---
id: "2026-10-14-gladstone-endgame-what-rsi-looks-like-inside-labs"
url: "https://postcutoff.com/e/2026-10-14-gladstone-endgame-what-rsi-looks-like-inside-labs/"
as_of: "2026-10-16T12:39:00+02:00"
date: "2026-10-14"
date_precision: day
category: policy-safety
importance: 3
confidence: medium
status: [Partly confirmed]
sources: 3
editor: human editor
human_review: null
version: "2026-10-16"
---

As of: 2026-10-16 12:39 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-14-gladstone-endgame-what-rsi-looks-like-inside-labs/

# Gladstone report: insiders expect RSI in 3–12 months

Full title: Gladstone AI's 'What recursive self-improvement looks like inside an AI lab': most insiders it spoke to expect RSI within 3–12 months

On Oct 14, 2026 Gladstone AI (Jeremie and Edouard Harris) published "What recursive self-improvement looks like inside an AI lab" on its Substack The Endgame, previewed by TIME. Based on six months of conversations, including more than a dozen in recent weeks, with anonymous researchers at Anthropic, OpenAI and Google DeepMind, it describes the AI R&D "scaling and selection funnel" that would have to be automated, says compute is the main bottleneck, and reports that most insiders expect RSI to start within 3–12 months absent US government intervention. The same day, Google DeepMind's Argon report found no measurable acceleration yet.

## Key facts
- Published Oct 14, 2026 on The Endgame (aiendgame.com), Gladstone AI's Substack; TIME had it in advance and quoted CEO Jeremie Harris and CTO Edouard Harris
- Method: six months of conversations with researchers at frontier labs, 'over a dozen conversations with lab insiders' in the final weeks; sources anonymous and unverified
- Headline claim: most researchers expect RSI to begin within 3 to 12 months, 'assuming there's no U.S. government intervention or slowdown'
- Model of AI R&D: a successive-halving 'scaling and selection funnel' (generate ideas → code → run experiments at growing scale → read scaling curves → fold winners into a big run); idea generation is still largely human
- Bottleneck: compute, not ideas: going from 50 to 10,000 ideas on the same first-round budget gives each idea 1/200 of the compute; 'compute buys you the right to try lots of bad ideas'
- Alignment is harder to put through the funnel: scheming can appear above a test-time-compute threshold and models may hide it from evaluations
- Also claims, with high confidence but no evidence shown, that Chinese intelligence has 'almost certainly penetrated' frontier labs
- Edouard Harris (TIME): the point is to make it 'possible to point to a part of this process and say: a human needs to be there'; he also described 'a fair amount of [disaster] prepping and bio-bunker purchasing' among such researchers

## What happened

Gladstone AI, a US AI-security consultancy, set out to describe the actual workflow
that recursive self-improvement would have to automate. It wanted lawmakers to have something concrete to regulate. The piece
describes how labs filter many research ideas through rounds of growing compute and sometimes skip straight to billion-dollar runs.
It argues that post-training results transfer to scale more reliably than pretraining results. It reports that most insiders it
spoke to expect RSI within a year.

## Why it matters

The report appeared the same week as two first-hand but opposite data points from Google. DeepMind's own measurement found no
significant acceleration of AI progress since April 2026. Sergey Brin, in a roundtable Google released on Oct 14, said Gemini is
"starting to get onto that RSI treadmill" (Business Insider). Gladstone's timeline rests on anonymous, unverified insider views,
not on measurements. Its main use is the step-by-step map of where human oversight could be required.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 167 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 106 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 197 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 136 days after its cutoff

## Sources

1. [The Endgame (Gladstone AI): What recursive self-improvement looks like inside an AI lab](https://aiendgame.com/p/what-rsi-looks-like) (aiendgame.com, official)
2. [TIME: A new report sheds light on the secretive race to build self-improving AI](https://time.com/article/2026/10/14/recursive-self-improvement-AI-report/) (time.com, press)
3. [Business Insider (via MSN): Sergey Brin says Google's AI models are reaching a critical moment of self-improvement](https://www.businessinsider.com/sergey-brin-google-gemini-ai-argon-recursive-self-improvement-rsi-2026-10) (businessinsider.com, press)

## Changes

- 2026-10-16 (filed): Created (which cites the report; Endgame post and TIME read)

## Related

- 2026-10-15: [Apollo: GLM-5.3 shows update-aware reasoning](https://postcutoff.com/e/2026-10-15-apollo-glm-5-3-update-aware-reasoning/index.md)
- 2026-10-14: [DeepMind: no measurable AI R&D acceleration yet](https://postcutoff.com/e/2026-10-14-deepmind-no-significant-ai-rd-acceleration-audit/index.md)
- People: [Sergey Brin](https://postcutoff.com/person/sergey-brin/)