Artificial intelligence safety: three initiatives in the same day

Three news items published in recent hours converge on the same subject, observed from different angles: the safety of the most advanced artificial intelligence systems.

Google DeepMind has published, within the safety report for the Gemini 4 Argon model, what the company describes as the first holistic empirical assessment of the automation and acceleration of machine learning research and development within its own operations. According to the company, the assessment found no signs of uncontrolled acceleration in this area. The news currently comes from a single source, the company’s own statement, which should be regarded as the position of the interested party: there is no independent confirmation of the methodology or the results described in the report.

On October 13, in parallel, more than 400 scientists signed a request addressed to companies developing artificial intelligence: to keep disconnected from the internet, at least in the initial phase, models capable of autonomously carrying out cyberattacks. In this case too, the news currently comes from a single source, the researchers’ open letter; no independent confirmation is available regarding the exact number of signatories or the companies targeted by the appeal.

On the funding front, OpenAI donated 5 million dollars to the Simons Foundation for a program - in which Anthropic also participates - dedicated to verifying mathematical proofs generated by artificial intelligence. The overall program, supported by both companies, is worth 15 million dollars in total. This information too currently comes from a single source, the joint statement from OpenAI and the foundation; there is no independent confirmation of the exact amount of funds contributed by each of the two companies.

The three initiatives do not stem from a single coordinated effort, but they describe a sector in which, on the same day, a company publishes a self-assessment of the risks posed by its own models, a scientific community calls for operational caution measures, and two of the industry’s largest companies fund a program designed to verify the reliability of the results produced by those same models.