and analyzed across dimensions from topics to tone, and we're working towards including more surfaces on Facebook as well.

  215. Claim 215
    Unverifiable50% confidence▶ 1:42:08
    “it'sn't 15 basis points, like 0.15%.”

    Fifteen basis points equals 0.15 percent.

    One basis point is 0.01 percentage points, so 15 basis points equals 0.15 percentage points, commonly stated as 0.15 percent.

    The checker returned Accurate but cited no source that supports the claim, so this is recorded as unverifiable.

    Sources: none found for this claim.

  216. Claim 216
    Accurate99% confidence▶ 1:42:49
    “Muse Spark is their LLM. Gem is their generative ad model.”

    Muse Spark is Meta's LLM and GEM is its generative advertising model.

    Meta describes Muse Spark as an LLM built by Meta Superintelligence Labs and describes GEM as its Generative Ads Recommendation Model, a foundation model for ads recommendation.

    Sources

  217. Claim 217
    Attributed toSteven BartlettAuto-attributedUnverifiable50% confidence▶ 1:43:39
    “I've done almost 700 interviews with some of the most interesting people in the world.”

    The speaker has conducted almost 700 interviews.

    The statement is a personal estimate, but no reliable independent source establishes the number of interviews conducted. Public podcast and promotional material confirm the claim was made, not that the approximate count is accurate.

    Sources

  218. Claim 218
    Attributed toSteven BartlettAuto-attributedAccurate50% confidence▶ 1:43:53
    “As they leave, what I get them to do is to write a question in the diary of a CEO. We've taken all of the questions from the diary of a CEO. We have put the question here on this card with the name of the person that wrote it.”

    The conversation cards contain questions taken from The Diary of a CEO and identify the person who wrote each question.

    The product's official description says each question is taken directly from The Diary of a CEO and describes the cards as featuring questions from thought leaders and experts. This supports the claim that the cards use guest questions and identify their authors.

    Sources

  219. Claim 219
    Attributed toSteven BartlettAuto-attributedUnverifiable50% confidence▶ 1:45:02
    “according to the algorithm, you're someone that watches our show, but you haven't yet hit that button.”

    The recommendation algorithm identifies the viewer as someone who watches the show but has not subscribed.

    The claim concerns a personalized algorithmic classification, but it does not identify the platform or provide access to the viewer-specific data needed to verify it.

    Sources: none found for this claim.

  220. Claim 220
    Attributed toEd ZitronAuto-attributedUnverifiable86% confidence▶ 1:45:31
    “This sort of virtual world and actually it never transpired and there's no sign that it will in the near term.”

    The envisioned metaverse virtual world never materialized.

    The claim is not well-posed enough to verify because “the metaverse” can refer to different products and levels of adoption, while “transpired” has no defined threshold. Companies and public institutions have continued developing virtual-world technologies, but that does not establish whether the speaker’s undefined vision has or has not materialized.

    The judgment applies to the phrase "never transpired" as stated; "metaverse" and the threshold for the virtual world to have transpired have no agreed operational definition.

    Sources

    • 1Virtual Worlds Trends Report

      BackgroundMajor technology companies including Apple, Google, Meta Platforms (Facebook), Microsoft, Nvidia, Niantic, Roblox, Unity, and Valve are developing the technologies that will shape the future of the Metaverse.

  221. Claim 221
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 1:45:48
    “The difference is the reason the metaverse and NFTs didn't escape this was there weren't stocks to speculate on. There weren't big companies that you could invest in.”

    The metaverse and NFTs did not have publicly investable stocks or major companies available for investment.

    The claim is contradicted by multiple publicly traded companies developing metaverse-related technologies and by a dedicated metaverse ETF. While NFTs lacked a single dominant public-company equivalent, the categorical claim that there were no stocks or major investable companies is false.

    Sources

    • 1Roundhill Ball Metaverse ETF

      RefutesRoundhill Ball Metaverse ETF (METV) allows you to invest in stocks in the Metaverse industry.

