London AI lab Inherent has released Faraday, an autonomous agent that outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing published scientific research, all while running on a model roughly 1 percent the size of those frontier systems. The announcement, published on August 22, marks one of the first concrete results from a growing wave of DeepMind alumni startups.
Inherent, a British research lab founded by four former Google DeepMind researchers, emerged from stealth earlier this month with a $50 million seed round. Faraday is the company's first public product, and it is built entirely for one purpose: replicating scientific papers without ever being told the answer in advance.
The Benchmark: Can AI Replicate Science?
Faraday was evaluated on blind paper replication tasks, where the agent received only the text of a published research paper and was asked to independently reproduce its findings. No source repositories or ground-truth outputs were provided. The agent had to read methodology, extract experimental design, write the necessary code, and validate results.
On this benchmark, Inherent says Faraday achieved higher replication fidelity than Claude Opus 4.8 and GPT-5.5. Chief scientist Edward Hughes told TechCrunch that the comparison was not the point. "What was most interesting to us about this was not so much the result of beating those frontier agents, which of course we liked, but was actually the way we went about building this," Hughes said.
Paper replication is a standard training exercise for human scientists. "Many PhD students actually start by doing this," Hughes noted. The task requires reading complex methodology, understanding experimental design, and executing procedures, which mirrors the foundational work every researcher does before attempting original discovery.
Building on Qwen 3.6: Size Is Not the Whole Story
Faraday runs on Qwen 3.6, an open-weight model from Alibaba with 27 billion parameters. To put that in context, frontier models from OpenAI, Anthropic, and Google operate in the hundreds of billions of parameters. The parameter count is a rough proxy for training expense, deployment overhead, and inference cost. Running a 27B agent versus a multi-trillion-parameter system changes unit economics fundamentally.
Rather than training a massive proprietary foundation model, Inherent applied specialized reinforcement learning algorithms to Qwen 3.6. The training objective focused on experimental design, hypothesis testing, and error correction, with a particular emphasis on what the company calls "research taste."
"Research taste" is Inherent's term for the ability to prioritize high-yield experiments, assess empirical validity, and propose follow-up work worth exploring. It is the kind of judgment that cannot be captured by static prompt templates or procedural rules. Inherent used reward-based reinforcement learning to teach the agent this skill directly, rather than showing it labeled examples of correct scientific reasoning.
"This mirrors how human PhD students learn science through iterative feedback, not lectures alone," wrote LLMs Blog in its analysis of the announcement. "The approach generalizes better to open-ended discovery work than supervised fine-tuning on static datasets."
Architecture: Pragmatic Choices
Inherent did not build every component from scratch. When Faraday needs to write or debug executable code, it delegates that task to OpenAI's GPT-5.5 Codex. This mirrors how human researchers rely on existing software utilities rather than writing custom tooling for every project.
The result is a system where a 27B reasoning model handles experiment design and interpretation, while a larger specialized model handles code generation. This hybrid approach lets Inherent focus engineering effort on the harder problem: teaching an agent how to think about science rather than how to write Python.
Inherent also designed Faraday to avoid sycophancy. The agent is trained to challenge user assumptions rather than blindly agree with inputs. "We're always guided by that north star of building an AI scientist agent and imbuing our agents with taste," Hughes said. The ideal interaction, he described, is one where the agent comes back and says, "I got curious about this, and I went off and I did these experiments. What do you think of these results?"
The DeepMind Spillover and London's AI Cluster
Inherent was founded by Edward Hughes alongside fellow DeepMind alumni Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins. The company operates from an office in King's Cross, London, just steps from Google DeepMind's global headquarters. The team currently has 12 employees and plans to expand to between 20 and 25 by the end of 2026.
London has quietly become one of the densest AI research hubs outside Silicon Valley. DeepMind's presence over the past decade anchored top-tier machine learning talent in the U.K., creating a rich hiring pool for spinouts. Hughes told TechCrunch, "We believe that London is the place to be."
The startup also highlighted a structural barrier to U.K. AI hiring: garden leave clauses, which bar departing employees from joining rivals for months after resigning. Hughes noted this practice affects his own hiring experience, while American startups face no comparable legal friction.
What This Means for Scientific AI
The broader signal is that specialized agents trained with reinforcement learning may deliver superior performance-per-dollar in narrow domains. If a 27B model can beat frontier-scale systems at a cognitively demanding task like paper replication, the market may be entering a phase where training methodology and task alignment matter as much as parameter count.
Inherent is pursuing one of AI's highest-value applications: accelerating scientific discovery. Success in this domain could unlock partnerships with pharmaceutical, biotech, and materials science companies. These verticals have funding and tolerance for specialized tools that generalist frontier labs are unlikely to build.
The company's longer-term ambition extends beyond replication toward generating novel scientific hypotheses. Paper replication is a validation step, not the end goal. Hughes told SingularityMoments that world models form another core element of Inherent's roadmap, allowing agents to simulate physical systems virtually before running experiments in physical labs.
Skepticism Is Warranted
The claims deserve healthy skepticism. Inherent's results rest on internal benchmarks. Independent verification by academic institutions or third-party evaluation suites would strengthen credibility. The company has not yet released its full replication dataset or benchmark code to the public.
Data contamination is another concern. Qwen 3.6 may have ingested research paper texts during its initial pretraining runs, which could give it an unfair advantage on replication tasks. Without third-party audit logs, industry observers cannot fully verify whether Faraday reasoned through unfamiliar papers or recognized patterns from its training data.
Faraday also depends on external tooling. When its code generation calls to GPT-5.5 Codex experience latency or downtime, operational throughput drops accordingly. The efficiency advantage is real for the 27B backbone, but the total inference cost includes the parameters of those external API calls.
Still, the result is notable. If small-footprint agents can reliably handle complex scientific verification tasks, automated research could become accessible to academic labs operating on modest budgets. Inherent's $50 million seed round gives it the runway to test this hypothesis at scale.
The company's progress also sits against a backdrop of shifting competition in frontier models. Just last week, Anthropic released Claude Opus 5, positioning itself as a safety-first alternative in an increasingly crowded market. Inherent's approach, specialized agents that beat generalist systems at defined tasks, represents a fundamentally different strategy from the brute-force scaling model that has dominated the field.