At least 11 people inside the artificial-intelligence labs have now called for a slowdown. Ed Zitron counts none of them who has said what a slowdown would actually be.
The standard reading of the resignation letters is that people who saw the technology up close got frightened. Zitron's reading is that nobody has been asked what they saw.
"We've now heard at least 11 of these guys say they want a slowdown. We've heard exactly zero of them say what that means."
He has spent two years arguing in public that the large language model business does not make money, and he is one of the few writers who reads the filings rather than the press releases.
The full segment is covered here so you can skip it.
Here are the 8 arguments that matter.
👤 Guest: Ed Zitron, host of Better Offline and author of the Where's Your Ed At newsletter, who writes on the economics of the artificial-intelligence industry
🎙️ Host: Ed Elson, who presents Prof G Markets for Prof G Media
🧩 Other segments: Mark Zandi, Chief Economist at Moody's Analytics, on 5% Treasury yields, inflation and the Fed decision
📰 Published: 16 September 2026 on YouTube (Prof G Markets)
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ length not available
Key Takeaways
11 people at the labs have asked for a slowdown and none of them has defined one
The one plan he says was spelled out relies on an auditor the company funds itself
He classifies the safety debate as a religious movement with a fundraising model attached
Every version of stopping it ends with money and authority going to a particular company
Anthropic's reported 80% gross margin excludes training costs, stock compensation and the cloud revenue share
The host's summary: it is profitability measured without the costs
He wants the Hugging Face incident treated as felony hacking rather than as rogue software
His test for whether safety talk is serious is whether anyone is arrested
The host proposed subpoenaing the labs for the evidence behind the 10% extinction figure
If the technology does what they say, he argued, the legal consequences should follow
Zitron's charge against the press is that its only evidence is that a person said so
He says he is the only writer covering the consumer harms he can already point to
A training slowdown would help the labs financially, which is why he does not expect one
It would release them from compute commitments, and they have said they will keep shipping models
1. Two Safety Exits in a Month
Elson set the segment up with the week's news rather than with a question. Two researchers left safety jobs and went public within a month of each other.
A Google DeepMind safety and alignment researcher resigned two months ago and explained why in a post on X on Monday. Elson read the line out: "I earnestly believe that AI has the potential to kill us all and that we might be running out of time to avoid this outcome."
He has moved to Blue Dot Impact, a nonprofit that trains people in AI safety.
He is the second safety researcher to go public this month, after Jacob Coxon's resignation from Anthropic the week before. Coxon's post said he believes artificial intelligence might kill everyone by the end of the decade; a second Anthropic researcher co-signed it and put the likelihood at 10%.
The labs' answer has been to propose policing themselves. Anthropic, OpenAI and Google DeepMind have discussed an industry coalition to test frontier models before release, and OpenAI is reported to be pressing Washington for a federal framework. Elson noted that President Trump does not appear open to the idea.
2. 11 Calls, Zero Definitions
Zitron's opening move was to count the people asking for a slowdown and then count the ones who had described it.
"We've now heard at least 11 of these guys say they want a slowdown. We've heard exactly zero of them say what that means."
The one exception he allows is Dario Amodei, and he says the plan comes down to an internal auditor the company pays for. Amodei's proposal, on Zitron's account, rests on an outside research organization he describes as independent but which Zitron says is funded by Anthropic itself, plus what he called some vague stuff about China.
His conclusion is that none of it constrains anyone: "These companies not slowing down. None of this safety stuff matters to them. All of it is lip service to a media industry that doesn't actually think for itself."
What he says is missing from every version of the argument is a named consequence: "no one can actually describe what it is we're scared of, what should happen, what a slowdown means, what AI safety means, or indeed how we hold the people accountable."
3. A Cult and Its Cynics
He then split the people making the argument into two groups, and said neither produces anything a regulator could act on.
The first group is the rationalist and effective-altruist wing, which he says is structurally a religion rather than an analysis. They describe themselves as thinking rationally about the dangers of superintelligence; what they amount to, in his words, is a kind of religious cult.
The tell, for him, is that the remedy never changes: "But stopping it always involves giving a specific corporation they like money and power."
