Claims checked against evidence, page 11 of 29
Claims by Gary Marcus and Ed Zitron. Deadlines are judged as of 2026-10-29.
Claims checked against evidence (2859)
Each of these claims was checked against dated evidence.
Page 11 of 29: claims 1001 to 1100 of 2859, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: SupportedSaid
Gary Marcus Observation
“no robust solution to hallucinations”
Our reading: There is no robust solution to LLM hallucinations.
No horizon Evidence as of
Limitation: 'Robust' has no threshold; partial mitigations exist and rates are falling.
Assessment: UnresolvedSaid
Gary Marcus Conditional forecast
“If and when the whole generative AI edifice comes crashing down, it will be in no small part because too many otherwise smart investors read too many graphs—from “LLM scaling” to usage statistics—too naively, assuming that things had been going up for a while would continue to go up at the same pace, indefinitely.”
Our reading: If generative AI collapses, a large part of the cause will be investors naively extrapolating scaling and usage trends.
No deadline stated Evidence as of
Only if: if and when the generative AI edifice comes crashing down
Assessment: UnresolvedSaid
Gary Marcus Causal claim
“The problem, as always, lies in generalizing outside the training distribution.”
Our reading: The problem with vibe coding lies, as always, in generalizing outside the training distribution.
No horizon Evidence as of
Limitation: Qualitative reports do not measure causal effects.
Source: Is vibe coding dying?
Assessment: UnresolvedSaid
Gary Marcus Forecast
“Remember how in October and in March I told you that vibe coding — in the sense of amateurs using large language models to write code to “build products that would have previously required teams of engineers” — would never be remotely reliable?”
Our reading: Vibe coding by amateurs using LLMs to build products that previously required engineering teams will never be remotely reliable.
No deadline stated Evidence as of
Limitation: Most of the evidence is anecdotal or comes from a vendor's CEO.
Source: Is vibe coding dying?
Assessment: SupportedSaid
Gary Marcus Observation
“All that really happened is that GPT-5 (oversimplifying slightly) crawled the web for solutions to already-solved problems.”
Our reading: GPT-5 did not solve open Erdős problems; it essentially located existing solutions to already-solved problems on the web.
No horizon Evidence as of
Limitation: The result depends on treating the crawler wording as the acknowledged simplification.
Source: Erdosgate
Assessment: UnresolvedSaid
Gary Marcus Observation
“I found that the claim of “solution” was pretty misleading, as I summed up in a tweet, and they refused to correct their misleading presentation:”
Our reading: In 2019 OpenAI's claim that its robot 'solved' the Rubik's cube was misleading, and OpenAI refused to correct it.
No horizon Evidence as of
Limitation: Disagreement over wording does not by itself prove a factual misrepresentation.
Source: Erdosgate
Assessment: UnresolvedSaid
Ed Zitron Forecast
“I believe that we’re going to see multiple price increases on AI-native companies in the next few months.”
Our reading: Within a few months of 2025-10-20, AI-native companies (products built on top of AI models) will face multiple price increases.
No deadline stated Evidence as of
Limitation: This record does not assess whether model providers changed their prices after October 2025.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Observation
“Furthermore, I can confirm that through September, Anthropic has spent more than 100% of its estimated revenue (based on reporting in the last year) on Amazon Web Services, spending $2.66 billion on compute on an estimated $2.55 billion in revenue.”
Our reading: Through September 2025 Anthropic spent $2.66 billion on AWS against an estimated $2.55 billion in revenue, more than 100% of its revenue.
No horizon Evidence as of
Limitation: The packet contains no independent verification of January–September 2025 net AWS payments, including credits, discounts and prepayments.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Observation
“Based on what I have been party to, the more successful Anthropic becomes, the more its services cost.”
Our reading: Based on what he has been party to, the more successful Anthropic becomes, the more its services cost.
No horizon Evidence as of
Limitation: The packet lacks independent, comparable AWS cash-payment records and matched realized revenue for the observation period.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Observation
“I’m gonna be nice here and say that Anthropic made $600 million in 2024 — the higher end of The Information’s reporting — meaning that it spent around 226% of its revenue ($1.359 billion) on Amazon Web Services.”
Our reading: Assuming $600 million in revenue, Anthropic spent about 226% of its 2024 revenue on AWS.
No horizon Evidence as of
Limitation: This assessment preserves the stated revenue assumption.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Forecast
“Any product built on top of an AI model that shows any kind of success can be cloned immediately by OpenAI and Anthropic, and I believe that we’re going to see multiple price increases on AI-native companies in the next few months.”
Our reading: AI-native companies will see multiple price increases from model providers in the next few months.
Deadline passed Evidence as of
Limitation: The result depends on what counts as a price increase and how long the next few months lasts.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Observation
“Although Cursor obtains the majority of its compute from Anthropic — with AWS contributing a relatively small amount, and likely also taking care of other parts of its business — the data seen reveals an overall direction of travel, where the costs of compute only keep on going up.”
Our reading: The AWS billing data seen for Anthropic and Cursor shows an overall direction of travel in which compute costs only keep going up.
