There is no machine on Earth that can ask a question it has never been taught to ask. Every AI system, no matter how powerful, waits for a human to type the prompt, to set the boundary, to define the shape of the answer. The question itself is the one thing that never comes from the silicon. It comes from the wet, messy, unpredictable biology of a human brain.
That absence is the starting point.
For four hundred years, the scientific method has followed a quiet rhythm. A researcher notices something strange. They wonder why. They form a hypothesis. They test it. They publish the result. The next researcher picks up the thread. The pattern repeats. The entire edifice of modern science rests on that first, fragile step: the moment someone decides a question is worth asking.
But what happens when the questions become too many? What happens when the data grows faster than any human can read? What happens when the space of possible hypotheses is larger than the number of stars in the sky?
The answer is that most questions never get asked at all.
The Team That Saw the Gap
In 2022, Tantum Collins and Edward Hughes, both then at DeepMind, collaborated on research into cooperative AI agents, focusing on games that required negotiation and trust. This work, documented in DeepMind’s publications, laid the groundwork for their later focus on hypothesis generation.
It was strange work for an AI lab. Most of the field was obsessed with making models bigger, faster, more accurate at answering questions. But Collins and Hughes kept circling back to a different idea. What if the hardest part of intelligence was not the answering? What if it was the asking?
Collins had a background that most AI researchers lack. He had worked on AI policy at the Biden White House. He had seen how governments think about technology, how they fund basic research, how they decide which scientific problems matter. He knew that the bottleneck in science was not computation. It was curiosity. It was the human ability to look at a pile of data and see something that no one else had seen.
Hughes brought the technical depth. He had spent years at DeepMind building systems that could learn from sparse rewards, that could explore environments without clear goals, that could improve themselves without human supervision [1]. His work on cooperative AI had shown that agents could learn to ask each other for help, to share information, to build on each other’s discoveries.
Louis Kirsch, also from DeepMind, contributed expertise in meta-learning, which enables models to accelerate their own learning. Kaloyan Aleksiev, formerly of Reka AI and Microsoft, added experience in deploying reliable large-scale AI systems.
The four of them looked at the landscape of AI and saw something missing. Every major lab was building tools to answer questions better. OpenAI had ChatGPT. Google had Gemini. Anthropic had Claude. All of them were designed to take a prompt and produce a response. All of them were passive. All of them waited for the human to make the first move.
What did not exist was a system that could say: “Stop. Look at this. This is interesting.”
The Faraday Platform
The team named their platform after Michael Faraday. He was not a mathematician. He had no formal training in science. He was a bookbinder’s apprentice who taught himself physics by reading the books he was supposed to bind. In 1821, he built a simple device that converted electrical energy into mechanical motion. It was the first electric motor. He did not know what it would be used for. He just thought the question was worth asking.
Faraday’s approach to science was open-ended. He did not start with a clear hypothesis. He started with a vague sense that something interesting might happen if he put a wire next to a magnet and ran a current through it. He was not trying to solve a specific problem. He was trying to see what the universe would show him.
Inherent’s platform is built on the same principle. It is called Faraday, and it does not answer questions. It generates them.
The system works by pairing human researchers with AI agents that are designed to improve themselves iteratively. The agents explore hypothesis spaces that are too large for any human to map. They run simulations. They find patterns. They flag anomalies. They come back to the human and say: “This is weird. Look at this.”
The human then provides judgment. They decide which threads to follow. They apply taste, intuition, and ethical reasoning. They tell the agent what matters and what does not. The agent goes back to work, now searching in a narrower space, guided by the human’s sense of what is worth knowing.
This is not automation. It is collaboration. The machine does what machines do best: explore vast spaces of possibility with inhuman speed. The human does what humans do best: recognize what is beautiful, important, or dangerous.
The company calls this “AI-native science.” It is a phrase that signals a break from the past. For four centuries, science has been done by humans asking questions that other humans could understand. The questions were limited by what a single brain could conceive. AI-native science removes that limit. The questions can now come from the machine. The human’s job is to choose which ones to pursue.
The $50 Million Bet
In February 2025, Inherent launched with a $50 million seed round led by Index Ventures and Radical Ventures, with participation from Nvidia’s venture arm, Ex/Ante, Metaplanet, Macroscopic Ventures, and Mythos Ventures. This ranks among the largest AI seed rounds in Europe.
Danny Rimer, a partner at Index Ventures, explained the logic [5]. “Most AI is built to answer questions,” he said. “What it can’t do yet is figure out which questions are worth asking, the open-ended curiosity that produced penicillin, the microwave, the GPU. That’s the gap Inherent is building into.”
