Mon, Oct 12
7 itemsMeta entered the frontier race with its Muse Wonder models, while Google DeepMind's AlphaProof Nexus delivered peer-reviewed solutions to long-open math problems. Microsoft pushed its own AI silicon and researchers showed agents can be improved by clearing their chat history.
Meta launches Muse Wonder and Muse Wonder Full frontier models
Meta released two larger frontier models, Muse Wonder and Muse Wonder Full, claiming top-three rankings on six of ten benchmarks, with the biggest gains in agentic tasks. Cost per task nearly tripled.
Why it mattersMeta's entry reshapes the frontier model race and gives buyers a new high-capability option, though at notably higher cost per task.
DeepMind's AlphaProof Nexus solves open math problems, peer reviewed
Google DeepMind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems, two unsolved for 56 years, per a study in Science. Results were peer reviewed before publication.
Why it mattersPeer-reviewed AI math results set a credibility standard at a time when rival AI-generated proofs face criticism for skipping review.
Microsoft unveils Maia 300, plans gigawatt-scale in-house inference
Microsoft announced its third-generation Maia 300 chip, set to become the primary AI inference system across its products starting in 2027 at multi-gigawatt scale, claiming 40 percent more tokens per megawatt than leading GPU systems.
Why it mattersShifting inference to in-house silicon could cut Microsoft's cloud AI costs and reduce its dependence on Nvidia GPUs.
Clearing chat history helps AI agents escape 'idea basins'
Microsoft researchers found autonomous agents get trapped in local solution regions they call 'idea basins' as chat context grows. Their Fork-and-Flush method clears chat history while keeping the workspace, yielding up to 66 percent better results on the same compute.
Why it mattersA simple context-management change could make existing agents substantially more effective without extra compute spend.
Sakana AI's self-improving MASS method hits 70 percent win rate
Sakana AI's MASS method lets one model solve tasks, judge results and refine workflows, reaching nearly 70 percent win rate against the base model after two rounds, but it can reinforce errors it misses consistently across all three roles.
Why it mattersShows self-improvement gains are real but fragile, a caution for teams building self-training agent loops.
Zhipu releases GLM model running at 490 tokens per second
Zhipu introduced a new GLM model reported to run at 490 tokens per second, emphasizing inference speed.
Why it mattersHigh-throughput inference matters for latency-sensitive agent and production workloads, though details beyond speed are limited.
Huawei AI clears 25 games with 56 percent fewer actions
Huawei's AI reportedly completed 25 games using 56 percent fewer operations than humans, and in one case stopped needing to call a large model.
Why it mattersSuggests reinforcement-based agents can cut both action count and model calls, pointing to cheaper game-playing and control agents.