The Collaboration That Built Modern AI

From our latest Radical Talks episode with two icons of AI: Geoffrey Hinton and Jeff Dean.

Some breakthroughs arrive through theory. Others through engineering. But the story of modern AI is ultimately the story of a partnership, one built over decades, where powerful ideas were paired with the infrastructure needed to realize its potential.

Earlier this month at NeurIPS 2025, in a packed room of founders, researchers, and engineers, two of the most influential figures in AI took the stage together: Geoffrey Hinton and Jeff Dean. Their conversation, hosted by Radical Ventures Co-Founder and Managing Partner Jordan Jacobs, reflected on a friendship and collaboration that helped shape the trajectory of the field itself.

The discussion traced how modern AI emerged through the sustained collaboration between theory and systems. Hinton brought ideas that were bold and far ahead of their time. Dean built the distributed infrastructure and compute that allowed those ideas to be tested, validated, and scaled. Together, they helped form the foundations on which today’s AI systems stand.

Their conversation surfaced a deeper truth beneath many of AI’s major inflection points: progress accelerates when strong ideas find the systems capable of carrying them forward. Below are three signals from their discussion that illuminate how AI reached its current turning point and what will define the breakthroughs ahead.

When Ideas Met Scale

Across the discussion, a single theme kept resurfacing: modern AI did not advance because entirely new algorithms appeared, but because existing ideas finally met the compute required to show what they could do.

Hinton describes how early demonstrations that larger neural networks worked better were easy to dismiss when the hardware made scaling impractical. Dean, experimenting with distributed systems in the early ’90s and later at Google, discovered firsthand that “bigger model, more data, more compute” was not just intuition, it was the beginning of a scaling law.

Their stories point to a simple, clarifying truth. AI did not leap forward because the ideas changed. It leapt forward because the systems finally caught up.

That convergence, algorithms meeting infrastructure powerful enough to run them, remains the engine of progress in the current era.

Choosing the Breakthroughs

Even in a field driven by exponential compute and automated learning, the most important turning points were shaped by human judgment.

AlexNet, Inception, sequence-to-sequence models, Transformers, Google’s Gemini effort, each milestone exists because someone made a high-stakes decision at the right moment. Decisions to unify fragmented teams, to ship research before consensus formed, to champion ideas early reviewers dismissed, or to redirect budgets toward hardware when the math demanded it.

Hinton’s framing is straightforward: breakthroughs don’t just happen, they’re chosen. Progress accelerates when leaders are willing to take intelligent risks, stand behind fragile ideas, and commit to long-term work in the face of uncertainty.

In many cases, it was trust, between collaborators, between research and infrastructure teams, that allowed ideas and systems to move forward together.

The Next Compute Frontier

If the last decade was defined by scaling existing architectures, the next will be defined by rethinking compute itself.

Dean recounts how the TPU program began with a simple question: what happens if hundreds of millions of people start speaking to their phones for minutes each day? The back-of-the-envelope math revealed a looming bottleneck that incremental improvements could not solve. The answer required custom silicon, purpose-built for the scale ahead, a decision that would become one of Google’s most significant strategic advantages.

Both Hinton and Dean emphasize that the constraints ahead are no longer conceptual; they’re physical. Memory bandwidth, thermals, energy efficiency, noise tolerance, and architectures capable of attending over billions of tokens will define what becomes possible next. Once again, the pattern holds. The next breakthroughs will arrive when new ideas encounter systems designed explicitly to carry them.

Together, Hinton and Dean make one thing clear: AI’s trajectory will be shaped not only by models and compute, but by safety, alignment, and the societal choices surrounding deployment. We are living in a moment defined by duality, systems of unprecedented capability, and an equally unprecedented responsibility to build and deploy them wisely.

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.