The investors are betting that the most valuable application of frontier AI is not automating existing workflows. It is enabling discoveries that human researchers could not reach alone. Index Ventures wrote in a blog post that “AI-native science will be messier, less legible, but capable of exceptional outcomes” [5].
The company is structured as a public benefit corporation. This is unusual for a venture-backed AI lab. Most startups are legally required to maximize shareholder value above all else. A public benefit corporation must also consider its impact on society. The founders believe this structure is a competitive advantage, not a constraint. It signals to researchers, governments, and the public that Inherent is building for the long term.
Matt Clifford, the former UK AI tsar and co-founder of Entrepreneurs First, has joined as an adviser. His presence gives the company credibility with policymakers who are wary of AI’s risks. Clifford has spent years thinking about how to govern powerful technologies. He knows that the biggest danger of AI is not that it will become too smart. It is that it will be used to answer the wrong questions.
The Context of European AI
Inherent’s $50 million round is part of a larger shift. European AI startups are increasingly raising at scales that were once reserved for Silicon Valley. Peec AI hit $10 million in annual recurring revenue within six months. Lovable generated $100 million in a single month. Mistral, the French AI lab, reached $300 million in annual recurring revenue. The gap between European and American AI funding is narrowing.
This is happening in categories where the technology is genuinely new. European investors are not just funding copycats of American products. They are betting on original ideas. Inherent’s focus on open-ended scientific discovery is one of those ideas. It is not a better chatbot. It is a new way of doing science.
The timing matters. The scientific establishment is facing a crisis of scale. The number of published papers has exploded. The amount of data has grown beyond what any human can process. The traditional method of reading, hypothesizing, and testing is breaking down. There are too many possibilities and too few researchers to explore them.
AI-native science offers a way out. Instead of asking humans to read every paper and form every hypothesis, the machine can do the first pass. It can find the connections that no human has seen. It can propose experiments that no human would think to run. The human’s role shifts from generating ideas to selecting them.
The Glasswing Precedent
Anthropic’s Glasswing project showed that advanced AI can identify security vulnerabilities faster and more comprehensively than human reviewers, scanning codebases for flaws that were previously missed.
Inherent’s bet is that the same dynamic applies to scientific discovery. The hypothesis space of a complex biological system is like a codebase with billions of lines. Human researchers can only check a tiny fraction of the possibilities. AI agents can explore the whole space. They can find the patterns that humans would never see.
But there is a key difference between Glasswing and Faraday. Glasswing was looking for known types of vulnerabilities. It had a clear target. Faraday is looking for unknown unknowns. It does not know what it will find. It is searching for questions, not answers.
This makes the problem harder. A system that searches for known vulnerabilities can be trained on examples. A system that searches for unknown discoveries must be trained on curiosity. It must learn to recognize when something is interesting, even if it does not know why.
The Human Element
The founders of Inherent are careful to emphasize that humans remain central to the process. The AI agents generate hypotheses, but humans provide the judgment. This is not a replacement of scientists. It is an augmentation.
Edward Hughes has spent years studying how AI agents cooperate with each other [1]. He knows that the hardest problems in multi-agent systems are not technical. They are social. Agents must learn to trust each other, to share credit, to communicate honestly. The same dynamics apply to human-AI collaboration. The machine must learn what the human values. The human must learn to trust the machine’s suggestions.
Tantum Collins brings the policy perspective. He knows that the biggest barriers to scientific progress are not technical. They are institutional. Universities are slow to adopt new methods. Funding agencies are risk-averse. Peer review rewards incremental results over bold ideas. Inherent’s public benefit corporation structure is designed to navigate these barriers. It signals that the company is aligned with the long-term interests of science, not just the quarterly returns of investors.
Louis Kirsch’s work on meta-learning is directly relevant to Faraday’s design. The platform must learn how to learn. It must improve its own ability to generate interesting hypotheses. This is a recursive problem. The system that improves itself must also improve its ability to improve itself. Kirsch’s research at DeepMind focused on exactly this kind of self-improvement [1].
Kaloyan Aleksiev’s experience at Reka AI and Microsoft gives the team practical expertise in building systems that work at scale [2]. Faraday must handle massive amounts of data. It must run simulations that take days or weeks. It must produce results that are reliable enough for scientists to trust. Aleksiev knows how to build infrastructure that meets these demands.
The Philosophical Shift
Inherent’s approach represents a philosophical shift in how we think about AI. Most AI research focuses on narrow tasks. A system is trained to play chess, or translate language, or diagnose diseases. The goal is to match or exceed human performance on a specific metric.
Faraday is different. It is not trained for a specific task. It is designed to explore. It is built for open-ended discovery. This is closer to the original vision of artificial intelligence, the one that Alan Turing imagined when he asked whether machines could think. Turing did not propose that machines should answer questions better than humans. He proposed that they should surprise us.
