Inherent, a London AI laboratory founded by former Google DeepMind researchers, says its agent Faraday has bested larger rivals on a task that mimics a core part of scientific work: independently reproducing the results of published papers. The benchmark pitted Faraday against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT‑5.5, both frontier‑scale systems, yet Faraday, built on the 27‑billion‑parameter Qwen 3.6 model, delivered higher accuracy and demonstrated what co‑founder Edward Hughes calls “research taste.”

Hughes explained that the test mirrors how many PhD students begin their careers—by verifying existing findings before attempting new experiments. While beating the bigger models was a pleasant surprise, the team was more interested in the training approach that got them there. Inherent relies on reinforcement learning, rewarding the agent for successful outcomes rather than hard‑coding procedural rules. That method, Hughes believes, will help Faraday generalize to future tasks that go beyond replication, such as designing novel experiments.

The startup’s strategy also includes leveraging existing tools rather than building everything in‑house. Faraday uses OpenAI’s GPT‑5.5 Codex for coding tasks, echoing how human scientists adopt established software. Hughes said the goal is to create an AI teammate that proactively runs experiments and reports back, saying, “I got curious, I did these experiments, what do you think of the results?”

Inherent emerged from stealth after securing a $50 million seed round and operates out of a King’s Cross office staffed by about a dozen employees. The company plans to expand to 20‑25 people by the end of the year, positioning itself as a potential landing spot for DeepMind talent, especially as Demis Hassabis’s recent role change has stirred uncertainty at the parent lab. Hughes also voiced personal frustration with the United Kingdom’s “garden‑leave” practice, which can delay departing engineers from joining new ventures—a restriction he noted American researchers typically avoid.

Looking ahead, Inherent’s ambition extends to building AI agents capable of original scientific discovery, not merely verification. The firm’s north‑star remains an “AI scientist” that can assess which experiments are worth pursuing and execute them autonomously. If Faraday’s early success translates to broader research domains, the startup could stake a claim in a field where larger, better‑funded competitors have yet to deliver concrete breakthroughs.

Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.