ML Coding

ML Coding is Yao Shunyu / 姚顺宇’s term in 140. 对姚顺宇的4小时访谈：请允许我小疯一下！在Anthropic和Gemini训模型、技术预测、英雄主义已过去 for AI systems helping with machine-learning research itself. It extends ordinary AI coding from application implementation into writing experiment code, running jobs, reading outputs, diagnosing failures, analyzing results, and proposing the next training or research hypothesis.

138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权 adds Luo Fuli / 罗福莉’s practical acceleration view. Agents can shorten the path from idea to code to evaluation, but that makes cards, evaluation, Research Taste, and Training Compute Allocation tighter bottlenecks rather than eliminating the need for model-team judgment.

136. 全球大模型季报第9集：和广密聊，Coding是AGI第二幕、硅谷御三家真相、模型正成为新一代OS makes ML Coding part of AGI Three Acts. The source argues that coding agents may have already started accelerating AI research by compressing implementation and data-iteration loops, and that the next act is automated AI researchers that can propose, run, and interpret experiments.

171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？ adds the Q2 2026 Auto Research and Recursive Self-Improvement bridge. The source treats coding as current revenue, user data, and model iteration substrate; if models can read papers, write experiments, run them, analyze results, and improve training recipes, ML Coding becomes part of the RSI path rather than only a productivity tool.

Key Claims

• Ordinary coding was the first breakout domain because code has clear feedback, executable tests, strong public data, and relatively shared standards.
• ML Coding raises the stakes: the code is not just product code, but part of model training, data processing, evaluation, and research-loop design.
• The concept connects to Recursive Self-Improvement without assuming full autonomy: a model that helps improve future models must still pass AI Verification, AI Coding Verification, and human research judgment.
• The hard part is not only writing code quickly; it is knowing which experiment is well-defined, which data and environment are meaningful, which failure mode matters, and which hypothesis is worth testing next.
• As ML Coding improves, the bottleneck can move toward Research Taste, verifier quality, compute allocation, and organization coordination rather than raw implementation speed.
• In an agent framework, ML Coding can involve multiple agents exploring variants and cross-checking results, which raises the importance of Agent RL and evaluation infrastructure.
• Episode 136 treats ML Coding as the bridge from coding agents to automated AI researchers, not only as a productivity aid for existing researchers.
• The LateTalk source adds that ML Coding is also a competitive weapon for frontier labs because better coding agents can accelerate model-company research itself.

Connections

• Yao Shunyu / 姚顺宇 — source speaker and practitioner.
• AI Coding Verification — operational constraint because weak tests or misleading metrics can corrupt the research loop.
• AI Engineering Thinking — habit of turning vague research goals into explicit experiment code, logs, metrics, and reviewable artifacts.
• Recursive Self-Improvement, Agent Self-Evolution, and Model Harness Co-Evolution — adjacent self-improvement loops that depend on code, tools, and verification.
• AI For Science, Discovery Model, and Research Taste — broader scientific-AI context where generated hypotheses must be worth testing.
• Frontier Model Scaling and Long-Horizon AI — model-training and extended-task directions that ML Coding is meant to support.
• Luo Fuli / 罗福莉, Open Claw, Agent Post-Training, and Training Compute Allocation — agent-accelerated research loop added by episode 138.
• AGI Three Acts, Model Harness Co-Evolution, Claude Code, and Codex — coding-to-research automation route added by episode 136.
• Auto Research, Recursive Self-Improvement, Recursive, and AI Investment Metrics — Q2 2026 coding-to-RSI update added by LateTalk.