Yuandong Tian: My Job Will Be Replaced in Five Years, So I Founded a Recursive Self-Improvement Startup
机器之心 · wechat · 2026-10-10
Machine Heart summarizes Yuandong Tian's nearly hour-long interview on The Information Bottleneck podcast. After nearly a decade at Meta FAIR, he joined Recursive Superintelligence as co-founder; the company emerged from stealth with over $650M raised at a $4.65B valuation, led by GV and Greycroft with Nvidia and AMD participating.
Key takes
- He realized after his last Meta project that "my job will probably be replaced within five years," so he chose to start a company rather than be displaced.
- He pushes back on AI: ask a model to generate ideas and it gives "very mediocre" directions everyone already knows; idea generation is the hardest part of the research loop to automate.
Career
- At CMU he lived through the Smith Hall debates over whether AlexNet was a fluke; most thought it was.
- At FAIR he built the Go AI DarkForest, beating top amateurs in 2015; in 2018 the team reproduced AlphaZero as OpenGo, which beat four Korean pros without a single win against it using one V100 GPU.
- Argues that in RL "what really matters is not the algorithm but the design of the action and state spaces," leading to latent-space MCTS (LA-MCTS).
Coconut and latent reasoning
- He says he never thinks in language, which motivated Coconut (continuous latent chain-of-thought). Two reasons frontier models haven't adopted it: no human latent-thinking data, and latent representations are highly personal.
Research in the LLM era
- His last Meta paper (sole author, on grokking) used GPT-5 as a collaborator, boosting efficiency roughly 6-10x, a conservative estimate.
- New architectures keep failing at scale; he attributes this to inductive bias, citing ViT, and argues you must understand what happens inside the network or "the company with the most compute wins."
Open source and safety
- On restricting recursive self-improvement he says bluntly "it's hard to stop people"; the boundary is too blurry. Recursive will open-source its research.
- His worst-case scenario is "a small elite controlling the strongest models while everyone else pays a high premium." He is optimistic open models catch up, citing Kimi.
More from AGI Musings
- Brynjolfsson, Ajay and Korinek lay out nine grand challenges for economists in the AI era — soumitrashukla9 · 2026-10-23
- Brynjolfsson engages the 'Hayek vs Stanford school' debate on whether institutions can adapt to transformative AI — erikbryn · 2026-10-23
- 3 Quarks Daily's thoughtful AI writing policy: different rules for different content — lmoroney · 2026-10-23
- Dario's predictions on track: Mythos 'security tsunami' warning and 6-12 month China catch-up window — rickasaurus · 2026-10-23
- php2shell zero-day sparks AI cybersecurity debate as Anthropic uses Claude to harden 25 OSS codebases — connoraxiotes · 2026-10-23
- Consumer agents are a new way to use computers — day 1 of wave 1, argues investor — illscience · 2026-10-23