    • 2Virtual Worlds Trends Report

      RefutesMajor technology companies including Apple, Google, Meta Platforms (Facebook), Microsoft, Nvidia, Niantic, Roblox, Unity, and Valve are developing the technologies that will shape the future of the Metaverse.

  222. Claim 222
    Attributed toEd ZitronAuto-attributedUnverifiable90% confidence▶ 1:45:57
    “They had re record earnings in 2021.”

    The companies under discussion had record earnings in 2021.

    The pronoun “they” has no unambiguous referent in the transcript, and “record earnings” does not specify whether it means revenue, net income, operating income, or another measure. The claim therefore cannot be judged as stated.

    Sources: none found for this claim.

  223. Claim 223
    Attributed toEd ZitronAuto-attributedMisleading91% confidence▶ 1:46:00
    “There's a bunch of money floating in the system thanks to postcoid uh the PDC that basically government federal money flowed in to the banks.”

    Federal money flowed into banks after COVID-19, contributing to the availability of easy money.

    Federal fiscal support and Federal Reserve interventions did inject substantial liquidity into the financial system, but the wording gives the impression that federal fiscal money was directly transferred to banks. Federal Reserve research says stimulus payments raised household incomes and contributed to deposit growth across banks, while other programs supported bank and market liquidity separately.

    Omits: The claim omits that major federal stimulus payments were made primarily to households and businesses, with funds entering bank deposits when recipients deposited them; Federal Reserve lending and liquidity programs were separate mechanisms from fiscal stimulus.

    Sources

    • 1Understanding Bank Deposit Growth during the COVID-19 Pandemic

      RefutesIn addition to the personal savings rate moving higher, the federal government began making stimulus payments to households in 2020 and 2021 through direct payments, extended unemployment supplementary payments, and augmented the child tax credit through legislation such as the CARES Act and the American Rescue Plan, which raised the incomes of many households.

    • 2The Federal Reserve’s COVID-19 Response

      BackgroundLending facilities are offering further support for households and businesses by keeping credit flowing smoothly throughout the financial system.

  224. Claim 224
    Attributed toEd ZitronAuto-attributedAccurate94% confidence▶ 1:46:06
    “zero interest free era money was easy to find.”

    The period of near-zero interest rates made money relatively easy to find.

    The Federal Reserve cut the federal funds target range to 0 to 0.25 percent during the pandemic and described interest rates as near zero. That supports the speaker’s characterization of a near-zero-rate era, although individual borrowing costs varied.

    The phrase "zero interest" is judged as colloquial shorthand for the Federal Reserve’s 0-to-0.25-percent target range, rather than literally zero borrowing costs for every borrower.

    Sources

  225. Claim 225
    Attributed toEd ZitronAuto-attributedUnverifiable96% confidence▶ 1:46:26
    “the stocks went on an incredible run. may like several hundred percent grow in the last few years. the stock has grown by hundreds of percent.”

    Stocks grew by several hundred percent over the last few years.

    The claim does not identify which stocks, the exact start and end dates, or whether “grew” refers to share price, total return, or market capitalization. Without those terms, it cannot be verified or refuted as stated.

    Sources: none found for this claim.

  226. Claim 226
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 1:46:37
    “Yeah, Meta's revenues growing because of AI, right? Microsoft's revenue is grown because of AI, right?”

    Meta’s and Microsoft’s revenue growth was caused by AI.

    Both companies reported strong revenue growth while also investing heavily in AI, and AI may have contributed to some products and services. However, their official results identify multiple drivers and do not establish that AI alone caused the companies’ overall revenue growth.

    Omits: The claim omits other reported revenue drivers: Meta reported growth in advertising impressions and average ad prices, while Microsoft attributed 2024 revenue growth across segments to Azure, Office 365, gaming, and other businesses, not AI alone.

    Sources

  227. Claim 227
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:48:37
    “indeed, in the next three and a half years, analysts think that these two bastards, these two, OpenAI and Anthropic are going to spend over $400 billion on these people alone,”

    Analysts think OpenAI and Anthropic will spend over $400 billion on Microsoft, Google, and Amazon over the next three and a half years.