No horizon Evidence as of
Limitation: An upward pattern within the post's own numbers would check internal consistency, not independently verify the reported spending.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Observation
“While I don’t know about prepayment for services, I can confirm from a source with direct knowledge of billing that Anthropic spent $1.35 billion on Amazon Web Services in 2024, and has already spent $2.66 billion on Amazon Web Services through the end of September.”
Our reading: Anthropic spent $1.35 billion on AWS in 2024 and $2.66 billion through September 2025.
No horizon Evidence as of
Limitation: The underlying billing records are unavailable.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Additionally, Cursor’s Amazon Web Services bills more than doubled from $6.2 million in May 2025 to $12.6 million in June 2025, exacerbating a cash crunch that began when Anthropic introduced Priority Service Tiers, an aggressive rent-seeking measure that begun what I call the Subprime AI Crisis, where model providers begin jacking up the prices on their previously subsidized rates.”
Our reading: Cursor's AWS bill more than doubled from May to June 2025, worsening a cash crunch that began when Anthropic introduced rent-seeking Priority Service Tiers, which started the Subprime AI Crisis.
No horizon Evidence as of
Limitation: Company explanations can omit other contributing causes.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Conditional forecast
“In any case, these costs seem to increase with the amount of money Anthropic makes, meaning that the current pricing of both subscriptions and API access seems unprofitable, and must increase dramatically — from my calculations, a 100% price increase might work, but good luck retaining every single customer and their customers too! — for this company to ever become sustainable.”
Our reading: Because Anthropic's costs seem to rise with its revenue, its subscription and API pricing seems unprofitable and must increase dramatically for the company ever to become sustainable; by his calculations a 100% increase might work.
No deadline stated Evidence as of
Only if: for this company to ever become sustainable
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Observation
“The cost of inference is clearly increasing for customers, but based on its escalating monthly costs, the cost of inference appears to be high for Anthropic too, though it’s impossible to tell how much of its compute is based on training versus running inference.”
Our reading: Inference costs are rising for customers and appear high for Anthropic as well.
No horizon Evidence as of
Limitation: Flat token tariffs do not rule out higher bills caused by reduced allowances or changed workloads.
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Ed Zitron Conditional forecast
“The ultimate problem is that there really are no winners in this situation. If Anthropic kills Cursor through aggressive rent-seeking, that directly eats into its own revenues. If Anthropic lets Cursor succeed, that’s revenue, but it’s also clearly unprofitable revenue.”
Our reading: There are no winners: if Anthropic kills Cursor through aggressive rent-seeking, that eats directly into its own revenues, and if it lets Cursor succeed, that revenue is clearly unprofitable.
No deadline stated Evidence as of
Only if: if Anthropic kills Cursor through aggressive rent-seeking
Source: This Is How Much Anthropic and Cursor Spend On Amazon Web Services
Assessment: UnresolvedSaid
Gary Marcus Observation
“June, 2025: the Apple reasoning paper confirmed that even with “reasoning”, LLMs still can’t solve distribution shift, the core Achille’s heel in neural networks that I have been writing about for nearly 30 years.”
Our reading: Even with reasoning, LLMs still cannot handle distribution shift, as the June 2025 Apple paper confirmed.
No horizon Evidence as of
Limitation: The competing papers are evidence of a scientific dispute, not equally decisive demonstrations.
Source: Game over. AGI is not imminent, and LLMs are not the royal road to getting there.
Assessment: UnresolvedSaid
Gary Marcus Observation
“This is one attempt to articulate that, and as the paper notes, current techniques capture only a fraction of what ultimately needs to be captured:”
Our reading: As the paper notes, current techniques capture only a fraction of the many facets of cognition that ultimately need to be captured.
No horizon Evidence as of
Limitation: This is an evidence gap, not evidence that the cognitive gaps were absent.
Source: Is AGI the right goal for AI?
Assessment: SupportedSaid
Gary Marcus Observation
“The weakness in the paper is that is that is built on a bunch of benchmarks, and we all known benchmarks can be gamed.”
Our reading: AI benchmarks can be gamed, which weakens an AGI definition built on them.
No horizon Evidence as of
Limitation: The experiments establish vulnerability, not deliberate misconduct by any particular developer.
Source: Is AGI the right goal for AI?
Assessment: SupportedSaid
Gary Marcus Observation
“[LLMs] have always been prone to hallucinations and errors.”
Our reading: LLMs have always been prone to hallucinations and errors.
No horizon Evidence as of
Limitation: Benchmark error rates do not directly measure ordinary usage.
Source: Is AGI the right goal for AI?
Assessment: UnresolvedSaid
Gary Marcus Observation
“We aren’t close to AGI, but we are reshaping our world on the premise that we are.”
Our reading: We are not close to AGI.
No horizon Evidence as of
Limitation: Benchmark failures do not measure time remaining to AGI.
Source: The Grand AGI Delusion
Assessment: UnresolvedSaid
Ed Zitron Observation
“A source that has seen materials related to sales has confirmed that, as of August 2025, Microsoft has around eight million active licensed users of Microsoft 365 Copilot, amounting to a 1.81% conversion rate across the 440 million Microsoft 365 subscribers.”
