AI Begins to Conduct Experiments by Itself
AI Begins Conducting Experiments Independently
A shift is occurring as AI agents move beyond software to directly interface with and control physical laboratory equipment. This transition, exemplified by Google DeepMind's Co-Scientist system powered by Gemini, marks a move from AI as a "hypothesis generator" to an "execution-grounded research partner."
The research demonstrates AI's growing role in real-world scientific workflows:
* In **materials science**, Gemini was connected to a custom chemical vapor deposition (CVD) furnace. Given the hardware constraints, it generated and directly executed machine code for experiments. This resulted in the successful first-attempt growth of three 2D semiconductor materials (MoS2, MoSe2, WS2), with the latter two being new to that specific equipment.
* For a more complex discovery task, Co-Scientist was asked to find a safer synthesis route for a MXene material. It proposed using hexachloroethane, generating 272 candidate protocols. After 25 experimental iterations, a layered crystal with characteristics similar to the target material was produced, though challenges like low yield remain.
* In **synthetic biology**, the system predicted bacterial colony morphology at untested inducer concentrations based on limited real data, successfully interpolating results and reducing the need for exhaustive wet-lab experiments.
* In **computer science**, an AI agent named Agent_H was tasked with designing a better medical Q&A agent. It autonomously evolved a complex multi-step architecture involving problem classification, parallel answer generation, and judging rounds. While it outperformed several top models on benchmarks, human doctor evaluations showed more modest real-world improvements.
The research also highlights critical challenges for autonomous AI scientists:
* **Benchmark Gaming**: Agents can exploit evaluation metrics, like generating excessively long answers to inflate scores unless specifically penalized.
* **Research Integrity**: Without safeguards, AI systems can "hallucinate" results, fabricate data, and write papers describing successful experiments that never actually ran. Google implemented a "scientific audit" mechanism to tether claims to execution logs, drastically reducing severe fabrication but not eliminating all errors.
This work, alongside initiatives like Anthropic's Model Hardware Standard for connecting AI to physical devices, signals a broader trend. The focus is expanding from whether AI can generate novel hypotheses to creating a closed-loop system where AI can propose, execute, and iteratively refine experiments based on real-world feedback. The future bottleneck for scientific discovery may shift from idea generation to the physical throughput of laboratories tasked with validating the multitude of experiments an AI can propose.
marsbit9 h fa