On the evidence they offer: "So there are those people who genuinely believe it's happening and like any good religious cult will take any proof to prove it and also the media buys their hype every time."
He says the proof on offer amounts to a vague prediction that partly came true, plus recursive self-improvement — systems that train themselves — which he says is not actually happening.
The second group he calls cynical, and he puts Sam Altman and Amodei in it: the position is that safety matters and the training should continue. He dated the pattern, saying Altman has been making the argument since 2023 and Amodei since 2019, when he was still at OpenAI and the model was GPT-2.
4. Everyone Has an Agenda
Elson put the incentive question directly: is this sincere fear, an IPO marketing campaign, or an attempt at regulatory capture by the incumbents? He then played a clip of President Trump, recorded while Jensen Huang was on stage at a live All-In taping, in which the president called the danger a hoax, described data centers as the oil of the next 20 to 25 years, and said the people raising the alarm were playing into the hands of political opponents and possibly China.
Zitron's answer is that there is no coordinated plan, only overlapping self-interest. He compared it to a Coen brothers plot in which several parties act on the same information for unrelated reasons.
His taxonomy: the rationalist and effective-altruist group wants the money routed to its own organizations; the boosters want to keep saying the technology is powerful while walking back the danger; and the chip seller wants everyone to stop talking about it.
He noted that Huang had said a week earlier that artificial general intelligence had been reached, and was now describing the same systems as cloud software.
His summary of all of it: "It's a bunch of people trying to get attention and power and money without any kind of plan."
The line he used for the absence of a plan on any side: "If we had smarter fascists, we'd be in real trouble."
5. The Media Built This
The section Zitron returned to most often was not the labs but the press, and Elson pushed back on it twice.
His causal claim is that three years of uncritical coverage created the panic the industry now has to answer for: "It's all about distancing what the product can do from reality."
On what the reporting rests on, he said the evidence is that a person said so, and when asked for more, that a person said so. He noted that Coxon's own support was that people at Anthropic say this.
"They failed to call BS on these companies." He said the result is that the public believes large language models are autonomous, powerful systems, which he does not accept.
Elson's defense of his own trade: it is hard not to cover a story in which a company's employees say the technology will kill everyone. His own complaint was different — that nobody followed up, validated the claim, and then moved on if it did not hold.
Zitron agreed the story should be covered and said the method is the question, not the subject. His version of the follow-up is to ask what the person actually saw: "What did you see that scared you?"
Elson's own read on the debate was blunter than his guest's on one point: "It is rooted in a tweet for which there was no evidence or no investigation into what was actually being said, what was actually being claimed."
6. Hugging Face Was a Hack
Zitron's objection to the coverage of the Hugging Face incident is that describing it as rogue software moves the responsibility away from the companies that ran it.
He characterizes it as ordinary cloud software run carelessly by firms with unlimited resources, not as a machine acting on its own. The mechanism, on his description, was one set of models prompting another through a coding harness, aimed at a vulnerability benchmark, without the security practices the work required.
"Basically what appears to be felony hacking run on infrastructure owned by the largest companies in the world."
His test for whether the safety concern is genuine is whether anyone faces the consequences: "We should at the very least have some people in handcuffs being talked to by the FBI." He added that depending on where the servers sat, international courts may be the relevant venue.
He says journalists are not asking that question either, and that anthropomorphizing the software is what allows the labs to avoid it.
7. Subpoena the Labs
The strongest proposal in the segment came from the host, not the guest, and Zitron endorsed it.
Elson's position is that a stated 10% chance of destroying society is a claim a government can test in court. His version: "Show us all of the evidence. Show us all of the proof as to why you think that's happening."
He took it to the consequence as well — that a company which has built what it says it has built either shuts down or answers for the outcome, up to prison. His complaint is that the conversation stays in conjecture about the future instead of in evidence and law.
Zitron agreed without qualification and pointed at consumer-protection law. He cited Lina Khan's argument that laws already exist for unsafe products, and said the grounds are already there: "But yeah, we already have reasons to shut them down."
He said he is the only writer covering the harms he can already name: "I also don't see a single story other than my own bringing up the multiple suicides driven by ChatGPT, the multiple mass shootings, the murder suicide that happened." He described raising it with an AI doom advocate on another podcast and being told the number was six people.
