Legal Quants Weekly Digest — Week 18

The community founder announced going full-time on Legal Quants, turning down standing offers from multiple AmLaw 20 firms, a prominent AI-native law firm, and leading legal tech startups — drawing an outpouring of support and crystallising a week of intense discussion about what an AI-native legal career path should look like. Crosby's $60M raise sparked a rigorous debate about whether "AI-native law firm" is a meaningful label or VC theater, with consensus forming around an MSO-style structure (tech co holds IP, licensed lawyers hold the service entity) parallel to US medical-practice rollups. On the technical side, an apparent Claude Code source leak surfaced over 50 functions named _DEPRECATED that are still actively called in production, plus an unreleased "coordinator mode" with worker agents and scratchpads. OpenAI released codex-plugin-cc, letting Codex act as a reviewer inside Claude Code, and members converged on a dual-model pattern — Claude for architecture, Codex for execution. Anthropic separately moved to end subscription-based auth for third-party harnesses like OpenClaw, forcing API billing or alternate models. One member documented NanoClaw's Docker-per-conversation architecture as a more robust answer to OpenClaw's reliability struggles. The Chinese open-source wave continued (GLM 5.1, upcoming Gemma 4), and debate on tiny-model usefulness landed on a clear verdict: temperature-zero classification and routing only. Yann LeCun's $1B "world models" raise prompted a philosophical detour into LLM reasoning, the Chinese Room, and the difference between cognitive offloading and cognitive surrender.