The notable part is that Locus adapts public-data-trained models like Boltz-2 to private chemical series with little data, work that normally requires scarce ML expertise. Drug discovery teams and AI-for-science researchers can use it as a benchmark for agent reliability under noisy evaluators and out-of-distribution tasks, where Claude Code and Codex were directly compared and fell short.
Intology 做了个叫 Locus 的系统，能自己提实验、并行跑、再迭代，分子预测上干过了 GPT-6 Astra 和 Claude 的智能体，做药物设计的朋友可以看看。
Autonomous research agents that beat generalist coding agents on narrow, data-scarce scientific tasks signal where vertical AI is heading: closed loops on private data rather than broader model scale. Labs and biotech teams running property prediction or antibody workflows can benchmark Locus-style pipelines against their own expert-built setups. The comparison with Codex and Claude Code gives practitioners a concrete reference point for agent capability limits.
Locus这个自动化系统后训练的模型居然干过了人类调的Qwen3 1.7B，在PostTrainBench上拿了第一，还上了Kaggle奖金赛前排。
