Claims checked against evidence, page 7 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 7 of 29: claims 601 to 700 of 2859, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Neither company is facing a financial shortfall as a result of being unable to provide their services (rather, they’re facing financial shortfalls because they’re providing their services to customers), and the only ones paying that price because of these “capacity limits” are the customers.”
Our reading: Neither Anthropic nor OpenAI faces a financial shortfall from being unable to provide their services; their shortfalls come from providing their services to customers, and the only ones paying the price of the capacity limits are the customers.
No horizon Evidence as of
Limitation: Management's counterfactual belief is not a measured financial effect.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Forecast
“I’m making the call that by the end of 2026, a majority of AI services will move some or all of their customers to token-based billing as they reckon with the true costs of running AI models.”
Our reading: By the end of 2026, most AI services will move some or all customers to token-based billing.
Not yet due Evidence as of
Limitation: Results could depend materially on which services are counted.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“Nevertheless, these services still burn tokens at whatever rate that it costs to pay for them — for example, $5 per million input and $25 per million output for Opus 4.7, as I mentioned previously — meaning that the company almost always loses money unless a person doesn’t use the subscription very much.”
Our reading: AI subscription services almost always lose money on users unless they use the subscription lightly.
No horizon Evidence as of
Limitation: Confirming a token price does not confirm the claimed prevalence of losses.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Pale Horse: Any further price increases or service degradations from Anthropic and OpenAI are a sign that they’re running low on cash.”
Our reading: Further price increases or service degradations from Anthropic and OpenAI would indicate they are running low on cash.
No horizon Evidence as of
Limitation: No contemporaneous cash and obligations assessment is matched to the October degradation.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“An aside/explainer: if you’re confused as to what “token-based billing” means, know that the vast majority of AI services currently subsidize their subscriptions, using another measure (such as “requests” or “rate limits”) to meter out how much a user can use the service.”
Our reading: The vast majority of AI services currently subsidize their subscriptions.
No horizon Evidence as of
Limitation: Service definitions and sampling choices could materially affect a prevalence estimate.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“As a result, adding more capacity simply makes the product work better for a much higher cost.”
Our reading: Because Anthropic's capacity issues are not affecting its growth, adding more capacity simply makes the product work better for a much higher cost.
No horizon Evidence as of
Limitation: The service evidence is a company announcement.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“So, to summarize, at least 4.6GW of the 15.2GW of data center capacity under construction is for OpenAI, with at least another 4GW of that reserved for Anthropic through partners like Microsoft, Google and Amazon.”
Our reading: At least 4.6GW of the 15.2GW of AI data center capacity under construction is for OpenAI and at least 4GW for Anthropic.
No horizon Evidence as of
Limitation: The packet lacks Sightline's underlying 15.2GW project inventory and matched customer allocations.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“And at the top of the pile sits NVIDIA, the largest company on the stock market, which is selling more GPUs than can be possibly installed, and very few people seem to notice or care.”
Our reading: NVIDIA is selling more GPUs than can possibly be installed.
No horizon Evidence as of
Limitation: The installation window is unspecified.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“Otherwise, everybody else is making single or double-digit millions of dollars and losing hundreds of millions of dollars to get there.”
Our reading: Every AI startup other than OpenAI, Cursor and Anthropic is making single- or double-digit millions of dollars in revenue and losing hundreds of millions of dollars to get there.
No horizon Evidence as of
Limitation: The annual period is inferred.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“This is a significant move, driven by (per my own reporting) Microsoft’s week-over-week costs of running GitHub Copilot nearly doubling since January.”
Our reading: Microsoft's GitHub Copilot changes are driven by its weekly running costs nearly doubling since January 2026.
No horizon Evidence as of
Limitation: The primary blog's cost discussion could not be recovered by read_page.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Venture capital funding allows it to massively subsidize its services.”
Our reading: Venture capital funding allows Anthropic to massively subsidize its services.
No horizon Evidence as of
Limitation: The packet lacks matched service receipts, actual utilization and comparable delivery costs.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Things have gotten so dire that even Deidre Bosa of CNBC agrees with me that AI demand is inflated by token-maxxing and subsidized services.”
Our reading: AI demand is inflated by token-maxxing and subsidized services.
No horizon Evidence as of
Limitation: Individual user accounts do not establish an industry-wide causal effect.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“If we then cut that up into GB200 or GB300 NVL72 racks with a power draw of around 140KW, that’s around 71,429 racks’ worth of hardware at an average of $4 million each, which gives us around $285.7 billion in revenue for NVIDIA.”
Our reading: By his estimate, the 15.2GW of data center capacity under construction through 2028 works out to around 71,429 NVL72 racks, or about $285.7 billion in NVIDIA revenue.
No horizon Evidence as of
Limitation: An arithmetic-only reading would warrant a more favorable assessment.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“The entire industry — including OpenAI and Anthropic’s theoretical revenues of $13.1 billion and $4.5 billion — hit around $65 billion last year, and that includes the revenues from providing compute generated by neoclouds like CoreWeave and hyperscalers like Microsoft.”
Our reading: The entire AI industry, including OpenAI's and Anthropic's theoretical revenues of $13.1 billion and $4.5 billion, hit around $65 billion in revenue last year.
