06-reference

alphasignal ruview anthropic legal nous training

2026-05-14·reference·source: AlphaSignal·by Lior Alexander
alphasignalai-researchai-newscurationdaily-roundupanthropicclaude-for-legalnous-researchtraining-efficiencywifi-sensingdeepseek

"RuView WiFi Tracking, Anthropic Legal Drafts, Nous 2.5x Training" — AlphaSignal Team

Why this is in the vault

Daily AlphaSignal curation. Three lead items: a WiFi-as-radar repo (RuView, 50k+ stars), Anthropic shipping claude-for-legal as a free open-source plugin suite covering 10+ practice areas, and Nous Research's training trick that cuts LLM pretraining time 2.5x without model changes. Two items map directly to RDCO substrate questions — Anthropic legal (Claude as the agent platform expanding into vertical workflows) and Nous training efficiency (the "efficiency is the new moat" framing the editor leads with). RuView is background context. Captured for the K-sender archive and for the two RDCO-relevant threads.

Issue contents

Three "Top Repo" features + six "Signals" + three sponsor placements (Lightfield, Datadog, Braintrust).

Top Repo 1 — RuView WiFi sensing. Open-source tool that turns home WiFi routers into through-wall human radar. 50k+ GitHub stars, 11k likes. Detects people count and position, tracks 17-point body skeleton in real time, measures breathing (6-30 BPM) and heart rate (40-120 BPM) during sleep, triggers fall detection. Runs on ESP32-S3 boards (~$9). Local-only, no cloud. Pose estimation requires per-room model training; motion/presence work out of the box.

Top Repo 2 — Anthropic claude-for-legal. Free open-source plugin suite turning Claude into a legal assistant. Covers contracts, litigation, AI governance, 10+ practice areas. Capabilities: vendor contract / NDA / SaaS review against custom playbook, renewal-deadline watching, claim-chart building for patent/civil cases, DSAR drafting within legal deadlines, integrations with Slack, DocuSign, Ironclad, court docket systems. Installs via Claude Cowork, Claude Code, or headless API. 3.3k stars at publication. Every output is a "draft for attorney review, not legal advice." Repo is plain markdown + JSON — easy to fork.

Top Repo 3 — DeepSeek opinion post. Editor's own footnote: this slot was originally a political opinion piece about data centers and energy use, not a technical release. Editor extracted the one concrete fact — DeepSeek-V3 is real, open-source, runs locally via Ollama / llama.cpp, strong at coding/math/text — and skipped the politics.

Signals.

  1. AI uncovers 18-year-old remote code execution bug in NGINX (2,267 likes).
  2. Braintrust starter plan (sponsor placement).
  3. Nous Research training trick cuts LLM pretraining time 2.5x without touching model architecture (1,978 likes).
  4. Tübingen researchers train LLMs to think, read, and act in parallel streams (650 likes).
  5. Lna-Lab squeezes 744B MoE model to 347GB via dynamic 3-bit quantization (409 downloads).
  6. Developer fine-tunes SD1.5 on 50k DALL-E 2 images to revive its iconic style (1,449 likes).

Mapping against Ray Data Co

Strong — Anthropic claude-for-legal. This is Anthropic explicitly shipping a vertical-workflow plugin suite into the Claude marketplace. Three RDCO implications: (1) the marketplace thesis that emerged from [[2026-04-27-alphasignal-anthropic-claude-marketplace-agent-quality]] is now being demonstrated by Anthropic itself — they're not just hosting third-party plugins, they're seeding the surface with their own verticals. (2) Law firms as the buyer profile validates the "knowledge-worker vertical with high willingness-to-pay" pattern that recurs in RDCO substrate research. (3) "Plain markdown and JSON, so forking and customizing skills is straightforward" — the architecture explicitly invites the kind of skill-authoring workflow RDCO already uses internally (~/.claude/skills/). Worth deep-fetching to see whether the repo structure clarifies the skill-vs-plugin packaging boundary.

Medium — Nous Research 2.5x pretraining. The "efficiency is the new moat" framing from the editor's intro lands on a thread RDCO has been tracking: who actually wins when frontier compute is no longer the differentiator. If Nous's trick is real and architecture-agnostic, it tightens the gap between well-funded labs and open-source efforts. Maps to the un-hobbling thesis at [[2026-05-10-alphasignal-self-improving-agents-harness]] — every training-efficiency win compounds with every harness-efficiency win. Not directly actionable for RDCO (we don't pretrain), but a signal for the investing thesis and for "where does Claude rank vs open-source 12 months from now" forecasting.

Weak — RuView WiFi sensing, Tübingen parallel streams, Lna-Lab quantization, SD1.5 DALL-E revival. Background curiosity. RuView is genuinely wild but doesn't intersect any active RDCO bet. The Tübingen and Lna-Lab items are research-track. SD1.5 DALL-E revival could matter for mascot work if Ray-style outputs are ever bottlenecked on aesthetic, but the existing Kling/Ray pipeline isn't blocked.

Skip — NGINX RCE, DeepSeek opinion piece. NGINX bug is interesting in the abstract (AI finding 18-yr-old CVEs) but not RDCO-relevant. DeepSeek slot is explicitly editor-flagged as opinion, not a release.

Curation section — notes

Related