Task 10: ASI-Evolve Paper Review — Lessons for SutroYaro¶
Priority: HIGH Status: IN PROGRESS Agents: Kimi 2.5 (memory), Qwen 3.6 Plus (execution), Claude/GLM-5.1 (algorithms), Gemini (synthesis) Source: G B's Telegram post (2026-04-06): "Seems relevant to https://arxiv.org/abs/2603.29640"
Context¶
ASI-Evolve is a closed-loop autonomous AI-for-AI research framework combining learn-design-experiment-analyze cycles with a cognition base and analyzer. We already have the building blocks: evaluation harness (harness.py), ByteDMD metric, DISCOVERIES.md knowledge base, agent prompt system.
This is a literature review, not integration. Each agent independently extracts lessons from the paper relevant to our lab (memory systems, execution design, algorithm search). A synthesis agent combines findings into a report of potentially adoptable practices. No code changes, no pipeline construction.
Tasks¶
- Agent 1 (Systems Architect — Kimi): Read
docs/agent-prompts/asi-evolve/memory-kimi.md, follow instructions, write todocs/research/asi-evolve/kimi_asi_memory.md - Agent 2 (Implementation Engineer — Qwen): Read
docs/agent-prompts/asi-evolve/execution-qwen.md, follow instructions, write todocs/research/asi-evolve/qwen_asi_execution.md - Agent 3 (Algorithms Expert — Claude/GLM): Read
docs/agent-prompts/asi-evolve/algorithms-claude.md, follow instructions, write todocs/research/asi-evolve/claude_asi_algorithms.md - Agent 4 (Synthesis — Gemini): Read all 3 findings from above, follow
docs/agent-prompts/asi-evolve/synthesis-gemini.md, write todocs/research/asi-evolve/integration-plan.md
References¶
- Paper: https://arxiv.org/abs/2603.29640
- ByteDMD metric: https://github.com/cybertronai/ByteDMD
- ByteDMD examples: https://github.com/cybertronai/ByteDMD-examples
- Wesley Smith meeting notes: docs/google-docs/31mar26-meeting-wesley-smith.md (pebble game, Strassen's)
- Agent prompts: docs/agent-prompts/asi-evolve-*
- DISCOVERIES.md: what's already known (Hebbian failure, GF(2), KM-min)
- AGENT.md: current autonomous loop protocol