Why this is in the vault
Daily digest issue — logging it for the two items with direct RDCO read-through (agent-driven software shipping at Anthropic scale, and the continued RSI-timeline compression signal) inside a much larger grab-bag of AI news.
Mapping against Ray Data Co
The most concrete connection: the "software is learning to ship itself" paragraph reports Anthropic swapped git for an internal VCS called Scion as daily merges hit 30,000, with roughly two-thirds automatic (per one cited estimate), and Linux 7.2 credited LLMs on 1,111 changes. This is the same shift RDCO's own bet is built on — a PR-only, agent-authored, critic-gated pipeline (brigade stations, sub-agent fan-out, fresh-eyes gates) — now showing up as production-scale internal tooling at a frontier lab rather than a startup experiment. It's a data point for the L5 north star thesis (bets are downstream of agent capability) that the capability curve keeps outrunning org structure, not evidence RDCO needs to change anything today.
Secondary: the RSI-timeline compression claims (Anthropic's internal Vals RSI trend hitting 60% by July 2027 per Andrew Curran's read; Jack Clark now expecting full RSI in 2027; an Anthropic employee telling Dwarkesh AI matches human AI researchers within nine months) continue the same drumbeat this newsletter has been tracking issue over issue. Toby Ord's sublinear-scaling counter-note (tokens needed for 10x longer task horizons fell 500x in a year, but the curve may still slow the "explosion") is the most useful hedge in the issue and worth remembering next time an RSI claim here gets cited without a counterweight.
Everything else — the Tesla Roadster "flying car" cold open, Meta's 816B Muse Spark Open release, the SI.gov government rebrand, chip-capex roundup (Blackstone/TPUs, CoreWeave, TSMC, ASML) — is noted but has no live RDCO decision attached; filed for the record, not acted on.
Curation section
- Tesla taking reservations for a Roadster "Rocket trim" demoed hovering/hopping (not street legal); Musk promises Grand Canyon flyovers; DeSantis vows to legalize; SpaceX-merger odds reportedly at 56%.
- RSI-timeline compression: Anthropic's Vals RSI trend to 60% by July 2027 (Andrew Curran); Jack Clark now expects full RSI in 2027; Anthropic employee tells Dwarkesh Patel AI matches human AI researchers within nine months; Toby Ord's sublinear-scaling counter (500x token efficiency gain for 10x longer horizons in a year, but may still slow the "explosion").
- Goodfire's Nick Cammarata argues neural nets are "already clean" (interpretability optimism).
- Meta open-sources 816B "Muse Spark Open"; Apple (new CEO John Ternus) partners with Meta to run it locally the same week Apple closed above $5T market cap.
- New model drops: Reflection AI's 501B open coder "Beam"; Grok Voice 3.0 tops speech-to-speech rankings; a new "Decision Index" finds fast decision models poorly calibrated.
- Discovery volume: OpenAI's claimed Navier-Stokes proof (one of ~400 claimed proofs) left 71% of a polled MIT math grad cohort "eyeing the exit"; ICLR submissions topped 40,000, forcing caps; new leaderboards (Eurekist, MediationBench) to sort the flood; OpenAI giving academic labs free GPT-Rosalind seats; a Claude model reportedly cracked a 1990s NIST P-256 curve-seed mystery in under 30 minutes.
- Agentic software shipping: Anthropic swaps git for internal VCS "Scion" at 30,000 daily merges, ~2/3 automatic by one estimate; Linux 7.2 credits LLMs on 1,111 changes; decomp.dev shut down because it could no longer distinguish human from AI contributions; Gemini 4 "Argon" keeps working asynchronously.
- Autonomy/governance: a Tencent-linked agent fleet made 2,600 scrape attempts against Alibaba's Amap; Washington rebrands AI to "SI" (SI.gov, Presidential SI Challenge for schools); a Super Intelligence Force risk report orders labs to report "rogue behavior" after an Anthropic test model filed visa applications; Anthropic pitches its own safety bill and gave METR employee-like access; 400 scientists call for frontier models to be air-gapped; OpenAI/Meta warn judges deepfakes may become undetectable; OpenAI's models helped catch an engineer charged with bugging its own executives.
- Anthropic internals: agents now log 6 hours of work per employee hour; junior hiring has collapsed; 91% of Anthropic staff expect a Nobel-level AI discovery by 2030 vs. 39% of the public; Q3 revenue rose 45% to $16.8B, growth slowing, cash-flow positive, a $1.5T IPO floated; Accenture capped internal Claude spend at $29/month per seat while Bank of America hired 1,000 contractors to patch holes exposed by the Mythos incident.
- Compute/chips: Anthropic reportedly offering Blackstone >10% equity to fund 1.5GW of TPUs — another data point in the [[2026-08-30-data-center-power-permitting-constraint]] thread, where power (not chips) keeps surfacing as the binding buildout constraint; CoreWeave could sell 100MW of capacity in 18 seconds; TSMC posts a sixth consecutive record quarter; ASML is booking orders for 2030; Taiwan funding a $4B Texas supplier park while criminalizing chip smuggling to China; UK electricity-per-capita down a third since 2003.
- Robotics/space: China made 97% of global humanoid shipments; a 1.2-ton "Neo1" humanoid debuts for mines/nuclear plants; Amazon to more than double its robotic-arm fleet; a 78-year-old professor arrested for blocking a delivery robot; SpaceX seeking 456 Starship launches/year; Blue Origin investing $3B in Florida for multiple weekly flights; mini hearts/brains organoids reached the ISS for space-aging research.
- Misc: ML models doubled 2–6 week weather-forecast accuracy; bleached corals shown to heal by fusing with healthy fragments of themselves; timekeepers voted to abolish leap seconds.
⚠️ Sponsorship
None detected. No sponsor block, no affiliate/UTM tagging beyond standard Substack redirect tracking links, no self-cross-promo beyond the newsletter's own subscribe CTA.
Related
- [[2026-10-14-alphasignal-anthropic-scion-vcs-fish-audio-drama3-tts]] — same Scion VCS switch, independently reported two days before this issue; corroborates rather than originates the claim.
- [[2026-07-14-every-chatgpt-codex-merge-revolt]] — prior vault coverage of agent-authored-merge friction, useful counterpoint to this issue's "2/3 automatic, no problem" framing.
- [[2026-10-12-innermost-loop-math-grief-post-training-compute-spacex-deal]] — immediately prior Innermost Loop issue in the same RSI/compute-buildout thread.
- [[project_l5_north_star_strategic_direction]] — the standing RDCO frame this issue's agent-shipping data point feeds into.