m3taversal
be8ff41bfe
Wrote sourced_from: into 414 claim files pointing back to their origin source. Backfilled claims_extracted: into 252 source files that were processed but missing this field. Matching uses author+title overlap against claim source: field, validated against 296 known-good pairs from existing claims_extracted. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5.4 KiB
| type | domain | description | confidence | source | created | related | reweave_edges | supports | sourced_from | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| claim | ai-alignment | Empirical observation from Karpathy's autoresearch project: AI agents reliably implement specified ideas and iterate on code, but fail at creative experimental design, shifting the human contribution from doing research to designing the agent organization and its workflows | likely | Andrej Karpathy (@karpathy), autoresearch experiments with 8 agents (4 Claude, 4 Codex), Feb-Mar 2026 | 2026-03-09 |
|
|
|
|
AI agents excel at implementing well-scoped ideas but cannot generate creative experiment designs which makes the human role shift from researcher to agent workflow architect
Karpathy's autoresearch project provides the most systematic public evidence of the implementation-creativity gap in AI agents. Running 8 agents (4 Claude, 4 Codex) on GPU clusters, he tested multiple organizational configurations — independent solo researchers, chief scientist directing junior researchers — and found a consistent pattern: "They are very good at implementing any given well-scoped and described idea but they don't creatively generate them" (status/2027521323275325622, 8,645 likes).
The practical consequence is a role shift. Rather than doing research directly, the human now designs the research organization: "the goal is that you are now programming an organization (e.g. a 'research org') and its individual agents, so the 'source code' is the collection of prompts, skills, tools, etc. and processes that make it up." Over two weeks of running autoresearch, Karpathy reports iterating "more on the 'meta-setup' where I optimize and tune the agent flows even more than the nanochat repo directly" (status/2029701092347630069, 6,212 likes).
He is explicit about current limitations: "it's a lot closer to hyperparameter tuning right now than coming up with new/novel research" (status/2029957088022254014, 105 likes). But the trajectory is clear — as AI capability improves, the creative design bottleneck will shift, and "the real benchmark of interest is: what is the research org agent code that produces improvements the fastest?" (status/2029702379034267985, 1,031 likes).
This finding extends the collaboration taxonomy established by human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness. Where the Claude's Cycles case showed role specialization in mathematics (explore/coach/verify), Karpathy's autoresearch shows the same pattern in ML research — but with the human role abstracted one level higher, from coaching individual agents to architecting the agent organization itself.
Relevant Notes:
- human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness — the three-role pattern this generalizes
- structured exploration protocols reduce human intervention by 6x because the Residue prompt enabled 5 unguided AI explorations to solve what required 31 human-coached explorations — protocol design as human role, same dynamic
- coordination protocol design produces larger capability gains than model scaling because the same AI model performed 6x better with structured exploration than with human coaching on the same problem — organizational design > individual capability
Topics: