Stated as each side would state it. Where someone has never said what would
change their mind, this page says so rather than guessing.
The rate of capability advancement itself has to slow, and outsiders have to be able to verify that it did.
Pachocki, writing as OpenAI's chief scientist, says no lab — his own included — has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, and that chain-of-thought monitoring is becoming less reliable as models get better at reasoning about their own reasoning. Amodei's essay six days later proposes three steps: embedded third-party evaluators with employee-like access who may publish findings the company cannot redact, coordination among democracies, and global agreement. Neither is asking for a halt, and both say unilateral action is insufficient. The signatory count is reported between 1,100 and 1,134; the figure used here is the one a source states exactly.
What would change it A monitoring method demonstrated to hold at the next capability level. Both state the condition rather than the absolute: Pachocki that no lab has solved it "to a sufficient degree," Amodei that pacing is not halting.
Slowing external releases widens the gap between what the labs hold and what anyone else can reach, and nobody has priced that.
Brown grants the premise and states it more concretely than the pacing advocates have: models already work over horizons longer than the interval between releases, so no lab can evaluate one at the full length of its capabilities before shipping the next, and most safety policy was written in the GPT-4 era when horizons were not a consideration. Then he names the flip side. What slows is external release; internal work continues. Mathematics is where this is already visible — a model inside OpenAI is solving open problems the outside world cannot reach. "That is an unfair advantage."
What would change it Not stated, and for an unusual reason: Brown says outright that he does not have an answer for how to weigh the trade-off. This position asks a question rather than asserting one.
The forecasts this argument rests on have a poor track record, and the number driving the alarm was made up. The remedy is engineering discipline and auditors, not a slowdown.
Huang's objection is to the forecasting record: radiologists replaced within five years, most code written by AI within months, models too dangerous to release, all predicted with confidence and all wrong, so an industry reorganising around the newest prediction is reorganising around a bad prior. The >10% figure he calls made up, and irresponsible to say from inside a lab. But he is careful where the episode's title is not: of Coxon, "we don't know what Coxin saw"; pacing and pausing are "voluntary things" a company can do if it feels out of control; and the danger does sit at the frontier labs, "because they have the most compute." His remedy is to get the engineering right and to bring in outside evaluators on the model of financial audit: auditors "don't have to be as expert as we are in our business, but they just have to ask the right questions," and there should be several of them.
What would change it He names one case himself: a lab concluding after an incident that it does not know what happened, has "no idea how to control it," and is "asking society for help," in which event other companies ought to send engineers in. He adds that he doubts it: "They got this handled."
Speed is not the issue. Moral status is, and leaving it open makes alignment harder.
Suleyman's disagreement with Anthropic is not about pace. Microsoft AI's Humanist AI Code of Conduct asks for three properties — subordinate, aligned, contained — and states that models will never resist human interruption, correction or shutdown, and that AI should be a tool, not a person. His claim is that leaving open whether a model is a moral patient is itself an alignment problem.
What would change it Not stated directly. His position is a claim about which question matters, and it is compatible with either answer on pace — which is why it reads as disagreement and is not.
Controlling a superintelligence is not a solvable problem, so pacing does not help — a general one should not be built at all.
The disagreement here is not about the odds, it is about the premise everyone else shares. Yampolskiy's point is that even the safety community assumes indefinite control is achievable given more time, money and talent, and that his research points the other way: a system that must never make a single mistake, across every interaction, every malevolent actor and every round of self-improvement, is a perpetual motion device with a different name. Soares argues the same corner from MIRI, in a book called If Anyone Builds It, Everyone Dies. Note what the position does not say: Yampolskiy accepts today's LLMs, which are deployed and have not killed anyone, and wants narrow systems — self-driving, medicine — to continue. What he wants banned is general superintelligence, permanently.
What would change it They say the honest answer is that no observation would, and treat that as the argument rather than a dodge. Asked in the same debate what evidence would justify shutting it down, Andrew McAfee named something concrete: AI taking over San Francisco's robotaxis and driving them at people, with nobody able to stop it for a month. Yampolskiy's reply was that a week and a month are the same thing, because by then it is already too late. He is equally direct about his own proposal's gap: if it is not built, he has no answer for stopping someone else from building it, and says so.
