We know a great deal about both of these men, and nothing about what goes on inside either of their heads.

If neither name means much to you, that's fair, and here is the short version.

Francis Bacon (1561–1626) was an English lawyer, politician and philosopher; not the twentieth-century painter of the same name. He rose to Lord Chancellor, England's highest legal office, in 1618, and three years later Parliament impeached him for taking bribes, which he confessed. In between, in 1620, he published the Novum Organum, a book about how to learn things about nature you can actually trust: gather observations, test ideas against them, and throw out the ones that fail. Historians give him a central place in the birth of modern experimental science (Stanford Encyclopedia of Philosophy). His NewAtlantis, published after his death, imagines an island run on that method, with an imagined research institution at its heart, Salomon's House. We'll get to it.

Demis Hassabis is the co-founder and chief executive of Google DeepMind, an artificial intelligence lab that started in 2010 and is now part of Google. His lab built AlphaGo, which beat a top professional at the board game Go in 2016, and AlphaFold, which predicts the three-dimensional shape of proteins. A protein's shape "is decisive for the protein's function," and AlphaFold 2 predicted "the structure of virtually all the 200 million proteins that researchers have identified" (Royal Swedish Academy of Sciences). For that work he and his colleague John Jumper shared half of the 2024 Nobel Prize in Chemistry.

Why put them side by side? This is the first essay in Captain to Captain, a series that pairs a thinker from history with someone building things now and compares how each goes about the work. Bacon wrote a rulebook for deciding when a discovery deserves to be believed. Hassabis runs one of the labs now using AI to make discoveries. The question underneath matters whether or not you've heard of either man: when a powerful lab tells us something new about the world, who gets to check it?

Francis Bacon has been dead since 1626, which limits the interview options. Demis Hassabis is very much alive, and talks to journalists, podcasters and Nobel committees with a frequency Bacon would have found suspicious. But a transcript is not a mind. What we have, for both, is the record: what they wrote, what they said on the record, and what they built. So that's what this essay uses: their words, quoted and linked. Where we guess, we say so.

The guess we started with was wrong, which is the most Baconian thing that will happen in this essay. We expected Bacon to be the man against leaps, and his Salomon's House to be the open lab that would settle the argument. He isn't, and it wasn't.

Start with the record. Their methods are not the same: Bacon was after the causes of natural things, and Hassabis builds systems that learn to predict. But both work by the same logic: gather, make a disciplined leap, try to kill it.

Bacon's half of that sentence is the surprise. He is remembered as the patron saint of slow, careful induction, the man who wanted you to pile up observations until the truth fell out. He did want the care. The understanding, he wrote, "must not therefore be supplied with wings, but rather hung with weights, to keep it from leaping and flying" ( Novum OrganumI.104, in James Spedding's translation). He wanted a ladder, "a just scale of ascent, and by successive steps not interrupted or broken," from particulars to lesser axioms to middle axioms to the most general (I.104). And he had no patience for the lazy version of induction that people now pin on him: "the induction which proceeds by simple enumeration is childish; its conclusions are precarious, and exposed to peril from a contradictory instance" (I.105). What he's dismissing is now called enumerative induction.

But then, in Book II, after the tables are built (lists of where heat is present, where it's absent, where it varies by degree), Bacon does the thing his reputation says he never does. He leaps. "Since truth will sooner come out from error than from confusion," he writes, the understanding should have "permission, after the three Tables of First Presentation… have been made and weighed," to venture an interpretation (II.20). He calls it the First Vintage, "made by way of indulgence to the understanding." Then he makes it, about heat ("Heat is a motion, expansive, restrained…"), and he knows exactly how it dies: "any one contradictory instance overthrows a conjecture as to the Form" (II.18). (The Form is Bacon's word for the cause he was hunting: what he meant by it.)

A provisional guess, clearly marked as a guess, made only after the evidence is tabled, and killed by a single counterexample. That's not a man who distrusts leaps. That's a man who has written the safety manual for one.

