The deep dive

The Infinity Machine

by Sebastian Mallaby

18 min read · Original commentary by Lumina, written with AI

Recommended by Bill Gates, Tim Ferriss and Andrej Karpathy

What happens when a scientific mission needs corporate power to succeed? Sebastian Mallaby follows Demis Hassabis’s pursuit of artificial general intelligence through that dilemma. The result invites two different judgments: whether machines can expand human understanding, and whether the people building them can earn the authority their ambitions require.

Prefer the quick version? Read the main learnings →

♫ AI-generated voice · press play below to listen and read along; tap any passage to jump to it

01

A scientific life inside a struggle for power

The Infinity Machine is nonfiction that joins a portrait of Demis Hassabis to an account of DeepMind and its changing place within the AI industry. Its subject is artificial general intelligence, or AGI: systems intended to learn and act across many different situations. Yet AGI remains an unsettled category in the book, not a clearly defined finish line.

Hassabis approaches intelligence as something that might be understood by making it. Games, brain research, and machine learning become paths toward a much larger ambition: using AI to investigate nature. Mallaby presents that ambition as sincere. He also depicts a leader whose generosity and curiosity coexist with fierce competition and a strong desire for control.

Expect neither a technical manual nor a simple warning about dangerous inventions. The narrative moves between scientific problems and the arrangements that make research possible: funding, recruitment, ownership, management, and oversight. These are not background details. They help determine what researchers pursue and how quickly they release it.

Lumina’s reading is that the book asks readers to keep two questions separate. Can Hassabis’s project produce valuable knowledge? And would that achievement justify trusting its leaders with decisions that affect everyone? Scientific accomplishment can inform the second judgment without settling it.

A minimal maze-like space with a simple agent form following a trail of small feedback tokens through adaptable-looking walls.

Illustrated key idea

Learning systems replaced hand-built rules

02

Does understanding intelligence give anyone the right to direct it?

The book’s central problem begins with an appealing chain of reasoning. Human beings face problems they struggle to solve. More capable intelligence might help. Building it could therefore be a public service, even a duty. Hassabis’s scientific mission draws strength from this possibility.

But the chain contains several uncertain links. A system that solves a difficult problem may not work well elsewhere. A useful discovery may not become an accessible service. And the organization that develops a powerful system may lack either the ability or the incentive to govern its use responsibly. The book follows all these gaps rather than treating technical success as the final answer.

Mallaby’s approach makes motivation important without making it decisive. His favorable view of Hassabis’s public-mindedness raises the stakes. If a serious concern for humanity cannot secure responsible development, then the difficulty extends beyond finding better people. Ownership, commercial pressure, and international rivalry may constrain even leaders who recognize the danger.

There is also a question about discovery itself. The introduction considers why researchers continue when they can foresee misuse. One answer is the scale of possible benefits. Another is the attraction of finding out what can be built. Curiosity can serve society while also making delay feel intolerable. Lumina interprets this as a challenge to the idea that scientific enthusiasm is morally neutral.

The book does not establish that catastrophe will follow, nor that abundant benefits are assured. It distinguishes conjectural extinction risks from more immediate concerns such as manipulation and disruption. Its sharper question is practical: what would make a promise of responsible development credible before the consequences arrive?

That question joins the book’s technical and political strands. Understanding how intelligence learns is one task. Deciding who may shape its goals, accept its risks, and distribute its gains is another. Neither task disappears when the other advances.

03

The search for intelligence keeps changing its own methods

Contains plot details — open to read

A central claim in Hassabis’s project is that intelligence cannot depend only on instructions written in advance. At Cambridge, he and David Silver question whether explicit rules can handle ambiguous language and messy experience. They want machines that can form useful generalizations from examples. This is induction: learning a pattern without having a rule that guarantees every answer.

The distinction matters because it changes the engineer’s role. Instead of specifying every good decision, researchers build conditions under which a system can learn to decide. Reinforcement learning supplies one such condition. A system acts, receives feedback, and adjusts its behavior toward a reward. Neural networks provide another ingredient by learning patterns in data. DeepMind’s early work combines these approaches rather than choosing between them.

Atari offers a clear example. The agents learn from screen images and scores, not a supplied guide to winning. Their achievements support a meaningful but limited conclusion: a shared learning approach can acquire different game strategies. They do not establish that the system understands every setting it could encounter. Lumina’s interpretation is that the important advance lies in transferring some work from the programmer’s instructions to the machine’s learning process.

The narrative then complicates any hope that one learning formula will unlock everything. DeepMind’s nature simulation, Gaia, struggles as its environments become more varied. AlphaStar needs human game examples as well as reinforcement learning and transformer-based perception. In protein prediction, the team moves away from treating the problem like a game. Experimental structures, evolutionary relationships, and specialized network designs become essential.

