Inherent's AI agent outperforms rivals in scientific paper replication

Inherent's AI agent Faraday has replicated published scientific papers more effectively than Anthropic's Claude Opus 4.8 and OpenAI's latest models in independent testing.

Inherent, a London-based AI startup founded by Google DeepMind alumni, has demonstrated that its AI agent Faraday outperforms competing models in independently reproducing published scientific research. Faraday operates using Qwen 3.6 with 27 billion parameters and was trained via reinforcement learning. The agent achieved superior replication results compared to Anthropic's Claude Opus 4.8 and OpenAI's current public models during testing. Inherent, which raised a $50 million seed round and employs 12 people, aims to scale to 20-25 staff by year-end to build AI capable of discovering new scientific knowledge rather than verifying existing results.

This capability represents a practical advancement in AI-assisted research replication. Faraday's performance demonstrates how specialized training could accelerate the validation of scientific findings without requiring human intervention. Inherent's focus on independent replication—rather than simply summarizing existing work—suggests potential for faster iteration in fields like medical research or climate science.

For the need of knowledge, this could reduce the time and resources required to verify scientific findings. If Faraday's replication efficiency scales, it might help researchers validate discoveries more quickly, potentially speeding up solutions for health or environmental challenges. However, the company explicitly states that Faraday's ability to replicate papers does not constitute independent scientific discovery, and performance metrics were not the primary focus of the testing. The real-world impact remains contingent on how well this approach translates to actual discovery rather than verification.

Source: TechCrunch