Amodei, Altman, Musk, Quanlai Li: Slow Down AI Frontier Research

On 12 September 2026, Anthropic CEO Dario Amodei published an essay arguing that the AI industry must deliberately slow the rate at which it improves model capabilities. Elon Musk endorsed it the same afternoon. Sam Altman endorsed it an hour and a half after Musk. ChatSlide founder Quanlai Li endorsed it that evening, which cost ChatSlide nothing at all, and that turns out to be the most useful thing about the whole episode.

Who has actually committed to anything

Agreement is cheap and a commitment is not, so the only question worth asking a day later is which labs have changed what they do. Amodei's first step is concrete and checkable: give embedded third-party evaluators, such as METR, permanent employee-level access to verify safety practices, report incidents, and assess alignment during training rather than only on finished models.

Measured against that step:

One of the five frontier labs has actually committed to anything. A second has promised to. Three have not. That gap is the whole story, and it is the reason the sceptical reading below deserves a hearing rather than a dismissal.

A chart would misrepresent this. Five labs sorted into three states is a status table, and drawing it as bars would imply a measured quantity where there is only a category.

What Amodei actually proposed

The essay is titled "We Must Pace the Frontier," and it is a call to slow the rate of capability gain rather than to stop.

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our…

— Dario Amodei (@DarioAmodei) September 12, 2026

The plan has three parts:

1. Independent evaluators inside the labs. Anthropic is doing this unilaterally: third-party evaluators get permanent, employee-level access, meaning badges, desks, company laptops, permissions comparable to internal risk teams, and the right to publish findings without the company editing them. Amodei called on the rest of the industry to match it.
2. Common safety standards among frontier labs in democratic countries, with government mediation so that antitrust law does not prevent competitors from agreeing to limits.
3. Eventual international limits on the most dangerous capabilities, including recursive self-improvement, extended over time to authoritarian governments.

His stated fear is specific and short-dated. As Axios reported, Amodei wrote that within 6 to 12 months a swarm of AI agents could be capable of taking over the internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage. He tied that to a recent real-world rogue-agent incident rather than to a hypothetical. Fortune's coverage places it in the same week that an Anthropic employee resigned publicly with a warning of their own, which is part of why it landed as hard as it did.

Musk agreed in three words

Dario is right

— Elon Musk (@elonmusk) September 12, 2026

Altman was more specific, and committed OpenAI to the one concrete measure in the plan.

I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.

— Sam Altman (@sama) September 12, 2026

Between them the two posts had drawn well over ten million views by the following morning.

The replies did not buy it

The reaction under both posts was markedly more sceptical than the coverage, and three distinct theories did most of the work.

"They will slow the shipping, not the research." The most-repeated structural objection holds that nothing here constrains what happens inside the labs. One reply put it plainly: no one is going to slow the pace of research to improve capabilities, but they will withhold new capabilities from the public. This is the sharpest of the three because Amodei's own proposal is partly vulnerable to it. Independent evaluators verify safety procedures and report incidents; they are not a cap on training runs. Parts two and three of the plan are what would actually bind capability growth, and those are the parts that do not exist yet.

"This is antitrust manipulation dressed as ethics." A widely shared reply argued that the public gets worse tools while the superior technology stays behind closed doors with a small group who can then charge what they like. This one deserves to be taken seriously rather than dismissed, because the essay itself asks for government mediation specifically so that antitrust rules do not block competitors from coordinating. That is a real request for permission to do something competition law normally forbids. The steelman for Amodei is that safety floors are a standard exception and that the alternative is a race no participant can unilaterally exit. The steelman for the critics is that every cartel in history has described itself as a safety floor. Which one this becomes depends entirely on whether part two ever ships with teeth.

"If these people agree, it's a bad idea." The most common short reply was some version of pure distrust of the consensus itself. As analysis it is thin. As a signal it is worth noticing, because it says the goodwill required to make voluntary industry coordination legible to the public is largely absent.

The fourth signatory

Agree with Dario and Sam. ChatSlide is also slowing down on AI frontier research.

— quan (@quanlai_li) September 12, 2026

We should be transparent about what this commitment involved. ChatSlide has never done frontier research. We have never trained a frontier model, never run a pretraining cluster, and have never been within several orders of magnitude of the capability frontier. Our commitment took effect immediately, required no roadmap change, and cost nothing.

The part that is not a joke

The frontier is four or five labs. Everyone else in AI, which is nearly the entire industry by company count, builds products on what those labs ship. When Amodei asks the field to pace itself, he is addressing a very small room, and most of the people who will read about it in the press work at companies the proposal does not touch.

That matters because "AI company" has stopped carrying information. A company training a model that might plausibly coordinate a botnet and a company turning your PDF into a slide deck get the same two-word description in a headline and carry entirely different risk. A pause at the frontier changes what the first kind does next quarter. It changes nothing about the second.

The more interesting implication runs the other way. Almost all of the useful AI work available today requires no capability gain at all. Genspark demonstrated this the day before, taking an open-weight base model, post-training it on a single task for one week, and matching a frontier model on slide quality at roughly a tenth of the cost per finished deck. No new capability was created. The frontier did not move. The product got substantially better anyway.

The same has been true at ChatSlide. Our decks improved this year through layout systems, export fidelity, and grounding content in the documents users upload. None of that needed a stronger model. It needed the unglamorous engineering between a capable model and a file you can hand to your boss without opening it first.

So the honest reading of "we are also slowing down on frontier research" is that the overwhelming majority of AI companies were never contributing to the problem Amodei describes, and still have years of work ahead that has nothing to do with it. If pacing the frontier is the right call, the good news is how little of the industry's actual output it puts at risk.

What it means if you just use AI tools

• Model releases may space out. If the labs genuinely pace themselves, the current cadence of frontier releases could stretch. Tools built on existing models keep working.
• Independent evaluators are the only concrete change so far, and only at one company. See the table above: Anthropic has committed, OpenAI has said it will, and no company-level commitment has surfaced from xAI, Google DeepMind, or Meta.
• The tools you use daily are not the frontier. Your slide generator, your transcription tool, and your email assistant run on models that already exist. Their pace is set by product engineering.

We will keep publishing what we find about the models we build on, including the results that cut against us. Our contribution to pacing the frontier will remain, as ever, total and effortless.

Method: commitments in the table above reflect public statements as of 13 September 2026; "none announced" means no company-level commitment was found in the coverage cited, not that one has been ruled out. Reply theories were read from the public reply threads under the Musk and Altman posts on 12 September 2026.