Superintelligence & RSI
Explore self-improvement methods, the evidence behind them, and the questions still open.
Research frontier ·
Definitions & the improvement loop
Superintelligence means broadly exceeding human expert capability, including research—not just winning one benchmark. Recursive self-improvement (RSI) is a feedback loop where an improved AI system helps build an even better successor or improvement process. No source below establishes broad superintelligence or an indefinitely accelerating loop.
- 01Propose a change
- 02Run an experiment
- 03Verify on held-out tasks
- 04Retain gains & repeat
Three different things often get called self-improvement: spending more compute on an answer, changing the agent’s code or memory, and training new model weights. Only the last changes the model itself. Each can help research, but gains must survive independent evaluation and repeated cycles.
Current directions, with evidence and limits
MASS2026-10-12Learn from a team, then improve the team
A shared model searches over team workflows, judges their outputs and trains on selected team traces. The updated model then repeats the process.
Sakana reports two cycles on a 27B model. This includes weight updates, but shared evaluator blind spots and generalization beyond the tested tasks remain open.
MiMo V2.62026Scale reinforcement learning in verifiable environments
Generate longer task trajectories, spend more compute on rewards and train the model on outcomes. Code execution and independent checks provide stronger feedback than self-confidence.
Xiaomi reports a six-day live RL run and held-out coding gains. This is a controlled training result; the vendor’s RSI framing is not proof of an autonomous, indefinitely improving system.
autoresearch2026Automate the experiment loop
A coding agent edits a small training program, runs a bounded experiment, compares the validation metric and keeps or discards the change.
The public repository makes the loop inspectable. It optimizes a human-defined training setup; improvement on one validation metric does not establish general research autonomy.
AlphaEvolve2025Search for algorithms with executable checks
Language models propose code; automated evaluators and evolutionary search retain stronger candidates. Better kernels and scheduling can make subsequent AI training more efficient.
Google reports deployed improvements in computing infrastructure. This demonstrates algorithm improvement, not a complete system autonomously redesigning its own intelligence.
Darwin Gödel Machine2025-05-30Let the agent improve its harness
A coding agent edits its tools and workflow, evaluates candidate agents and keeps an archive of useful variants. The foundation-model weights remain separate from these code changes.
Published coding-benchmark gains support bounded harness improvement. Passing more coding tests is not evidence of general superintelligence or mathematically guaranteed self-improvement.
Absolute Zero2025-05-06Generate a curriculum through self-play
The model proposes and solves code-reasoning tasks. Execution validates the tasks and answers, supplying rewards for both task generation and learning.
The paper reports coding and math gains without a supplied task dataset for this stage. Pretraining, the executor and human-designed rewards still matter; ‘zero data’ does not mean no prior knowledge.