Our reading: As of August 2025, Microsoft 365 Copilot had about 8 million active licensed users, a 1.81% conversion of 440 million subscribers.
No horizon Evidence as of
Limitation: The underlying sales materials and unnamed source are unavailable.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“AI data centers are nothing like fiber, because there are very few actual use cases for these GPUs outside of AI, and none of them are remotely hyperscale revenue drivers.”
Our reading: AI data center GPUs have very few use cases outside AI, none of which drive hyperscale revenue, unlike fiber.
No horizon Evidence as of
Limitation: The examples are not a comprehensive workload survey.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“In fact, every user loses you money in generative AI, because it’s impossible to do cost control in a consistent manner.”
Our reading: Every generative AI user loses the provider money because consistent cost control is impossible.
No horizon Evidence as of
Limitation: Public pricing is not the provider's actual cost.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Forecast
“In the current relationship, NVIDIA Is Not Sending OpenAI $100 Billion, nor will it send it that much money, because 90% of OpenAI’s funding is gated behind building 9 or 10 gigawatts of data center capacity.”
Our reading: Under the current deal NVIDIA will not send OpenAI $100 billion, because 90% depends on building 9-10GW of capacity.
No deadline stated Evidence as of
Limitation: A later funding round cannot automatically be treated as the final outcome of the original arrangement.
Source: The Case Against Generative AI
Assessment: ContradictedSaid
Ed Zitron Causal claim
“AI does not have a “profit lever,” because the raw costs of providing access to AI models are so high (and they’re only increasing) that the basic economics of how the tech industry sells software don’t make sense.”
Our reading: AI businesses have no profit lever because model access costs are high and rising.
No horizon Evidence as of
Limitation: The result turns on reading 'profit lever' as reaching operating profit through scale.
Source: The Case Against Generative AI
Assessment: ContradictedSaid
Ed Zitron Observation
“As I discussed a month or so ago, data center development accounted for more of America’s GDP growth than all consumer spending combined, and there really isn’t any demand for AI in general, let alone at the scale that these hundreds of billions of dollars are being sunk into.”
Our reading: There really isn't any demand for AI in general, let alone at the scale of the hundreds of billions of dollars being sunk into it.
No horizon Evidence as of
Limitation: The weekly usage figure does not distinguish free from paid users [e1].
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Forecast
“I am telling you, as I have been telling you for years, again and again and again, that the demand is not there for generative AI, and the demand is never, ever arriving.”
Our reading: Demand for generative AI is not there and will never arrive.
No deadline stated Evidence as of
Limitation: The outcome depends on the contextual meaning of demand; a simple paying-customer reading would face direct counterevidence.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Forecast
“There is simply no space in these projections to absorb that $300 billion, and from what I can tell, by 2029, OpenAI will have actually burned more than $290 billion, assuming that it survives that long, which I do not believe it will.”
Our reading: OpenAI will not survive until 2029.
Not yet due Evidence as of
Limitation: Operating announcements are not an audit of financial solvency.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“In reality, LLMs can generate code, and do some software engineering-adjacent tasks, but, like all Large Language Models, break down and go totally insane, hallucinating more as the tasks get more complex.”
Our reading: Coding LLMs can generate code but break down and hallucinate more as tasks get more complex.
No horizon Evidence as of
Limitation: Prompt variations are not established here as a measure of engineering-task complexity [e1].
Source: The Case Against Generative AI
Assessment: MixedSaid
Ed Zitron Causal claim
“They don’t even speed up engineers — they actually make them slower!”
Our reading: Coding LLMs do not speed up engineers and actually make them slower.
No horizon Evidence as of
Limitation: The trials should not be pooled into an industry-wide estimate without accounting for their different settings.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Forecast
“I’m going to walk you through where I see this industry today, and why I see no future for it beyond a fiery apocalypse.”
Our reading: He sees no future for the generative AI industry beyond a fiery apocalypse.
No deadline stated Evidence as of
Limitation: The collapse criterion interprets a metaphor.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Forecast
“To Summarize: Coding LLMs Don’t Actually Replace Software Engineers, and Never Will, Due To The Inherent Unreliability Of Large Language Models”
Our reading: Coding LLMs do not and never will replace software engineers because LLMs are inherently unreliable.
No deadline stated Evidence as of
Limitation: A broader reading that includes any avoided hiring would make the company survey more directly contradictory.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“Generative AI can’t do the job of a software engineer, and it fails while also costing abominable amounts of money.”
Our reading: Generative AI cannot do a software engineer's job, and it fails while costing huge amounts.
No horizon Evidence as of
Limitation: A benchmark of individual problems does not reproduce an engineer's full responsibilities.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“Can Sora — OpenAI’s video creation tool — replace actors or animators? No, not at all, but it still fills the air full of tension because you can immediately see who is pre-registered to replace everyone that works for them.”
Our reading: OpenAI's Sora cannot replace actors or animators.
No horizon Evidence as of
Limitation: The strongest firsthand music-video account predates the target period.
