Inherent AI Outperforms OpenAI, Anthropic Rivals
A London-based artificial intelligence laboratory, Inherent, founded by Google DeepMind alumni, recently announced that its AI agent, Faraday, has outperformed larger models from industry leaders Anthropic and OpenAI. This achievement is notable as Faraday accomplished the feat using a significantly smaller model, signaling a potential shift in the competitive landscape of advanced AI development.
After emerging from stealth mode with a substantial $50 million seed round just weeks prior, the British startup has begun to unveil its innovations. Its newly released AI agent demonstrated its capabilities by independently reproducing the findings of published scientific papers, a task performed without prior knowledge of the answers. This specific challenge, while seemingly a niche application, serves as a fundamental training exercise for human scientists, as noted by cofounder and chief scientist Edward Hughes, who commented, "Many PhD students actually start by doing this."
Inherent's Faraday: A New Benchmark in AI Research Replication
By Decode Today News
The core of Faraday's breakthrough lies not merely in its ability to replicate results, but in the efficiency and methodology behind its success. While Inherent relished the outcome of surpassing frontier agents, Hughes emphasized to TechCrunch that the process of building the system was the most compelling aspect. This focus on the developmental approach rather than just the end-result highlights a differentiated strategy in the crowded AI sector.
Understanding Faraday's Breakthrough
In a direct comparison, Faraday was measured against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. These competing models are recognized as large, frontier-scale systems, implying significant computational demands and extensive training costs. In contrast, Faraday operates on a comparatively tiny model known as Qwen 3.6, which possesses just 27 billion parameters. Parameters serve as a proxy for a model's size and, crucially, its associated training expenses and resource requirements. This dramatic reduction in scale, coupled with superior performance, suggests considerable advancements in AI infrastructure and cost efficiency for complex tasks.
The ability of a smaller model to outperform larger counterparts holds significant implications for the broader industry. It could lead to more accessible, less resource-intensive AI solutions, potentially democratizing access to advanced AI capabilities for a wider range of organizations. For businesses, this could translate into substantial savings on computational resources and energy, making sophisticated enterprise integration of AI more viable. It also challenges the prevailing notion that sheer scale is the sole determinant of AI prowess.
The Philosophy of 'Research Taste'
Inherent's ambition extends beyond mere accuracy in replication. The company set a higher bar for Faraday, aiming for the agent to demonstrate "research taste"—an intuitive understanding of which experiments are worthwhile and how to design them effectively. This intangible quality, crucial for human scientists in discovering new knowledge, is notoriously difficult to teach to an AI system.
To instill this elusive "taste," Inherent leverages reinforcement learning, a training method that rewards an AI system for positive outcomes rather than relying on predefined rules. Unlike conventional approaches that might heavily focus on the study of how science itself is conducted, Inherent's reward-based methodology is a strategic bet. The company believes this approach will foster greater generalization, enabling its agents to contribute effectively across diverse scientific fields, ultimately aligning with its long-term goal of building an AI scientist agent. Hughes reiterated this guiding principle, stating, "We're always guided by that north star of building an AI scientist agent and imbuing our agents with taste."
Strategic Development and Collaborative Spirit
The company's strategic choices are also defined by what it opts not to develop in-house. Rather than creating its own coding tools, Inherent chose to integrate OpenAI's GPT-5.5 Codex for Faraday's coding needs. This decision mirrors how human scientists often rely on existing software rather than building every tool from scratch, promoting efficiency and focusing internal resources on core innovations.
Furthermore, Inherent actively avoids developing agents that merely echo user expectations. Instead, Hughes envisions an AI teammate that exhibits curiosity and initiative, much like his preferred human collaborators. He described this ideal as a teammate who returns saying, "I got curious about this, and I went off and I did these experiments. What do you think of these results?" This emphasis on independent thought and collaborative exploration is central to Inherent's vision for its AI agents.
This collaborative ethos extends to the company's operational structure. Its dozen employees work in-person from an office in King's Cross, London. This once-underdeveloped London neighborhood has transformed into one of the world's leading AI hubs, largely due to the presence of Google DeepMind. Hughes expressed confidence in the location, stating, "We believe that London is the place to be," underscoring the strategic importance of London's dense concentration of AI talent for the startup's growth and innovation.
Navigating Talent and Growth in the UK AI Landscape
While bullish on London's talent pool, Hughes has also voiced concerns regarding "garden leave," a common practice in the U.K. that prohibits departing employees from joining or starting rival companies for several months post-resignation. This restriction, generally absent for American researchers, grants U.S. startups a distinct advantage in rapidly recruiting talent. Hughes, having been personally affected by this policy, told TechCrunch, "This is a personal view rather than a company view, but I was affected by the garden leave problem."
Despite these challenges, Hughes successfully navigated the constraints, founding Inherent alongside three other co-founders: Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins, all fellow DeepMind alumni. The startup shows no signs of slowing its expansion, with plans to grow its headcount to between 20 and 25 by the end of the year. This ambitious hiring drive, coupled with Inherent's aspirations in world models and the reported unsettling effect of Demis Hassabis's new role on some DeepMind staff, positions Inherent as a potentially appealing destination for experienced DeepMind employees considering new opportunities.
Key Takeaways from Inherent's Advancement
- Inherent's Faraday, an AI agent, outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 in scientific paper replication.
- Faraday achieved this using a significantly smaller model, Qwen 3.6, with only 27 billion parameters, suggesting improved cost efficiency and AI infrastructure utilization.
- The startup secured a $50 million seed round and was founded by DeepMind alumni, including co-founders Louis Kirsch, Kaloyan Aleksiev, Tantum Collins, and Edward Hughes.
- Inherent emphasizes developing "research taste" in its AI agents, using reinforcement learning to foster intuitive scientific discovery.
- The company's strategic approach includes leveraging existing external tools like OpenAI's GPT-5.5 Codex and fostering a collaborative, in-person work environment in London's King's Cross AI hub.
- Inherent plans to expand its team to 20-25 employees by year-end, potentially attracting talent from larger AI firms amidst industry shifts.
