From our latest Radical Talks episode with Tantum Collins (“Teddy”), Co-Founder and CEO of Inherent
Recursive self-improvement (RSI) is usually framed as a model problem: an AI system using AI to build better AI. Tantum Collins, (“Teddy”) Co-Founder and CEO of Inherent, thinks that framing misses something important: the organization.
Inherent is an AI lab building general-purpose inventive AI to accelerate scientific discovery. Prior to starting the company and raising a $50M seed round, Teddy led the meta-research team at DeepMind and later worked on AI policy inside the Biden White House.
His perspective is clear: the real bottleneck to safe, effective AI-driven science isn’t the model. It’s the organization wrapped around it. At Inherent, Teddy and his team are betting that redesigning the company for innovation is equally as important as using AI to build better AI.
Collins doesn’t disagree with the conventional definition of RSI, but in his view, scientific discovery isn’t produced by a model in isolation. That’s why he and his co-founders are building Inherent to structure itself as a live experiment in human-AI collaboration.
Faraday, Inherent’s AI agent, has been designed to operate both in the research and in the organization itself. It doesn’t just help generate hypotheses or design experiments, it sources papers, allocates compute, and acts on ideas raised in a meeting, without waiting to be prompted.
Living Inside the Experiment
The company’s internal culture is treated as part of the research agenda. Collins calls it “living within the experiment”: every employee, including the founders, has to be willing to show up on a Monday and find that reporting lines, decision-making systems, or internal forecasting tools have changed completely. They do this so that they can continue to uncover what actually works.
This deliberately reflects a pattern seen in past general-purpose technologies which delivered only marginal gains until someone was willing to tear down the factory and rebuild it around the new tool instead of bolting it onto the old one. Collins doesn’t think any one individual or organization, including Inherent, yet knows what the “factory” should look like for AI-driven science. He’s confident only that it will look strange relative to any company operating today.
Betting Wide, Not Deep
The same experimental approach shapes Inherent’s scientific strategy. Where the major labs have made large, specific bets on particular scientific domains, Inherent is deliberately staying horizontal for now, applying Faraday first to its own technical development before expanding into external domains through design partnerships. Their logic is to expose the system to a wide range of scientific problems, not just one, and that is what will teach it how to discover things it hasn’t seen before.
Collins is candid about the fact that some of what emerges from this process may not look intelligible to humans at first: new model architectures, new organizational structures, results shaped more like the alien-looking, ultra-efficient joints produced by generative design tools than anything a person would have sketched by hand. Part of Inherent’s bet is that better interfaces between humans and AI systems will eventually make that alien-looking work legible, rather than simply accepting that it can’t be understood.
Why Structure First
That same innovative approach extends to how Inherent has structured itself as a company. Inherent is structured as a Public Benefit Corporation, which gives the company legal room to prioritize, for example, fairness in internal decision-making and the day-to-day experience of the people doing the work, even when that isn’t the most shareholder-friendly path. Collins points to the Industrial Revolution and the gig economy as cautionary tales: efficiency gains that came at the cost of making work worse for the people inside it. Avoiding that outcome, he argues, has to be a design constraint from day one.
That same instinct shows up in how Inherent thinks about compute and infrastructure. Instead of forcing AI systems to operate under permissions built for human users, the company is endeavoring to build infrastructure where AI systems are “first-class citizens” with their own, adjustable set of responsibilities, reassessed on a regular basis as trust is earned.
For Collins, that’s the real test of RSI, not just building a model that gets smarter, but building an organization committed to keep rebuilding itself around what the model teaches it.
This post is based on insights from Radical Talks, a podcast from Radical Ventures exploring innovation at the frontier of AI. For more conversations with leaders in AI, subscribe wherever you get your podcasts.