DeepMind Alumni Startup Inherent Beats Frontier Models at Scientific Replication

London-based AI lab Inherent has released Faraday, an AI agent capable of independently replicating findings from published scientific papers. The startup, founded by Google DeepMind alumni, claims its system outperformed the industry's largest frontier models in this specialized research task.

Faraday surpassed both OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 in its ability to reproduce scientific results. Notably, Faraday achieves these results using the Qwen 3.6 model, which contains 27 billion parameters—a fraction of the size of the massive systems it beat. To handle the technical requirements of the process, Faraday integrates OpenAI's GPT-5.5 Codex for its coding tasks.

The Pursuit of 'Research Taste'

Inherent emerged from stealth with a $50 million seed round. The company is moving beyond simple verification, aiming to build an "AI scientist" capable of genuine discovery. To achieve this, Inherent utilizes reinforcement learning to reward successful outcomes and develop what they call "research taste," rather than relying on rigid, rule-based instructions.

This approach mimics the early stages of academic training. The process of replicating existing research allows the agent to build foundational skills, a method common among PhD students.

Implications for AI Scaling

The success of Faraday suggests a potential shift in how the industry views model scale. The fact that a 27B parameter model can outperform frontier-scale systems in complex R&D tasks indicates that specialized training methods—specifically reinforcement learning focused on scientific intuition—may be more critical than raw parameter count for discovery.

If specialized training can substitute for massive compute, the barrier to creating highly capable AI agents for research and development could drop significantly. This would allow smaller labs to develop potent tools for scientific discovery without needing the infrastructure of a tech giant.

The Road to Discovery

While Faraday has proven its ability to replicate known findings, the ultimate goal for Inherent is the discovery of new knowledge. The company is positioning Faraday as an "AI teammate" that can operate with the autonomy of a human researcher. Future benchmarks will likely focus on whether this "research taste" can translate from replicating existing papers to proposing and executing original, successful experiments.