Faraday’s early results suggest that the path toward useful AI scientists may depend less on maximizing model size and more on cultivating the curiosity, experimental judgment and specialized behavior that make human researchers effective. (Source: Image by RR)

Faraday Tests Whether AI Can Develop Curiosity and Experimental Judgment

London-based AI startup Inherent, founded by former Google DeepMind researchers, says its newly unveiled scientific agent Faraday has outperformed significantly larger frontier systems from Anthropic and OpenAI at independently reproducing results from published scientific papers. Replication, as noted in the article at techcrunch.com, is a foundational part of scientific training, and Inherent views the task as an intermediate step toward a much more ambitious goal: developing AI agents capable of discovering genuinely new scientific knowledge rather than simply retrieving or verifying what humans already know.

What makes Faraday particularly notable is its relatively modest underlying model. The agent operates on Qwen 3.6, a 27-billion-parameter model, yet Inherent says it surpassed Claude Opus 4.8 and GPT-5.5 on its research-replication evaluation. Rather than attempting to compete by training an enormous frontier foundation model, Inherent is concentrating on the agentic system built around a smaller model—using reinforcement learning to teach Faraday how to conduct research, select useful experiments, investigate promising leads and develop what the company calls “research taste.”

That concept is central to Inherent’s strategy. The company doesn’t simply want an AI system capable of executing instructions; it wants something resembling an intellectually curious scientific collaborator—an agent willing to independently investigate an unexpected result, conduct follow-up experiments and return to its human teammate with potentially useful discoveries. Inherent is also deliberately using existing tools where appropriate: rather than developing its own programming system, Faraday can use OpenAI’s Codex much as a human scientist might rely on specialized software created by someone else.

The achievement remains narrow, and successfully replicating existing scientific research is substantially different from originating important new discoveries. But Inherent’s approach highlights a potentially important direction for AI development: specialized agents built around smaller models may outperform far larger general-purpose systems when trained intensively for specific kinds of reasoning and behavior. If that approach scales from replication toward genuine discovery, the next major advance in AI-assisted science may depend less on building ever-larger models and more on teaching machines how good scientists actually think and explore.

read more at techcrunch.com