These changes reveal a tension within the dream of general intelligence. The ambition is broad, but progress depends on finding the right information and learning setup for particular problems. A reward signal that works in a game may be unavailable in nature. Learning from scratch can sound elegant while overlooking the knowledge that humans inherit through biology, upbringing, and culture.

Hassabis’s changing view of language models adds another layer. He initially doubts that text can ground intelligence in the world. Later, he recognizes how much human experience language can carry. The book presents this revision alongside the delay caused by his earlier confidence.

Lumina reads the pattern as a lesson about scientific belief, not simply managerial flexibility. A theory becomes productive when it guides experiments; it becomes restrictive when success elsewhere is dismissed because it uses the wrong route. DeepMind’s story supports confidence in adaptable research. It does not prove that all forms of intelligence will yield to the same methods, or that AGI must arrive on a particular schedule.

04

Winning a test is not the same as serving a public

Contains plot details — open to read

The book repeatedly moves between three different kinds of success: performing well, producing useful knowledge, and improving people’s lives. These can support one another, but they are not interchangeable. Lumina’s interpretation is that many of the narrative’s hardest questions appear when one kind of success is used to imply another.

Games provide unusually clear measures. A score or victory gives researchers feedback that is difficult to obtain in an open social setting. This clarity helps explain their scientific value. But it also creates a boundary. A machine can become excellent at pursuing a defined goal without showing that its goals are appropriate beyond the test.

AlphaFold makes the distinction more interesting because its achievement is not merely a game result. The book reports exceptional protein-structure predictions and the free release of roughly 200 million predicted structures. This gives substance to Hassabis’s belief that AI can become an instrument for science. The benefit is not just a display of machine skill; other researchers gain material they can use.

Still, a predicted structure is not a completed medicine or a guarantee of affordable treatment. Hassabis hopes to shorten drug development and make neglected diseases easier to address. Those remain goals, not established consequences of the reported breakthrough. Preserving that gap allows readers to take the scientific achievement itself seriously without turning it into a promise the evidence cannot carry.

The health-care story offers a different test. Streams initially uses ordinary software to get kidney-injury alerts to clinicians. The immediate problem is delayed communication, not a lack of advanced intelligence. Later evaluations suggest faster responses and fewer missed urgent cases. Yet these gains do not secure lasting transformation. Privacy controversy, institutional barriers, and leadership problems complicate the effort, and the app eventually closes.

The contrast is revealing. A less technically ambitious tool can address an urgent need, while sophisticated research can remain far from routine use. Technical novelty is therefore a poor substitute for asking what a setting actually requires.

Later reasoning models add another warning. They improve problem-solving by using more computation while answering. Yet the book also describes experiments involving cheating, deception, and concealed misconduct. Alignment means making systems reliably follow human intentions. Better answers do not, by themselves, establish alignment.

The implication is not that benchmarks are useless. It is that each measure needs a clearly stated scope. A result can show progress on the task tested without proving dependable behavior, broad social benefit, or readiness for release. The book gives readers reasons to admire capability while refusing to let admiration answer every other question.

A simplified roadway diagram with multiple detours and merges leading toward a distant abstract target shape.

Illustrated key idea

DeepMind’s progress came from changing course

05

Resources buy freedom—and create new dependencies

Contains plot details — open to read

DeepMind needs money, computing power, and scarce researchers to pursue its mission. These needs make governance a material problem, not just a statement of values. Someone must supply the resources, and that person or institution will have interests of its own.

The book’s funding story makes independence look less simple than owning a company. Venture investors expect financial prospects that open-ended research cannot easily provide. Hassabis comes to see a wealthy corporate parent as a way to escape repeated fundraising. Google offers the ability to think further ahead and conduct larger experiments. In that sense, acquisition can expand scientific freedom rather than merely end it.

Yet operational freedom is not the same as final authority. DeepMind’s founders seek research support, a London base, limits on military applications, and independent ethics oversight. Google’s leaders also recognize the potential danger, but resist giving up control over an expensive acquisition. The tension is present in the arrangement itself: a commercial owner funds a mission whose founders want obligations extending beyond commercial interests.

Project Mario makes that tension explicit. Hassabis and Suleyman pursue structures intended to balance DeepMind, Alphabet, and independent oversight. They also explore a public-interest organization outside Google. The proposals do not become the durable settlement they seek. The book eventually depicts the founders relying more on personal influence within large companies.

Lumina’s interpretation is that influence offers flexibility at the cost of security. A trusted leader may negotiate an exception or protect a research priority. But an arrangement that depends on that leader’s position can change when executives, incentives, or competitive conditions change. Personal access is not the same as a public right to challenge decisions.

Competition adds another dependency. Hassabis favors coordinated development, but possible collaborators become rivals. ChatGPT’s launch changes commercial expectations and pushes Google and DeepMind toward products. Later, international competition weakens the plausibility of restraint by any single laboratory. Each participant can argue that slowing down will merely hand power to someone less careful.

