A few weeks ago, I finished reading Jimmy Soni’s The Infinity Machine, which charts the story of Demis Hassabis and DeepMind. In a landscape dominated by AI hype and speculation, Soni’s book offers valuable historical grounding. Understanding where we are going requires looking closely at the tightly intertwined DNA of the small cohort that built this path.
A historical irony lies at the heart of this story. Demis Hassabis co-founded DeepMind with a pure, scientific mission: solve intelligence, and then use that to solve everything else. In direct response to DeepMind's early breakthroughs and its subsequent acquisition by Google, Elon Musk, Sam Altman, and a few others founded OpenAI. They built it specifically as a non-profit counterweight to prevent Google from monopolizing AGI.
Today, the tables are completely turned. OpenAI is now a for-profit commercial engine, functioning as the corporate force it was built to counter. Meanwhile, the talent pool has splintered into the rival houses that define our current landscape: Mustafa Suleyman, Demis’s original partner at DeepMind, is running AI at Microsoft, and Dario Amodei broke away from OpenAI’s safety compromises to found Anthropic.
This technological moment is not a sudden magic trick; it is the result of a long, friction-filled evolution between competing philosophies of building.
Of course, Soni’s book is a biography of Demis Hassabis, so it naturally favors Hassabis’s safety-conscious, science-first philosophy. But in the real world, the tension between DeepMind’s research focus and OpenAI’s speed-to-market wasn’t just a clean intellectual debate. It was a high-stakes corporate escalation. And that escalation is exactly what ended up changing the physics of entrepreneurship for the rest of us.
The Exploration vs. Execution Tension
During my six years in the AI startup space, spanning the quiet years before ChatGPT was launched, the explosion of its release, and the chaos that followed, I had a front-row seat to the constant tug-of-war between research and product. In deep tech, the temptation to explore the horizon always competes with the pragmatism of shipping. Even DeepMind struggled to balance the two when large language models first emerged.
Hassabis’s core hypothesis was rooted in grounded intelligence. He believed that true understanding could not be achieved by merely processing static symbols. To build AGI, an AI had to be an active agent operating within an environment, learning cause, effect, and feedback from its own actions. DeepMind focused heavily on reinforcement learning, training systems from scratch to master complex rules and environments, from Atari games to the near-infinite states of Go. They believed that an intelligence needed to construct its own mental model of the world's mechanics, whether in physics simulations or on chessboards, before it could truly understand concepts.
OpenAI took the opposite bet, leaning heavily into the scaling hypothesis. They bypassed the need for active, environmental feedback loops, choosing instead to scale a theory developed natively inside Google. In a masterstroke of historical irony, OpenAI took the architecture from Google’s own seminal 2017 paper, "Attention Is All You Need," and put it into practice at a scale that Google itself had not yet dared to try. By feeding the entire written history of human text into a massive transformer model, OpenAI bet that if you scaled compute and parameters far enough, reasoning and logic would emerge naturally from next-token prediction.
The book details how Hassabis was humbled by how effective large language models turned out to be. Even if language models only manipulate symbols that lose meaning in translation, mastering the relationships of words proved more than enough to create an incredibly powerful tool. OpenAI’s singular focus on execution allowed them to capture the public market while DeepMind remained in research mode.
This grounded approach is why systems like AlphaFold are more than just technical milestones. They prove that different categories of problems require different AI architectures. While language models excel at processing text, solving the protein-folding problem required structural reinforcement learning models rather than simple next-token prediction. Some of the deepest scientific mysteries cannot be solved by scaling up transformers.
This suggests a warning to those who write off Google as having lost the AI race. As transformer scaling hits the limits of public human text, the next frontier of AGI will likely depend on hybrid systems. Success will belong to the labs that can combine LLMs with planning, reinforcement learning, and structural reasoning. By maintaining their multi-paradigm approach rather than doubling down solely on text prediction, Google may hold the critical keys to the future of AI.
The release of AlphaFold also highlights a critical human element in this landscape. When DeepMind cracked the protein-folding challenge, they chose not to lock the output behind a paywall to maximize margins. Google and DeepMind gave the entire database to global scientists for free, accelerating research in health, chemistry, and biology.
