Recursive Superintelligence

1. Investors

GV's thesis post: Why Self-Improving AI is the Next Frontier. Wilson Sonsini was legal advisor on the round.

2. Co-founders (the 8)

3. The technical thesis

Tian's RSI framework (ICLR April 2026 workshop)

Tian's framing: "Model + Harness = Self-Improvement" — same template as AlphaGo (policy-net + MCTS), now with an LLM where the Go-net was. Four engineering pillars:

1. Data, data, more data — synthetic environments via Solver↔Challenger self-play (Language Self-Play / LSP), regularized to avoid reward hacking
2. Memory + long context + continuous learning — three primitives: retrieval, evolution (state updating, causal inference), condensation (state abstraction). TTT-Discover for continuous fine-tune. Watch for "Lost in the Middle" failures
3. Better search — representation of action space matters more than search algo. LaMCTS partitions latent space then searches inside. AlphaEvolve and training-free GRPO challenge "is RL even the right paradigm?"
4. Cost — scaling ladders + surrogate models + interpretability ("open the blackbox") all required for RSI to be economical

"Keep your prompt prefix stable. Make your context append-only."

— Tian, RSI workshop slide on context engineering

"With Search (harness) >> Without search. Maybe Agent Harness is still needed even if the model is strong."

— Tian, contrasting Grandmaster-Level Chess Without Search

"Can the research ideas presented in this talk [be] automatically discovered by AI?"

— Tian, closing slide

Socher quotes (TechCrunch, May 14)

"Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved."

"Our main focus is to build truly recursive, self-improving superintelligence at scale, which means that the entire process of ideation, implementation, and validation of research ideas would be automatic."

"There will be products, and you'll have to wait quarters, not years."

"Compute is not to be underestimated. I think in the future, a really important question will be: How much compute does humanity want to spend to solve which problems?"

4. Reading list

Tier 1 — must-read before any conversation

1. MUST Clune, "AI-GAs: AI-Generating Algorithms" — arXiv:1905.10985 — the manifesto paper for the whole company
2. MUST AlphaEvolve (Novikov et al., DeepMind 2025) — arXiv:2506.13131 — closest published analog to "Level 1" system
3. MUST Sakana, "The AI Scientist" (Clune senior author) — sakana.ai/ai-scientist
4. MUST POET / Enhanced POET (Clune + Stanley) — arXiv:1901.01753
5. MUST MAP-Elites (Mouret & Clune) — arXiv:1504.04909 — foundational quality-diversity algorithm
6. MUST OMNI (Zhang, Lehman, Stanley, Clune; ICLR 2024) — openreview
7. MUST OMNI-EPIC (Faldor, Zhang, Cully, Rocktäschel) — arXiv:2405.15568
8. MUST SWE-agent (Yang et al., NeurIPS 2024) — arXiv:2405.15793
9. MUST "A Self-Improving Coding Agent" (Robeyns et al., 2025) — arXiv:2504.15228 — cleanest small-scale RSI demo
10. MUST Tian's RSI deck — PDF

Tier 2 — should read

11. SHOULD Live-SWE-agent — arXiv:2511.13646 — runtime-self-evolving agents; 77.4% on SWE-bench Verified
12. SHOULD Darwin–Gödel Machine (Sakana / Clune) — sakana.ai/dgm
13. SHOULD Quality-Diversity through AI Feedback (Bradley et al., ICLR 2024)
14. SHOULD World Models (Ha & Schmidhuber 2018) — arXiv:1803.10122
15. SHOULD ViT (Dosovitskiy et al., ICLR 2021) — arXiv:2010.11929 — Dosovitskiy's signature paper
16. SHOULD RAG (Lewis et al., NeurIPS 2020) — arXiv:2005.11401
17. SHOULD NAS with RL (Zoph & Le, ICLR 2017) — arXiv:1611.01578

Tier 3 — bonus depth signals

18. BONUS Rainbow Teaming / MADRID (Samvelyan, Rocktäschel et al., AAMAS 2024)
19. BONUS "Abandoning Objectives" (Lehman & Stanley 2011) — philosophical foundation for open-endedness
20. BONUS Toolformer (Schick et al., 2023) — arXiv:2302.04761
21. BONUS OpenEvolve — HuggingFace blog + repo — open-source AlphaEvolve reimpl. Best codebase to fork.

5. How to apply

The email to talent@recursive.com

Keep it ≤ 250 words. Successful applications to OpenAI/DeepMind/Anthropic from non-publishing engineers share four moves:

1. Name a specific paper / talk from a co-founder and explain why a specific follow-up is non-obvious
2. Link a personal artifact that re-implements or extends a piece of their stated agenda
3. State one concrete thing you'd build in the first 60 days
4. Compress credentials into one line

Subject line = a concrete artifact (e.g., "OpenEvolve fork: 3× cheaper inner loop via X — would love to contribute"). Attach: one PDF résumé + one GitHub link. Nothing else. Skip the "love your mission" framing — they have 25 researchers and know the mission.

Co-founder DMs (matched to mid-senior SWE with ML exposure)

Side projects to build (priority order)

1. Fork OpenEvolve — add one defensible improvement (smarter island migration, programmatic eval cache, MAP-Elites archive instead of flat population). Ship a writeup with before/after numbers on a public benchmark. Highest-signal thing you can do in a week.
2. Tiny "AI Scientist" loop (ideate → code → eval → write) on a toy domain (e.g., evolve sklearn pipelines on OpenML tasks)
3. Self-improving SWE-agent variant that rewrites its own scaffolding per arXiv:2504.15228 — show measurable gain on a 20-task SWE-Bench subset
4. PRs to jennyzzt/awesome-open-ended, OpenEvolve, or pyribs (QD library) — visible, low-ego, attention-getting

6. Comparable labs (backup options)

7. Recommended action plan (this week)

The window matters — they emerged from stealth yesterday, hiring is wide-open, and being one of the first ~50 applicants with a credible artifact is a much better position than being applicant #500 in two months.

8. Custom findings & new info

9. Sources

Primary launch coverage

• TechCrunch — What happens when AI starts building itself? (May 14)
• SiliconANGLE launch coverage
• Tech.eu launch coverage
• TheNextWeb launch coverage
• TechFundingNews — UK $4.65B valuation
• Wilson Sonsini — legal advisor announcement
• GV thesis post on RSI
• CNBC — Jensen Huang bets on the British startup

Founder homepages

• Richard Socher
• Yuandong Tian
• Tim Rocktäschel
• Josh Tobin
• Jeff Clune
• Alexey Dosovitskiy — Google Scholar
• Caiming Xiong — Salesforce author page
• Tim Shi — LinkedIn