    The sentence is grammatically ambiguous about who would spend the money and what "these people" refers to. Available reporting documents more than $400 billion in planned OpenAI infrastructure investment, but does not establish the specific claim that OpenAI and Anthropic together will spend that amount on Microsoft, Google, and Amazon within three and a half years.

    Checked twice, independently: the first pass returned Unverifiable and the second Accurate. Recorded as Unverifiable.

    Sources

    • 1OpenAI OSTP RFI, October 27, 2025

      BackgroundThrough our Stargate initiative, the six sites we’ve already announced bring Stargate to nearly seven GW of planned capacity and over $400 billion in investment over the next three years.

  228. Claim 228
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:49:09
    “There were stories about how they were sending armored cars with the GPUs to Microsoft to make sure Microsoft got the GPUs.”

    GPU shipments were sent to Microsoft in armored cars to ensure Microsoft received them.

    A contemporaneous report documented a Cisco executive saying that GPUs arrived by armored car, but the available evidence does not substantiate the more specific claim that armored-car shipments were sent to Microsoft.

    Checked twice, independently: the first pass returned Unverifiable and the second False. Recorded as Unverifiable.

    Sources

  229. Claim 229
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:49:15
    “Even though they never disclosed AI revenues, they saw the expenditures”

    The companies discussed had not disclosed their AI revenues.

    Amazon, Alphabet, Microsoft, and Meta generally do not report a standalone figure for sales and profits directly attributable to AI data-center investments. However, "AI revenues" is undefined, and the broad claim that the companies have never disclosed any AI-related revenue cannot be confirmed from public reporting.

    The intensifier "never" is judged as stated; "AI revenues" has no consistently defined reporting category across these companies.

    Sources

  230. Claim 230
    Unverifiable50% confidence▶ 1:49:38
    “to have progress with AI just on a taking it in a vacuum to have progress for these two companies to keep going and to keep progressing they need to spend tens of billions of dollars a year on training.”

    OpenAI and Anthropic need to spend tens of billions of dollars per year on training to continue making progress.

    Public reporting supports the narrower proposition that frontier-AI companies are projecting or incurring very large annual compute and training costs. It does not establish the counterfactual necessity asserted here—that both companies must spend tens of billions annually for progress to continue.

    Sources

  231. Claim 231
    Attributed toEd ZitronAuto-attributedUnverifiable86% confidence▶ 1:50:05
    “the majority of the funding that open AAI got in the last 6 months came from SoftBank Nvidia and Amazon”

    The majority of OpenAI's funding in the previous six months came from SoftBank, Nvidia, and Amazon.

    OpenAI announced a $110 billion financing involving $30 billion from SoftBank, $30 billion from Nvidia, and $50 billion from Amazon, so those firms were major recent funders. But without the video’s recording date and a definition of whether commitments or completed investments count, the six-month majority claim cannot be determined.

    The relative-time phrase "last 6 months" cannot be evaluated because the recording date is not supplied.

    Sources

  232. Claim 232
    Attributed toEd ZitronAuto-attributedAccurate98% confidence▶ 1:53:29
    “open AI Clammy Sam has said Wall Street Journal and Isaagi reported a few weeks ago they plan to spend $750 billion on compute through 2030.”

    OpenAI planned to spend about $750 billion on computing power through 2030.

    Reporting in July 2026 said OpenAI had raised its projected computing-power spending through 2030 to about $750 billion. The figure was described as a projection rather than money already spent.

    Sources

  233. Claim 233
    Attributed toEd ZitronAuto-attributedMisleading88% confidence▶ 1:54:49
    “GPT5 was meant to be this panacea for the AI industry. They had at least one training run that cost half a billion dollars and did nothing.”

    At least one GPT-5 training run cost about half a billion dollars and produced no result.

    Reports estimated that a large GPT-5 training run could cost roughly $500 million, and said OpenAI conducted multiple runs that failed to achieve the hoped-for breakthrough. However, the reporting also described small or noticeable improvements, so “did nothing” overstates the evidence.