No horizon Evidence as of
Limitation: No comprehensive, scope-matched industry aggregation or defensible bound is supplied.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“I’m talking about hundreds of billions of dollars of GPUs sitting in warehouses that aren’t being installed, with it taking six months to install a single quarter’s worth of GPU sales.”
Our reading: Hundreds of billions of dollars of GPUs are sitting uninstalled in warehouses, with a quarter's GPU sales taking six months to install.
No horizon Evidence as of
Limitation: Customer-owned equipment may be absent from NVIDIA's inventory balance.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“It is blatantly, alarmingly obvious that Anthropic cannot afford to provide a stable and reliable service to its customers”
Our reading: Anthropic cannot afford to provide a stable and reliable service to its customers.
No horizon Evidence as of
Limitation: One incident does not establish overall reliability during the target period.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“Even if said numbers were real, the majority of OpenAI, Cursor and Anthropic’s revenues come from subsidized software subscriptions.”
Our reading: Most of OpenAI's, Cursor's and Anthropic's revenue comes from subsidized software subscriptions.
No horizon Evidence as of
Limitation: No matched January–March 2026 revenue and delivery-cost breakdown is supplied for any of the three companies.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Observation
“NVIDIA Is Selling Years’ Worth of GPUs In Advance And Warehousing Them”
Our reading: NVIDIA is selling years' worth of GPUs in advance of installable capacity, and those GPUs are being warehoused.
No horizon Evidence as of
Limitation: The original Business Insider article could not be read; its later findings are included only through an identified secondary summary.
Source: Four Horsemen of the AIpocalypse
Assessment: UnresolvedSaid
Ed Zitron Forecast
“And we’re going to find out fairly quickly, because the world of token subsidies is going away.”
Our reading: Subsidized AI token pricing is ending soon, exposing customers to the true costs.
No deadline stated Evidence as of
Limitation: The October counterevidence is secondary reporting because the original research was marked PARTIAL.
Source: Four Horsemen of the AIpocalypse
Assessment: MixedSaid
Ed Zitron Observation
“This does not really make sense given the fact that all support documents and the Claude website reflect that Pro users do not have access to Claude Code.”
Our reading: Anthropic's claim that this was a small test on about 2% of new signups does not fit the fact that all support documents and the Claude website showed Pro users without Claude Code access.
No horizon Evidence as of
Limitation: The counterexample relies on contemporaneous firsthand reporting because a sufficiently complete archived product page was not recovered.
Assessment: SupportedSaid
Ed Zitron Observation
“This might sound like a good thing, but it actually means that each request using Opus 4.7 is actually 7.5 of them.”
Our reading: Under GitHub Copilot's 7.5x multiplier, each Opus 4.7 request counts as 7.5 premium requests.
No horizon Evidence as of
Limitation: The evidence establishes published accounting rules, not an audit of individual account deductions.
Source: Exclusive: Microsoft To Shift GitHub Copilot Users To Token-Based Billing, Tighten Rate Limits
Assessment: ContradictedSaid
Ed Zitron Observation
“The standard version of Claude Opus 4.6 has a premium request multiplier of three — meaning that, even with the promotional pricing, Claude Opus 4.7 is around 250% more expensive to use.”
Our reading: Even with promotional pricing, Claude Opus 4.7 on GitHub Copilot costs about 250% more in premium requests than Opus 4.6.
No horizon Evidence as of
Limitation: The verdict concerns the percentage increase in premium-request units, not changes in subscription bills or cost per completed task.
Source: Exclusive: Microsoft To Shift GitHub Copilot Users To Token-Based Billing, Tighten Rate Limits
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“This is a significant move, reflecting the significant cost of running models on any AI product.”
Our reading: Microsoft's move to token-based billing reflects the significant cost of running models on any AI product.
No horizon Evidence as of
Limitation: The result depends on retaining the source's broad scope.
Source: Exclusive: Microsoft To Shift GitHub Copilot Users To Token-Based Billing, Tighten Rate Limits
Assessment: UnresolvedSaid
Ed Zitron Observation
“The documents reveal that the weekly cost of running Github Copilot has doubled since the start of the year.”
Our reading: According to internal documents, the weekly cost of running GitHub Copilot doubled between the start of 2026 and April 2026.
No horizon Evidence as of
Limitation: The search did not obtain authenticated internal cost records.
Source: Exclusive: Microsoft To Shift GitHub Copilot Users To Token-Based Billing, Tighten Rate Limits
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, although OpenAI thought or claimed that it would be.
No horizon Evidence as of
Limitation: Support for one clause is insufficient to support the whole claim.
Source: Peak absurdity, Part II
Assessment: SupportedSaid
Gary Marcus Observation
“Today, even after a very strong week, Oracle stands at just $170.”
Our reading: As of the post date, even after a very strong week, Oracle stock stands at just $170, compared with about $308 on September 11.
No horizon Evidence as of
Limitation: Marcus gives no quotation timestamp; the assessment uses daily closing prices.
Source: Peak absurdity, Part II
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“As I’ve said for years, AI companies are inherently unsustainable due to the unreliable and inconsistent outputs of Large Language Models and the incredible costs of providing the services.”
