SAN FRANCISCO – Two proposals for keeping artificial intelligence safe were made at one conference keynote this week. Neither of them is a law, and nobody in the 63 minutes on the recording named one.
The two men were not on a stage together and this page will not pretend they were. The Dreamforce keynote on 15 September ran them as separate interviews with one host, with product demonstrations between. CNBC reported it the same way on the day, and CNBC is also where the host’s name comes from: Salesforce’s Marc Benioff. The recording never says his surname. Two of the three guests address him as Mark.
This page reports what each man proposed. It does not say they agreed. Neither was asked about the other by name, and neither answered the other.
Contents
What Anthropic’s chief executive proposed#
Dario Amodei was asked for “your state of the AI industry and pacing”. He answered with an analogy.
“let’s say you’re running a car company and another car company, not yours …”
That other company has a safety incident. The tempting response, he says, is to attack them: “These guys are unsafe. We’re safe.” He proposes something else.
His answer has three steps, which he says come from “the essay”. The host calls it “the essay this weekend on pacing”.
“let’s look let’s look at our own record, right? …” and “Let’s make our practices better.”
“Then let’s organize the rest of the industry and say, what can we do to set standards for everyone? …”
“And then finally, there should be an international component to it.”
The mechanism is companies agreeing standards among themselves. No law, no agency, and the international part is not described.
He also limits it himself.
“the first of which we committed to and the the others of which we’re going to have a dialogue …”
Later in the same interview, answering a different question, he names his own mechanism.
“even if we freeze the technology in place, which we’re not doing, pacing the frontier is not not doing that at all. The technology will keep getting better.”
So his proposal is what he calls pacing the frontier, and it is not a halt. The host’s opening question named pacing too. He does not say pause, slow or stop anywhere in his interview on the recording.
His argument is that the gap is in use, not capability.
“we’re making use of maybe only 5 or 10% of what the possible value of the technology is.”
What Nvidia’s chief executive proposed#
Jensen Huang came on last of the three. The host asked him about “an incredible de debate for the last day on pacing and safety”, a debate, not one man’s argument. The transcript is machine captioning and garbles names throughout, so “de” is either a caption glitch or a false start. The phrase is printed as the transcript renders it.
Huang begins where people do not expect him to.
“safety is paramount in a lot of ways. It’s job one. However, safety is an engineering problem.”
His test sits inside the company.
“if you build a product or a service and you’re not confident in its functionality, capability, or safety, then don’t release it …”
And: “You pace yourself until you are confident you’re releasing something that the market would appreciate.” If it feels out of control, he says, “take a pause and make sure you get it right.”
Then the line that travels.
“The market forces are already there. We don’t need any new laws. We don’t need new regulations.”
He also rejects the idea that speed and safety trade off: “innovation speed and safe products are not it’s a false choice.”
What the third man was asked, which was something else#
Siemens’ Roland Busch was interviewed between them and was never asked about rules. Searched his segment for safety, pacing, regulation, rule and law: zero for all five.
He was asked about AI inside industry, and his answer is about a different problem: reliability on a factory floor.
“This is bringing a probabilistic technology into a deterministic world …”
And: “Hallucination does not really work on the shop floor as you can imagine.”
He takes no position on rules because nobody asked him for one.
He credits Amodei twice, both times as “Dario said it”. Once about how much the models have not yet seen, once about the speed of change. A Salesforce executive in the gap said “Daario was right” about something else. So the segments were not sealed off from each other, even though none of the three men answered another.
Nobody named a law#
The 63 minute recording was searched with word boundaries for law, laws, bill, regulation, regulations, rule, rules, statute, executive order, congress and legislation. It opens mid-sentence, so there is keynote before it that this page has not seen.
Five hits exist and every one is accounted for. “Laws” twice: Amodei’s “scaling laws” about model performance, and Huang’s “we don’t need any new laws”. “Regulations” twice: Huang’s same sentence, and a product demonstration. “Rules” once, meaning business rules in software.
So across the recording no existing or proposed law is named once. Two of the five hits are not about rules at all.
That is the finding. Of the mechanisms actually described, one is an industry agreeing standards with itself. The other is each company judging itself, with the market as the check.
Neither mechanism carries a failure case. Nobody says who decides a company is not confident enough to hold a release, or what happens when the commercial pressure runs the other way. And “market forces are already there” arrives with no mechanism, no example and no instance of those forces failing. Searched the recording for audit, audited, auditable, oversight, accountable, enforce and independent: zero for all seven. The one outside element anybody named, an international component, was never described, and this page cannot say what would run it. Statute, rule and guidance are not the same thing, and neither man proposed any of the three.
What happens next, and what this page could not see#
Amodei’s second and third steps are a dialogue he says is still to come. Others have responded to the essay, though not on this stage. CNBC reported the same day that several named industry figures backed his proposal.
CNBC also reports that OpenAI’s Sam Altman appeared with the same host later that day, in a session not in this recording. By CNBC’s account he said safety and monitoring must come before capabilities. So this page sees Amodei, Busch and Huang, and not Altman.
Near the end of the keynote the chief executive of the staffing group Adecco reached back to Huang’s remark about jobs.
“it was great to hear Jensen saying that uh you know jobs are going to remain in the world …”
The thing to watch is not another keynote. It is whether a standard appears anywhere with a name attached to it. Until one does, what the industry says about its own safety is the whole of the record. A claim is worth what its source is worth, and here is how to check one.
Source: the Dreamforce 2026 keynote, published 15 September 2026, 63 minutes, transcript pulled and read in order. The recording is a third-party upload carrying the note “Livestream courtesy of Salesforce”, and it begins mid-sentence, so the keynote before that point was not available to this page. The CNBC report of 15 September, linked above, is what the claims about Sam Altman and about support for the essay rest on. Amodei’s essay itself was not read. Nobody from this masthead was in the room.