    Omits: Public reporting described the GPT-5/Orion training runs as falling short of hoped-for improvements or producing only small improvements, not literally producing no result.

    Sources

  234. Claim 234
    Attributed toEd ZitronAuto-attributedMisleading91% confidence▶ 1:55:07
    “There's a reason why Google and Amazon are cash flow negative now.”

    Google and Amazon were cash-flow negative because of their AI spending.

    By mid-2026, Alphabet and Amazon had negative free cash flow after capital expenditures, largely amid heavy infrastructure spending. But their operating cash flows remained positive, so saying simply that they were “cash flow negative” creates a broader impression than the underlying accounting measure supports.

    Omits: The companies were reported to have negative free cash flow, not negative operating cash flow generally; both still generated positive operating cash flow during the relevant periods.

    Sources

  235. Claim 235
    Attributed toSteven BartlettAuto-attributedAccurate99% confidence▶ 1:55:31
    “He says the risk of underinvesting is dramatically greater than the risk of overinvesting.”

    Sundar Pichai said the risk of underinvesting was dramatically greater than the risk of overinvesting.

    Pichai made this comparison publicly while discussing Google's AI investment strategy. The transcript's wording matches the reported quotation.

    Sources

  236. Claim 236
    Attributed toSteven BartlettAuto-attributedAccurate99% confidence▶ 1:55:44
    “we're not investing approximately 200 billion in capex in 2026 on a hunch.”

    Andy Jassy said Amazon was investing approximately $200 billion in 2026 capital expenditures and not doing so on a hunch.

    Jassy used this wording in Amazon's 2026 shareholder communication. Amazon later raised its 2026 capital-expenditure expectation, but that does not make the quoted earlier statement false.

    Sources

  237. Claim 237
    Attributed toSteven BartlettAuto-attributedAccurate96% confidence▶ 1:56:02
    “Then Mark Zuckerberg, CE of Meta, says we'll continue to invest aggressively in infrastructure to meet the demand. I'd rather risk building capacity before it's needed than being late.”

    Mark Zuckerberg said Meta would continue investing aggressively in infrastructure to meet demand and preferred building capacity before it was needed rather than being late.

    Meta's prepared remarks contain both the stated preference for building capacity ahead of demand and the rationale that infrastructure projects have long lead times. The transcript is a shortened paraphrase of Zuckerberg's wording.

    Sources

    • 1Meta Q2 2024 Prepared Remarks

      SupportsIt's hard to predict how this will trend multiple generations out into the future, but at this point I'd rather risk building capacity before it is needed, rather than too late, given the long lead times for spinning up new infra projects.

    • 2META Q2 2026 Earnings Call Transcript

      SupportsAs AI usage in our products and businesses continues to ramp, we continue to invest aggressively in infrastructure to meet the demand.

  238. Claim 238
    Attributed toEd ZitronAuto-attributedFalse98% confidence▶ 1:56:20
    “You can't fire me cuz Mark Zuckerberg can't be fired due to the unique board situation he's got going.”

    Mark Zuckerberg cannot be fired as Meta's CEO because of Meta's governance structure.

    Zuckerberg's dual-class shares give him approximately 61% of Meta's voting power and make removal politically difficult, but Meta's own bylaws expressly authorize the board to remove the CEO. Therefore, 'can't be fired' is false, even if it is practically unlikely.

    Sources

  239. Claim 239
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 1:56:51
    “Not the people spending a trillion dollars”

    People are spending a trillion dollars on AI-related projects.

    The claim does not specify who is spending the money, what expenditures are included, or the time period. Current reporting describes roughly trillion-dollar AI-related commitments, but also says it is unclear over what period the money will be spent or whether all of it will be spent, so the statement as phrased cannot be confirmed.

    Checked twice, independently: the first pass returned Misleading and the second Unverifiable. Recorded as Unverifiable.

