Two themes stand out: longer-context models are getting real product names, and infrastructure companies are still pulling in large checks. At the same time, Washington and the G7 are treating AI access and safety as policy issues, not just lab issues.
Model releases & updates
- DeepSeek-V4 is previewing million-token context, which raises the ceiling for long-document work and agent workflows that break on shorter windows.
- GLM-5.2 is being positioned for long-horizon tasks, which points to more models optimized for multi-step work instead of single-turn chat.
- OpenAI’s reasoning model is being used to help physicians diagnose rare childhood genetic diseases, which shows frontier models moving into higher-stakes clinical support.
- OpenAI’s near-autonomous AI chemist improved a medicinal chemistry reaction, which suggests the lab-to-model loop is starting to produce concrete chemistry results.
Money & moves
- Baseten is reportedly raising $1.5 billion just months after its last mega-round, which is another sign that inference infrastructure is still getting premium pricing.
- OpenAI is adding senior leaders ahead of an IPO, which looks like preparation for a more formal public-market story.
- Amazon wants to sell more of its AI chips directly, which puts more pressure on Nvidia’s hardware lock-in and gives buyers another negotiating lever.
- Elastic agreed to buy CRV-backed DeductiveAI for up to $85 million, which shows larger software vendors still buying AI capability instead of building everything in-house.
- Yann LeCun called xAI a “failure” and warned about a “big bubble explosion,” which adds fresh skepticism around current AI valuations.
Policy & governance
- Anthropic asked for regulation, and Washington went further, which means U.S. policy may now move faster and broader than the companies want.
- OpenAI, Anthropic, and Google joined G7 talks with Trump and other leaders, which shows frontier AI has become a geopolitical input, not just a product category.
- China pushed AI safety proposals as the G7 wrapped without Beijing, which underscores how split the major powers remain on who sets AI rules.
- Bernie Sanders unveiled a $7 trillion plan to give Americans control of the AI industry, which puts ownership and redistribution at the center of the next U.S. AI fight.
- World leaders want American AI but not a U.S. kill switch, which is a direct warning for exporters building state-level and enterprise deals outside the U.S.
Research
- Reliability without Validity argues that LLM-as-a-judge systems can agree with each other while still being biased or invalid, which is a reminder that benchmark scores can flatter the wrong models.
- Beyond Static Leaderboards pushes predictive validity for LLM agents, which matters because static rankings often miss whether an agent actually performs in real tasks.
- OpenAI’s rare-disease work and its chemistry result both point to applied research that is judged by outcome, not benchmark vanity scores.
- Actionable Activation Directions proposes activation-space methods to detect and mitigate emergent misalignment across model families, which gives labs another tool for model safety work before deployment.