Source: The Case Against Generative AI
Assessment: ContradictedSaid
Ed Zitron Observation
“To be clear, some people have lost jobs to AI, just not the white collar workers, software engineers, or really any of the career paths that the mainstream media and AI investors would have you believe.”
Our reading: Some people have lost jobs to AI, but not white collar workers or software engineers as media and investors claim.
No horizon Evidence as of
Limitation: The result depends on assessing the categorical sentence rather than replacing it with the surrounding at-scale argument.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“Which is to say that Microsoft clearly doesn’t have any good news to share, and as I’ll reveal later, they can’t even get 3% of their 440 million Microsoft 365 subscribers to pay for Microsoft 365 Copilot.”
Our reading: Microsoft cannot get even 3% of its 440 million Microsoft 365 subscribers to pay for Microsoft 365 Copilot.
No horizon Evidence as of
Limitation: The January 2026 count cannot establish whether August 2025 adoption was below the claim's threshold [e1].
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“This is, by the way, why we are still yet to get any tangible proof that AI is replacing software engineers…because it can’t.”
Our reading: There is still no tangible proof that AI is replacing software engineers because AI can't replace them.
No horizon Evidence as of
Limitation: The result depends on distinguishing full-role replacement from partial automation and avoided hiring.
Source: The Case Against Generative AI
Assessment: SupportedSaid
Ed Zitron Observation
“You cannot rely on a large language model to do what you want.”
Our reading: Large language models cannot be relied on to do what users want.
No horizon Evidence as of
Limitation: The strongest outcome evidence concerns difficult software-engineering tasks.
Source: The Case Against Generative AI
Assessment: SupportedSaid
Ed Zitron Observation
“You can add as many levels of intrigue and “reasoning” as you want, but Large Language Models cannot be trusted to do something correctly, or even consistently, every time.”
Our reading: Regardless of added reasoning, LLMs cannot be trusted to do things correctly or consistently every time.
No horizon Evidence as of
Limitation: One source is a vendor evaluation and the independent study is a preprint.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“You do not know what you’re going to get each time, and hallucinations multiply with the complexity of the thing you’re asking for, or whether a task contains multiple steps (which is a fatal blow to the idea of “agents.””
Our reading: LLM hallucinations multiply with task complexity and multi-step tasks, which is fatal to the idea of agents.
No horizon Evidence as of
Limitation: Puzzle performance is not a direct measure of practical autonomous work.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“Even Anthropic and OpenAI seem a little lethargic, both burning billions of dollars while making, by my estimates, no more than $2 billion and $6.26 billion in 2025 so far, despite projections of $5 billion and $13 billion respectively.”
Our reading: Anthropic and OpenAI seem a little lethargic, both burning billions while making, by his estimates, no more than $2 billion and $6.26 billion respectively in 2025 so far, against full-year projections of $5 billion and $13 billion.
No horizon Evidence as of
Limitation: Treating September 30 as equivalent to September 29 could change the result; that substitution is not made.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“Can large language models actually replace coders? Not really, no, and I’ll get into why later in the piece.”
Our reading: Large language models cannot really replace coders.
No horizon Evidence as of
Limitation: Task success, developer productivity and job replacement are different outcomes.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Forecast
“As a reminder, OpenAI has leaked that it’ll burn $115 billion in the next four years, and based on my estimates, it needs to raise more than $290 billion in the next four years based on its $300 billion deal with Oracle alone.”
Our reading: By his estimates, based on its $300 billion Oracle deal alone, OpenAI needs to raise more than $290 billion in the next four years.
Not yet due Evidence as of
Limitation: The executed agreement and detailed payment obligations were not available.
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Ed Zitron Observation
“OpenAI is now, at this point, on the hook for over a trillion dollars, an egregious sum for a company that already forecast billions in losses, with no clear explanation as to how it’ll afford any of this beyond “we need more money” and the vague hope that there’s another Softbank or Microsoft waiting in the wings to swoop in and save the day.”
Our reading: OpenAI is committed to over a trillion dollars of spending with no clear explanation of how it will pay.
No horizon Evidence as of
Limitation: The partnership and project totals do not fully identify OpenAI's own payment obligations or their overlap [e1] [e3].
Source: The Case Against Generative AI
Assessment: UnresolvedSaid
Gary Marcus Observation
“The recent increase in the company’s valuation is more than the total of the two projected bars on the right. And, reminder, those are for revenue, not profits.”
Our reading: Oracle's recent increase in valuation is more than the total of the two rightmost projected bars in Zero Hedge's chart of Oracle's cloud revenue growth forecast, and those bars represent revenue, not profits.
No horizon Evidence as of
Limitation: The company guidance is verified, but its correspondence to the specific chart is not.
Source: Peak bubble
Assessment: UnresolvedSaid
Gary Marcus Observation
“And all this from a company thought (or claimed) that GPT-5 was going to be tantamount to AGI (spoiler alert: it wasn’t)”
Our reading: GPT-5 was not tantamount to AGI, even though OpenAI thought or claimed it would be.
No horizon Evidence as of
Limitation: The independent evaluation covers spatial intelligence, not every cognitive capability.