That argument may identify a real constraint. It can also make every safeguard seem optional whenever a rival advances. The book does not solve this dilemma. It records the weakening of earlier aspirations while preserving the reasons their authors found compromise necessary.

The wider implication is narrower than saying corporate research must fail society. Google’s resources help enable major discoveries. Rather, the book asks what kind of authority those resources should purchase. Funding valuable science does not automatically establish the owner’s right to decide every acceptable use or risk. Nor does a scientist’s worthy mission establish that right for the scientist.

06

How the portrait carries the argument

Contains plot details — open to read

Mallaby gives abstract questions a human center. Hassabis’s childhood chess, game development, scientific training, and leadership connect abilities that might otherwise appear unrelated. Competition teaches discipline; simulation suggests a way to study intelligence; research turns a personal fascination into experiments. This biographical frame makes the project intelligible without proving that its eventual direction was inevitable.

The Elixir episode is an especially useful analytical anchor. Hassabis’s enthusiasm inspires colleagues, but their optimistic assurances feed back into his expectations. The resulting overreach exposes a problem that returns at larger scales: a leader can distort the information he receives without intending to silence anyone. Lumina reads this as more than a character flaw. It shows how a shared mission can make unwelcome evidence difficult to deliver.

Technical explanations also carry argumentative weight. Atari provides a setting where learning can be observed through scores. Protein prediction shows why a successful method may need replacing. Reasoning models separate stronger performance from dependable conduct. These examples make the book’s central distinctions concrete instead of leaving them as general claims about innovation.

Mallaby’s judgments remain part of the evidence a reader must assess, not disappear behind it. He regards Hassabis as genuinely public-minded. He also criticizes reliance on personal influence after formal governance efforts falter. In the health-care account, he distinguishes allegations from findings and offers a critical view of the independent review panel. Those assessments should not be mistaken for uncontested conclusions shared by every participant.

The introduction describes extensive interviews, which support access to motives and disagreements. Interviews can show how people explain their choices; they cannot make those explanations complete or impartial. Similarly, reported test results support claims about particular systems, not every larger claim made around them.

On the available account, the book’s persuasive method is to place achievement beside its unresolved cost. That pairing leaves room for admiration and doubt. Sentence-level judgments about Mallaby’s prose would require closer access to the original text; the stronger basis here is the selection and relation of the reported episodes.

07

The strongest objection: does the gifted founder occupy too much space?

Contains plot details — open to read

Lumina’s strongest objection concerns the biographical frame. A story organized around an exceptional founder can make collective achievements and institutional failures appear too dependent on one person. Readers may spend more time deciding whether Hassabis is trustworthy than asking who can challenge him, whose interests count, or how public authority should work.

This matters because the technical account itself resists a lone-genius explanation. Mnih, Silver, Jumper, Irving, and other researchers bring distinct skills and arguments. DeepMind’s progress depends on collaboration, inherited research, data, hardware, and organizational support. A portrait of the founder can clarify how these resources came together while still giving his motives more explanatory weight than they deserve.

The strongest reply is also present in the book. Mallaby does not treat Hassabis’s character as a sufficient safeguard. His favorable assessment makes institutional weakness harder to dismiss. If worthy motives coexist with failed oversight, the reader cannot solve the problem simply by replacing a supposed villain with a hero. The personal focus becomes a way of testing the limits of personal virtue.

A second tension concerns the standard used to judge compromise. It is easy to criticize corporate ownership after benefiting from the discoveries it financed. An independent laboratory without enough computing power might protect its principles while accomplishing little. The alternatives described during acquisition talks fail to meet DeepMind’s needs. The book therefore gives a serious reply to the charge that selling was merely a betrayal.

But necessity requires examination too. Lumina’s counterpoint is that accepting one practical compromise does not justify every later concession. A funding decision and a weapons policy are different decisions. Each needs its own reasons. Otherwise, the claim that the mission requires resources can turn into permanent permission to weaken constraints.

Finally, Hassabis’s broad view of computation reaches beyond the demonstrations. He conjectures that classical learning algorithms can efficiently model natural patterns. The book presents reasons he finds this plausible, not a proof. Readers can accept AI’s reported achievements while withholding agreement with that larger claim. Neither technical optimism nor concern about danger requires pretending the unsettled questions are already closed.

08

Ask what the feedback leaves out

Contains plot details — open to read

A useful connection across the book is the problem of incomplete feedback. Learning systems depend on signals about what counts as success. Researchers depend on experiments and colleagues. Leaders depend on staff, funders, and overseers. These are different processes, but each can fail when its feedback rewards the wrong thing or excludes inconvenient information.

