OpenAI And Devs Unite To Chart AI's Next Unknown Horizon
By 813 Staff
Under the hood, a significant change is emerging — OpenAI And Devs Unite To Chart AI's Next Unknown Horizon, according to Google DeepMind (@GoogleDeepMind) (this afternoon).
Source: https://x.com/GoogleDeepMind/status/2052011542707630461
The European Union’s AI Office, in a regulatory brief circulated to member states late last week, flagged Google DeepMind’s latest partnership as requiring enhanced scrutiny under the bloc’s General-Purpose AI Code of Practice. That partnership, officially announced on May 6 via a tweet from @GoogleDeepMind, involves a collaboration with the developers of an undisclosed frontier model to explore what the lab called “the next frontier of” reasoning and autonomous agent behavior. While the tweet itself omits the name of the partner, internal documents obtained by 813 Morning Brief indicate the recipient is Paris-based startup Mistral AI, with which DeepMind has had exploratory talks since late 2025. Engineers close to the project say the joint work centers on a new architecture that mixes large language models with real-time tool use, essentially giving an AI system the ability to browse the web, execute code, and control APIs without human hand-holding at every step.
The rollout, however, has been anything but smooth. Regulatory filings show the EU’s AI Office requested additional safety documentation from both companies on May 3, two days after a well-timed leak of internal benchmarks suggested the hybrid model performed unexpectedly well at breaking cryptographic puzzles—a capability not included in its original training objectives. Google DeepMind declined to comment on those benchmarks, but a spokesperson confirmed the company is “fully cooperating with all regulatory requests.” The timing matters: the EU’s AI Act imposes tiered obligations on “high-impact” general-purpose AI systems, and any model that can autonomously interact with external infrastructure automatically triggers a systemic risk designation. If the unnamed model indeed clears that bar, both Mistral and DeepMind could face real-time reporting obligations and mandatory independent audits before any commercial deployment.
What happens next remains uncertain. The partnership’s first public milestone is a closed-door demonstration tentatively scheduled for the AI Summit in London on June 12, where engineers plan to show the system completing a multi-step hiring workflow—from scraping job boards to drafting offer letters. Privacy advocates have already flagged concerns about data provenance in that demo. For now, the industry’s attention stays fixed on Brussels: the EU’s AI Office has until June 1 to decide whether to open a formal investigation. If it does, this collaboration will become the highest-profile test case yet of how Europe intends to police the next generation of autonomous AI.
Source: https://x.com/GoogleDeepMind/status/2052011542707630461
