GLM has built its own inference infrastructure in-house. Specific specs or performance figures have not been detailed yet.
A model provider owning its inference stack could shift API pricing and latency, which matters when developers choose which model to build on.
OpenAI shared a framework for tracking, investigating, and disclosing model misalignment. Alongside it, the company released six reports of unexpected or concerning model behavior.
Understanding how misalignment is tracked and disclosed gives developers a reference point for designing safety monitoring and risk checks in their own systems.
Newly unsealed court filings reveal a Microsoft executive privately called OpenAI's data practices "theft," while both companies scraped paywalled New York Times content to build datasets. Internal emails show the companies feared this practice would gut news publishers in a "doom loop."
This exposes legal and reputational risk around training data provenance that directly affects licensing deals and data pipeline decisions for anyone building on these models.
Claude Code has revamped its Projects feature so users can run multiple AI agents together in the cloud, with a shared memory, goals, and library of files and artifacts. A coordinator directs parallel threads, each its own Claude Code cloud session on a separate branch and repo copy, with overlapping work resolved as a merge conflict like any other PR.
This shapes how developers structure parallel agent workflows and manage merge conflicts within existing PR-based review processes.
Researchers at Google and Google DeepMind have launched the DeepMind Institute (DMI) to study and debate the safe development of AGI and its societal impact. The institute will address not only technical development but also jobs, economic policy, human values, institutions, and governance, with plans to expand participation from outside researchers, the humanities, arts, and government.
This signals AGI safety and governance work becoming a formalized internal research function at a major AI lab, which matters for teams tracking how policy and ethics considerations may shape future model releases and access.
A Hacker News post highlights research aiming to push ternary LLMs beyond the 1.58-bit limit. No further details or numbers were shared.
Developers working on model quantization and inference cost reduction should watch this attempt to surpass ternary LLM compression limits.
OpenSpec is presented as a lightweight and configurable AI spec framework. It appeared on Hacker News, drawing attention from developers.
Developers evaluating tools for defining and managing AI specs may want to check out this new lightweight option.