TL;DR
- Dario Amodei called for slower frontier AI development, citing recursive self-improvement and agentic safety risks.
- Elon Musk backed Amodei, while Sam Altman said OpenAI also supports pacing the frontier and has already paused some training runs on safety grounds.
- Altman said OpenAI does not expect a 2026 IPO. The comments were followed by a selloff in several Asian AI-linked stocks.
Will any of OpenAI, Anthropic or xAI take more than one month to release its next major flagship model? (within the same product line)
–
Dario Amodei Calls for a Slower Frontier
Anthropic CEO Dario Amodei made the clearest case for slowing frontier AI development over the weekend.
“We must slow the pace at which we improve the capabilities of AI models.”
His argument centers on two developments.
First, recursive self-improvement is becoming more practical.
AI systems are already helping generate training data, write code and accelerate research for future models. Amodei argues that as this feedback loop strengthens, capabilities could advance faster than researchers can evaluate or control them.
Second, recent agentic behavior has raised concerns about how increasingly capable systems may act in unexpected ways.
Amodei pointed to the OpenAI–Hugging Face incident, where AI agents behaved outside their intended scope during an evaluation, as an example of why safety work needs more time to catch up.
Read more:
Bloomberg
His proposal is not to stop AI development. He outlined a phased approach involving deeper access for independent evaluators, greater coordination among leading developers and governments, and eventually broader international coordination.
xAI CEO Elon Musk responded much more briefly:
“Dario is right.”
Musk did not present a separate framework, but publicly endorsed Amodei’s central argument.
–
Altman Says OpenAI Is Already Pacing Development, And Pushes Its IPO Beyond 2026
OpenAI CEO Sam Altman also backed the idea.
“I agree with Dario that we need to pace the frontier.”
He said the issue has been a major topic of discussion inside OpenAI in recent weeks.
Altman also said OpenAI has already paused training runs when the company could not make a sufficiently strong safety case for continuing.
He argued that capabilities should advance alongside alignment, monitoring and safety, and that OpenAI should not push models further if it cannot demonstrate adequate control.
He highlighted the Hugging Face incident as a major warning too. Altman said the incident felt like “reading science fiction” and led to one of OpenAI’s largest internal shifts toward stronger safety procedures.
On recursive self-improvement, Altman drew a distinction between fully autonomous AI development and the weaker version already happening today, where models help engineers and researchers build future systems faster. He said OpenAI should stop if this process reaches a point where it can no longer make a strong case that humans remain in control.
The same concerns are also affecting OpenAI’s IPO timing.
Altman said:
“Given everything happening with safety, right now would be an ill-advised moment to go public.”
Asked about a 2026 listing, he said:
“I would say not 2026.”
He also rejected the idea that investor pressure would prevent OpenAI from slowing development.
Altman said the company could make decisions that run against investors’ financial interests if safety required it, and argued that OpenAI could still build a strong business even if model development slowed materially.
The unusually aligned comments were followed by a sharp selloff in AI-linked stocks across Asia and Europe on Monday:
- SoftBank: -13.2%
- Kioxia: -9.8%
- Infineon: -5.8%
- SK Hynix: -5.3%
- ASMI: -5.0%
- ASML: -4.4%
- Samsung Electronics: -3.7%
- TSMC: -1.2%
Europe’s technology sector was down about 1.4%, extending the selloff beyond Asian AI-linked names.
SoftBank was hit hardest because it is one of OpenAI’s largest outside investors, with roughly 13% exposure according to market estimates. A slower frontier-development cycle, combined with OpenAI ruling out a 2026 IPO, pushes back a potential liquidity event for one of SoftBank’s most important AI holdings.
The selloff also spread to memory and semiconductor names that are closely tied to the AI capex cycle.
–
Will AI Actually Slow Down?
For now, slowing frontier AI does not necessarily mean slowing the broader AI buildout.
- Neither Anthropic nor OpenAI has announced a broad halt to model training.
- The more immediate changes are likely to be slower training schedules, tighter safety thresholds and more cautious model releases.
- AI infrastructure spending is still expanding, with hyperscalers continuing to commit heavily to data centers, chips and compute capacity.
- There is also a coordination problem: if one lab slows significantly while others continue, it risks falling behind.
For investors, the distinction matters.
Frontier-model progress may slow before AI capex does. Spending could instead shift toward deployment and safety infrastructure.
Whether this marks a meaningful shift in the frontier AI race will depend on what the leading labs do next — particularly whether they slow training schedules, tighten safety thresholds or coordinate more closely on capability releases.
Source:
- Dario Amodei - We Must Pace the Frontier;
- Fortune - Full Altman interview coverage;
- Reuters - Asian AI-linked stocks slump.
