London AI Lab Faraday Beats Claude and GPT-5.5 at Science Tasks Using a 27B-Parameter Model

Aug 230 views

Key takeaways

Faraday agent runs on 27-billion-parameter Qwen 3.6, far smaller than the Anthropic and OpenAI models it outperformed.
Inherent raised a $50 million seed round and employs around 12 people, with plans to reach 20–25 by end of year.
The startup uses reinforcement learning to teach 'research taste,' aiming for agents that can discover new scientific knowledge.

London-based AI startup Inherent announced this week that its Faraday agent has outperformed frontier models from both Anthropic and OpenAI on a scientific paper replication benchmark, doing so on a model roughly a fraction of the size of its competitors. The company, which only emerged from stealth weeks ago after closing a $50 million seed round, says Faraday successfully reproduced the findings of published scientific papers without being given the answers in advance — a benchmark that pits it directly against Claude Opus 4.8 and GPT-5.5. What makes the result especially notable is that Faraday runs on Qwen 3.6, a 27-billion-parameter model, while both Anthropic and OpenAI's compared systems operate at a significantly larger, frontier scale.

Inherent was founded by four co-founders including Edward Hughes, who serves as cofounder and chief scientist, along with fellow Google DeepMind alumni Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins. The team operates entirely in person from an office in King's Cross, the London neighborhood that DeepMind's long-term presence helped transform into one of Europe's most prominent AI research hubs. Despite being relatively low-profile compared to other DeepMind spinouts, Inherent is now beginning to show the public what it has been quietly developing.

The benchmark task — independently replicating the findings of scientific papers — was chosen deliberately. Hughes noted that PhD students routinely begin their research careers with exactly this kind of exercise, making it a meaningful, if early-stage, proxy for autonomous scientific reasoning. Beyond raw accuracy, Inherent set a higher bar for Faraday: the agent was expected to demonstrate 'research taste,' meaning an intuitive sense of which experiments are worth running and how to structure them effectively. Teaching that kind of judgment, Hughes explained, is where reinforcement learning becomes central to Inherent's approach.

Rather than encoding detailed rules about how science is conducted, Inherent uses reward-based training that incentivizes good experimental outcomes. The company believes this generalizes better across scientific fields than approaches that rely heavily on curated demonstrations of scientific method. Faraday also uses OpenAI's GPT-5.5 Codex for coding tasks rather than a proprietary tool, a decision the team frames as analogous to how human scientists adopt existing software infrastructure rather than building every layer from scratch.

Inherent currently employs around a dozen people and plans to scale to 20 to 25 staff by year's end. Hughes has publicly commented on the UK's 'garden leave' practice — which can prevent departing employees from immediately joining or founding rival companies — calling it a personal obstacle he experienced firsthand. With Demis Hassabis taking on new responsibilities at Google leaving some DeepMind staff reportedly unsettled, Inherent's hiring momentum could make it an attractive destination for researchers considering a transition.

The bigger picture

Inherent's Faraday result is worth taking seriously, but context matters. Beating larger models at a narrowly defined benchmark — paper replication without answer leakage — is not the same as demonstrating general scientific discovery capability, which remains the company's stated long-term goal. That said, doing so with a 27-billion-parameter model rather than a frontier-scale system has real implications: it suggests that architectural choices, training methodology, and task-specific reinforcement learning can close the gap against raw compute in at least some research contexts. For Anthropic and OpenAI, whose products command significant enterprise pricing partly on the premise that bigger models do more, this is an early but pointed counterexample worth monitoring.

The reinforcement learning angle is where Inherent's longer-term bet becomes clearest. The lab is explicitly wagering that reward-based training generalizes across scientific domains better than approaches trained heavily on demonstrations of scientific process. This mirrors ongoing debates in the broader AI research community about the relative merits of imitation learning versus outcome-optimized training — a question that frontier labs including DeepMind itself have grappled with in systems like AlphaFold and AlphaProof. If Inherent's approach scales as intended, it could become a credible rival not just to commercial AI assistants but to more specialized scientific AI efforts from well-capitalized players.

On the talent and geopolitical side, Hughes's vocal criticism of UK garden leave restrictions adds an interesting layer. London's AI scene depends on researchers being able to move freely between institutions and startups, and that friction is a structural disadvantage relative to Silicon Valley. Inherent's King's Cross base, its DeepMind pedigree, and its growth plans position it to benefit if the UK eventually softens those rules — and to keep making noise about why it should. Readers should watch whether Faraday gets released for broader external testing and how Inherent's benchmark claims hold up under independent scrutiny.

LagPing's take

We're covering Inherent because it represents something genuinely interesting in a landscape that often rewards hype over substance: a small team making a specific, testable claim rather than announcing a vague roadmap. The $50 million seed round is notable, but what caught our attention was the benchmark methodology — pitting a 27-billion-parameter model against Claude Opus 4.8 and GPT-5.5 at an academically grounded task is a bold move, and the kind of thing that can be checked. We're also interested in the London angle here. King's Cross has become a real cluster for serious AI research, and debates around garden leave restrictions speak directly to whether that ecosystem can sustain momentum. For readers following AI science tools or the broader question of whether small efficient models can punch above their weight, Inherent is now a name worth tracking.

X / Twitter WhatsApp

View at TechCrunch AI

Find "OpenAI" on Amazon

As an Amazon Associate, LagPing earns from qualifying purchases. Product links are affiliate links.