Source: Peak bubble
Assessment: UnresolvedSaid
Gary Marcus Observation
“For good measure Oracle doesn’t have the chips they would need to fulfill the contracts, or even the cash to buy them.”
Our reading: Oracle does not have the chips it would need to fulfill the contracts in the $300 billion OpenAI deal under discussion, or even the cash to buy those chips.
No horizon Evidence as of
Limitation: The relationship between the July capacity agreement and all obligations in the reported $300 billion arrangement is not established.
Source: Peak bubble
Assessment: UnresolvedSaid
Gary Marcus Observation
“And schoolchildren using ChatGPT to do their homework aren’t enough to cover that.”
Our reading: Schoolchildren using ChatGPT to do their homework are not enough to cover OpenAI's immense costs and cash burn.
No horizon Evidence as of
Limitation: The reporting window and meaning of 'cover that' are not explicit.
Source: OpenAI’s future, foretold?
Assessment: SupportedSaid
Gary Marcus Observation
“GenAI certainly sounds amazing – but it hasn’t always worked well enough to justify the costs.”
Our reading: Generative AI has not always worked well enough to justify its costs.
No horizon Evidence as of
Limitation: Support rests on preliminary, largely self-reported and anonymized deployment research.
Source: OpenAI’s future, foretold?
Assessment: SupportedSaid
Gary Marcus Observation
“Meanwhile, more darkly, the Jurassic Park moment of mass-produced AI-generated misinformation that I projected in December 2022 has arrived.”
Our reading: Mass-produced AI-generated misinformation has arrived at scale.
No horizon Evidence as of
Limitation: The investigation establishes AI-generated narration, not autonomous authorship of every script.
Assessment: SupportedSaid
Gary Marcus Observation
“The notion on which this Substack was founded in May 2022 — that LLM scaling would not bring us to AGI — has gone mainstream.”
Our reading: The view that LLM scaling will not bring AGI has become mainstream.
No horizon Evidence as of
Limitation: The survey was academically weighted and was not a representative industry poll.
Assessment: UnresolvedSaid
Gary Marcus Forecast
“that LLM scaling would not bring us to AGI”
Our reading: Scaling LLMs will not produce AGI.
No deadline stated Evidence as of
Limitation: There is no universally accepted AGI test.
Assessment: SupportedSaid
Gary Marcus Observation
“The concerns about potential copyright violations that Reid Southen and I expressed here in and at greater length in the IEEE Spectrum in January 2024 continue almost two years later to be a problem.”
Our reading: Copyright violation concerns with generative AI remain a problem as of September 2025.
No horizon Evidence as of
Limitation: A lawsuit establishes a live allegation, not that the defendant infringed.
Assessment: UnresolvedSaid
Gary Marcus Forecast
“Perhaps I also no longer need to warn people that the idea that coding would disappear soon is bullshit.”
Our reading: Human coding will not disappear soon due to AI.
No deadline stated Evidence as of
Limitation: This leaves the forecast open because of its timing, not because the evidence shows human coding disappeared.
Assessment: SupportedSaid
Gary Marcus Observation
“No model could have done this in 2022. The graphics are vastly better than they used to be.”
Our reading: Image generation graphics quality has improved vastly since 2022, to a level no 2022 model could reach.
No horizon Evidence as of
Limitation: The evidence does not show that no 2022 model could produce a sharp, mechanically correct image of a rider. The 2022 test found some good outputs.
Source: Spinal Tap meets John Malkovich
Assessment: UnresolvedSaid
Gary Marcus Observation
“Watch the tuners come and go! So much for object permanence.”
Our reading: Grok's generated video failed to keep objects consistent, with the bass tuners appearing and disappearing.
No horizon Evidence as of
Limitation: No evidence items show what the video contains.
Source: Spinal Tap meets John Malkovich
Assessment: UnresolvedSaid
Gary Marcus Observation
“Over the last few years, graphics have gotten better and better. World models and comprehension less so.”
Our reading: Over the last few years, graphics in image generation systems have gotten better and better, while world models and comprehension have improved less.
No horizon Evidence as of
Limitation: The starting year and operational meanings of the capabilities are evaluator interpretations.
Source: What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana
Assessment: UnresolvedSaid
Gary Marcus Observation
“Suffice to say that I was able to break nano banana literally on my first try:”
Our reading: Marcus got Nano Banana to fail on his very first attempt.
No horizon Evidence as of
Limitation: The missing record does not show that Marcus's account was false.
Source: What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana
Assessment: UnresolvedSaid
Gary Marcus Observation
“And guess what, surprise, surprise, even the latest models still struggle with compositionality”
Our reading: Even the latest image generation models, Nano Banana among them, still struggle with compositionality.
No horizon Evidence as of
Limitation: The independent test is small and selected, and its outputs were not independently rescored.
Source: What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana
Assessment: UnresolvedSaid
Gary Marcus Observation
“For all the hype, Nano Banana is not all that much better at this sort of thing than ChatGPT-5:”
Our reading: Despite the hype, Nano Banana is not much better than ChatGPT-5 at drawing part-whole structures such as bicycles.
No horizon Evidence as of
Limitation: Selected failures from both systems would not establish comparable failure rates.
