Headlines

1. OpenAI and Apple Ship On-Device GPT-5.2 to iPhone 18, Promising Private Agentic Siri
OpenAI and Apple announced that a distilled version of GPT-5.2, code-named “Sovereign,” will run natively on the iPhone 18’s neural engine, enabling on-device long-horizon task planning without cloud round-trips. The rollout begins next month for U.S. users, with a focus on privacy-preserving personal automation. Source: Bloomberg

2. EU AI Act Second Pillar Takes Effect: High-Risk Foundation Models Face Mandatory Stress Tests
On September 27, EU regulators activated the new “systemic risk” regime for large general-purpose AI models, requiring red-team stress tests, incident reporting, and cross-border audits. Several U.S. frontier labs are scrambling to align their European deployments before the October 14 compliance deadline. Source: Reuters

Model Releases

3. Mistral AI Releases Mistral Large 3 with 2.5M-Token Context and Native Code Execution
Mistral Large 3 brings a 2.5 million-token context, a new hybrid sparse attention architecture, and a built-in Python interpreter for agentic coding tasks. In internal benchmarks, it surpasses its predecessor by 18% on long-context retrieval and 12% on multi-step tool use, with European enterprise users as the initial target. Source: The Verge

4. Google DeepMind Opens Gemini 3 Ultra API with Autonomous Research Agent Mode
Gemini 3 Ultra, released in public preview, now includes an “Agentic Research” mode that can independently form hypotheses, search proprietary and web sources, and generate structured reports. Early testers from pharmaceutical firms report that the model reduces literature-review time from days to under an hour, though occasional hallucination remains flagged in the safety card. Source: TechCrunch

Industry

5. Nvidia and Microsoft Announce $30B “Aurora Grid” for 2027 AI Supercomputing
The two companies unveiled a multi-year partnership to build a network of 20 liquid-cooled, gas-turbine-assisted data centers across the U.S. Southwest, targeting 4 exaflops of dense AI compute by Q4 2027. The announcement includes a joint power-sharing agreement with regional utilities to stabilize grid loads during training peaks. Source: CNBC

6. Anthropic Reaches $10B Annualized Revenue, Launches “Claude Enterprise Agents”
Anthropic disclosed that Claude product revenue hit a $10 billion run rate, driven by healthcare and legal automation, ahead of its new “Enterprise Agents” product for compliance-heavy industries. The platform includes audit logs, human-in-the-loop guardrails, and a new SLA guaranteeing a 99.95% uptime for critical agent deployments. Source: The Information

Research Highlights

7. Stanford and Meta Release “VerifierNet” Paper: Self-Correcting Reasoning Loops Improve Factuality by 41% on Long-Form QA
A new preprint from Stanford and Meta demonstrates a hierarchical verifier model that checks each reasoning step against external knowledge graphs, then triggers targeted re-generation only for failed sub-chains. The method, tested on state-of-the-art LLMs, cuts unsupported claims in 10,000-word scientific summaries from 15% to under 2%. Source: arXiv (arXiv:2609.20178)

8. DeepMind’s “MemoryMap” Models Emergent Planning in Agents Without Explicit Reward Shaping
Researchers at Google DeepMind introduced MemoryMap, an architecture that stores compressed episodic memories in a continuous latent space, enabling agents to reuse past trajectories for new tasks. In the NetHack Challenge and a robotic manipulation benchmark, it outperformed state-of-the-art baselines by 34% without any task-specific reward engineering. Source: Nature Machine Intelligence


Compiled from multiple media reports.