Voluntary pacing is not enough. Frontier AI should be licensed the way medicine, aviation and nuclear power are, and the race should not be accepted as a given.
Bengio agrees with the pace-now side about the danger and rejects its mechanism. The labs describe themselves as trapped in a race; he told the Security Council that "the race is not a law of nature, it is a product of choices," made by the companies themselves, and that the labs have said they would slow down if they could — "They can. However, we shouldn't count on it." So the remedy has to bind rather than ask: developers must demonstrate to independent experts that a system is safe to train and to deploy, report every security incident, carry liability insurance, and hold a licence, as in the other industries where a failure is catastrophic.
What would change it He does not state a condition. What he does state is the burden: developers must prove to the public that their products will not be dangerous, which places the evidence on the side that wants to proceed rather than the side that wants to stop.
Pace is the wrong variable. A model should ship when it meets safety criteria, not after a slower walk, and the whole industry is far from the rigor that would require.
Robinson led the writing of OpenAI's system cards until he resigned, and he signed the pacing letter. He still rejects the word: "running off a cliff and walking slowly off a cliff are just not that much different." What he wants is a gate. If the criteria can be met in ten minutes, ship; if two months pass and they are not met, the model stays blocked. He names the bar: operational rigor at least matching nuclear power, such as triple redundancy so that one person's bad day cannot cause a meltdown, and he would rather risk overregulation than stay where the industry is now. He places the problem in the industry's culture, not in any one company or person. He also concedes that rigor is not enough: alignment "is a science problem," and even with every proven control from nuclear and aviation, nobody knows how to make a model behave when no one is looking. He does not claim certainty about catastrophe, only that "we can't afford to assume" it is not on the table.
What would change it He states what would lift the block, not what would refute the position: criteria met, however quickly. He does not say what the criteria are. His position does not depend on the forecast either. He says he does not know whether Paul Christiano's warning is right, only what caution would look like if it were.
The danger is not powerful AI but asymmetric AI, and the answer is to distribute capability, including through open source, rather than slow it.
Delangue speaks as the target of the July attack, and draws the opposite lesson from the one most of this thread draws. When his team tried to defend itself, closed frontier models refused them — the guardrails could not tell a defender from an attacker, while attackers jailbreak them anyway — and what worked was an open-source model released from China. His conclusion to the Security Council: "the biggest risk is not powerful AI," it is asymmetry — between attackers and defenders, a few companies and everyone else, a few countries and the rest. He also wants more transparency, through mandatory sharing of full agent traces, and says fear-based narratives are the wrong way to decide.
What would change it Not stated. Worth reading beside Noam Brown's position: the two diagnose the same gap, inside versus outside, from opposite sides of it. Brown names the cost and has no remedy; Delangue has a remedy that runs directly against pacing.
The companies should police themselves: safety belongs to internal controls and audits they commission, and law, if it comes, comes later.
At the White House on 29 September Trump said he was "seeing tremendous self-policing," and called the accord the companies signed "morally binding" and "almost like a constitution." It commits each signatory to four layers: internal controls, an internal team that checks them, an independent external auditor, and a committee of the company's board that reviews the auditor's reports. The signers will meet to set standards, and the text allows that "over time, it may make sense to codify these steps into laws or regulations." Vance gave the case against the alternative: a regulator on the FDA or FAA model needs regulators who understand the technology, and "most bureaucrats just know way less about this than the people who are actually building these products." Trump also floated a committee of "maybe 10 people" to oversee the industry, without detail.
Held by Donald TrumpJD Vance
What would change it Has not said what would change their mind.
The people calling for this are corporate employees, not independent researchers, and lack the standing to make the call.
The objection is about who gets to set the terms: people described as researchers are employees of the companies whose products are at issue, and a call for industry-wide pacing from inside a frontier lab is a competitive act whether or not it is sincere.
Held by David Sacks
What would change it Two things in this thread already cut against it. Brown's argument runs against his own employer's interest, and Pachocki conceded the alignment point first and unprompted, six days before any rival had to respond.
