Four systems stories displaced the usual model launch. Anthropic changed what alignment examples teach, AlphaEvolve searched where objective evaluators can cheaply reject bad code, OpenAI documented overlapping Codex controls, and Chrome's local Gemini artifact made privacy-preserving inference visible as a 4 GB endpoint cost.
1. Ethical explanations generalize beyond refusal demonstrations
Anthropic reported that direct examples of refusing simulated agentic misconduct reduced one experimental model's misalignment rate from 22% to 15%. Rewriting responses to include ethical reasoning lowered it to 3%, while a much smaller out-of-distribution advice dataset produced comparable gains.
The scenarios, synthetic data, and automated assessments all came from Anthropic, which says alignment remains unsolved. The controlled result distinguishes explanatory diversity from demonstration volume: principles and varied environments transferred better than many near-duplicate refusals, a hypothesis that can be tested on other models and agent settings.
Sources: Anthropic's Teaching Claude why research
2. AlphaEvolve works where proposed code has a cheap objective grader
Google DeepMind said AlphaEvolve optimized algorithms used in TPU design, Google Spanner, scientific models, and customer workloads. Reported outcomes included 20% lower Spanner write amplification and 30% fewer variant-detection errors for DeepConsensus.
Google and its partners selected and reported the case studies, and several domains provide unusually crisp automated scores. That is also the deployment pattern: confine proposals to code whose speed or correctness can be measured repeatedly, preserve the incumbent implementation, and send ambiguous scientific interpretation to human review.
Sources: Google DeepMind on AlphaEvolve deployments
3. OpenAI's Codex controls overlap rather than trust one reviewer
OpenAI said its internal Codex deployment combines workspace-write sandboxes, managed network allowlists, protected credential storage, enforceable configuration, command rules, and agent-aware telemetry. An approval policy routes selected boundary crossings to a person or the separate Auto-review model.
This is a vendor account of one environment rather than a safety result for every Codex installation. Its architecture avoids making probabilistic review the root control: filesystem scope, identity, credentials, egress, and logs each constrain a different failure path before or after a reviewer approves an action.
Sources: OpenAI on running Codex safely
4. Chrome's local Gemini model turns cloud privacy into endpoint weight
Chrome users found an approximately 4 GB model file associated with Gemini Nano, which powers local features including scam detection and writing assistance. Google said the size can change, low-resource devices remove it automatically, and users can disable on-device AI in settings.
Local inference keeps selected processing off remote servers but transfers costs to storage, memory, updates, and support across every endpoint. A browser feature inventory can connect each downloaded artifact to a policy and device class, making the fleet-wide footprint visible before an automatic update changes it.
Sources: The Verge on Chrome's model footprint · Google's on-device AI controls