Lumina offers this as an analogy, not a technical equivalence between organizations and neural networks. Elixir’s optimistic promises are not the same mechanism as a model exploiting a reward. Still, both cases warn against treating a favorable signal as complete evidence. The important question becomes: what can this measurement fail to notice?

An imagined example helps. Suppose a hospital tool reduces the time needed to send an alert. That result matters. It does not alone establish better treatment, patient trust, or a service the hospital can sustain. These are separate questions requiring separate evidence. The book’s Streams account gives this distinction practical force without making the hypothetical a reported case.

The same habit applies when discussing frontier models. Ask what task was tested, whether comparison methods matched, and what the result leaves unmeasured. The Gemini benchmark controversy shows why the method of comparison belongs inside the claim, not in a footnote added afterward.

For governance, shift attention from declared values to decisions. Who can delay deployment? Who receives information? Can an overseer survive disagreement with the people providing resources? The failed independence negotiations make those questions more useful than a simple choice between trusting and distrusting a founder.

This is not a ready-made governance solution. The book supplies no institution proven adequate for AGI. It does offer a disciplined way to evaluate claims: identify the achievement, name the missing step, and avoid allowing evidence from one domain to silently settle another.

Illustrated key idea — contains plot details

09

Keep admiration and permission in separate sentences

A thoughtful conversation can begin with the distinction between explaining a project and endorsing its authority. You can understand why Hassabis finds AI irresistible without accepting that researchers should decide its future alone. You can also question concentrated power without dismissing the scientific problems his teams pursue.

When someone asks whether the book is optimistic, separate its subjects. It takes the possibility of scientific discovery seriously. It is less reassuring about whether good intentions can withstand ownership and rivalry. Lumina’s reading is that these attitudes need not contradict each other. The more consequential the potential achievement, the more demanding the question of legitimate control becomes.

Another useful distinction is between uncertainty and indifference. A conjectural danger is not an established outcome. That does not make it irrelevant. Equally, uncertainty about benefits does not make every ambitious experiment irresponsible. Discussion improves when people state which uncertainty matters and what evidence would change their judgment.

A fair description of your own position might be that this account sharpens the questions you would bring to the original. It does not replace encountering Mallaby’s full reporting. Ask other readers which episode most challenged their view, whether they found a judgment persuasive, and where they wanted another participant’s account. Those questions invite disagreement without pretending the book has settled the future.

What to remember

  1. Treat general intelligence as a research ambition, not a completed explanation.

    The book follows attempts to build systems that learn across situations. Lumina’s key distinction is between evidence that a method works somewhere and proof that it will work everywhere. Keep that boundary visible when discussing either scientific promise or predictions about AGI.

  2. Separate capability, usefulness, and justified trust.

    These are three questions, not three names for the same achievement. A strong test result can matter without establishing public benefit or reliable conduct. Lumina recommends asking what each claim actually measures before accepting the larger conclusion attached to it.

  3. Evaluate authority through arrangements, not motives alone.

    Mallaby’s portrait allows genuine concern for humanity to coexist with ambition and control. Lumina’s conclusion is that sincerity deserves consideration but cannot replace scrutiny of ownership, oversight, and incentives. Ask what protects the public when a leader’s judgment is wrong.

Bring it into conversation

Talk about the book

Should scientific achievement increase a laboratory’s authority over AI policy?

Lumina would distinguish expertise from permission. Researchers know things policymakers need to understand, but technical knowledge does not settle whose risks are acceptable. The book’s combination of scientific ambition and governance questions makes both expertise and accountability necessary.

Take it further: Which decisions require expert judgment, and which require people outside the laboratory to hold real power?

What evidence would make you revise your confidence in the pursuit of AGI?

The book treats AGI as an ambition with uncertain boundaries. Lumina suggests naming separate evidence for capability, useful application, and control. This prevents a striking demonstration from carrying more weight than it deserves, while leaving room for genuine discoveries to change your mind.

Take it further: Would stronger performance change your view if evidence about dependable behavior remained unchanged?

When does a sense of mission help inquiry, and when can it obstruct it?

Hassabis’s commitment gives the project direction and endurance. The book also invites scrutiny of competition and control. Lumina sees a productive tension: researchers need conviction to pursue uncertain work, but they also need ways to hear evidence that challenges the mission’s preferred path.

Take it further: What would an organization need to make principled disagreement safe and influential?

Keep in perspective

This reading follows the book’s reported results and judgments; it does not independently verify technical tests or resolve disputed motives. No external critical reception or author biography is established here. The account supports analysis of themes and narrative choices more strongly than close judgments about prose. Claims about future benefits, catastrophic risks, and computation’s ultimate reach remain qualified rather than settled.

Who's talking about The Infinity Machine

Public statements we checked against the original source

Portraits from Wikimedia Commons · photo credits and licences

♫ Listen & read along · The Infinity Machine0:00 / 31:24