The artificial intelligence race has entered a new phase — one where raw computing power is no longer enough. After years of exponential progress driven by ever-larger models and ever-bigger datasets, the world's AI giants may be reaching the limits of what scale alone can achieve. The next breakthrough, argue a growing number of scientists, investors, and technologists, will depend not on how much data an AI system can consume — but on how well it can learn from experience.
For UK investors and founders building in the AI ecosystem, this shift marks a critical inflection point: a move from data-hungry language models to adaptive, embodied intelligence.
The Scaling Era: Running Out of Road
Much of the extraordinary value creation in the US technology sector — from Nvidia's trillion-dollar market cap to the sky-high valuations of OpenAI, Anthropic, and Google DeepMind — rests on a shared belief that artificial general intelligence (AGI) is within reach.
AGI, loosely defined as a machine's ability to learn and reason across domains like a human, remains more aspiration than reality. Yet the idea has fuelled an unprecedented build-out of AI infrastructure, with the largest firms pouring hundreds of billions into data centres, GPUs, and cloud capacity.
The results have been remarkable — but also revealing. Each new model, from GPT-3 to GPT-5, delivers smaller performance gains at exponentially higher cost. "Scaling is now producing diminishing returns," notes one industry researcher. Even those who once evangelised the "bigger is better" doctrine are beginning to question whether the deep learning paradigm can take us much further.
From Data to Experience
The turning point may already be here. In a recent paper titled Welcome to the Era of Experience, computer scientists Rich Sutton and David Silver — two of the most influential figures in AI — argue that the field has exhausted most of the high-quality human data available online.
"The pace of progress driven solely by supervised learning from human data is demonstrably slowing," they write. "AI is at the cusp of a new period in which experience will become the dominant medium of improvement."
Put simply: AI systems have already read, watched, and listened to nearly everything humanity has produced in digital form. Future progress depends not on more training data, but on interaction — on systems that learn from their environment the way humans and animals do.
This shift mirrors the trajectory of the field itself. Sutton's earlier essay, The Bitter Lesson, argued that the most powerful AI advances came not from clever algorithms but from scaling compute and data. Now, even he seems to concede that scaling has hit a ceiling.
Rethinking Intelligence: Lessons from Biology
Among the most vocal proponents of a new approach is Karl Friston, professor of neuroscience at University College London and chief scientist at Canadian AI company Verses. Friston's theory — known as active inference — proposes that true intelligence requires not just perception, but agency: the ability to act in the world, make predictions, and adapt to change.
"Generative AI models are astounding," Friston concedes, "but they have no agency under the hood." A model might predict traffic flows in a city, he says, but would fail when something unexpected — say, a Taylor Swift concert — disrupts the data. "For true AGI, you have to be active, embodied, and situated in a world you can act upon."
Gabriel René, Verses' chief executive, adds: "AI breaks when it hits the real world because the real world keeps changing. Intelligence is about adaptation — not about compressing historical knowledge and memory."
This line of thinking is gaining momentum. The emerging thesis: the next generation of AI won't come from ever-larger monolithic systems, but from billions of smaller, adaptive agents — lightweight models specialised for specific tasks that learn continuously through experience.
Investors Bet on the Next Paradigm
For investors, this marks a profound shift in what AI innovation — and therefore AI value creation — might look like. The last cycle rewarded scale: those with the largest compute budgets and access to the biggest data pools. The next may reward efficiency, adaptivity, and real-world application.
This is good news for UK and European AI start-ups, which have struggled to compete with the resource intensity of US tech giants. The emerging "experience-first" approach plays to different strengths: applied AI, robotics, simulation, and edge computing — all areas with deep UK research roots.
Already, early-stage investors are scouting opportunities in "embodied AI" — systems that learn by doing, from drones and autonomous vehicles to warehouse robotics and digital twins. London-based firms such as DeepMind, Wayve, and Secondmind are among those pursuing this hybrid model of learning through interaction rather than imitation.
As one London-based VC partner told The Innovative Times: "The next wave of AI won't be about who has the biggest model — it'll be about who can make it think on its feet."
The End of the Free Data Boom
Behind the shift lies an uncomfortable reality: the AI industry's dependence on free human data may be unsustainable. Most high-quality text, image, and video data has already been harvested — and new training runs increasingly scrape the digital barrel.
This has legal as well as practical implications. Global media organisations and creators are pushing back on unlicensed data use, while regulators explore compensation frameworks for copyrighted material. That means the next frontier — experiential data generated by AI interacting with the physical or virtual world — is both a technical and commercial necessity.
But gathering this kind of data is far harder than scraping websites. It requires infrastructure — robots, simulations, and feedback loops — that bridge the gap between digital learning and real-world adaptation.
Risks of the Experience Era
While the shift to adaptive AI opens new frontiers, it also raises new risks. Systems that learn from experience can be unpredictable, evolving in ways developers didn't intend.
There are also economic challenges. The infrastructure to enable experiential learning — from embodied robotics to digital twin environments — remains capital-intensive. Unlike cloud-based data training, it doesn't scale neatly.
Moreover, generative AI still offers immense commercial value today. Most business use cases — from coding assistants to marketing tools — don't require AGI. Investors betting too heavily on speculative, biologically inspired systems could find themselves years ahead of monetisation.
As one AI fund manager put it: "We're still in the 'dot-com' phase of AI. There's real innovation, but also a lot of noise. Experience-driven learning could be transformative — or just the next hype cycle."
A UK Opportunity — If Seized
For the UK, however, the pivot toward experiential AI could play to national strengths. British universities — particularly UCL, Cambridge, and Edinburgh — remain at the forefront of neuroscience, robotics, and computational cognition.
Government policy, too, is starting to recognise the strategic importance of AI infrastructure. Initiatives under the Department for Science, Innovation and Technology (DSIT) aim to position the UK as a global hub for "frontier AI safety" and next-generation research.
If UK start-ups can harness these resources to develop adaptive, energy-efficient AI systems, they could carve out niches distinct from the hyperscale US giants. "This is the moment for Europe and the UK to define what responsible intelligence looks like," says one venture investor focused on deep tech. "It's not just about catching up to OpenAI — it's about doing things differently."
Investor Takeaways: The Next Phase of AI
Emerging Themes
- From scale to adaptation: Diminishing returns from model size mean AI must evolve through real-world experience.
- Hybrid architectures: Expect convergence between symbolic reasoning, reinforcement learning, and traditional deep learning.
- Embodied intelligence: Robotics, autonomous systems, and edge AI offer the most natural routes to "experiential learning."
Risks to Watch
- Overhyped AGI timelines: True artificial general intelligence may remain decades away.
- Capital intensity: Real-world data collection and robotics training require major investment.
- Regulatory uncertainty: IP ownership of experiential data remains unresolved.
The Bottom Line
Artificial intelligence has reached a crossroads. The era of limitless scaling — more data, bigger models, faster chips — is giving way to something subtler and arguably more human: learning by experience.
For investors, the lesson is clear. The next phase of AI won't be won in data centres alone. It will be shaped in the messy, unpredictable world where algorithms meet reality — where intelligence stops being simulated and starts being lived.