Claude hand-writes a legal-citation detector net, 1.8M cites/s on Apple ANE
paul_cal · x · 2026-11-05
A user had Claude write every weight of a legal-citation checker neural network from scratch — zero training, running on Apple's Neural Engine. On the LePhantomCite held-out set it caught all 31 fake citations with only 1 of 983 real ones flagged; on an M5 Pro it runs at 50µs per citation, 1.8M cites/s on ANE + GPU, scanning 3.3GB of case law in 2.3s.
The follow-up notes the catch: this kind of generation only works when a ruleset exists to follow — LePhantomCite fakes can be identified via mismatches like volume vs. year and reporter vs. court. Tasks without such rules may not replicate the result.
More from coding & agent
- Raindrop engineer makes the case for building your own internal Slack agent — zeeg · 2026-11-05
- Genkit Go, Eino, ADK Go, tRPC-Agent-Go, LangChainGo compared with real code — rseroter · 2026-11-05
- How do you connect cloud AI agents to tools behind a customer's firewall? — JoleneAtCynapsa · 2026-11-05
- Elixir guide for OO programmers by Bruce Park goes free online with an AI tutor — burkov · 2026-11-05
- MCP listing on Claude brought 24,000 users in 5 weeks, founder says launch there first — ElevatorFriendly648 · 2026-11-05
- User reports OpenAI's Sol 6.1 running at 'quarter speed' for days — ruslansv · 2026-11-05