Our reading: AI companies are inherently unsustainable because of unreliable LLM outputs and high service costs.
No horizon Evidence as of
Limitation: The financial metric and accounting basis are unspecified.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Forecast
“On June 1 2026, any article or post about Codex’s efficacy must be rewritten, as rate limits will be halved.”
Our reading: On June 1 2026, Codex rate limits will be halved.
Deadline passed Evidence as of
Limitation: The strongest retrospective evidence concerns one subscription tier.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“These actions are morally wrong, and are also the direct result of the AI industry’s deceptive and manipulative scare campaign, one promoted by men like Altman and Amodei, as well as doomer fanfiction writers like Yudowsky, and, of course, Daniel Kokotajlo of AI 2027 — both of whom have had their work validated and propagated via the New York Times.”
Our reading: The attack on Sam Altman's home was a direct result of the AI industry's deceptive scare campaign.
No horizon Evidence as of
Limitation: The affidavit contains allegations supporting probable cause, not a final determination of motive or responsibility.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Observation
“I’m in a research-heavy business, and I have tried the most-powerful LLMs and highest-priced RAG/post-RAG research tools, and every time find them bereft of any unique analysis or suggestions.”
Our reading: In his research-heavy business, the author has tried the most powerful LLMs and the highest-priced RAG/post-RAG research tools and every time found them bereft of any unique analysis or suggestions.
No horizon Evidence as of
Limitation: Failure to locate records does not show that the sessions did not occur.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Observation
“It’s also clear, at this point, that Anthropic and OpenAI have both offered subscriptions that were impossible to provide at scale at the price and availability that they were leading up to 2026, and that they did so with the intention of growing their revenue to acquire more customers, equity investment and attention.”
Our reading: It is clear that Anthropic and OpenAI both offered subscriptions that were impossible to provide at scale at the price and availability they had leading up to 2026, and did so intending to grow revenue to acquire more customers, equity investment and attention.
No horizon Evidence as of
Limitation: The reported Pro losses lack a defined accounting basis or matched subscriber revenue and cost data [e1].
Source: I Will Never Respect A Website
Assessment: ContradictedSaid
Ed Zitron Causal claim
“OpenAI and Anthropic need to show continued revenue growth, which means that they must have capacity available for new customers, which means that old customers will always be the first to be punished.”
Our reading: Because OpenAI and Anthropic must show revenue growth and keep capacity for new customers, existing customers will always be punished first.
No horizon Evidence as of
Limitation: The counterexample was a small, temporary experiment at Anthropic.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Conditional forecast
“And if it doesn’t make that much revenue — and also doesn’t get funding or debt to support it — OpenAI will run out of money, much as Anthropic will if that capacity gets built and it doesn’t make tens of billions of dollars a month to pay for it.”
Our reading: If OpenAI misses its projected revenue and lacks funding or debt, it will run out of money; Anthropic will too if its capacity is built and it doesn't earn tens of billions a month.
No deadline stated Evidence as of
Only if: OpenAI misses projected revenue and lacks funding or debt
Source: I Will Never Respect A Website
Assessment: ContradictedSaid
Ed Zitron Observation
“Every “AI agent” story is either about code generation, summarizing some sort of information source, or generating something based on an information source that you may or may not be able to trust.”
Our reading: Every AI agent story is about code generation, summarization, or generating content from an information source of uncertain trustworthiness.
No horizon Evidence as of
Limitation: The classification depends on the meaning of 'generating something'; a reading encompassing all tool commands would change the result.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Observation
“Every single LLM product is being sold under false pretenses about what’s actually sustainable and possible long term.”
Our reading: Every LLM product is sold under false pretenses about long-term sustainability and possibility.
No horizon Evidence as of
Limitation: Terms of service do not capture the whole sales presentation.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Forecast
“Anyone using a subscription to OpenAI or Anthropic’s services needs to wake up and realize that their way of life is going away — that rate limits will make current workflows impossible, that prices will increase, and that the product they’re selling even today is not one that makes any economic sense.”
Our reading: For OpenAI and Anthropic subscribers, rate limits will make current workflows impossible and prices will increase.
No deadline stated Evidence as of
Limitation: The OpenAI help page is mutable; its displayed update age places the accessed version on October 27.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Forecast
“LLMs will always hallucinate, their outputs are not trustworthy as a result, they cannot be deterministic, and any chance of any mistakes of any kind are unforgivable.”
Our reading: LLMs will always hallucinate, so their outputs are not trustworthy.
No deadline stated Evidence as of
Limitation: The forecast has no deadline.
Source: I Will Never Respect A Website
Assessment: ContradictedSaid
Ed Zitron Observation
“LLMs will always hallucinate, their outputs are not trustworthy as a result, they cannot be deterministic, and any chance of any mistakes of any kind are unforgivable.”
Our reading: LLMs cannot be deterministic.
No horizon Evidence as of
Limitation: The evidence is the implementing lab's report, not an independent replication.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Forecast
“The con of the Large Language Model era is that any subscription you pay for is massively subsidized, and that any product you use can and will see its service degraded as these companies desperately try to either ease their capacity issues or lower their burn rate.”