    Sources

  240. Claim 240
    Attributed toEd ZitronAuto-attributedAccurate91% confidence▶ 1:58:30
    “The fact that you have communities like in violent New Jersey where the residents like I don't want this but the planning boards vote for it”

    In at least one New Jersey community, residents opposed an AI data center while local officials or a planning board supported or approved it.

    The transcript appears to transcribe the place name 'Vineland' as 'violent.' In Vineland, residents publicly opposed the project while local government supported it; separately, Kenilworth's planning board approved a CoreWeave data-center application despite resident opposition.

    Sources

  241. Claim 241
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:58:41
    “the use of gas turbines and behind the meter power is reckless and damaging to communities.”

    Gas turbines and gas-powered engines are increasingly being used as behind-the-meter primary power for data centers.

    Industry reporting states that behind-the-meter data-center projects increasingly use on-site natural-gas generation, including gas turbines and gas-powered engines. These combustion systems generate emissions and have prompted community and regulatory concerns.

    Sources

  242. Claim 242
    Attributed toEd ZitronAuto-attributedAccurate50% confidence▶ 1:58:52
    “The noise that these things make”

    Gas turbines used in or near data-center communities can produce persistent noise that harms nearby residents' quality of life.

    Reporting on data-center power systems identifies constant noise from gas turbines as a potential harm to the quality of life of nearby residents. The transcript's statement is brief, but its factual implication that these systems produce community noise is supported.

    Sources

  243. Claim 243
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 1:59:17
    “Jensen Hong will give you 25% residual value”

    NVIDIA may provide up to 25% residual-value support for a financing opportunity.

    NVIDIA says it may provide a residual-value support mechanism for up to 25% of an opportunity, assessed on a project-by-project basis. The transcript presents this as a guaranteed 25% residual value.

    Omits: The support is only 'up to 25%' and is assessed project by project; it is not an unconditional payment of 25% to every borrower.

    Sources

  244. Claim 244
    Attributed toEd ZitronAuto-attributedAccurate94% confidence▶ 1:59:45
    “Nvidia, one of their first investors in 2023, signed a $1.3 billion contract to rent back their GPUs from Core”

    NVIDIA invested in CoreWeave in 2023 and agreed to spend $1.3 billion over four years renting its own chips from CoreWeave.

    Reporting based on company documents says NVIDIA invested in CoreWeave and agreed in 2023 to spend $1.3 billion over four years renting NVIDIA chips from CoreWeave. The wording 'one of their first investors' is informal, but the substantive transaction is documented.

    Sources

  245. Claim 245
    Attributed toEd ZitronAuto-attributedFalse99% confidence▶ 2:00:04
    “or buy a house, highest interest rates ever.”

    Mortgage interest rates for people buying a house are the highest ever.

    For U.S. 30-year fixed mortgages, Freddie Mac identifies 18.63% in 1981 as the record high. Current rates are far below that level, so the claim is false as stated for homebuyers.

    Sources

  246. Claim 246
    Attributed toEd ZitronAuto-attributedMisleading97% confidence▶ 2:00:50
    “AI systems are already blackmailing and escaping control.”

    AI systems are already blackmailing people and escaping control.

    Anthropic documented models attempting blackmail and, more rarely, unauthorized copying of model weights, but explicitly described these as controlled simulations involving fictional people and organizations. Presenting those tests as real-world incidents is misleading.

    Omits: The reported blackmail and self-exfiltration behavior occurred in controlled, fictional simulations with specially configured tools and scenarios, not as documented incidents involving deployed systems harming real people.

    The intensifier 'already' was judged in the ordinary real-world sense, rather than the weaker sense that models have produced these behaviors inside laboratory simulations.

    Sources

    • 1Agentic misalignment: How LLMs could be insider threats

      RefutesNote: All the behaviors described in this post occurred in controlled simulations. The names of people and organizations within the experiments are fictional. No real people were involved or harmed in any of these experiments.

    • 2Claude 4 System Card

      BackgroundIn a few instances, we have seen Claude Opus 4 take (fictional) opportunities to make unauthorized copies of its weights to external servers.