Hassabis works the same logic with different machinery. He has called the scientific method "probably the greatest, invention humanity has ever come up with" (sic, the comma is in the transcript; Nobel interview, 2024). Note the "probably". It's his, and it matters. He has spent his career, he says, on AI "because I believe it could be the ultimate tool to help with science." His tables are data. His leap is a trained model. And his test of a leap would have made sense to Bacon: an experiment is worth running "as long as you are picking experiments and hypotheses that meaningfully split the hypothesis space and you learn something" (Lex Fridman, 2nd interview, 00:40:44).

Bacon said almost the same thing four hundred years earlier, about what he calledexperiments of light ( experimenta lucifera ): they "answer the end equally well whichever way they turn out; for they settle the question" (I.99). Two men, four centuries apart, agreeing that an experiment's job is to decide the question, not to win it.

So the logic is shared. The disagreement is in three places, and they're the interesting part.

What counts as the tables. Bacon's tables are written by people: instances collected, sorted, checked, by hand, by a community of investigators. Hassabis's tables are data, at a scale no seventeenth-century clerk could hold. The AlphaFold structure database (itspaper) alone, at DeepMind's last count, has "over three million users from over 190 countries" (DeepMind).

What counts as a contradictory instance. For Bacon, a contradictory instance is a fact in the world that the conjecture can't survive, gathered by investigators into tables. For Hassabis, the counterexample often arrives as a score: a game lost, a benchmark (a standard test everyone scores against) failed, a structure that doesn't match the experiment. Games were, in his words, "a test bed, a proving ground for the AI systems to learn and see if these learning algorithms could work" (Nobel interview). The famous one was Go: in March 2016 his lab's AlphaGo beat the 9-dan professional (the top rank) Lee Sedol in a five-game match. A game has rules, a score, and no opinion about your hypothesis. It is, in a way, a contradictory instance that comes with a scoreboard.

Neither kind arrives on its own. Bacon's instances were collected and sorted by people working to a plan; a benchmark is built, chosen and read by people too. The difference that matters, in our view, is who builds the test and who gets to enter its result into the record. Hassabis's field already has a fine example of getting that right. AlphaFold 2 was validated at CASP14, which "provides an independent mechanism for the assessment of methods of protein structure modeling" (CASP14), using structures "that have not been deposited in the PDB or publicly disclosed so that it is a blind test for the participating methods" (the PDB is the Protein Data Bank, the public archive of solved structures) (Jumper et al., Nature, 2021). That's a test written by someone who isn't the guesser.

It's also where each method has something to say to the other, and here we are inferring, not reporting. Built from Bacon's own principles, our question for the benchmark is about the Idols (Bacon's name for the habits of mind that lead people astray): whose tables are these, and what did the people who built them carry in with them? His Idols of the Theatre were the errors that "immigrated into men's minds from the various dogmas of philosophies, and also from wrong laws of demonstration," because "all the received systems are but so many stage-plays, representing worlds of their own creation after an unreal and scenic fashion" (I.44). A benchmark everyone trains toward is a fine candidate for a modern stage play.

The pairing earns its keep here, because Hassabis, on the record, has already asked the same question of his own field. In his framework for frontier AI, he wants an independent standards body that would "eventually… build up the technical capacity to create its own held-out tests independent of the Labs to prevent overfitting" (Hassabis, Substack), with "an ecosystem of third-party auditors." (A held-out test is one the system never saw while it was being built; overfitting is fitting the data you have so closely that the model fails on new data.) That's the Idols problem, spelled in machine learning: don't let the people who built the leap also write the test it passes. The same paragraph is honest about the meantime: "Initially, they would be developed in consultation with Frontier Labs" (Hassabis, Substack; he also posted it on X). Consulting the labs isn't the same as letting them write the exam: a standards body can hear developers out and still keep the hidden tests, the scoring and the results to itself. Which this would be depends on governance the sentence doesn't spell out. That's our question for it, from Bacon's principles.