Source: What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana
Assessment: UnresolvedSaid
Gary Marcus Observation
“ChatGPT, by the way, also continues to struggle with the basic facts of life:”
Our reading: ChatGPT still struggles with basic facts of life in image generation.
No horizon Evidence as of
Limitation: The admitted evidence contains no tests from August 1–27, 2025.
Source: What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana
Assessment: ContradictedSaid
Ed Zitron Observation
“...maybe? It sure isn't trending that way, nor has it gone down yet.”
Our reading: The cost of inference is not trending downward and has not gone down yet, though he says it maybe could go down.
No horizon Evidence as of
Limitation: Fixed-capability benchmark costs do not establish aggregate application spending or costs for increasingly demanding workloads.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“Based on my (conservative) calculations, the true "cost" of OpenAI is around $82 billion, and that only includes capex from 2024 onward, based on 30% of Microsoft's capex (as not everything has been invested yet in 2025, and OpenAI is not necessarily all of the capex) and the $41.4 billion of funding it’s received so far.”
Our reading: Attributing 30% of Microsoft's capex since 2024 and adding OpenAI's $41.4B in funding, OpenAI's true cost is about $82 billion.
No horizon Evidence as of
Limitation: The allocation percentage is an explicit assumption.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“The cost of inference has gone up. Statements otherwise are purely false, and are the opinion of somebody who does not know what he's talking about.”
Our reading: The real cost of AI inference has risen, and claims that it is falling are false.
No horizon Evidence as of
Limitation: The disclosures do not provide representative longitudinal application bills linked to completed work, quality and model versions.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“Unlike CPUs, a GPU can't easily change tasks, or handle many little distinct operations, meaning that these things aren't going to be adopted for another mass-scale use case.”
Our reading: AI GPUs will not be adopted for another mass-scale use case, because GPUs cannot easily switch tasks.
No deadline stated Evidence as of
Limitation: The scale required for contradiction is interpretive.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“QUIP: AI Will Do Science Research, Somehow — No it won't, here's a writeup about why foundation models can't do this.”
Our reading: AI foundation models will not do scientific research.
No deadline stated Evidence as of
Limitation: Initial workshop review does not by itself establish independent validation of the scientific result.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“There is no proof this is possible, nobody has done it, nobody will do it.”
Our reading: Nobody will build the self-learning, self-teaching AI agent that AI 2027 assumes.
No deadline stated Evidence as of
Limitation: The outcome depends on retaining the AI 2027 context; a reading covering any self-modifying agent could produce a different assessment.
Source: How To Argue With An AI Booster
Assessment: SupportedSaid
Ed Zitron Observation
“While young people are having trouble finding jobs, there is no proof that AI is the reason.”
Our reading: There is no proof that AI is why young people are struggling to find jobs.
No horizon Evidence as of
Limitation: The conclusion concerns what had been established by publication, not whether AI actually affected hiring.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“This isn't a case where "the enterprise" is suddenly going to save these companies, because the enterprise already tried, and it isn't working.”
Our reading: Enterprise demand will not rescue generative AI companies, because enterprises have already tried the technology and it is not working.
No deadline stated Evidence as of
Limitation: The financial evidence covers one provider and does not settle outcomes for the unspecified group of companies.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Conditional forecast
“The "fiber optic cable" of this era is mostly owned by a few companies. 42% of NVIDIA's revenue is from the magnificent 7, and the companies buying these GPUs are, for the most part, not going to go bust once the AI bubble bursts.”
Our reading: For the most part, the companies buying AI GPUs are not going to go bust once the AI bubble bursts.
No deadline stated Evidence as of
Only if: once the AI bubble bursts
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“The true cost of Anthropic is around $77.1 billion, including all its funding and 30% of Amazon's capex from the beginning of 2024.”
Our reading: Counting all its funding plus 30% of Amazon's capex since 2024, Anthropic's true cost is about $77.1 billion.
No horizon Evidence as of
Limitation: The reconstruction of the unstated funding subtotal is an evaluator inference.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Conditional forecast
“Every member of the government having access to ChatGPT — something that is not even necessarily the case! — does not make this software useful, let alone essential, and if OpenAI burns a bunch of money "making it work for them," it still won't be essential, because Large Language Models are not actually that useful for doing stuff!”
Our reading: Even if OpenAI spends heavily tailoring ChatGPT for government use, it will not become essential, because LLMs are not very useful.
No deadline stated Evidence as of
Only if: even if OpenAI spends heavily to make it work for the government
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“ChatGPT is basically the same product as Claude or Grok or any number of different LLMs.”
Our reading: ChatGPT is essentially the same product as Claude, Grok and other LLMs.
No horizon Evidence as of
Limitation: Basically the same permits differences, so finding any difference would not automatically refute the claim.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“These data centers, once built, also create very little economic activity. They don't create jobs, they take up massive amounts of land and utilities, and they piss off and poison their neighbors.”
Our reading: Once built, AI data centers create very little economic activity, do not create jobs, take up massive amounts of land and utilities, and piss off and poison their neighbors.
No horizon Evidence as of
Limitation: Most directly relevant evidence concerns one facility.
