The Thinking Game: Inside DeepMind's Decade-Long Quest for Artificial General Intelligence

01 A Lab on Camera, Not on a Press Release

The Thinking Game is a feature documentary that received its premiere at the Tribeca Film Festival and was co-produced by Google DeepMind itself. It follows Demis Hassabis and the DeepMind team over roughly five years of work toward artificial general intelligence, and the remarkable thing about the film is not any single revelation but the accumulation of ordinary scenes: planning meetings, failed experiments, arguments about research direction. Frontier labs are usually visible to the public only through announcement posts and staged demos, so a sustained moving-image record of one is genuinely rare.

That access comes with an obvious caveat, and this analysis treats it as central rather than incidental. A documentary produced by its own subject is a curated artifact. DeepMind had every incentive to show the moments that flatter its narrative and none to show the ones that do not. The film is best read the way analysts read a company's annual report: rich in verifiable detail, selective in emphasis, and silent by design about certain costs and failures.

What the film does document, and what the rest of this article examines, is a specific thesis: that AGI arrives as the output of an industrial research process, not as a single eureka. Understanding that process, and where the evidence for it is strong versus thin, is more useful than arguing about whether the goal is imminent.

02 Games as the First Assay: The AlphaGo Era

DeepMind's early strategy used games as a measurement instrument. A game like Go has perfect information, unambiguous rules, and a clean win condition, which makes it possible to know whether a learning system is actually improving. The founding insight of the lab's first decade was that this kind of controlled environment lets you iterate quickly on learning algorithms before you trust them anywhere else.

The measurable record from that era is unusually crisp. In October 2015 AlphaGo beat European champion Fan Hui five games to none, the first time a Go program defeated a professional player in an even match. In March 2016 it beat Lee Sedol four games to one in a match watched by tens of millions, with the second game's shoulder hit, move 37, becoming one of the most discussed moves in modern AI. In May 2017 it beat Ke Jie, then the world's top-ranked player, three games to none, after which DeepMind retired the program from competitive play. A year earlier AlphaGo's underlying approach had already been generalized, and in 2017 AlphaZero learned chess, shogi, and Go from the rules alone, surpassing the strongest dedicated programs in each domain.

The film uses these matches as narrative scaffolding, and they hold up to scrutiny: every score above is a matter of public record, not studio spin. What is interpretation rather than fact is what the wins meant. Beating the best human at Go demonstrated that learned search could exceed human expertise in one closed domain; it did not, and could not, demonstrate anything about generality.

03 AlphaFold: The Pivot From Games to Science

The film treats AlphaFold as the moment the lab's bet on general learning systems paid a dividend in the real world. The measurable record here is stronger than for almost anything else the lab has claimed. At CASP13, the protein-structure prediction assessment held in 2018, DeepMind's entry was the top performer on the hardest free-modelling targets, with a score around 58.9 on the GDT-TS scale. Two years later, at CASP14, AlphaFold 2 achieved a median GDT-TS of about 92.4 across all targets, a score DeepMind characterized as competitive with the accuracy of experimental methods in many cases. Independent assessors at the CASP conference effectively declared the 50-year-old protein-folding problem largely solved for many practical purposes.

The aftermath converted a benchmark result into scientific infrastructure. AlphaFold's predictions were compiled into a public database now covering hundreds of millions of predicted protein structures, and the method has since been cited by thousands of papers across structural biology and drug discovery. In 2024 the Nobel Prize in Chemistry was awarded jointly to Demis Hassabis and John Jumper for AlphaFold, alongside David Baker for computational protein design. These are checkable facts, not promotional claims, and the documentary is careful to ground its AGI narrative in them.

The interpretive leap the film invites, and that deserves scrutiny, is that mastery of protein folding demonstrates a general method. AlphaFold was a specialized architecture for a specific scientific problem, not a general agent. The honest reading is narrower: the lab showed it could take a hard real-world problem, define a measurable target, and beat every existing approach at it. That is evidence of research discipline, and it is the same discipline the lab points to when it argues it can repeat the trick in more general settings.

04 How a Frontier Lab Actually Runs

The film's most analytically valuable material is organizational, not intellectual. What emerges across its scenes is a picture of frontier research as an industrial process: research agendas planned in multi-year horizons, compute allocated like a strategic resource, and a deliberate policy of pursuing multiple technical bets in parallel rather than a single grand design. The AlphaGo program, the AlphaFold line, and the later Gemini model family are best understood as different instances of the same underlying method, applied to domains with progressively less structure.

