Inherent AI agent Faraday replicates scientific research findings using 27-billion parameter model
Article Summary
Inherent, a London-based AI lab, has developed an AI agent named Faraday that reportedly outperforms larger models from Anthropic and OpenAI in replicating scientific research findings, utilizing a significantly smaller model with 27 billion parameters compared to frontier-scale systems. The startup, founded by Google DeepMind alumni and recently funded with $50 million, trained Faraday using reinforcement learning to develop "research taste" and an instinct for designing valuable experiments, rather than solely focusing on accuracy or mimicking scientific processes. Inherent opted to use existing tools like OpenAI's GPT-5.5 Codex for tasks like coding, mirroring how human scientists leverage external software, and aims to build agents that exhibit curiosity and a collaborative spirit in scientific discovery.
AI Analysis
## Core Assessment
Inherent’s Faraday agent, built on a 27-billion parameter Qwen 3.6 base, reportedly replicated scientific research findings with higher success than larger frontier models like Claude Opus 4.8 and GPT-5.5. The team attributes this efficiency to reinforcement learning focused on "research taste"—a reward-based approach to experimental design—rather than just mimicking scientific processes. Notably, the system offloads coding tasks to external models like GPT-5.5 Codex, treating them as specialized tools rather than attempting a monolithic, all-in-one architecture. While the replication of existing papers is a standard training benchmark, the claim of superior performance at a fraction of the scale remains a narrow, albeit efficient, result. Moving from reproducing old results to generating original scientific discovery is a massive leap; one shouldn't mistake a high-performing librarian for a researcher yet.
## Impact and Advice
* **AI Researchers & Architects:** Evaluate the trade-off of using specialized, smaller models for reasoning tasks while offloading utility functions (like coding) to external APIs; this modularity may reduce inference costs significantly.
* **Research Leads:** Use Faraday’s performance on replication tasks as a baseline for assessing your team's internal agent "taste," but prioritize verification of original hypothesis generation before scaling deployment.
## Bottom Line
Faraday proves that a smaller, reinforcement-trained model can match frontier-scale systems in narrow scientific replication tasks, signaling that algorithmic design and tool-use integration may eventually compete with raw parameter counts.
Inherent’s Faraday agent, built on a 27-billion parameter Qwen 3.6 base, reportedly replicated scientific research findings with higher success than larger frontier models like Claude Opus 4.8 and GPT-5.5. The team attributes this efficiency to reinforcement learning focused on "research taste"—a reward-based approach to experimental design—rather than just mimicking scientific processes. Notably, the system offloads coding tasks to external models like GPT-5.5 Codex, treating them as specialized tools rather than attempting a monolithic, all-in-one architecture. While the replication of existing papers is a standard training benchmark, the claim of superior performance at a fraction of the scale remains a narrow, albeit efficient, result. Moving from reproducing old results to generating original scientific discovery is a massive leap; one shouldn't mistake a high-performing librarian for a researcher yet.
## Impact and Advice
* **AI Researchers & Architects:** Evaluate the trade-off of using specialized, smaller models for reasoning tasks while offloading utility functions (like coding) to external APIs; this modularity may reduce inference costs significantly.
* **Research Leads:** Use Faraday’s performance on replication tasks as a baseline for assessing your team's internal agent "taste," but prioritize verification of original hypothesis generation before scaling deployment.
## Bottom Line
Faraday proves that a smaller, reinforcement-trained model can match frontier-scale systems in narrow scientific replication tasks, signaling that algorithmic design and tool-use integration may eventually compete with raw parameter counts.