Our reading: Any LLM subscription you pay for is massively subsidized, and any LLM product you use can and will see its service degraded as these companies desperately try to ease their capacity issues or lower their burn rate.
No deadline stated Evidence as of
Limitation: The financial assertion lacks a stated accounting basis.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Observation
“The so-called “agentic AI” that is supposedly powerful and autonomous is actually incredibly demanding of its human users — you must set it up in so many different ways and connect it to so many different services and check that every “agent” (different chatbot) is instructed in exactly the right way, and that none of these agents cause any problems (they will) with each other.”
Our reading: So-called agentic AI, supposedly powerful and autonomous, is actually incredibly demanding of its human users, who must set it up in many ways, connect it to many services, and check that every agent is instructed correctly, and its agents will cause problems with each other.
No horizon Evidence as of
Limitation: Neither example is a representative workload study.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Forecast
“These conversations are also nonsensical and misleading when you compare them to what Large Language Models can do, and this rhetoric is a blatant attempt to scare people into paying for software today based on what it absolutely cannot and will not do in the future.”
Our reading: LLMs cannot and will not do what AI executives' job-loss and superintelligence rhetoric claims.
No deadline stated Evidence as of
Limitation: The labor evidence covers a U.S. payroll sample through June 2026, not every workforce through the evaluation date.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Conditional forecast
“In fact, I think that as capacity limits get hit, both Anthropic and OpenAI will degrade the experience for the user (either through model degradation or rate limit decay) as much as they can.”
Our reading: As capacity limits are hit, Anthropic and OpenAI will degrade user experience as much as they can via model degradation or rate limits.
No deadline stated Evidence as of
Only if: as capacity limits get hit
Source: I Will Never Respect A Website
Assessment: ContradictedSaid
Ed Zitron Forecast
“I see no scenario where costs come down, or where rate limits are eased.”
Our reading: The author sees no scenario in which costs for Anthropic's and OpenAI's services come down or their rate limits are eased.
Settled before its deadline Evidence as of
Limitation: The finding concerns the rate-limit branch.
Source: I Will Never Respect A Website
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Not only must Anthropic and OpenAI find a way to make their unsustainable and unprofitable services burn less money, but they must also constantly dance with metering out whatever capacity they have to their customers, because the more extra capacity they buy, the more money they lose.”
Our reading: Anthropic's and OpenAI's services are unsustainable and unprofitable, and buying more capacity makes them lose more money.
No horizon Evidence as of
Limitation: The contract's accounting basis is unknown.
Source: I Will Never Respect A Website
Assessment: SupportedSaid
Gary Marcus Observation
“On the darker side, it really does arm attackers to a greater degree than Mythos’s predecessors.”
Our reading: Claude Mythos Preview gives cyber attackers more capability than earlier Claude models did.
No horizon Evidence as of
Limitation: The finding concerns controlled offensive testing, not observed criminal deployment.
Source: Claude Mythos, evaluated
Assessment: UnresolvedSaid
Gary Marcus Observation
“On the happy side, in its current form, Myth is nowhere near as scary as Tom Fridman (who worries about schoolchildren accidentally taking down power grids) and others made it out to be.”
Our reading: In its current form, Claude Mythos Preview is nowhere near as dangerous as Thomas Friedman and others portrayed it.
No horizon Evidence as of
Limitation: Simulated network compromise and individual exploits do not directly measure disruption of a real power grid.
Source: Claude Mythos, evaluated
Assessment: SupportedSaid
Gary Marcus Observation
“Second, they show that a neurosymbolic hybrid model, which integrates both a neural networks (for pattern recognition) and a symbolic planner generalizes far better.”
Our reading: A neurosymbolic hybrid with a symbolic planner generalizes far better than VLAs on the Tower of Hanoi task.
No horizon Evidence as of
Limitation: This establishes the reported result for one simulated task and the tested configurations. It does not establish a general advantage across robotics.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“LLMs have persistent weaknesses around planning, reasoning and generalization.”
Our reading: LLMs have persistent weaknesses in planning, reasoning and generalization.
No horizon Evidence as of
Limitation: Generalization experiments use restricted settings.
Source: Even more good news for the future of neurosymbolic AI
Assessment: UnresolvedSaid
Gary Marcus Observation
“LLMs are an efficient way to pattern recognition where perfect results are not required, but an inefficient way to reason a plan.”
Our reading: LLMs are efficient for imperfect pattern recognition but inefficient for reasoning out a plan.
No horizon Evidence as of
Limitation: Different efficiency baselines could produce different judgments.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“Third, they show that the neurosymbolic hybrid was vastly more efficent in terms of energy use - by nearly two orders of magnitude.”
Our reading: The neurosymbolic hybrid used nearly two orders of magnitude less energy than the VLAs in the Tufts study.
No horizon Evidence as of
Limitation: The outcome depends on interpreting the statement as a training comparison.
Source: Even more good news for the future of neurosymbolic AI
Assessment: UnresolvedSaid
Gary Marcus Observation
“You can read the full paper by Timothy Duggan, Pierrick Lorang, Hong Lu, and Matthias Scheutz, which very much vindicates the Apple reasoning paper, here.”
