Billion-Dollar Bets on RSI; Hard Limits Persist

Both Anthropic and OpenAI made unusually clear statements saying recursive self-improvement is underway at their labs. Researchers including John Schulman identify hard limits. AI can speed up well-specified research but cannot discover new training paradigms. They argue that "thinking alone can't generate new bits." Zhipu AI dedicated 60% of its $5 billion raise to RSI loops. Oriol Vinyals says RSI will happen but will be slow. He argues the real bottlenecks are research taste and reliably evaluating whether self-improvement actually occurred.

The debate sharpened this week. Recursive is a new startup with a $4.65 billion seed round. Founder Richard Socher showed an AI beating humans on optimization tasks in under two days. It also found GPU kernel improvements without CUDA experts. Socher says reward hacking is the core unsolved alignment problem and that AI research could compress from years to weeks.

Labs and builders pursuing RSIRSI is already underway and worth betting billions on; capability compression from years to weeks is within reach

Ethan Mollick · The Wharton School · One Useful ThingAnthropic and OpenAI both signaled early recursive self-improvement achieved, with anxious tones suggesting genuine concern over compounding leads.Read this signal

Azeem Azhar · Exponential ViewZhipu AI's $5B raise dedicates 60% to next-gen GLM models and fully self-training RSI loops, signaling a scaling-plus-autonomy pivot.Read this signal

TETeortaxes · IndependentXi's BRICS open-source push may be compatible with pacing deals, while Chinese models pursue recursive self-improvement and market undercutting.Read this signal

swyx · Cognition · Latent SpaceRecursive ($4.65B seed) demos AI that beat humans on optimization tasks in <2 days and found NVIDIA GPU kernel improvements without CUDA expertsRead this signal

Researchers seeing hard limitsAI can accelerate constrained research but cannot generate paradigm-level breakthroughs; RSI will be gradual, bottlenecked by research taste and evaluation

Dwarkesh Patel · Dwarkesh PodcastJohn Schulman, Beren Millidge, and Charlie O'Neill argue RSI faces hard limits: AI can 10x well-specified research but cannot think its way to new training paradigms.Read this signal

The Decoder · The DecoderVinyals argues AI self-improvement will be gradual, not explosive — bottlenecked by 'research taste' and evaluation, not computeRead this signal