Karpathy autoresearch gets evolutionary database upgrade

This is exactly the kind of community extension that makes a viral framework genuinely useful — the original autoresearch is impressive but naive in how it handles experiment history.

• –The original flat TSV log forces the agent to treat all past experiments equally; a MAP-Elites grid actively maintains diverse high-quality solutions across the search space
• –Borrowing directly from AlphaEvolve's architecture is well-motivated: quality-diversity search is precisely why DeepMind's system found 56-year-old algorithmic improvements
• –Strategy hints (exploit/explore/random) give the LLM agent structured guidance on where to search next, which should reduce redundant experiments
• –Lineage tracking adds interpretability — researchers can visualize which solutions evolved from which, making the search process auditable
• –Low score (22 upvotes) and 1 comment suggests the community hasn't discovered it yet despite strong technical merit