Gonçalo Esteves / building capable systems Field Notes →

The Vault / Concepts

Concept

Recursive self-improvement (RSI)

AI that improves the AI that builds the next AI — the show's name for the loop it believes it is watching close, cited in 69 indexed episodes with no page until EP308 made it the centrepiece.

status · researched added 5 Nov 2026 updated 8 Nov 2026 from Moonshots #306 · Moonshots #307 · Moonshots #308 · Moonshots #309 ai see it in the graph →

Created 2026-11-05, promoted off the Moonshots #308 decomposition.

What it is (one line): AI that improves the AI that builds the next AI — the show's name for the loop it believes it is watching close, cited in 69 indexed episodes with no page until EP308 made it the centrepiece.

Why it is in this archive

Gate: cited in 69 episodes (second-highest unpromoted concept after "abundance"). On EP308 the panel declared a threshold crossed, which makes the claim history worth holding in one place: who said the loop closed, when, and on what definition.

The definitional ladder — what "RSI" has meant on tape

The word stays constant; the referent moves. In order of strength:

  1. Workflow RSI — a company's AI measurably improves the company's own process of building AI. Salim Ismail's coaching frame: "recursive self-improvement at the level of workflow within the company" (Moonshots #308 0:54:08).
  2. Capex RSI — a lab spends more compute/capital on AI "researchers" than on human ones. Alexander Wissner-Gross, EP308: "according to with capital or operating costs, the answer is yes, we are there. We have reached this part of the singularity" (0:52:34), reading the Anthropic amended S-1 (internal Claude use 250× Jan→Sep 2026, above monthly payroll; Claude >90% of code contributions in September — press readings of the filing).
  3. Model-edits-model — one model surgically edits another's weights. MIT CSAIL's weight-space programming (report 2 Nov 2026, Khattab group): Astra edited 0.016% of Qwen3.5-0.8B's weights, MATH-500 7.9%→48.1% with no retraining. ⚠️ The authors say the method removes obstacles and cannot add abilities (the gain came from suppressing a repetition loop; AIME stayed ~0–2%). The panel's "learning new capabilities without backprop" overstates it.
  4. Physical RSI — "True RSI only occurs when robots or nanobots create better copies of themselves" (Alex, EP308 0:53:15).

The pattern to watch: on EP308 Alex declares level 2 crossed and simultaneously moves "true" RSI up to level 4 — the threshold is announced and re-armed in the same answer. Every future "we've reached RSI" claim on the pod should be placed on this ladder before it is repeated.

Claim history

date episode who claim ts
2026-10-28 Moonshots #306 Dave Blundin Fable 5.1 was "the first model that was capable of improving itself" — ❓ unsourced 0:59:50
2026-10-31 Moonshots #307 Alexander Wissner-Gross self-replicating data centres will go to the Moon to escape regulation —
2026-11-04 Moonshots #308 Alexander Wissner-Gross capex threshold crossed ("we are there"); true RSI = nanobots 0:52:34, 0:53:15
2026-11-04 Moonshots #308 Dave Blundin "over 90 percent sure" AIs will edit their own weights [attr?] 1:08:38
2026-11-04 Moonshots #308 panel [attr?] Claude wrote "10% → 40% → 60%" of successive Claudes — 🔴 panel-only trajectory, no source; only the S-1's Sept 2026 >90% figure is sourced 0:53:47
2026-11-07 Moonshots #309 Alexander Wissner-Gross ★ level 4 dropped: "with just software only improvements… Probably without any new data centers any new hardware advances at all"; "could stop all data center construction tomorrow"; "super intelligence escape velocity". This reverses his EP308 nanobot condition four days later, without acknowledgement 0:33:58–0:34:22, 0:40:44, 0:41:38
2026-11-07 Moonshots #309 Alexander Wissner-Gross definition: "Substantially all research… conducted by… machine researchers"; "to the right of 50 of compute on margin being devoted to ai researchers" 0:44:52, 0:58:28
2026-11-07 Moonshots #309 Dave Blundin definition: if the AI has "any good ideas at all" (Noam 6/10 vs AI 1/10), "You're way past rsi" 0:45:08–0:45:57
2026-11-07 Moonshots #309 Alexander Wissner-Gross dated: frontier releases become continuous, "early 2027 or so" 0:59:39, 1:00:20
2026-11-07 Moonshots #309 Salim (reading METR) ✅ METR 5 Nov: Anthropic AECI 1.9× its earlier pace (CI 1.5–2.8×), ~80% AI-attributed, which METR calls "highly uncertain" 0:38:47

EP309 note (2026-11-08): METR is the first independent measurement on this page; everything before it was a filing or a mate. It measures acceleration of a benchmark index, not a closed loop, and its attribution share is hedged. In the same episode the panel's ladder collapses downward: level 4 (physical) was the "true" bar on 4 Nov, and by 6 Nov software-only RSI is enough to finish the job with no new data centres. The trigger is political, not technical: the claim answers the midterm threat to construction. Use METR's number with its CI and caveat, and do not use the panel's "escape velocity".

What is actually verified

  • Anthropic's amended S-1 figures (90% of September 2026 code contributions, 250× internal-use growth, >payroll from April, 45× per-engineer vs 2024) — via secondary readings of the filing; EDGAR not yet opened by this desk.
  • The MIT result, with its own authors' narrow conclusion.
  • Nothing on tape verifies a self-accelerating loop; what is verified is spend reallocation and tool adoption, which is also what any operator can replicate at small scale.

EP311 (Moonshots #311, aired 2026-11-13): open vs closed loop

  • Demis Hassabis (clip): today's RSI is human-in-the-loop, 3–5× productivity; closed-loop RSI "carries the most risk", and the result will be "quite alien" (0:58:03–0:59:54).
  • Alexander Wissner-Gross: "everyone's doing it already including his lab"; "there's a human pressing a button approving a poll request, but… we're pretty far into the RSI loop" (1:04:10–1:04:33). On the ladder: claims the closed loop is near, with no artefact.
  • Emad Mostaque: human-in-the-loop is impossible at "10,000 token a second… and there's 100,000 of them" (1:04:43); 🔴 a DeepSeek-V4-class model trains "in about an hour" on an Elon-scale cluster (1:05:04), unsourced; RSI inside a cluster "that nobody picks up" through distillation (1:06:41).

Related

Demis Hassabis · Paul Christiano · Anthropic · Alexander Wissner-Gross · Dave Blundin · Salim Ismail · Orbital compute (data centres in space) · Moonshots: the disclosures ledger

How this page stays current

On any episode where "RSI", "self-improvement" or "models editing models" fires in the index, place the claim on the ladder above and append to the claim history with the timestamp.

Reading rule. Anything marked unverified or on-air is a claim made on the podcast and has not been checked against a primary source yet. It stays marked until it has. Sourced numbers link to where they came from.