Our reading: The Tufts paper strongly vindicates the Apple reasoning paper's findings.
No horizon Evidence as of
Limitation: A weaker reading of vindicates as merely consistent with would receive more support.
Source: Even more good news for the future of neurosymbolic AI
Assessment: UnresolvedSaid
Gary Marcus Forecast
“But anybody who thinks LLMs are a direct route to the sort AGI that could fundamentally transform society for the good is kidding themselves.”
Our reading: LLMs are not a direct route to the sort of AGI that could fundamentally transform society for the good, and anyone who thinks they are is kidding themselves.
No deadline stated Evidence as of
Limitation: The forecast's central terms remain underspecified.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“Claude Code (in particular version 4.6), which does seem genuinely impressive to me and a number of friends I trust, still makes plenty of mistakes, and still can’t be trusted.”
Our reading: Claude Code (in particular version 4.6), which seems genuinely impressive to Marcus and a number of friends he trusts, still makes plenty of mistakes and still can't be trusted.
No horizon Evidence as of
Limitation: The controlled evaluation covers one programming language and a particular task suite.
Source: Even more good news for the future of neurosymbolic AI
Assessment: UnresolvedSaid
Gary Marcus Observation
“It’s obviously not AGI.”
Our reading: Claude Code is not AGI.
No horizon Evidence as of
Limitation: This is an evidentiary finding, not a finding that Claude Code was AGI.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“Still, Claude Code ain’t perfect, or even close.”
Our reading: Claude Code is not perfect or even close to it.
No horizon Evidence as of
Limitation: The evidence does not establish a representative failure rate across Claude Code users or tasks.
Assessment: UnresolvedSaid
Gary Marcus Observation
“Instead, the way Anthropic built that kernel is straight out of classical symbolic AI. For example, it’s in large part a big IF-THEN conditional, with 486 branch points and 12 levels of nesting — all inside a deterministic, symbolic loop that the real godfathers of AI, people like John McCarthy and Marvin Minsky and Herb Simon, would have instantly recognized.”
Our reading: Claude Code's print.ts kernel is classical symbolic AI, largely a big IF-THEN conditional with 486 branch points and 12 nesting levels inside a deterministic symbolic loop.
No horizon Evidence as of
Limitation: No independent syntax-tree count was completed.
Assessment: SupportedSaid
Gary Marcus Observation
“Putting things differently, Anthropic, when push came to shove, went exactly where I have said for 25 years that the field needed to go: to Neurosymbolic AI.”
Our reading: Anthropic adopted neurosymbolic AI in building Claude Code.
No horizon Evidence as of
Limitation: The result depends on Marcus's broad contextual definition of neurosymbolic AI.
Assessment: UnresolvedSaid
Gary Marcus Observation
“The source code leak proves it. Tucked away at its center is a 3,167 line kernel called print.ts.”
Our reading: The leaked Claude Code source shows a 3,167-line kernel called print.ts at its center, proving it is not a pure LLM.
No horizon Evidence as of
Limitation: A longer file alone does not disprove a shorter function inside it.
Assessment: SupportedSaid
Gary Marcus Observation
“And the thing is, Claude Code is NOT a pure LLM. And it’s not pure deep learning. Not even close.”
Our reading: Claude Code is not a pure LLM and not pure deep learning.
No horizon Evidence as of
Limitation: The independent study is a preprint.
Assessment: UnresolvedSaid
Gary Marcus Forecast
“The first part has already come to pass. In time, other three will, too.”
Our reading: The other three parts of Marcus's 2020 agenda (knowledge, reasoning, world models) will eventually be adopted, as neurosymbolic AI already has been.
No deadline stated Evidence as of
Limitation: The terms are vague, and current industry uses of 'reasoning' and 'world models' may not match the paper's meaning.
Assessment: UnresolvedSaid
Gary Marcus Causal claim
“Claude Code isn’t better because of scaling. It’s better because it is neurosymbolic.”
Our reading: Claude Code's superiority comes from being neurosymbolic, not from scaling.
No horizon Evidence as of
Limitation: The comparator and meaning of scaling materially affect the test.
Assessment: UnresolvedSaid
Gary Marcus Observation
“The demo was definitely proof of concept that we need to get our regulatory and technical house in order, but not the immediate threat the media and public was lead to believe.”
Our reading: The Mythos demo was not the immediate threat the media and public were led to believe.
No horizon Evidence as of
Limitation: The principal technical evidence is Anthropic's own disclosure.
Source: Three reasons to think that the Claude Mythos announcement from Anthropic was overblown
Assessment: UnresolvedSaid
Gary Marcus Observation
“The model itself is incrementally better than previous recent models, but certainly not an off-the-chart breakthrough:”
Our reading: Mythos is incrementally better than recent previous models and certainly not an off-the-chart breakthrough.
No horizon Evidence as of
Limitation: The original Mythos system card could not be read through the required tool.
Source: Three reasons to think that the Claude Mythos announcement from Anthropic was overblown
Assessment: UnresolvedSaid
Gary Marcus Observation
“ChatGPT can’t reliably run a timer but it has still been implicated in delusions, suicides, cognitive surrender, mass disinformation, and so much more.”
