Conjecture Machines

2025-08-05 · Source: AI Policy Perspectives · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Intermediate, long

Summary

AI agents, dubbed "Conjecture Machines," are rapidly transforming scientific research by accelerating hypothesis generation, experimental design, and algorithm discovery. Google DeepMind's Co-Scientist, for instance, identified a complex superbug hypothesis in two days that a human team spent a decade on, and found two effective drug candidates for liver fibrosis where human picks failed. These LLM-based systems, unlike chatbots, plan, execute, and self-correct, leveraging stronger frontier models, "scaffolding" for enhanced capabilities, and customisable "skills." While agents excel at ideation and finding optimal solutions, exemplified by AlphaEvolve's role in designing Google's TPU chips and solving Erdős problems, they face a significant validation bottleneck. This challenge, particularly outside mathematics where *in silico* validation is possible, necessitates urgent attention from policymakers and funders to ensure widespread access, agent-ready data, robust experimental infrastructure, and updated peer review processes.

Key takeaway

For policymakers and science funders, you must urgently address the widening validation bottleneck created by AI agents. Prioritize investments in experimental infrastructure and automated labs to match the rapid pace of AI-generated hypotheses. Additionally, ensure widespread access to agents and agent-ready data, while updating peer review to accommodate AI-assisted research and prevent de-skilling new generations of scientists. Your strategic choices now will shape future scientific discovery.

Key insights

AI agents are accelerating scientific discovery by automating ideation and solution finding, but validation remains a critical human-dependent bottleneck.

Principles

• Agents excel at generating diverse hypotheses and candidate solutions.
• Validation, especially physical, remains a slow, costly human process.
• Scaffolding and custom skills enhance agent capabilities.

Method

An AI agent, given a goal and tools, plans steps, runs subagents, detects errors, and interacts with external databases or tools to achieve its objective.

In practice

• Use Co-Scientist for rapid hypothesis generation.
• Develop agent "skills" to capture tacit lab knowledge.
• Employ AlphaEvolve for optimizing solutions in vast spaces.

Topics

• AI Agents
• Scientific Research
• Hypothesis Generation
• Research Validation
• Science Policy
• Peer Review

Code references

• google-deepmind/science-skills

Best for: AI Scientist, Research Scientist, Policy Maker

Related on AIssential

• Advancing Scientific Discovery in the Agentic AI Era — Accelerated computing and AI agents are dramatically speeding up scientific discovery beyond human processing limits.
• How Far Are We From True Auto-Research? — Current auto-research agents generate polished papers but critically lack experimental rigor and result integrity.
• What’s your hypothesis? — AI can generate and validate scientific hypotheses significantly faster than humans, accelerating discovery.
• TAI #192: AI Enters the Scientific Discovery Loop — AI models are transitioning from tools to active participants in scientific discovery, particularly when coupled with robust verification.
• The AI Scientist: Towards full automation of the research life cycle — An AI system autonomously navigates the entire scientific research pipeline, from ideation to peer-reviewed publication.
• 17 AI Models Tested on REAL Scientific Research — Advanced AI models currently demonstrate limited capability in independent scientific discovery, even with curated data and extensive tools.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Policy Perspectives.