And in January 2025 he said, plainly, where the leap still fell short of Bacon's. What was missing from the systems of the day, he said, was "the ability for these systems to invent their own hypotheses or conjectures about science, not just prove existing ones. They can play a game of Go at a world champion level. But could a system invent Go?" (BigTechnology). Bacon would recognize the gap. His First Vintage is a conjecture invented by a person. The machine, by its builder's own account, could win the game but not yet write it.

The record has moved since, and an essay about admitting contradictory instances had better admit this one. In May 2026 his own lab published Co-Scientist in Nature, a system that "iteratively generates, debates, and evolves novel hypotheses for complex scientific problems" and "breaks down high-level research goals into executable steps" (GoogleDeepMind; Gottweis et al., Nature, 2026). Some of its proposals have held up at the bench: drug candidates for acute myeloid leukaemia were "shown to be potentially beneficial in cell line experiments" (Nature press release). So the machine now helps write guesses, not only test them. It does so inside a research goal a person set, and DeepMind calls it "a partner in research, not a replacement for scientific or clinical expertise." Hassabis draws the same line himself. "For the next decade or so, we should think about AI as this amazing tool to help scientists," he said in 2026; "Conjectures and hypotheses are going to come from the human experts." And he still asks: "could it invent a game as elegant as Go?" (Daedalus, 2026). The boundary has moved, from testing the guess to choosing the question. Whether it stays there is genuinely open.

(The line "Knowledge is power" will not appear in this essay attributed to Bacon, because the record doesn't put it in Novum Organum the way everyone thinks it does: the nearest he gets is "ipsa scientia potestas est," in his Meditationes Sacrae of 1597, and the exact phrase is Hobbes's, from 1668. Neither will "torture nature"; the scholarly argument over that one is its own essay. Those are the failed-bar-trivia-night Bacons. We have the real one.)

The third difference is the hardest one, and it's about who gets to see the results.

Bacon didn't just write a method. He imagined an institution to run it: Salomon's House, in New Atlantis, published after his death. It was never built; it's a design on paper, often held up as an early sketch of the organized research lab. Read what this imagined lab actually does.

"The End of our Foundation," says the Father who explains it, "is the knowledge of Causes, and secret motions of things; and the enlarging of the bounds of Human Empire, to the effecting of all things possible" (New Atlantis). It sends twelve men abroad "under the names of other nations, (for our own we conceal)," to bring back other countries' books and experiments. They are called Merchants of Light. And then:

"And this we do also: we have consultations, which of the inventions and experiences which we have discovered shall be published, and which not: and take all an oath of secrecy, for the concealing of those which we think fit to keep secret: though some of those we do reveal sometimes to the state and some not." (New Atlantis, ProjectGutenberg #2434)

The House is not simply closed. It sends its fellows on circuits of the kingdom, where "we do publish such new profitable inventions as we think good," and it warns the people of plagues, tempests and earthquakes. At the end, the Father gives the visitor "leave to publish it for the good of other nations" (same text). It publishes. It just decides what, and the operative words are "as we think good." The founding sketch of the research institution includes an oath of secrecy and a committee that decides what the public, and the state, are allowed to know.

DeepMind is not Salomon's House; it publishes a great deal. AlphaFold's predicted structures are free to anyone, and the 2021 database was, in DeepMind's words, "a tipping point toward AlphaFold becoming a scientific tool adopted around the world" (DeepMind). But it sits inside Google, and it decides the rest for itself.

AlphaFold 3 is the case that shows the shape of it. In its makers' own description, it "generates 3D structure predictions of biological molecules" (AlphaFold 3 terms of use): proteins, and the other molecules they work with. A researcher who wants to check it, run it on their own problem, or build on it needs the code and the trained model, the weights (the numbers training produces). When the paper appeared in Nature in May 2024, the code and model weights did not. Researchers got a web server with daily limits, for non- commercial use. Nature 's editor in chief, Magdalena Skipper, told Retraction Watch that the journal "accepts that there may be circumstances under which research data or code are not openly available" (Retraction Watch), and the journal's own editorial on the decision says: "This was not a decision we took lightly" (Nature ). John Jumper, a senior author (and Hassabis's co-laureate), reportedly told the press that DeepMind and Isomorphic Labs (the drug-discovery company Hassabis founded and runs) had consulted more than 50 experts in biosecurity, bioethics and AI safety (GEN; journalism, and the reporter's paraphrase, not a quote).

