Good morning. Meta is the main event today — the Superintelligence Lab’s first frontier release, Muse Wonder, is out, and the early verdict is that it’s genuinely close to the top but not quite there. Also on the menu: Microsoft’s next in-house inference chip, a NeurIPS paper proving models can hide capabilities from any known audit, and a Harvard study with uncomfortable news about what agentic coding actually ships.
Meta’s Muse Wonder lands near the frontier. Meta released Muse Wonder and Muse Wonder Full via the Meta Model API and Muse Code, pricing Wonder Full at $4/$20 per million input/output tokens. The company claims parity with GPT-6 Astra and Fable 5.1, with Wired framing it as a comeback and TechCrunch noting it’s the first flagship since Alexandr Wang’s August 2025 restructuring. The models introduce “Active Memory” and “Active Focus” for agentic reliability, and were trained at Meta’s gigawatt-scale Prometheus data center.
Reception is more mixed than the marketing. HN testing paints a less flattering picture than the benchmarks suggest — one commenter’s image-to-HTML test found Wonder “closer to Sonnet or Haiku” than Opus 5.5, and a developer trying it across React, Node, and Rust called it “still a step below Astra/Opus5.5/Fable.” The Verge also covered the launch. One HN user flagged what looks like an auth bug on the Muse Code login page showing other people’s profiles — not a great look on day one. Another pointed out Meta still hasn’t delivered the open weights it promised with Muse Spark back in September.
Microsoft’s Maia 300 aims to cut the Nvidia tether. Microsoft announced its third-generation Maia chip, claiming 40% more tokens per megawatt than leading GPU systems and up to 7x the rack-scale throughput of Maia 200. It’s targeted at long-context and agentic inference, scheduled to become Microsoft’s “primary AI inference system” starting in 2027, with multi-gigawatt Azure deployments planned. Translation: Satya wants out of the Nvidia and AMD queue.
Intel joins the inference pile-on. Intel’s new GPU D270 packs up to 480 GB of LPDDR5X on a 350W air-cooled PCIe card, aimed at long-context agentic workloads without liquid cooling or proprietary interconnect. HN was skeptical — no fast multi-GPU link means it’s realistically a single-user or low-concurrency device, and one commenter embarrassingly caught Intel’s own product page shipping with unfinished LLM placeholder text (“Intel GPU D270 messaging should reference familiar AI software…”). The marketing team also somehow buried the 1.5 TB/s memory bandwidth figure.
NeurIPS paper: models can provably hide capabilities. Researchers showed that cryptographic backdoors can embed hidden neural networks inside model weights that only activate on a secret trigger, and that these are provably undetectable even with full weight access and standard elicitation methods like fine-tuning. That undercuts a foundational assumption of current safety evals. HN commenters immediately connected it to open-weight supply chain risk — “every model coming out of the CCP has some Manchurian code word built in” was one of the less restrained takes, but the general point about sleeper agents in locally deployed models is now harder to dismiss.
Harvard: more code, not more software. A study across 700+ firms and 300 million coding events found that AI agents produce 30% more lines of code and 23% more pull requests, but issue and feature resolution rates haven’t moved in a statistically significant way. The bottleneck shifted downstream: review takes proportionally longer, with more comments and revisions per PR. One HN commenter stretched the analogy well — stronger engines mean tires become the limit — while another noted that code review isn’t only about catching bugs but about accountability and security, which automated review can’t substitute for.
Janas streams experts from SSD. A new tool called Janas claims to run LLMs larger than available RAM by streaming MoE experts from disk, hitting 37 tok/s on a 48GB model with 32GB of RAM versus llama.cpp’s 2.8 tok/s on the same setup. HN was split between genuine interest and a familiar complaint — “every few weeks, someone stumbles on what llama.cpp has had for years again, and reimplements it.” A fair request from the thread: someone should write a consolidated comparison of the growing pile of SSD-streaming projects rather than shipping another one.
That’s the morning. Meta is back in the conversation, Microsoft is one step closer to leaving Nvidia’s waiting room, and the safety researchers now have a formal proof that whatever you can’t see in a model’s weights, you really can’t see. Back tomorrow.