Demis Hassabis, a chess prodigy before he was a neuroscientist before he was a founder, appears in the film less as an inventor than as a research executive: someone whose core skill is choosing what the organization should learn next. The film shows him repeatedly arguing that progress comes from systems that learn rather than systems that are programmed, which has been DeepMind's consistent position since before its 2014 acquisition by Google. Watching that thesis get tested at scale, with real budgets and real setbacks, is the film's contribution.

An analyst should distinguish what the camera verifies from what it merely dramatizes. The existence of disciplined long-horizon planning, of internal skepticism, and of failure as routine is visible in the footage and consistent with what former employees have said publicly. Exact figures for compute spending, staffing, and abandoned projects are things the film does not provide, and public estimates of them remain exactly that, estimates.

05 What Would AGI Even Mean

The documentary's framing term, artificial general intelligence, is one that even its strongest proponents define loosely. In the standard encyclopedic framing, AGI refers to a hypothetical system able to understand, learn, and apply knowledge across the full range of tasks a human can, in contrast to narrow systems that excel in one domain. That definition is a target, not a specification, and it is precisely the vagueness of the target that makes the term so useful in marketing and so hard to falsify.

The film positions DeepMind's work as a bet that generality comes from combining learning with search and planning, not from scale alone. Large language models, the technology most visible to the public since 2022, are statistical systems trained on text to predict the next token; the encyclopedic definition emphasizes both their power and their distance from the planning-heavy systems DeepMind spent its first decade building. The Gemini era has in practice blurred this line, folding language-model capability into a lab whose founding bets were elsewhere. Whether that combination is a path to generality or a very good integration of narrow capabilities is an open technical question, and the film honestly presents it as such.

Two claims should be kept separate when evaluating the film's thesis. The first, well supported, is that DeepMind has repeatedly converted a general learning approach into state-of-the-art results across very different domains. The second, unverifiable, is that the same approach scales to human-level generality. The documentary asks you to find the second credible because the first is true. That is a reasonable argument, but it is an argument, not evidence.

06 Safety Debates From the Inside

The film gives meaningful screen time to safety, and the way it does so is informative. DeepMind maintains dedicated research groups on alignment, evaluations, and frontier-risk assessment, and its leadership has signed the public statements, alongside peers, warning that mitigating AI extinction-level risk should be a global priority alongside pandemics and nuclear war. The documentary shows these debates happening inside the lab as research decisions, not as an external compliance function bolted on afterward.

A skeptical reading is available and should be stated plainly. A frontier lab that both races toward AGI and warns about its dangers is in a position of structural tension: the incentives that fund the race also depend on the race continuing. Internal safety teams give a lab the ability to slow down in principle; whether they can in practice, against competitive pressure from rivals working on the same timelines, is a question no documentary produced by the lab itself can be expected to answer.

What is verifiable is that the public debate has moved in the direction the film depicts. Governments now operate AI safety institutes, frontier models are subject to published evaluation regimes, and leading labs publish safety frameworks with named capability thresholds that would trigger stronger safeguards. Whether those commitments bind in practice is the live question, and the film's own footage of schedule pressure inside the lab is, read carefully, one of the better pieces of evidence on the side of the skeptics.

07 The Limits of the Camera

The film cannot show several things that matter most to an outside evaluation. It cannot show the internal economics of frontier research: the compute budgets, the abandoned lines of work, the projects that quietly failed. It cannot show geopolitics except at the margins, even though the AI race is now a central feature of great-power competition. And it cannot show the future, despite closing, as such films do, on an optimistic note about the decade ahead.

The correct way to use a documentary like this is as a primary source about a self-presentation, cross-checked against independently verifiable facts. The match results, the CASP scores, the database, the Nobel Prize: all check out. The claim that all of this constitutes progress toward AGI is a framing the evidence supports only partially, in that the trajectory of the last decade is real and impressive; the destination remains defined too loosely to confirm or deny.

That is the honest summary of DeepMind's decade as the film presents it, and as the public record permits us to say. The lab has demonstrated, repeatedly and measurably, that a small number of general learning techniques can be industrialized into results competitive with human expertise across domains. Whether that observation is the first chapter of a path to AGI or a description of a very powerful family of specialized tools is the question the film wants you to leave with, and it is the one that remains open.