AI News 2026-10-17

$ cat top-story.md

Claude reveals the origins of the NIST ECDSA curves

Asked where the NIST ECDSA curve seeds came from, Claude showed they are not hashes of a joke phrase as rumored, but outputs of the 1990s lrand48 generator seeded by Unix timestamps. It identified the weak generator from statistical patterns and recovered its state via lattice reduction; some base points trace to a small-seed generator in FreeLIP, while a few origins remain unexplained.

Why it matters: It is not a cryptographic break and makes hidden backdoors less plausible, but it is a striking example of a model doing real, verifiable research work. Top of Hacker News with 750+ points.

// source: Anthropic Research

## 📦 Models & Releases

▸ Muse Spark Open on four Mac Studios with MLX

Apple ML Research runs Meta's 816B-parameter MoE (42B active, NVFP4) across four M5 Ultra Mac Studios using MLX with RDMA over Thunderbolt, reaching over 36 tokens/s at 8K context. Why it matters: Shows that near-trillion open-weight models are becoming feasible on consumer-class clusters, not just GPU racks. // Apple ML Research

## 🛠️ Tools & Repos

▸ PS5LM: llama.cpp on a jailbroken PS5

Runs GGUF models on the console GPU and serves them on your network through OpenAI- and Anthropic-compatible APIs (387 stars, GPL-3.0). Why it matters: A fun hack, but the OpenAI-compatible endpoint means any client that speaks that API can use it as a backend. // GitHub / HN

## 🤖 Agents & MCP

▸ Claude Code v2.1.302

Adds CLAUDE_CODE_AUTO_MODE_MODEL for the auto-mode classifier, enables fork subagents by default in claude -p, the Agent SDK, desktop and VS Code (disable with CLAUDE_CODE_FORK_SUBAGENT=0), and stops plugin marketplaces from re-cloning on every refresh. Why it matters: Headless and SDK pipelines change behavior with the fork-subagent default, so check your scripts. // GitHub

▸ How the Claude Code team does code review

Claude reviews every PR with verifier agents and a blocking security review, while human approval scales with risk tier; merged PRs more than doubled June to September and the revert share fell over 40%. Why it matters: A concrete, copyable blueprint for agent-first review: risk tiers, merge queue, and turning incidents into eval cases. // claude.dev / HN

## 🏠 Self-Hosting & RAG

▸ Open WebUI v0.12.3

Adds end-to-end encrypted temporary chats (admin-enabled), separate Redis keys for streaming replies, live collaborator avatars and workspace search across chats, channels and notes; v0.12.2 added self-hosted MinerU 4.x support and fixed SCIM deleting admin accounts. Why it matters: The Redis change reduces load in multi-worker setups, and the SCIM fix matters if you use provisioning. // GitHub

## 📰 Worth Reading

▸ Benchmark Map: ~90% of US working hours are barely covered by AI benchmarks

Mercor estimates that most work, especially middle-income knowledge work, is poorly measured by existing benchmarks; results are provisional and based on 29 benchmarks. Why it matters: A useful caveat when leaderboard scores drive model choice: build your own evals for your domain. // Mercor / HN

▸ Gemini 4 Argon plays StarCraft

A Show HN benchmark pitting Gemini 4 Argon against StarCraft scenarios (199 points). Why it matters: A long-horizon agentic test that says more about planning than static QA benchmarks. // Show HN

▸ Designing more efficient quantum circuits with Claude

Anthropic describes using Claude to find more efficient quantum circuit designs. Why it matters: Another data point on models contributing to technical research workflows. // Anthropic Research

AI News — October 17, 2026