Fig. · Two 2009 theses converge into one physical verifier — an original visualization.
See the relational graph →STRUCTURAL-THESIS · CONTEXT JAMMING · JUNE 2026
The Physical Convergence.
Two doctoral theses from 2009 quietly became the operating system for modern AI. Lila Sciences is the first company to install both — in a building, on a robot arm, against the laws of physics.
By Bret Kerr / Context Jamming · June 2026
For fifteen years, frontier artificial intelligence has been running on two operating systems that were both written in 2009 — and almost nobody noticed they were the same year. One was a neuroscience thesis about how the brain builds imagined scenes. The other was a physics thesis about how a boundary can encode an entire universe. Neither mentioned artificial intelligence. Both became it. Lila Sciences, a Cambridge company that emerged from Flagship Pioneering with $550 million and a plan to automate the scientific method, is the first organization to install both operating systems at once — and to run them not in a data center, but against the physical laws of nature. This is the story of how that happened, told through a lens I call Architectural Determinism.
- —Demis Hassabis, UCL — The Neural Processes Underpinning Episodic Memory
- —Jared Kaplan, Harvard — Aspects of Holography (AdS/CFT correspondence)
- —Fifteen years of parallel divergence before a single point of convergence
THE 2009 YEAR ZERO
The thesis I want to advance is simple: a frontier lab's foundational technology is an isomorphic projection of its founders' doctoral work. Call it Architectural Determinism. The clearest evidence sits in a single year.
In 2009, Demis Hassabis submitted a UCL thesis built on Scene Construction Theory — the argument that the hippocampus is not a passive memory drive but an active construction engine, generating coherent spatial scenes to remember the past, imagine the future, and plan. That bottom-up, simulate-to-master instinct became the architectural signature of DeepMind: Monte Carlo Tree Search in AlphaGo, latent-space world models in Dreamer, structure prediction in AlphaFold.
The same year, Jared Kaplan completed a Harvard thesis on the AdS/CFT correspondence — a formulation of the holographic principle, in which the complex dynamics of a higher-dimensional "bulk" are completely governed by constraints on a lower-dimensional boundary. Kaplan carried that instinct into AI: first the 2020 scaling-laws paper, then Constitutional AI at Anthropic, where a small set of written rules (the boundary) governs the behavior of an enormous parameter space (the bulk).
Two doctorates. One builds scenes from the bottom up. One constrains a system from a boundary down. For fifteen years they defined two separate empires of machine intelligence — and both, by 2024, had run into the same wall.
Note: The framing here (Architectural Determinism, 2009 Year Zero) represents Bret's analytical thesis, not Lila Sciences' official positioning.
- —Hassabis lineage: bottom-up agentic simulation, digital-first
- —Kaplan lineage: top-down boundary constraints, digital-first
- —Lila: bottom-up construction bounded by physical verifiers — physical-first
DEEPMIND vs ANTHROPIC vs LILA
Set the three labs side by side and the convergence becomes obvious. DeepMind models reality in a latent space to find optimal paths, then hands its predictions to human scientists to verify in a wet lab. Anthropic accelerates the thinking — literature synthesis, hypothesis generation — but stops at the laboratory door. Both are extraordinary. Both are still digital, and both are now starved for new high-quality data.
Lila's move is the synthesis: use a Hassabis-style construction engine to generate vast hypothesis spaces, then bound those hypotheses not with a text constitution but with the ultimate boundary condition — the physical laws of chemistry, biology, and thermodynamics, enforced by an autonomous robotic lab. The Kaplan boundary, made physical. The interactive matrix below lets you sort the comparison along any dimension; the short version is that Lila is the only one of the three whose verifier is reality itself.
Comparative matrix demonstrating three expressions of the same 2009 inheritance.
- —Lila Iris™ (the brain): converts hypotheses into executable experimental workflows
- —AI Science Factory™ (the body): globally distributed autonomous robotic labs
- —Whole-network feedback: every experiment, anywhere, becomes training signal everywhere
THE LILA INTELLIGENCE FLYWHEEL
The data wall is real: the supply of high-quality human text is finite, and the frontier labs are scraping the bottom of it. Lila's answer is to stop scraping and start manufacturing data. It trains on what the company calls "evergreen tokens" — brand-new data points generated continuously by real physical experiments. Every hypothesis, design, result, and failure is a token no competitor possesses.
The architecture is "One Mind. One System.": a reasoning model (Lila Iris) wired directly into a network of autonomous labs (the AI Science Factory) in a continuous loop — hypothesis, design, robotic execution, result, new token, improved model, better hypothesis. The laws of physics are the boundary constraint; a bad pipetting action and a bad idea both fail against the same unforgiving wall. CEO Geoffrey von Maltzahn has argued that public models will soon produce an "avalanche of similarly plausible hypotheses" — and that the durable moat will not be generating the hypothesis but owning the physical infrastructure that can test millions of them at speed. Reality, in other words, becomes the proprietary training set.
The Flywheel of Intelligence: Hypothesis → Experimental Design → Robotic Execution → Result → New Training Token → Improved Model.
- —Year Zero / Architectural Determinism in physical AI
- —Evergreen tokens & escaping the data wall
- —The Bitter Lesson meets the physical laboratory
- —Open-endedness as a defense against premature optimization
HIGH-SIGNAL ANGLES
Four frames make Lila a narrative, not a press release. First, the 2009 convergence — Lila as the first company to operationalize both the Hassabis construction engine and the Kaplan boundary inside one physical system. Second, evergreen tokens — the next durable moat in AI will not be a better model but proprietary physical data nobody else can generate. Third, the Bitter Lesson, applied to the bench: Rich Sutton's observation that general computation beats human heuristics, now aimed at lab execution itself — Lila reports a small team of non-experts produced an mRNA construct with a stated 3× expression improvement over a commercial benchmark in four months, and platinum-free electrodes claimed at 1,000–5,000× lower cost. Fourth, open-endedness: hiring Kenneth Stanley (author of Why Greatness Cannot Be Planned) as SVP of Open-Endedness signals a deliberate defense of serendipity inside a heavily capitalized company — protecting the system's "taste for what important novelty looks like."
Note: Lila's performance numbers (3× mRNA, 1,000–5,000× electrode cost) are Lila's stated results, not independently verified facts.
Building the Community a Physical-AI Company Actually Needs
A company selling physical outcomes can't run the standard developer-relations playbook of hackathons and SDK tutorials. Its ecosystem strategy has to look more like scientific co-creation: deep, joint-venture-style relationships with enterprise R&D, national labs, and the "bits-to-atoms" talent who can speak both software and the bench. The hard part isn't reach; it's credibility in skeptical technical rooms where "AI for science" still draws eye-rolls. The way through is not louder messaging — it's empirical proof, surfaced where researchers actually argue: the 3× and the 1,000–5,000× numbers, the open-sourced rigor, the willingness to be specific and falsifiable. Network quality over vanity reach. Earn the technical room first; the audience follows the trust.
That, read all the way down, is the real continuity from 2009: the structure of the training determines the structure of the result. It was true for a hippocampus thesis and a holography thesis. It is true for the communities a company like Lila chooses to build in.