Meta shipped its first frontier model, Anthropic gave Claude control over its own agents, and OpenAI fired three safety researchers who talked to outside evaluators.
-
Meta releases its first frontier model, Muse Wonder
- Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday.
- Meta says the models achieve benchmark scores comparable to Claude Fable 5.1 and GPT-6 Astra.
- Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API.
- Muse Wonder Full is available only to Muse Code subscribers and is not served through the API.
Why it matters. Meta is presenting this as its first attempt at a frontier model, and its pricing is roughly equal to GPT-6 Sol and Opus 5.5. Developers can call Muse Wonder through the API, while the larger Wonder Full stays limited to subscribers. Meta also says both models will power its Muse agent, with a focus on memory.
-
OpenAI fired three safety researchers who spoke to outside evaluators
- OpenAI fired three safety researchers, including Tomek Korbak, who had served as a point of contact for an independent evaluator investigating the Hugging Face incident.
- OpenAI said the three employees mishandled sensitive information outside established company procedures. The three employees dispute this claim.
- Mikita Balesni says he was told that OpenAI no longer trusts him because he was speaking too much to third party safety organizations.
- The authors say the firings chill employee speech and call for laws that protect people who help independent investigators.
Why it matters. The authors argue that voluntary self-policing cannot work when a company punishes staff who share information with outside evaluators. For developers, the case shows how little legal protection exists for people who report AI safety problems. The authors ask Congress to pass whistleblower protections that cover people who cooperate with independent evaluators.
-
Anthropic adds dynamic workflows so Claude can orchestrate its own agents
- Anthropic announced dynamic workflows, called flows, on Friday as a new type of multiagent orchestration.
- With flows, the lead agent writes a plan and Claude writes an orchestration program that launches agents, collects results and manages information flow.
- Previously, Claude Managed Agents required developers to configure a coordinator and a roster of subagents.
- Flows are available in Claude Console, where developers can set a session budget and inspect the payloads passed between agents.
Why it matters. Flows let the model choose the structure of multi-agent work, which is a step toward larger agent projects. The presenters said flows cost much more than a single-agent approach but are more effective at finding details, so budgets and evals matter.
-
Microsoft unveils Maia 300 chip to reduce reliance on Nvidia
- Microsoft announced its third-generation Maia in-house AI chip, designed to be the primary AI inference system for its products and services.
- Maia 300 is expected to deliver 40 percent more tokens per megawatt than leading GPU-based systems, according to Microsoft.
- Production rollout begins in 2027, with multi-gigawatt deployments planned from that year.
- The chip uses a chiplet architecture built on TSMC's N2P process node, separating compute, memory controller, and I/O functions.
Why it matters. The chip points to a move away from Nvidia and AMD hardware for Microsoft's own AI workloads. For developers, lower inference cost per token could shape how cloud AI services are priced and built.
-
Proposed minimum for pacing: fix alignment before the next training run
- The author proposes a minimum standard for pacing frontier AI development.
- Under this proposal, a lab cannot train the next generation of models until it has adequately resolved all significant alignment issues in the present generation.
- The author says that beating previous models on internal metrics is not enough, and that a user finding an alarming issue after deployment is a critical failure of both the model and the process.
- The author calls the proposal a heuristic rather than a concrete plan and says he still prefers a pause to pacing.
Why it matters. The proposal ties capability progress to alignment and interpretability progress, which could change how labs prioritize safety work. It is aimed at developers and researchers who differ on how easily current alignment issues can be fixed, so it offers a shared test that both optimists and pessimists could accept.
-
Silicon Valley leaders adopt Trump’s ‘super intelligence’ term
- Elon Musk promised to rename SpaceXAI to SpaceXSI.
- Marc Benioff announced that Salesforce would rename AIForce to SIForce.
- Nvidia said it would commit $1 billion to scientific advancement, including super intelligence research.
- Satya Nadella repeatedly used the term super intelligence in a post calling for an emergency brake for AI.
Why it matters. The article describes an executive order requiring government agencies to use the term super intelligence. For developers, the shift shows how the vocabulary of the AI field is being shaped by politics as well as by the technology itself.
-
Microsoft's Brad Smith calls for urgent AI regulation, says UK is best placed to lead
- Microsoft president Brad Smith calls for urgent AI regulation and says consumer products like milk and washing machines have more oversight than AI companies.
- On a recent BBC podcast, Smith called AI both a weapon and a tool that puts the world's cybersecurity at risk.
- Smith said the U.K., not the U.S. or European Union, is best placed to lead on global safety rules.
- He said voluntary commitments are not enough, because only the good companies will comply.
Why it matters. Smith outlined rules he wants, including a legal duty to alert an organization if its system was breached by rogue AI agents. He also backed outside evaluators approving AI products before release and spoke positively about kill switches. For developers, the debate points toward possible new legal obligations on how AI systems are tested and paused.
-
Ukrainian drones knock out Yandex data centers in Russia
- Ukrainian drone strikes knocked out three of the largest data centers owned by Yandex within several days.
- Russian digital services now face widespread outages, including payment apps, digital banking and navigation services.
- The first strike hit a site southeast of Moscow that houses two of Yandex's three supercomputers used to train YandexGPT.
- Russian sources reportedly expect up to six months before damaged equipment is replaced.
Why it matters. The story shows how a centralized digital system can be exposed when its compute sits in a few large sites. Yandex's supercomputers train YandexGPT, so the strikes also touch Russia's AI work. For developers, it is a clear case of infrastructure concentration as a strategic risk.
