Good morning. Meta’s follow-up to last week’s Muse Wonder is here — an open-weight 816B monster called Muse Spark Open that technically runs on your desk if your desk holds two Mac Studios and ten thousand dollars. Also today: Anthropic hands METR a company laptop, Sergey Brin casually mentions recursive self-improvement, and a Goodfire researcher argues the field has been wrong about polysemanticity for years.
Meta’s Muse Spark Open is open-ish, and runs on Macs if you have two of them. Meta released Muse Spark Open, an 816B-parameter (42B active) multimodal model with a 1M-token context window, pitched as the most capable US open-weight model to date. The headline trick is that it runs locally on two Apple M5 Ultra Mac Studios via MLX — the product of several months of quiet Meta-Apple cooperation that The Verge notes had Zuckerberg and new Apple CEO John Ternus publicly complimenting each other, with Alexandr Wang joking that people are “excited for mommy and daddy to be playing nice again.”
The community is not impressed. Artificial Analysis has Muse Spark Open scoring 39, roughly tied with DeepSeek v4.1 Flash and behind both GLM 5.3 Flash and Qwen3.8-Flash-Next — models that fit on a single B300. The “open weight” license also requires large inference providers to negotiate separately with Meta, which one HN commenter noted usually translates into less price competition from public hosts. TechCrunch’s writeup takes the democratization framing at face value; HN mostly didn’t.
Anthropic gives METR employee-level access. Anthropic is embedding METR researchers inside the company with laptops, office space, and internal tools to independently audit models and safety practices, with METR free to publish findings without Anthropic’s sign-off. TechCrunch reports the structure is three-tier, with METR at the top — modeled loosely on how bank examiners embed inside major financial institutions. It follows a White House accord that OpenAI, Google, Meta, Nvidia, and xAI also signed, and comes after a run of disclosed agent-misbehavior incidents since late August.
Sergey Brin claims Google models are approaching recursive self-improvement. In a Business Insider interview, Brin said Google’s models are nearing RSI, with little supporting detail. HN was unkind: one commenter noted it’s not established that next-token predictors can recursively self-improve at all, another pointed out Argon still trails Opus 5.5 and 6.1 Sol on real tasks, and a third summarized it as “guy who sells AI says AI is cool.”
A Goodfire researcher argues neural networks were clean all along. Nick at Goodfire published a long piece arguing that polysemanticity and superposition — the idea that neurons tangle unrelated concepts together — were largely artifacts of weak tools and limited architectures, and that networks can be interpreted directly without sparse autoencoders as a decomposition layer. HN reaction is split: some call it a genuine mechinterp breakthrough, others say it rhymes with past cycles of overconfidence, and Yudkowsky is reportedly unconvinced. One commenter linked the first non-trivial network Nick claims to mostly understand as the cleanest version of the argument.
OpenCode 2.0 rebuilds around a persistent daemon. The AI coding agent shipped a major rewrite with a background service that the TUI, web, and desktop apps all share, plus a rebuilt API, a new @opencode/client package, and the ability for the agent to edit its own config and plugins. Reception is mostly warm — automatic plugin re finally works, and the “everything’s a plugin” direction is popular — but several HN users questioned why coding harnesses are increasingly defaulting to always-on daemons. One commenter noted the installer takes six minutes to hit 10% because of progress-bar rendering.
Open data for Indic languages, and IBM’s multi-task endpoint. NVIDIA and IIT Madras’ AI4Bharat released Nemotron-pretraining-Indic-v1, a 2.8T-token corpus covering all 22 constitutionally recognized Indian languages in both native and romanized script, along with the full GPU-accelerated generation pipeline. Separately, IBM launched Granite Switch 4.1, bundling chat, decisions, and speech-to-text into one endpoint using Activated LoRAs that share the base model’s KV cache across adapter switches.
Google uses Antigravity to hunt antimicrobial peptides. DeepMind published a walkthrough of using its Antigravity agentic science platform plus ProtNLM to scan millions of uncharacterized proteins in UniProt and identify a novel antimicrobial peptide candidate against respiratory pathogens. Wet-lab validation is next; antimicrobial resistance is projected to cause 10 million annual deaths by 2050, so the problem is worth the agent cycles.
That’s the brief. If anyone actually gets Muse Spark Open running on their dual-Mac-Studio rig today, send pelicans.