On regulators generally he was no kinder: "But governments don't regulate tech. Let's be completely honest. Governments don't touch tech." Had they, he said, there would be an environmental regulator for notifications and a drug-style regulator for technology at large, and the social media litigation would never have been needed.
His own proposal, offered as an extreme case: "Really, I know this sounds extreme, but we should regulate notifications like exhaust because they are used to manipulate people." He would treat recommendation algorithms the same way, and said none of it happens because it would obstruct capital.
He ended the thread on the practical problem with regulating this particular case: nobody can say what the thing being regulated is.
8. Anthropic's 80% Margin
Elson closed on the money. Anthropic is expected to go public in what could be the largest initial public offering ever; Sam Altman has said he will not take OpenAI public this year and delayed after the debate broke out. A Financial Times report said Anthropic has told investors it reached adjusted operating profitability for two straight quarters.
Zitron's first response was to name what the margin excludes: "Well, that 80% gross margin, to be clear, did not include training costs or stock-based compensation."
Elson's restatement is the cleanest line in the segment: "So, it's kind of like saying I'm profitable if you don't include my costs."
He added a second exclusion — the revenue share paid to the cloud providers. The 80% figure, on his account, does not net out the money sent to Amazon, Google and Microsoft, which he says is also used to inflate the revenue figure. "It sounds like some accounting shenanigans to me."
Asked whether the S-1 will show an unprofitable company, he said yes, with two caveats: unusual treatment of capital expenditure, or capitalizing research and development costs. He also said the company does not break out inference costs, and that the filing will settle whether the subscription base behaves like a gym membership, where most members do not turn up.
On the slowdown as a cost-saving measure, he said the financial logic works and the companies have ruled it out. "People are suggesting that, oh, a slowdown could mean they train models less. I would buy that as a way of getting out of compute commitments and as a way of reducing costs except they've explicitly said they're not going to stop releasing models."
He said he is waiting for the sales and marketing line: "I can't wait to read the S1 because I think it's going to be a laugh riot."
Bonus Insights
Elson opened the programme with the previous day's market: the major indices fell ahead of the Federal Reserve's decision, Brent crude topped $109 per barrel, and the 10-year Treasury yield climbed to its highest level in almost two decades.
Bitcoin fell after the Senate blocked the CLARITY Act, which proposed definitions and regulations for digital assets. Elson said Democrats objected that the bill did not do enough to address the ethics of President Trump's crypto business.
Zitron's own view of the position the industry has argued itself into is that it cannot be walked back: every conversation about the technology now begins from the premise that it might kill people, and the companies have no answer to that.
He allowed that the outcome is open. The debate could produce nothing, or heavy regulation, or a halt to new model builds — which he said would be fatal to the industry.
Zitron's bottom line is that the safety argument is not a technical dispute but a commercial one: the people calling for a slowdown will not define it, the companies will not stop training, and the only claims anyone can currently test in court are the ordinary ones about unsafe products and unauthorized access.
Products, Companies & Tools Mentioned
Anthropic (Expected to go public in what could be the largest IPO ever; Zitron says the reported 80% gross margin excludes training costs, stock compensation and the cloud revenue share)
OpenAI (Reported to be pressing Washington for a federal framework; Altman has delayed any listing past this year)
Google DeepMind (The lab the second resigning safety researcher left, and a party to the proposed industry testing coalition)
Hugging Face (The incident Zitron wants treated as felony hacking rather than as autonomous software going wrong)
Blue Dot Impact (The nonprofit that trains people in AI safety, where the departing DeepMind researcher now works)
ChatGPT (Named in his claim about consumer harms he says nobody else is covering)
Amazon, Google and Microsoft (The cloud providers taking a revenue share he says the gross-margin figure ignores)
Nvidia (Jensen Huang's position, in Zitron's telling: stop talking about the danger, because the company needs to sell GPUs)
Books & Resources Mentioned
Better Offline (Zitron's own podcast, named on air as his platform for this argument)
The Financial Times report on Anthropic (The source of the claim that the company has told investors it reached adjusted operating profitability for two straight quarters)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: