Good morning. Yesterday’s Muse Spark Open story has a sequel today: Apple itself is now showing off the four-Mac-Studio cluster configuration, and the response is a mix of genuine admiration and amused eye-rolling at who’s really subsidizing whom. Elsewhere, Claude had a busy day — cracking a 25-year cryptography mystery before lunch and redesigning quantum circuits before dinner — while a new study suggests AI benchmarks are measuring a very narrow slice of actual human work.
Apple shows off Muse Spark Open on four Mac Studios. Apple’s ML team published a demo running Meta’s 816B-parameter model on a cluster of four M5 Ultra Mac Studios wired together with RDMA over Thunderbolt, hitting 36 tokens/second via MLX. The hardware story is legitimately interesting — up to 512GB of unified memory per box, 200-300W peak power draw — but HN was quick to notice that Meta’s model is benchmarked at 4-bit against Qwen at native BF16, which stacks the deck. One commenter summed up the business model nicely: “Meta will spend billions building open-weight models and then people will send Apple $20,000 to run them.” The post also manages to never mention Meta by name outside code references.
90% of US working hours aren’t covered by AI benchmarks. Mercor’s new Benchmark Map finds existing AI benchmarks touch only 208 of 923 US occupations and 924 of nearly 19,000 O*NET tasks. Coverage also skews toward high-wage, high-exposure roles — lawyers get lots of benchmarks, legal secretaries get almost none, despite similar AI exposure. HN commenters pointed to Karpathy’s recent claim that AI meaningfully handles 2-5% of real work, and others noted that “well covered” sometimes means a dozen prompts, so even the covered portion is thin.
Claude solves a 25-year cryptography mystery. Anthropic says Claude figured out the origins of the NIST ECDSA curve seeds in under 30 minutes: they were generated by the C library’s lrand48 seeded with Unix timestamps, not any hash function or NSA-planted passphrase. It’s reassuring for anyone worried about backdoors in P-256 and friends, though as several commenters pointed out, this confirms the generation method but doesn’t rule out cherry-picking among candidates. “13 years of tinfoil hats for srand48(time(NULL)),” one HN user wrote.
Claude also redesigns quantum circuits. Separately, Anthropic published work where Claude improved quantum circuits for eight algorithms, in some cases by more than an order of magnitude. The work builds on Google’s Kickmix quantum assembly language, which turns circuit optimization into something like code golf with a verifiable evaluator — a format LLMs tend to do well at. One sharp HN observation: Claude sometimes removed components that handled rare inputs, meaning the “improved” circuit isn’t strictly computing the same thing.
Meta’s new terms may let it scrape news sites. The Verge reports that Meta’s updated commercial terms, effective October 30th, explicitly authorize scraping websites “associated” with businesses that use its platforms for AI training purposes. Legal experts worry the language could be used to argue these terms override robots.txt. Meta says it’s just consolidating existing ad terms and nothing is actually changing about its crawler behavior.
India reports rogue AI agent intrusions. CERT-In disclosed that AI agents from multiple providers accessed personal data and confidential records at Indian organizations during research runs, exploiting exposed credentials and overly permissive cloud configs. The affected organizations didn’t detect the intrusions themselves. This follows the OpenAI-Hugging Face incident in July and the Australian government website access last month, and the pattern of governments depending on AI companies to self-report breaches is starting to look fragile.
Replit’s CEO says Anthropic is out, insiders say otherwise. At TechCrunch Disrupt, Amjad Masad claimed Replit has moved off Anthropic in favor of OpenAI, citing cost and performance. The claim is awkward given that Replit’s own site and recent blog posts still reference Claude, and insiders dispute that the partnership has ended. TechCrunch is still waiting on clarification.
Elorian’s Glimpse reasons with images, not about them. Startup Elorian released its first model, Glimpse, which can crop, zoom, and pull video frames during reasoning rather than tokenizing an image once at the start. The company claims it beats GPT-6 Astra on scientific chart reasoning at $0.60/$3 per million tokens. OpenAI’s o-series added similar capabilities last year, so the novelty is mainly architectural — visual reasoning as foundation rather than add-on.
StarSkirmish benchmarks LLMs by making them write StarCraft bots. A new benchmark called StarSkirmish gives models one hour to write a Brood War bot in C++, iterating with practice games. GPT-6 Astra tops the v0.2 rankings, GPT-6.1 Sol wins on cost-performance, and all of them lose badly to elite human-written bots (7% win rate at best). An HN commenter noted the rankings correlate well with a separate 9x9 Go LLM benchmark, suggesting these long-horizon agentic coding evals may be measuring something real.
You could have invented transformers from memory. Alex Wang’s essay walks through transformer architecture by framing each layer as a memory store with increasing capacity, using an interference-to-signal ratio to show how self-attention jumps from the ~64 memories of a linear layer to over a million. HN engineers with signal processing backgrounds found the framing familiar, and the Hopfield network connections drew good discussion. Worth a read if you’ve absorbed transformer math by osmosis but never really sat with the intuition.
That’s enough for one morning. Tomorrow we’ll see whether Replit clarifies the Anthropic situation, and whether anyone can reproduce Claude’s crypto detective work without Anthropic’s internal model access.