Our reading: ChatGPT cannot reliably run a timer, yet it has been implicated in delusions, suicides, cognitive surrender and mass disinformation.
No horizon Evidence as of
Limitation: Legal allegations establish implication, not proven causation.
Source: What should we take from Anthropic’s (possibly) terrifying new report on Mythos?
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“AI progress isn’t hampered by a lack of access, talent, resources, novel approaches, or industry buy-in, but by a single-minded focus on Large Language Models, a technology that has been so obviously-limited from the very beginning that Gary Marcus was able to call it in 2022.”
Our reading: AI progress is not hampered by a lack of access, talent, resources, novel approaches, or industry buy-in, but by a single-minded focus on Large Language Models, a technology that has been obviously limited from the very beginning.
No horizon Evidence as of
Limitation: The cited experiments predate the target period and serve as mechanism evidence only.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“Said data centers make up 90% of NVIDIA’s revenue, which means that 8% or so of the S&P 500’s value comes from a company that makes money selling hardware to people that immediately lose money on installing it.”
Our reading: Data centers make up 90% of NVIDIA's revenue, so about 8% of the S&P 500's value comes from a company selling hardware to buyers who immediately lose money installing it.
No horizon Evidence as of
Limitation: No verified index weight around April 8, 2026 is supplied.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“As I discussed a few weeks ago, despite everybody talking about the hundreds of gigawatts of data centers being built “to power AI,” only 5GW are actually “under construction,” with “under construction” meaning anything from “we’ve got some scaffolding up” to “we’re about to hand over the keys to the customer.””
Our reading: Only about 5GW of the hundreds of gigawatts of announced AI data centers are actually under construction.
No horizon Evidence as of
Limitation: The source supports a narrower statement about 2026 scheduled openings.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“As I wrote a few weeks ago, LLMs are good at writing a lot of code, not good code, and the more people you allow to use them, the more code you’re going to generate, which means the more time you’re either going to need to review that code, or the more vulnerabilities you’re going to create as a result.”
Our reading: LLMs write a lot of code, not good code; broader use means more review time or more vulnerabilities.
No horizon Evidence as of
Limitation: Security tasks and maintenance tasks measure different aspects of quality.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“That “growth” is only possible because all of the companies in question heavily subsidize their products, spending $3 to $15 for every dollar of revenue.”
Our reading: The growth of the AI companies in question is only possible because all of them heavily subsidize their products, spending $3 to $15 for every dollar of revenue.
No horizon Evidence as of
Limitation: The comparison uses reported operating results, not a defined measure of the claimed spending or product subsidies [e1].
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“Back in November 2025, I reported that OpenAI had made — based on its revenue share with Microsoft — $4.329 billion between January and September 2025, despite The Information reporting that it had made $4.3 billion in the first half of the year based on disclosures to shareholders.”
Our reading: Based on its Microsoft revenue share, OpenAI made $4.329 billion in January–September 2025.
No horizon Evidence as of
Limitation: The netting explanation relies on an unnamed source, and the payment amounts reported by TechCrunch originate with Zitron [e1].
Source: AI Is Really Weird
Assessment: SupportedSaid
Ed Zitron Observation
“The above quote has been used repeatedly to suggest that Anthropic has 50% gross margins and is “profitable,” which is extremely weird in and of itself as that’s not what Dario Amodei said at all.”
Our reading: Amodei's remarks don't actually claim Anthropic has 50% gross margins or is profitable.
No horizon Evidence as of
Limitation: This assesses the supplied normalized proposition, not the separate allegation that third parties repeatedly misrepresented the interview.
Source: AI Is Really Weird
Assessment: ContradictedSaid
Ed Zitron Causal claim
“The revenues are terrible, Large Language Models are yet to provide any meaningful productivity improvements, and the only reason that they’ve been able to get as far as they have is a compliant media and a venture capital environment borne of a lack of anything else to invest in.”
Our reading: AI revenues are terrible, LLMs have yet to provide any meaningful productivity improvements, and the only reason they have got as far as they have is a compliant media and a venture capital environment that comes from having nothing else to invest in.
No horizon Evidence as of
Limitation: The positive study uses a staggered rollout and causal identification assumptions rather than a large randomized deployment.
Source: AI Is Really Weird
Assessment: SupportedSaid
Ed Zitron Observation
“To be clear, LLMs can absolutely write code, and can absolutely create software, but neither of those mean that the code is good, stable or secure, or that the same can be said of the software they create.”
Our reading: LLMs can write code and create software, but that doesn't mean the code is good, stable, or secure.
No horizon Evidence as of
Limitation: This does not establish how often generated software is defective or compare it with human-written software.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“I have confirmed with a source at that there is no actual metric or tracking of any return on investment involved in token burn at Meta, meaning that TokenMaxxing’s only purpose is to burn more tokens to go higher on a leaderboard, and is already creating bad habits across a company that already has decaying products and leadership.”
Our reading: Meta does not track any ROI on token burn; tokenmaxxing there exists only to climb a leaderboard.
No horizon Evidence as of
Limitation: Internal measurement systems are not fully public.