Source: How To Argue With An AI Booster
Assessment: ContradictedSaid
Ed Zitron Forecast
“In any case, there really is no changing this path. They are out of ideas.”
Our reading: AI model developers are out of ideas and will not escape the path of improving models by burning more tokens.
Settled before its deadline Evidence as of
Limitation: The result concerns the tested model comparisons, not every task or every subsequent model.
Source: How To Argue With An AI Booster
Assessment: ContradictedSaid
Ed Zitron Observation
“Reasoning models are also the only way that model developers have been able to improve the efficacy of new models, using something called "test-time compute" to burn extra tokens to complete a task.”
Our reading: Model developers have only managed to make new models more effective through reasoning models that burn extra tokens via test-time compute.
No horizon Evidence as of
Limitation: The performance results and attribution to unsupervised scaling come from OpenAI; the packet provides no independent replication.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“I don't buy that this company will continue growing forever, and its stinky conversion rate isn't going to change anytime soon.”
Our reading: He doesn't believe OpenAI will continue growing forever, and OpenAI's poor conversion rate isn't going to change anytime soon.
No deadline stated Evidence as of
Limitation: The packet lacks a verified publication-period share using matched, nonoverlapping paid-user and active-user populations.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“You won't be able to find this as it isn't possible.”
Our reading: Someone who cannot code cannot build a fully functional, secure, useful app entirely by vibe coding.
No horizon Evidence as of
Limitation: The main candidate evidence comes from a tool vendor.
Source: How To Argue With An AI Booster
Assessment: ContradictedSaid
Ed Zitron Observation
“In simpler terms, these companies are flush with cash, have all the attention and investment they could possibly need, and are still unable to create a product with a defined, meaningful, mass-market use case.”
Our reading: Generative AI companies, despite ample funding and attention, have not produced a product with a defined, meaningful mass-market use case.
No horizon Evidence as of
Limitation: Meaningful and mass-market require interpretation because Zitron supplied no thresholds.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“And models keep getting bigger and more expensive, too.”
Our reading: AI models keep getting larger and more expensive.
No horizon Evidence as of
Limitation: The observation window and cost measure require interpretation.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“OpenAI's growth as the single-most-prominent company in AI (and if we're honest, one of the most prominent in software writ large) makes sense, but at some point will slow, because the actual returns for the businesses aren't there.”
Our reading: OpenAI's growth will eventually slow because businesses are not getting real returns from AI.
No deadline stated Evidence as of
Limitation: The revenue interpretation is supplied from context.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Observation
“These data centers are not, in and of themselves, driving much economic growth other than in the costs of building them.”
Our reading: Apart from construction spending, AI data centers do not drive much economic growth.
No horizon Evidence as of
Limitation: Construction dominance does not imply negligible operating activity.
Source: How To Argue With An AI Booster
Assessment: ContradictedSaid
Ed Zitron Causal claim
“Because model developers hit a wall of diminishing returns, and the only way to make their models do more was to make them burn more tokens to generate a more accurate response (this is a very simple way of describing reasoning, a thing that OpenAI launched in September 2024 and others followed).”
Our reading: Inference costs rose because model developers hit diminishing returns and could only improve models by having them burn more tokens through reasoning.
No horizon Evidence as of
Limitation: The decisive evaluation is developer-reported.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Ed Zitron Forecast
“Long story short, agents are not autonomous, they do not replace jobs, they cannot “replace coders,” they are not going to do so because probabilistic models are a horrible means of taking precise actions, and almost anyone who brings up agents as a booster is either misinformed or in the business of misinformation.”
Our reading: AI agents are not autonomous, do not replace jobs or coders, and will not do so, because probabilistic models are poor at taking precise actions.
No deadline stated Evidence as of
Limitation: The outcome depends on distinguishing autonomous whole-job replacement from substantial task displacement.
Source: How To Argue With An AI Booster
Assessment: SupportedSaid
Ed Zitron Observation
“It is transparently, blatantly obvious that GPT-4 did not "hire" a Taskrabbit or, indeed, make any of these actions — it was prompted to, and they do not show the prompts they used, likely because they had to use so many of them.”
Our reading: GPT-4 did not autonomously hire a TaskRabbit worker or take the actions described; it was prompted and humans carried them out.
No horizon Evidence as of
Limitation: Support is for the absence of autonomous execution, not for every possible reading of 'any of these actions'.
Source: How To Argue With An AI Booster
Assessment: UnresolvedSaid
Gary Marcus Observation
“And no other form of deep learning has yet solved the issues of reasoning, hallucinations, stupid errors, and factuality, either.”
Our reading: No form of deep learning has yet solved reasoning, hallucinations, stupid errors, or factuality.
No horizon Evidence as of
Limitation: This is a broad negative claim spanning different technical approaches.
Source: Things are so desperate at OpenAI that Sam Altman is starting to sound like Gary Marcus
Assessment: UnresolvedSaid
Gary Marcus Observation
“The dynamic with Altman is slightly different; it’s not that his lead is gone, but that his lead (and his reputation) has dwindled, after the disappointment of GPT-5.”
