Five stories defined the AI conversation this week: Meta open-sourced its most capable US model yet, Claude solved a decades-old cryptography mystery in minutes, a Goodfire essay argued that neural networks are far more legible than feared, new research found that AI benchmarks barely map onto real American work, and Anthropic’s Claude Code team revealed how it reviews code at the volume agents now produce.
Meta open-sources Muse Spark Open, its most capable US model to date
Meta released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date. The natively multimodal model packs 816B total parameters with 42B active and a 1M-token context window, built for coding, agentic work, and long-horizon reasoning. It is the first model in its class to run on a desktop, leveraging two Apple M5 Ultra Mac Studios via the MLX framework — 512 GB of unified memory per machine gives the pair a terabyte of RAM reading at 1.2 TB/s. Meta says it demonstrated full on-device agentic workflows, including shipping real iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve end to end. Weights are available on Hugging Face, Amazon SageMaker, and Google Cloud, with API access through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI. The model ships under the Muse Model License with limited commercial terms, and Meta released custom Helion GPU kernels delivering 1.03–3.95× kernel-level speedups on NVIDIA and AMD hardware. The release follows Meta’s Superintelligence Scaling Framework and includes additional safety post-training.
Claude cracks the ECDSA seed mystery — in under half an hour
Anthropic researchers revealed that Claude has solved a decades-old open question in cryptography: the origin of the “seed” parameters behind the widely used ECDSA elliptic curves adopted by NIST in 2000. Long suspected of hiding a backdoor, the provenance of the seeds supplied by the NSA had been a source of controversy for years. In October 2026, a curious Anthropic employee asked an internal research model to figure out how the seeds were generated — and Claude did it in under half an hour. The seeds for NIST curves P-192 through P-521 turned out to be the output of the C library’s lrand48 pseudo-random number generator, seeded with a Unix timestamp — not the output of any hash function, and certainly not a backdoor. P-256, for example, required 1,842 iterations of drawing 20-byte candidates before matching the standardized seed. Anthropic stressed this is not a cryptographic break and introduces no weakness; rather, it closes a chapter of years of speculation, largely prompted by the late NSA engineer Jerry Solinas’s 2015 recollection of a humorous passphrase.
“Neural networks are already clean” rethinks interpretability
Goodfire’s Nick Cammarata (co-author of early mechanistic interpretability work at OpenAI) published an informal research essay arguing that neural networks are far more legible than the field believed. The essay contends that the canonical evidence for “polysemanticity” and “superposition” — the idea that networks pack unrelated concepts into single neurons — was misread. What looked like arbitrary packing is, Cammarata argues, a motif he terms “recursive generalization,” which is far less concerning. Drawing on months of studying image and language models up to trillions of parameters, Goodfire says most neurons can now be understood directly, without first untangling them. Cammarata frames interpretability as potentially the most important scientific problem of the moment: “I don’t think we can solve alignment if we don’t know what we’re building.” The 65-minute essay marks a notable shift in tone from the pessimism that followed early obstacles in the field.
90% of US working hours fall outside AI benchmark coverage
New research from custody and AI labor-data firm Mercor maps AI benchmarks against the actual US economy and finds a striking gap: 90% of American working hours are spent on tasks with little or no coverage from existing AI benchmarks. From an index of 924 benchmarks mapped to O*NET’s 923 occupations and 18,797 tasks, only 208 occupations have a task covered by at least five benchmark prompts, and six of 22 job families have no occupations covered at all. Coverage is also concentrated unevenly, skewed toward the largest labor markets and highest wage-bill occupations — healthcare practitioners, education, and science are notably under-covered. The researchers warn this makes it very hard to know how good AI models actually are at the work most Americans do, and call the Benchmark Map a first step toward a continuously updated resource for measuring AI capabilities across the economy.
Inside Claude Code’s volume-era approach to code review
Anthropic’s engineering team published a detailed look at how it reviews code when agents write roughly 90% of the code. Between June and September, merged pull requests to Claude Code’s codebase more than doubled while the share that had to be reverted fell by more than 40%. The team moved most of the “reading” to machines and optimized for the scarce resource of human time and attention. Compounding the pressure, changes landing per day in Anthropic’s monorepo grew from about 2,500 in May to more than 30,000 in September. Review now runs continuously — more like spellcheck while you write than a gate at the end of the line — with most changes already reviewed and fixed by Claude before a human is asked to look. Humans stay focused on higher-level questions and novel challenges, and the team embraces a “you land it, you own it” culture built around fast, easy undo. The authors (including Addy Osmani) frame it as a blueprint other teams could adopt.
That’s this week’s roundup. From open-weight model releases and a closed cryptographic mystery to a rethinking of what interpretability research actually requires — the pace of AI’s real-world footprint keeps accelerating.