Source: AI Is Really Weird
Assessment: ContradictedSaid
Ed Zitron Observation
“If that’s the case, I truly have no idea how it could’ve possibly accelerated so aggressively, and as I’ve mentioned before, there is no way to reconcile having made $5 billion in lifetime revenue as of March 9, 2026, having $14 billion in annualized revenue on February 12 2026, and having $4.5 billion in revenue for the year 2025.”
Our reading: There is no way to reconcile Anthropic having made $5 billion in lifetime revenue as of March 9, 2026, $14 billion in annualized revenue on February 12, 2026, and $4.5 billion in revenue for 2025.
No horizon Evidence as of
Limitation: The result depends on assessing the authentic reported statements with their qualifications.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Forecast
“Coding LLMs are popular only because of their massive subsidies and corporate encouragement, and in the end will be seen as a useful-yet-incremental and way too expensive way to make the easy things easier and the harder things harder, all while filling codebases full of masses of unintentional, bloated code.”
Our reading: Coding LLMs are popular only because of massive subsidies and corporate encouragement, and in the end will be seen as a useful-yet-incremental and far too expensive way to make easy things easier and hard things harder, while filling codebases with bloated, unintentional code.
No deadline stated Evidence as of
Limitation: The research tests narrower settings than the claim.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“In simpler terms, Anthropic is cherry-picking four-week windows of API spend — ones that are pumped by big announcements and new model releases — and annualizing them.”
Our reading: Anthropic cherry-picks four-week API spend windows boosted by launches and annualizes them.
No horizon Evidence as of
Limitation: Coinciding announcements and launches are not sufficient evidence of cherry-picking.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Worse still, LLMs allow shitty software engineers that would otherwise be isolated by their incompetence to feign enough intelligence to get by, leading to them actively lowering the quality of code being shipped.”
Our reading: LLMs let incompetent engineers get by, lowering the quality of shipped code.
No horizon Evidence as of
Limitation: The cited study measured immediate learning in a small experimental setting.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“It’s unclear what the actual economic or productivity effects are, other than an abundance of new code that’s making running companies harder.”
Our reading: The actual economic or productivity effects of coding LLMs are unclear, other than an abundance of new code that is making running companies harder.
No horizon Evidence as of
Limitation: METR's study population does not represent all companies.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“Combined with mandates to “use as much AI as possible,” this naturally increases the cost of having software engineers, which — and I say this not wanting anyone to lose their jobs — does the literal opposite of replacing workers with AI.”
Our reading: Model providers reportedly pushing enterprise clients onto API billing, combined with mandates to use as much AI as possible, naturally increases the cost of having software engineers, the opposite of replacing workers with AI.
No horizon Evidence as of
Limitation: Public billing descriptions cannot substitute for employer invoices and staffing records.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“That “fast-growing” part is only possible because both Anthropic and OpenAI subsidize the compute of their subscribers, allowing them to burn $3 to $15 for every dollar of subscription revenue.”
Our reading: OpenAI and Anthropic's fast growth is only possible because they subsidize subscribers' compute, burning $3-$15 per dollar of subscription revenue.
No horizon Evidence as of
Limitation: The available estimate concerns a different period and company-wide finances, rather than matched subscriber costs and subscription revenue [e1].
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“At this point, I estimate Anthropic’s revenue split to be more in the region of 75% API and 25% subscriptions, based on its supposed $2.5 billion in annualized revenue (out of $14 billion, so a little under 18%) in February coming from “Claude Code” (read: subscribers to Claude, there’s no “Claude Code” subscription).”
Our reading: Anthropic's revenue is roughly 75% API and 25% subscriptions, by the author's estimate.
No horizon Evidence as of
Limitation: No matching early-April annualized API/subscription breakdown is supplied.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“This leads to context bloat that absolutely rinses your token budget.”
Our reading: Anthropic's default 1 million token context window causes context bloat that greatly raises token spend.
No horizon Evidence as of
Limitation: The announcement does not establish that every request contains one million tokens [e1].
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Forecast
“Furthermore, TokenMaxxing is exactly the kind of revenue that disappears first.”
Our reading: TokenMaxxing spending is exactly the kind of revenue that disappears first.
No deadline stated Evidence as of
Limitation: Policy announcements do not measure realized revenue.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“It’s also weird that people discussing agents don’t seem to want to discuss that OpenAI’s Operator Agent does not work, that AI browsers are fundamentally broken, or that agentic AI does not do anything that people discuss.”
Our reading: OpenAI's Operator does not work, AI browsers are fundamentally broken, and agentic AI doesn't do what people say.
No horizon Evidence as of
Limitation: One independent evaluation is small and informal.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“Stargate was a lie! The whole time it was a lie! That time that Sam Altman and Masayoshi Son and Larry Ellison stood up at the white house and talked about a $500 billion infrastructure project was a lie! They never formed the entity!”
Our reading: The $500B Stargate project was a lie; the entity was never formed.
No horizon Evidence as of
Limitation: The formation statement is available through a third-party transcript.
Source: AI Is Really Weird
Assessment: UnresolvedSaid
Ed Zitron Observation
“And with that incredibly easy access, only 3% of households pay for AI.”
