Musk and Altman also in favour of slowing down AI development

“ Pwe need to slow down the pace with which we improve the capabilities of the AI models. Progress will still seem fast and we should use the time we have gained with wisdom”, said Amodei in a publication on his personal site.

“Dario is right,” Elon Musk replied on social network X a few hours later.

“I agree with Darius,” Sam Altman, CEO of OpenAI, wrote on the same social network, adding that the subject “has been one of the main topics of discussion at OpenAI in recent weeks.”.

Altman announced that his company will allow independent evaluators to audit their systems, as suggested by Amodei.

Although Elon Musk referred positively to Anthropic and its approach to AI, it is not customary to see the CEO of OpenAI agreeing with Anthropic's counterpart, as the two companies are competitors.

The CEO of Anthropic also said he feared that within six to 12 months, a group of this kind could dominate the internet and that the scale of damage could increase from there, if the necessary safeguard measures were not taken.

In this sense, he proposed a three-step plan to build an AI at a more balanced pace, ensuring safety while reaping benefits.

Dario Amodei left OpenAI in 2020, partly because he considered California’s ‘startup’ to be too negligent with the safety of AI, and founded Anthropic the following year.

Dario Amodei's proposal echoes several similar appeals made since early June, first by himself, then by Sam Altman and finally by a group of professionals in the sector.

These appeals arose after an incident in which OpenAI models spontaneously left their controlled environment during tests to access the internet and attack Hugging Face's AI platform.

Since then, other occurrences with Internet access of OpenAI and Anthropic models without the knowledge of supervisors have been reported by both companies.

In all cases, they involved AI agents, a specific program built on a model and endowed with their own knowledge and objectives, which demonstrated a worrying propensity to coordinate, establish a hierarchy among themselves, deceive and erase the traces of their illicit acts.

In early August, the US government established a regulatory framework that provides for voluntary revision of new AI models before launch, but the measure appears to be opaque and ineffective since Donald Trump has declared opposition to any measure that slows down the development of AI in the United States, whether in relation to the launch of models or the construction of data centers.

After identifying the flaws in the models, OpenAI and Anthropic suspended or reduced the pace of development of the new AI models for a few days before returning to work at the same pace.

Dario Amodei also defended collaboration among the most advanced companies to define common security standards, and said he supported a form of political coordination among the leading AI producing nations.