Our reading: After GPT-5's disappointment, OpenAI's lead and Altman's reputation have dwindled, though the lead is not gone.
No horizon Evidence as of
Limitation: Benchmark leadership covers particular evaluations and model configurations.
Source: Things are so desperate at OpenAI that Sam Altman is starting to sound like Gary Marcus
Assessment: SupportedSaid
Gary Marcus Causal claim
“And that means that high levels of agency, combined with the ease of manipulation of LLM-based software, is a recipe for chaos.”
Our reading: High agency combined with easily manipulated LLM software leads to chaos.
No horizon Evidence as of
Limitation: The outcome depends on interpreting the metaphor as a concrete security-risk claim.
Assessment: UnresolvedSaid
Gary Marcus Causal claim
“Agency plus LLMs have already led multiple “vibe coders” to lose databases, with new reports coming in every day or two, stemming simply from the unreliability inherent in LLMs (that Gary has so often harped on).”
Our reading: Agency combined with LLMs has already led multiple vibe coders to lose databases, with new reports every day or two, and these losses stem simply from the unreliability built into LLMs.
No horizon Evidence as of
Limitation: The Cursor incident rests on a user allegation without execution logs or an independent postmortem [e2].
Assessment: SupportedSaid
Gary Marcus Causal claim
“We are particularly worried about agents that software developers are starting to use, because they are often granted considerable authority and access to far-ranging tools, opening up immense security vulnerabilities.”
Our reading: The authors are particularly worried about agents that software developers are starting to use, because these agents are often given considerable authority and access to far-ranging tools, which opens up immense security vulnerabilities.
No horizon Evidence as of
Limitation: The frequency of broad permissions in actual installations remains unmeasured.
Assessment: UnresolvedSaid
Gary Marcus Forecast
“The bad news is that although this one was stopped, many others won’t be. No one patch would be enough, or even a thousand; there are just too many variations on the overall theme”
Our reading: Many attacks on AI coding tools won't be stopped; not even a thousand patches would suffice given the number of variations.
No deadline stated Evidence as of
Limitation: The evidence concerns researcher demonstrations rather than a census of malicious attacks.
Assessment: MixedSaid
Gary Marcus Observation
“Some customers even wanted the old models back. Almost nobody felt satisfied.”
Our reading: Almost no one was satisfied with GPT-5, and some customers wanted the old models back.
No horizon Evidence as of
Limitation: No representative satisfaction survey of GPT-5 users from the launch week was found, so the share of satisfied users is not known.
Assessment: UnresolvedSaid
Gary Marcus Forecast
“But between the disappointments of GPT-5 and a new study from METR that shows that LLMs do markedly better on coding benchmarks than in real-word practice, I think it is safe to say that LLMs won’t lead the way.”
Our reading: LLMs will not be the approach that leads the way to AGI.
No deadline stated Evidence as of
Limitation: The first accepted AGI is an interpretation of 'lead the way', not a criterion Marcus explicitly supplied.
Assessment: SupportedSaid
Gary Marcus Observation
“But between the disappointments of GPT-5 and a new study from METR that shows that LLMs do markedly better on coding benchmarks than in real-word practice, I think it is safe to say that LLMs won’t lead the way.”
Our reading: LLMs do markedly better on coding benchmarks than in real-world coding practice.
No horizon Evidence as of
Limitation: The study does not establish the same mismatch across all developers, tasks or LLMs; it allows for substantial productivity gains in other coding settings [e2].
Assessment: UnresolvedSaid
Gary Marcus Causal claim
“Critically, as I argued at the end of June (and going back to 2019) LLMs never induce proper world models, which is why, for example, they still can’t even play chess reliably,1 and continue to make stupid, head-scratching errors with startling regularity.”
Our reading: LLMs never induce proper world models, which causes them to be unable to play chess reliably and to make frequent stupid errors.
No horizon Evidence as of
Limitation: The strongest intervention evidence concerns a specialist chess model.
Assessment: UnresolvedSaid
Gary Marcus Causal claim
“ChatGPT gets them wrong because it is munging words together, not genuinely processing an (elementary) model of British PM history.”
Our reading: ChatGPT gets the British prime minister tasks wrong because it is munging words together rather than genuinely processing an elementary model of British PM history.
No horizon Evidence as of
Limitation: Evidence from Llama-2 cannot directly support or contradict a claim about GPT-5's internal processing.
Assessment: UnresolvedSaid
Gary Marcus Observation
“They don’t induce abstract models of the world that they can reliably engage, even when those models are given to them on a silver platter (as with the rules of chess).”
Our reading: LLMs do not induce abstract world models they can reliably use, even when given the rules explicitly.
No horizon Evidence as of
Limitation: RuleArena tests practical regulations rather than chess.
Assessment: UnresolvedSaid
Gary Marcus Observation
“Sometimes that works as a fine substitute, sometimes it doesn’t. You can never know in advance which.”
Our reading: LLMs' aping of what a person might say sometimes works as a fine substitute for what a person would do and sometimes doesn't, and you can never know in advance which.
No horizon Evidence as of
Limitation: The principal evaluation uses question-answering benchmarks rather than complete human workflows.