When AI Research Starts Moving Faster Than Human Research - Zhengyao Jiang

Weco let an AI coding agent rewrite the harness around another agent for eight days: its code, prompts and tools, while the underlying language model stayed fixed. Tim Scarfe asks Weco co-founder Zhengyao Jiang what the reported gains over two years of human engineering actually demonstrate.The discussion examines AIDE 85's generated code, held-out evaluation and the difficulty of separating useful discoveries from reward hacking. Jiang explains Weco's four levels of recursive self-improvement and compares the experiment with AlphaEvolve and the Darwin Gödel Machine.The limits matter as much as the gains. Jiang explains why the experiment did not establish that the system had become a better improver. The conversation closes with open-ended search, human-designed primitives and Parameter Golf: where does the next useful idea come from when the agent is searching inside a space that people designed?---TIMESTAMPS:00:00:00 Eight days of self-improvement: what counts?00:03:25 AIDE and the puzzle of useful spaghetti code00:08:38 Four levels of recursive self-improvement00:12:02 What AIDE 85 changed and how it was tested00:20:04 AlphaEvolve, Darwin Gödel Machine and the RSI claim00:26:21 Reward hacking and the limits of detection00:33:09 Open-ended search, harness tuning and creativity00:39:43 Parameter Golf and the limits of self-improvement---REFERENCES:organization:[00:00:30] Weco AI00:00:33]" rel="nofollow noopener ugc" target="_blank">weco.ai/other:[00:00:33] AIDE²: The First Evidence of Recursive Self-Improvement00:14:11]" rel="nofollow noopener ugc" target="_blank">weco.ai/blog/first-evidence-of-recursive-self-… Faulty reward functions in the wild00:29:59]" rel="nofollow noopener ugc" target="_blank">openai.com/index/faulty-reward-functions/[00:2… The Hugging Face incident and the road ahead00:03:29]" rel="nofollow noopener ugc" target="_blank">openai.com/index/hugging-face-incident-and-the… AIDE00:04:29]" rel="nofollow noopener ugc" target="_blank">github.com/WecoAI/aideml[00:04:29] MLE-bench00:04:33]" rel="nofollow noopener ugc" target="_blank">github.com/openai/mle-bench[00:04:33] ALE-Bench00:04:52]" rel="nofollow noopener ugc" target="_blank">github.com/SakanaAI/ALE-Bench[00:04:52] WeatherBench 200:08:18]" rel="nofollow noopener ugc" target="_blank">github.com/google-research/weatherbench2[00:08… ReAct00:39:43]" rel="nofollow noopener ugc" target="_blank">react-lm.github.io/[00:39:43] Parameter Golf00:20:08]" rel="nofollow noopener ugc" target="_blank">github.com/openai/parameter-golfpaper:[00:20:08] AlphaEvolve: A coding agent for scientific and algorithmic discovery00:21:35]" rel="nofollow noopener ugc" target="_blank">arxiv.org/abs/2506.13131v1[00:21:35] Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents00:23:45]" rel="nofollow noopener ugc" target="_blank">arxiv.org/abs/2505.22954v3[00:23:45] Hyperagents00:27:01]" rel="nofollow noopener ugc" target="_blank">arxiv.org/abs/2603.19461v1[00:27:01] SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agents00:33:14]" rel="nofollow noopener ugc" target="_blank">arxiv.org/abs/2605.21384book:[00:33:14] Why Greatness Cannot Be Planned: The Myth of the Objectivelink.springer.com/book/10.1007/978-3-319-15524…