DeepMind Alumni's Inherent AI Outperforms Giants with Smaller Model
London-based AI lab Inherent, founded by DeepMind alumni, has unveiled Faraday, an AI agent that outperformed larger models from Anthropic and OpenAI at replicating scientific research, leveraging a significantly smaller model. This breakthrough emphasizes the potential for efficient, specialized AI agents to advance scientific discovery and foster a new collaborative paradigm between humans and AI.
Key Takeaways
- Inherent's Faraday agent replicated scientific research more effectively than larger frontier models.
- Faraday achieved this using a much smaller model (27 billion parameters), highlighting efficiency.
- The AI was trained to develop "research taste" through reinforcement learning, aiming for scientific discovery.
- Inherent promotes a collaborative "AI teammate" model, integrating with existing tools like OpenAI's Codex.
- This demonstrates that advanced AI capabilities aren't solely dependent on massive model sizes.
Inherent, an AI laboratory established in London by former Google DeepMind team members, has announced a significant breakthrough: its AI agent, Faraday, has surpassed the performance of much larger models from industry leaders like Anthropic and OpenAI. This achievement is particularly notable because Faraday accomplished the feat using a fraction of the computational resources, running on a comparatively compact model.
Emerging from stealth mode recently with a substantial $50 million seed funding round, Inherent has begun to reveal its innovative work. According to TechCrunch AI, the British startup's Faraday agent independently replicated the findings of published scientific papers, a task it completed without prior knowledge of the outcomes. While this might seem like a niche capability, co-founder and chief scientist Edward Hughes emphasizes its foundational importance, likening it to the initial training exercises for human PhD students.
Redefining AI Research Capabilities
Hughes clarified that the primary objective wasn't merely to outperform other frontier agents, but rather the unique methodology employed to achieve this. Faraday operates on Qwen 3.6, a model with just 27 billion parameters. In contrast, it competed against systems such as Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, which are significantly larger in scale. This efficiency in a smaller model presents a compelling case for investors and developers alike, challenging the notion that only massive models can deliver cutting-edge performance.
Beyond simple accuracy in replication, Inherent set a higher bar: for Faraday to demonstrate "research taste." This intangible quality involves discerning which experiments are most valuable to conduct and how to design them effectively. To instill this discernment, Inherent leverages reinforcement learning, a training paradigm that rewards the AI for successful outcomes rather than prescribing explicit rules. This approach, they believe, will generalize more effectively towards their long-term goal of creating AI agents capable of contributing across diverse scientific domains.
A Collaborative Approach to AI Development
Inherent's strategic focus also dictates what it chooses not to build. Instead of developing its own coding tools, Faraday utilizes existing solutions like OpenAI's GPT-5.5 Codex, mirroring how human scientists integrate readily available software into their workflows. The company aims to cultivate AI agents that act as proactive, curious teammates. Hughes envisions an AI that returns with: "I got curious about this, and I went off and I did these experiments. What do you think of these results?" This ethos promotes a collaborative relationship between humans and AI.
The company itself operates with a similar collaborative spirit, with its dozen employees working in person from their London office in King’s Cross, a hub for AI innovation. Inherent plans to expand its team to 20-25 by year-end, potentially attracting talent from larger AI labs looking for new challenges.
Why This Matters for AI Product Managers
AI Product Managers should closely observe Inherent's advancements, particularly the demonstration of high performance from a smaller model. This challenges the prevalent "bigger is better" paradigm, suggesting that efficient, specialized agents can deliver significant value. This insight can influence product roadmaps towards optimizing for specific, high-impact tasks with more constrained models, potentially reducing computational costs and improving latency.
The concept of "research taste" and using reinforcement learning to cultivate it is a critical consideration for designing future AI agents. Product leaders should think about how to embed similar intangible qualities—like curiosity, strategic thinking, or user empathy—into their AI products. This moves beyond mere task automation towards creating truly intelligent, proactive collaborators that enhance user experience and outcomes.
For GTM strategies, Inherent's focus on scientific discovery highlights a powerful vertical for AI agents. AI PMs can explore how agents capable of complex problem-solving and independent experimentation can unlock new markets in R&D, drug discovery, materials science, or even creative industries. The "AI teammate" model also offers a compelling UX framework, shifting from tool-centric to partnership-centric product design.
Related Articles
Reflexio: AI Agents That Learn and Improve Over Time
Reflexio, featured on Product Hunt, introduces behavioral learning to enable AI agents to continuously improve their performance and adaptability over time through experience and feedback.
Meta's Muse: A New Personal AI Agent for Getting Things Done
Meta has unveiled "Muse," a personal AI agent designed to streamline tasks and enhance user productivity, marking its entry into the growing field of intelligent digital assistants.
Frigade Assist API: AI Agents That Visually Guide Users in UI
The Frigade Assist API, featured on Product Hunt, allows AI agents to provide visual, step-by-step guidance within application interfaces, showing users exactly where to click. This enhances user experience by reducing friction and simplifying complex workflows.