DeepMind alumni bring a smaller model into the AI scientist race

A smaller system takes aim at a difficult research task

A relatively quiet London startup is now making a claim in one of the most ambitious areas of AI: scientific research. Inherent, founded by former Google DeepMind employees, says its Faraday agent has outperformed much larger systems from Anthropic and OpenAI at independently reproducing findings from published papers.

The claim should be treated with appropriate caution. The supplied report does not provide the full benchmark, sample count, task selection, scoring methodology, or independent verification. Faraday’s result is therefore best read as a company-reported milestone rather than conclusive evidence that a smaller model is broadly more capable than frontier systems.

What Faraday is being asked to do

• Work from a published paper without being given the expected answer.
• Decide which experiments to run and how to implement them.
• Use code and tools to test whether the reported findings can be reproduced.
• Go beyond numerical accuracy by demonstrating what Inherent calls scientific taste.
• Explore a training approach based on reinforcement learning rather than simple imitation of research workflows.

The distinction matters. A system that summarizes a paper or writes a plausible explanation is not necessarily doing science. Reproduction requires a chain of decisions involving literature interpretation, experimental design, software execution, debugging, and analysis of unexpected results. It is also a familiar starting point for human researchers, many of whom learn by checking and extending existing work.

Training for judgment, not just completion

Inherent’s longer-term ambition is to build an AI scientist that can contribute to new discoveries. Faraday is positioned as an intermediate step: an agent that can investigate an existing result, identify useful follow-up experiments, and return with evidence rather than simply telling a user what they want to hear.

The company is using reinforcement learning to pursue that goal. Instead of specifying every rule for good research, the system receives feedback based on outcomes. In theory, that could help an agent learn which experiments are informative and which are a waste of time. In practice, scientific quality is difficult to score. A failed reproduction may reveal a real flaw, a missing dependency, an environmental mismatch, or an implementation mistake. Distinguishing those cases requires careful benchmarks and human review.

One striking design choice is what Inherent is not building. Faraday reportedly relies on OpenAI’s GPT-5.5 Codex for coding rather than developing a proprietary programming model. That mirrors how human scientists use existing software: the scarce capability may lie less in writing every tool from scratch than in framing questions and closing the loop between hypothesis, experiment, and interpretation.

Why the claim matters

Faraday’s reported performance points to a possible advantage for smaller, specialized systems. With a model of roughly 27 billion parameters, Inherent says it can compete with much larger frontier-scale models when the agent architecture, tools, and training objective are aligned to a specific workflow. That does not mean model size has stopped mattering, but it does suggest that capability can be distributed across the model, the surrounding tools, and the reward process.

The larger test will be generalization. Reproducing known papers is far removed from finding genuinely new knowledge. The next questions are whether Faraday can work across scientific fields, cope with unfamiliar code and incomplete instructions, and produce conclusions that outside researchers can independently verify.

Inherent is also betting on London as a dense talent hub. Its team currently numbers about a dozen people and plans to grow to roughly 20 to 25 by year-end. As the company recruits among DeepMind alumni and other researchers, its progress will offer a useful case study in whether a small European lab can turn agentic research workflows into a durable advantage.

Source: TechCrunch AI