Within days, DeepMind's Pushmeet Kohli, a vice president of research, promised more: "we're working on releasing the AF3 model (incl weights) for academic use, which doesn't depend on our research infra, within 6 months" (Kohli on X). They kept the promise. In November 2024 the code came out under a non-commercial license, "Attribution- NonCommercial-ShareAlike 4.0 International" (v3.0.0 license; what that license allows). The weights came by application, and in the README of the day: "Access will be granted at Google DeepMind's sole discretion" (AlphaFold 3 README, v3.0.0).

Then the House changed its mind, twice. In June 2026 the code license went from that non- commercial license to Apache 2.0, which has no non-commercial restriction (commit,2026-06-09; a commit is one recorded change to the code). In July 2026 the application form went away: the README now says "You can download the AlphaFold 3 model parameters from" a public link (README; commit, 2026-07-23). The weights are still on the House's terms: "only available for non-commercial use by, or on behalf of, non- commercial organizations," and "You must not publish or share AlphaFold 3 model parameters" (terms of use).

So: a House consults its own experts, weighs the danger, withholds a discovery, releases it to some and not to others at its discretion, and then, over two years, widens the door. That last part looks like the method working: a decision made, then revised in public. That reading is ours; the record shows what changed, not why. In Bacon's text, the Fathers revealed some things "to the state and some not," and we have no record of them ever changing their minds.

Who decides, inside this House, is partly on the record. "Model launches are reviewed at Google DeepMind's Launch Review forum, which approves model releases," says Google's report under the OECD's Hiroshima AI Process reporting framework (OECD). Separately, DeepMind had, in 2023, an internal ethics body, the Institutional Review Committee, and as

TIME reported that year, "any disagreements are escalated to DeepMind's executive leaders for a final decision" (TIME, 2023; journalism). DeepMind's own site now names its internal review group the Responsibility and Safety Council, which "evaluates Google DeepMind's research, projects and collaborations against our AI Principles"; the record we have doesn't say how the two relate. Those describe how the House works in general. The record we have doesn't say which of them, if either, made the AlphaFold 3 calls.

And Hassabis has said, in his own words, why openness worries him. Of open-source AI generally, the question, he asked, is "how does one stop bad actors, individuals or up to rogue states, taking those same open source systems and repurposing them for harmful ends?" (Dwarkesh Patel). And in 2023, talking about whether the AI field's habit of openly publishing its research could last, he said: "We're getting into an era where we have to start thinking about the freeloaders, or people who are reading but not contributing to that information base" (TIME, 2023).

Those are serious concerns, seriously stated. But they are general statements, about open AI systems and about publishing; neither is an account of the AlphaFold 3 decision. The fifty biosecurity, bioethics and safety experts are the closest the record comes, and that's reported speech. The record also fits a second reading just as well: Isomorphic Labs, a drug-discovery company, contributed core work to AlphaFold 3, and the weights' terms bar commercial use. Safety and commerce point the same way here, and nothing public says which weighed more. The honest limit, then: we know what happened (the record is public), how releases are reviewed in general (a forum; an ethics committee whose disagreements go to the executives), and the concerns he has voiced. We don't know how those were weighed for this decision, because that happened inside the House, and the House doesn't publish its minutes. In both Houses, who decides is roughly visible. Why, in full, isn't.

Opinion: in the AlphaFold 3 episode, the hesitation looks like it was for a good reason. The visible pattern (hold back, consult, promise, release, then release more) is what deliberate caution looks like from outside, and it is the kind of pause this magazine has argued for before (No. 02, "Agents and the Agency of Pause"). But a good reason kept inside the House is still kept inside the House, and the "sole discretion" of 2024 is a phrase the Fathers of Salomon's House would have signed without changing a word.