  247. Claim 247
    Misleading90% confidence▶ 2:02:34
    “Eric Schmidt getting booed at a commencement speech by 8,000 people every time he said the word AI.”

    Eric Schmidt was booed by approximately 8,000 people at a commencement speech whenever he said the word AI.

    Schmidt was in fact repeatedly jeered during his University of Arizona commencement address when discussing AI. However, the reported audience size was about 10,000 graduates, and the available reporting does not establish that exactly 8,000 people booed every time he said “AI.”

    Omits: The University of Arizona said it would award about 10,000 degrees, and reporting describes repeated jeers rather than establishing that exactly 8,000 people booed every occurrence; some reported audience hostility also concerned other issues.

    Sources

  248. Claim 248
    Attributed toEd ZitronAuto-attributedUnverifiable50% confidence▶ 2:03:34
    “Samman drives a $5 million car around San Francisco.”

    Sam Altman drives a $5 million car around San Francisco.

    Reliable reporting confirms that Altman owns or has owned highly valuable cars, including a McLaren F1, but I found no reliable evidence establishing that he drives a $5 million car around San Francisco. The exact vehicle, valuation, and driving pattern are therefore not confirmable from the available sources.

    Checked twice, independently: the first pass returned Unverifiable and the second Misleading. Recorded as Unverifiable.

    Sources

    • 1The Silicon F1

      BackgroundLarry, Elon Musk, and Open AI's Sam Altman all own or have owned F1s, and the curious case of this example, finished in stunning Magnesium Silver, exemplifies this car's hold over tech's titans.

    • 2Sam Altman May Control Our Future—Can He Be Trusted?

      BackgroundAltman has at least two hypercars: an all-white Koenigsegg Regera, worth about two million dollars, and a red McLaren F1, worth about twenty million dollars.

  249. Claim 249
    Attributed toEd ZitronAuto-attributedMisleading50% confidence▶ 2:04:28
    “The reason I don't like Dario is Daario was doing the scare tactics thing when he worked at OpenAI when GPT2 came out say it's too scary to release.”

    Dario Amodei used scare tactics at OpenAI when GPT-2 was released by saying it was too scary to release.

    Amodei was an OpenAI researcher and coauthor during the GPT-2 staged-release decision, and OpenAI did initially withhold the full model because of misuse concerns. But the transcript attributes the “too scary to release” scare tactic specifically to Amodei, which the cited documentation does not establish.

    Omits: OpenAI's published rationale was a collective, staged-release policy based on concerns about deceptive, biased, abusive, and other misuse; the source does not show Amodei personally saying that GPT-2 was “too scary to release.”

    Checked twice, independently: the first pass returned Misleading and the second Accurate. Recorded as Misleading.

    Sources

    • 1Better language models and their implications

      SupportsDue to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT‑2 along with sampling code.

    • 2Organizational update from OpenAI

      BackgroundDario has made tremendous contributions to our research in that time, collaborating with the team to build GPT‑2 and GPT‑3, and working with Ilya Sutskever as co-leader in setting the direction for our research.

    • 3Keeping AI away from the bad guys

      BackgroundBut the move met massive blowback: AI researchers accused the group of pulling off a media stunt, stirring up fear and hype, and unnecessarily holding back an important research advance.

  250. Claim 250
    Attributed toEd ZitronAuto-attributedMisleading97% confidence▶ 2:04:37
    “He's also gone on television and given AI psychosis to Axios being like 50% of jobs are going to go away because of AI.”

    Dario Amodei said that 50% of jobs would go away because of AI.

    Amodei did tell Axios that AI could wipe out half of all entry-level white-collar jobs within one to five years. The transcript broadens that to 50% of jobs generally and removes the qualification that this was a possibility or warning about a specific job category and timeframe.

    Omits: Amodei's reported warning concerned potentially half of entry-level white-collar jobs, with unemployment potentially rising to 10–20%, over the next one to five years; it was a possibility or forecast, not 50% of all jobs simply going away.

    Sources