The surprise is the key. A system that can only produce expected results is not intelligent. It is just a calculator. A system that can produce unexpected results, that can find patterns that no human has seen, that can ask questions that no human has thought to ask—that is something new.
Inherent’s founders believe that this kind of AI will transform science. They point to the history of discovery. Penicillin was discovered by accident. The microwave was discovered by accident. The GPU was invented by people who were trying to solve a different problem. The greatest scientific breakthroughs often come from unexpected places. They come from questions that no one thought to ask.
Faraday is designed to produce those accidents on purpose.
The Governance Question
The public benefit corporation structure is not just a legal formality. It is a signal about how the company intends to operate. Inherent’s founders believe that the biggest risks of AI come from misaligned incentives. A company that is legally required to maximize shareholder value will make different decisions than a company that must also consider its impact on society.
The governance question is particularly important for AI-native science. The discoveries that Faraday enables could be transformative. They could lead to new drugs, new materials, new sources of energy. They could also lead to new risks. A discovery that is published too quickly could be dangerous. A discovery that is kept secret could be unethical.
Inherent’s public benefit corporation structure gives it the flexibility to make decisions that are not purely profit-driven. It can choose to publish results that are not commercially valuable. It can choose to delay publication if the risks are too high. It can choose to collaborate with universities and government labs, even if that means sharing credit.
Matt Clifford’s involvement as an adviser reinforces this commitment. Clifford was the UK government’s AI tsar. He knows how to think about the societal implications of powerful technologies. He has written extensively about the need for governance structures that can keep pace with technological change.
The Competition
Inherent is not the only company working on AI for scientific discovery. Google’s DeepMind has made breakthroughs in protein folding with AlphaFold [1]. Microsoft has invested heavily in AI for chemistry and materials science. Anthropic is building systems that can reason about complex problems [4]. OpenAI has tools that can help researchers analyze data.
But most of these efforts are focused on answering specific questions. AlphaFold answers the question: “What shape will this protein fold into?” It does not ask: “Which protein should we study?” The question is given by the human. The machine just computes the answer.
Inherent’s bet is that the next frontier is not better answers. It is better questions. The company is building a platform that can generate hypotheses, not just test them. This is a fundamentally different approach.
The difference matters because the number of possible scientific questions is infinite. There will never be enough human researchers to ask them all. The only way to explore the full space of possibilities is to use machines that can generate their own questions. Inherent is building those machines.
The Challenge of Evaluation
How do you measure the success of a system that is designed to produce surprises? This is a difficult question. Standard metrics do not apply. You cannot measure Faraday’s performance by counting how many questions it answers correctly. It does not answer questions. It asks them.
The company will need to develop new ways of evaluating its platform. One approach is to measure the novelty of the hypotheses it generates. Another is to track how many of its suggestions lead to published papers or patents. A third is to ask domain experts whether the questions are interesting.
The founders acknowledge that Faraday’s success will take years to evaluate, but the $50 million seed round provides the necessary runway for long-term development.
The Broader Implications
If Inherent succeeds, the implications go beyond science. The ability to ask good questions is the foundation of all human progress. It is what drives art, philosophy, and politics. It is what makes us curious, creative, and alive.
A machine that can ask good questions is not just a scientific tool. It is a new kind of intelligence. It is a partner in the act of discovery. It challenges our assumptions about what it means to be smart.
The founders are careful not to overpromise. They describe Faraday as a platform for human-AI collaboration, not a replacement for human scientists. They emphasize that the human remains in control. The machine suggests. The human decides.
But the line between suggestion and decision is blurry. As the machine gets better at generating interesting questions, the human’s role may shift from active participant to passive curator. The question becomes: who is really doing the science?
The Open Question
The $50 million is a bet on a future that does not yet exist [5]. It is a bet that the most important scientific discoveries of the next century will come from machines that can ask their own questions. It is a bet that the human role in science will shift from generating ideas to selecting them. It is a bet that curiosity can be coded.
Whether that bet pays off is an open question. The team has the right background. The investors have the right incentives. The governance structure is designed for the long term. But the challenge is enormous. Building a machine that can ask good questions is harder than building a machine that can answer them. It requires a different kind of intelligence, one that is not yet fully understood.
The founders are undeterred. They have seen the gap. They have built a platform to fill it. They have raised the money to keep building. Now they have to prove that a machine can do what no machine has ever done: look at the universe and ask, “What if?”
That is the question that matters.
Sources
1. DeepMind
2. Reka AI
3. Google
4. Anthropic
6. Nvidia
7. Metaplanet