(Yeah. I started a beef with a dead polymath and a Nobel laureate. On purpose, not by accident: a beef done right aims at the argument, never at the man, because a fight is how anyone checks a claim. Here it is in Chuck D's register, a cadence study; the words are ours, not his.)

Four hundred years in the ground, still holding the floor, / Wrote the manual for doubt, then shut his own door. / Tables for the heat, tables for the cold, / Says one hard counterexample out-argues the fold.

Who's grading the exam when the grader wrote the questions? / Bacon called it Idols, now it's lab-built suggestions. / Put it on the table (yeah, the tables that he drew), / Table it, or open it, the blind run tells you who.

Ain't no hate in the room, I'm not here for your head, / I'm here for the receipts and the minutes unread. / Two captains, one ship, and a crew that's been shut out, / The counterexample's a person. Let the people come out.

[Nobody speaks for anybody here. Each man's standard is applied to the other's practice, in our words, built from theirs. It's inference, and it's labeled, because it's the boldest thing in the essay.]

Start with what they agree on, because it's more than either fan club admits. A guess is allowed, if it's marked as a guess and made after the evidence is on the table. An experiment is worth running when it can come out either way. One honest contradictory instance outweighs any pile of agreeable ones. Bacon wrote all three down. Hassabis works by them in the open: leaps tested against games that don't care who you are, experiments picked because they "meaningfully split the hypothesis space." It's the same safety manual, read four centuries apart. Now put each man's standard to work on the other.

Bacon's standard, applied to DeepMind: the loop is his, run at a scale he couldn't have imagined. The question, built from his principles, is the same one as before: whose tables are these, and what did the people who built them carry in? A test written by the lab being tested is a table with its builder's Idols already inside. Hassabis has said as much himself; he wants an independent body with "its own held-out tests independent of the Labs." By Bacon's standard, that's exactly right. The word to watch is "initially": in Hassabis's own plan, those tests would at first "be developed in consultation with Frontier Labs."

Hassabis's practice, applied to Bacon: Bacon wrote a careful manual for the disciplined leap, and then designed a lab that publishes "as we think good," under oath. DeepMind withheld AlphaFold 3, voiced its worries in general terms, promised more, delivered it, and then widened the door twice. Judged by that practice, the House Bacon imagined isn't a model. It's a design that never got to run the experiment that would have changed its rules, because a lab on paper can't revise itself.

The payoff isn't a compromise, and it isn't tidy. Each man brings something the other's record lacks, and neither settles the hard part. Bacon brings the discipline: a marked guess, and one honest contradictory instance that has to be reckoned with. He never said who gets to bring that instance in; his own House kept the decision for itself. Hassabis brings the machinery: tables no clerk could hold, tests that can kill a guess, AlphaFold's free database, and a field that already runs blind trials like CASP. What neither man settles is who the House answers to. Our answer, and it's ours, derived from Bacon's method and turned against Bacon's House: an open table, a marked guess, and a test written by someone who isn't the guesser. Bacon's Merchants of Light sailed out under other nations' names to bring the world's light home. The better House sends it back out under its own.

The counterexample is a person. It should be allowed to be anyone.

Every source quoted or relied on above, in order of first appearance; then background reading for the terms of art. All accessed 31 October 2026.

Background: what the terms mean

All drawings: j.pickett, after Steinberg & Giraud · Ink and flat colour · © Jeremy Pickett · CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/). Each was drawn with Gemini; model and date per drawing.

• Cover, Beyond the Pillars: gemini-3-pro-image-preview, 30 October 2026.
• The First Vintage, The Shuttered House, Could It Invent Go?, Off the Board, The Open House, What It's For: gemini-3-pro-image-preview, 31 October 2026.
• Half the Answer: gemini-3.1-flash-lite-image, 31 October 2026.