It is easy to look at the massive capital requirements of AI and demonize big tech as a monolith of corporate greed. But the landscape is more complex. The leaders driving these companies are constantly weighing the tension between building commercial monopolies and using AI to solve real-world problems. While OpenAI represents a shift toward aggressive commercialization, Demis Hassabis's insistence on giving away AlphaFold suggests that scientific idealism is not entirely dead in the corporate labs. The motivations of the individuals behind the code still matter, and they suggest that the race for AGI can still be directed toward public good.
While AlphaFold validated DeepMind's scientific depth, OpenAI's rapid execution proved that markets move at the speed of shipping, not research. To keep pace, Google merged its Brain division with DeepMind. Microsoft backed OpenAI with billions. Anthropic split off to run its own safety-focused race, and Meta open-sourced its models to drive the cost of intelligence down.
This escalation between the major labs had a massive downstream effect. In their battle for dominance, they accelerated the release of increasingly capable, cheap, and compressed models. They accidentally democratized capabilities that once required a team of PhDs and millions of dollars in compute, placing those tools directly into the hands of individual builders.
And that is what inverted the very physics of starting a company.
The Dopamine Loop & The Premium on Judgment
This brings us to what I call the new physics of entrepreneurship.
Historically, starting a company was constrained by the friction of building. You needed capital, developers, and months of prototyping. AI has collapsed these constraints. Today, a single founder can write code, build databases, and spin up functional applications in minutes.
But this ease introduces a new trap, one I’m calling the Build-Dopamine Loop.
It is not that this building is mindless; rather, the feedback loop is simply too powerful. It feels better to build and iterate at lightspeed than it does to stop, think, and challenge whether you are building the right thing for the right customer. The loop pulls us back to execution, tempting us to skip the hard work of customer validation, challenging our own assumptions, and developing stronger hypotheses.
The Anatomy of the Loop:
- The Cue: The cost of execution falls to zero; a founder can build anything instantly.
- The Routine: The builder prompts an AI agent to generate code and assets.
- The Reward: A functional software prototype compiles and runs, triggering a hit of productivity.
- The Trap: The loop repeats, bypassing the human discovery and validation required to ask why or for whom.
This is a fundamental shift in the economics of starting a company. Historically, the physics of scarcity forced validation. Because building was incredibly costly, founders had to spend time and cycles figuring out what was worth building. Decades of work by practitioners, including Steve Blank, Jeff Bussgang, and other pioneers of the Lean Startup and customer-discovery movements, turned this into a rigorous science. The goal was to avoid wasting time and money building things for customers who didn't want them. Founders were drawn to these approaches because they had to build the cheapest version possible, the MVP, to show progress with limited resources.
Those rules changed almost overnight. A new generation of founders is emerging without these constraints. When you can build instantly with minimal friction, there is less pressure to make sure you know exactly what to build and who you are building it for.
As a student of entrepreneurship and a practitioner who has experienced both sides of this coin (wasting time and money building the wrong things, but also seeing the massive wins that come from truly understanding a customer’s struggling moment), I find myself asking how this game will play out.
The physics have changed in ways that are both encouraging and concerning. On one hand, founders can get much farther with fewer resources. On the other, they are more likely to get caught in the build-dopamine cycle and bypass the validation required to build the right thing for the right customer.
It is early, and the rules of who succeeds and how are still being written. But this shift forces a critical question for universities, educators, and mentors stepping back in to guide the next generation: Do we keep using the playbooks from Steve Blank and others, or do we change them? These are the questions we are exploring in our living laboratory at the Center for Applied Entrepreneurship & Innovation, Mays Business School, and the broader Texas economic landscape we operate in.
When execution becomes free, human judgment becomes the ultimate multiplier. The hardest part of entrepreneurship is no longer engineering; it is the manual, often painful work of finding struggling moments in human lives and identifying the exact problem worth solving.
Reclaiming the Path in the Lab
This is the exact set of physical laws we are testing in our "living laboratories" at the Center for Applied Entrepreneurship & Innovation at Mays Business School at Texas A&M University.
Our responsibility as entrepreneurs returning to the university to guide the next generation is to learn with them, co-creating the playbooks required to navigate these altered physics together. We cannot teach status-quo business plans. We must learn how a single founder can orchestrate automated systems to test ideas at a pace that used to take years, while cultivating the shared discipline to slow down, exercise judgment, and do the hard work of human discovery before writing a single line of code.
Our goal is to help them use these altered physics not to chase venture cap-table clout, but to build sustainable businesses that solve real problems and reduce human suffering, mirroring the scientific idealism that started DeepMind in the first place.