OpenAI Writes the Playbook for Disclosing Misaligned Models

A new framework commits the lab to flagging failures early — even before it can fully explain or fix them.

OpenAI is turning one of its thorniest problems — what to do when a model misbehaves in ways no one yet understands — into a published process. The company's new framework sets explicit criteria and timelines for tracking, investigating, and disclosing instances of model misalignment, including cases where the behavior has not been fully explained or mitigated.

The shift matters because disclosure has historically lagged understanding. Labs have tended to stay quiet until a complete post-mortem was ready. The new framework starts the clock earlier, committing OpenAI to tell the public what it is seeing as it is seeing it, so researchers and users are not left in the dark while an investigation is still running.

More complex cases may take longer to resolve. The framework acknowledges that some failures resist quick explanation, and sets out how OpenAI will handle the gap between detection and understanding — a gap that has quietly become one of the defining risks of building frontier systems.

Cohere and Aleph Alpha Merge Into a Transatlantic Model Maker

The two labs sign a definitive agreement to run as a single company, Cohere, with more than 1,000 employees across North America and Europe.

The deal creates the first foundational AI model developer anchored on both sides of the Atlantic, pairing Cohere's enterprise footprint in Canada and North America with Aleph Alpha's presence in Germany and Europe. Under the unified Cohere brand, the combined company will scale past 1,000 employees and consolidate a rare continental bridge in a market otherwise defined by coastal competition.

Claude Design, Slides, and Docs Move Into Claude Code

Claude Design, Claude Slides, and Claude Docs now work inside Claude Code. Developers can ask for a design review deck or a UI mockup and point it at the actual files and RFCs in a repository, edit the result themselves, or keep iterating in the conversation and share a link when it is ready.

TypeSafe AI Ships Its 'System One' Model, Jev

Jev forgoes text generation entirely. It is built for fast decisions inside software: unstructured data in, type-safe structured results with calibrated probabilities out — a category the company, founded by a ChatGPT co-inventor, calls a "System One" model.

DeepMind Opens an Institute for AGI's Economic and Social Questions

Demis Hassabis and Shane Legg are expanding interdisciplinary research on how AGI could reshape the economy, science, and society. The new DeepMind Institute is meant to spur the discussions they believe are needed to get the next steps right, drawing on two decades of shared work.

Meta Makes the Case for Personal Superintelligence

Meta argues that superintelligence should broadly benefit individuals rather than a few institutions. The company is building personal agents and creation tools — emphasizing privacy and the balancing of power — in a vision modeled on the personal computer and the internet rather than on centralized control.