Our reading: Only 3% of households pay for AI despite easy access.
No horizon Evidence as of
Limitation: Restricting the claim to Bank of America households would support the numerical statement.
Source: AI Is Really Weird
Assessment: SupportedSaid
Gary Marcus Observation
“Right now he is in trouble. People are increasingly concerned about the economics of his company,”
Our reading: As of April 2026, Sam Altman is in trouble, with people increasingly concerned about OpenAI's economics.
No horizon Evidence as of
Limitation: The result depends on reading in trouble as facing pressure and scrutiny.
Assessment: SupportedSaid
Gary Marcus Observation
“Third, there is the well-known tendency of generative AI to be sycophantic; to tell users that their ideas were the greatest ever, even when they are not.”
Our reading: Generative AI tends to be sycophantic, telling users their ideas are excellent even when they are not.
No horizon Evidence as of
Limitation: The decisive March evidence is an independent benchmark, not a peer-reviewed representative survey.
Source: In the Iran war, it looks like AI helped with operations, not strategy
Assessment: UnresolvedSaid
Gary Marcus Causal claim
“First, strategy, more than e.g., writing requisition memos, requires a broad, deep understanding of the world, but generative AI has never had robust models of the world.”
Our reading: Generative AI has never had robust world models, so it is poorly suited to strategy, which requires a broad, deep understanding of the world.
No horizon Evidence as of
Limitation: The research concerns restricted experimental domains.
Source: In the Iran war, it looks like AI helped with operations, not strategy
Assessment: UnresolvedSaid
Gary Marcus Observation
“AI is fine for writing memos; nobody should rely on it to plan a war or to guess its likely outcomes.”
Our reading: AI is adequate for writing memos.
No horizon Evidence as of
Limitation: The selected latest-month period prevents older studies from serving as decisive outcome evidence.
Source: In the Iran war, it looks like AI helped with operations, not strategy
Assessment: UnresolvedSaid
Ed Zitron Observation
“In any case, nowhere are the subsidies higher than those of Anthropic and OpenAI, who use their tens of billions of dollars of funding to allow users to burn anywhere from $3 to $13 per every dollar of subscription revenue to outpace their competition.”
Our reading: Anthropic and OpenAI provide the highest subsidies, letting users burn $3–$13 of compute per subscription dollar.
No horizon Evidence as of
Limitation: The selected monthly observation period is an evaluator interpretation.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Observation
“This means that AI labs are often having to pony up massive amounts of up-front capital on top of their incredibly high ongoing costs.”
Our reading: AI labs often must pay large up-front sums for future compute on top of very high ongoing costs.
No horizon Evidence as of
Limitation: This is an evidence gap, not a finding that prepayments were rare.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“The media campaign worked, and Anthropic was able to close a $30 billion round on February 12, 2026.”
Our reading: Anthropic's Claude Code media campaign succeeded and helped it close a $30 billion round on February 12, 2026.
No horizon Evidence as of
Limitation: Public funding announcements may omit influences discussed privately.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Causal claim
“AI labs are thus forced to raise prices on AI startups, either through tolls (priority processing) or raw cost increases.”
Our reading: Cash and compute pressure forces AI labs to raise prices on AI startups via priority tolls or direct cost increases.
No horizon Evidence as of
Limitation: The dated pricing evidence concerns one provider.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Observation
“As I discussed back in AI Is A Money Trap, there really is no liquidity mechanism for the billions of dollars sunk into most AI companies.”
Our reading: There is no real exit/liquidity path for the billions invested in most AI companies.
No horizon Evidence as of
Limitation: A listing does not mean every existing shareholder can immediately sell.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Observation
“The AI industry has done a great job in obfuscating exactly how brittle its economics really are”
Our reading: The AI industry has obscured how brittle its economics are.
No horizon Evidence as of
Limitation: The research focuses on foundation-model developers rather than every layer of the AI industry.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Observation
“AI labs have, this entire time, been massively subsidizing their own products.”
Our reading: AI labs have been heavily subsidizing their own products all along.
No horizon Evidence as of
Limitation: The OpenAI evidence is an executive statement reported by a news outlet, without a financial breakdown.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Forecast
“The problem is that a subsidy is always a losing proposition, which means that at some point Anthropic and OpenAI will have to massively reduce the amount of tokens that people use on their accounts.”
Our reading: Anthropic and OpenAI will eventually have to massively cut how many tokens subscribers can use.
No deadline stated Evidence as of
Limitation: Published limits do not provide a consistent token allowance across models and workloads.
Source: The Subprime AI Crisis Is Here
Assessment: UnresolvedSaid
Ed Zitron Forecast
“At some point, the incredible, toxic burn-rate of generative AI is going to catch up with them, which in turn will lead to price increases, or companies releasing new products and features with wildly onerous rates — like the egregious $2-a-conversation rate for Salesforce’s “Agentforce” product — that will make even stalwart enterprise customers with budget to burn unable to justify the expense.”
Our reading: Generative AI's burn rate will eventually force price increases or onerous new pricing that enterprise customers cannot justify.
No deadline stated Evidence as of
Limitation: Buyer research was commissioned by a vendor.
Source: The Subprime AI Crisis Is Here