--- title: "AGI as scientific discovery" publication: deslop.media byline: deslop.media editorial provenance: editorial_synthesis provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it." issue: 1 issue_title: The Scaling Bet grade: draft issue_version: 3 issue_version_date: 2026-10-11 piece_version: 1 first_published: 2026-10-11 reporting_current_through: 2026-10 canonical_url: https://deslop.media/feature/demis-hassabis/agi-as-scientific-discovery/ cite_as: "\"AGI as scientific discovery,\" deslop.media editorial, Issue 1 (Draft v1, October 11, 2026)" changes: https://deslop.media/issues/1-the-scaling-bet/changes/ flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags" --- When the call from Stockholm came on October 9, 2024, Demis Hassabis had spent the better part of three decades organizing his life around a single question: whether intelligence was the kind of thing that could be reverse-engineered well enough to do science with. He answered the Nobel committee's telephone interview from a London office, and two months later, on December 8, 2024, he stepped onto the Aula Magna stage at Stockholm University and gave the prize lecture under the title *Accelerating scientific discovery with AI*. A year later, on December 16, 2025, on the season finale of the Google DeepMind podcast, he condensed the program into a sentence for its host, the mathematician Hannah Fry: *"In the past, we trained answerers; now, we need to train researchers."* ## The definition before the products A model is released; benchmarks are reported; the question of whether the model is, or is approaching, AGI is adjudicated against the leaderboard. Hassabis's framing inverts the order. His personal AGI bar, repeated across the Lex Fridman interview of July 23, 2025 and a sequence of venues since, is scientific creativity — the generation of novel hypotheses and conjectures, not the solving of existing ones. In April 2025, on the *TIME100* cover, he gave the externalized version: a system that could not only solve existing problems but come up with new explanations for the universe. What he means by a research closed loop became explicit on the December 16, 2025 podcast with Fry. Two prerequisites, joined: *"world models,"* which capture intuitive physics and the spatial dynamics of the things being studied; and automated experimentation, which closes the perception–action loop by acting in the physical world and observing the consequences. AGI, on this account, is the joining. The system that knows how to imagine an experiment and the system that knows how to run one are the same system. The object is more demanding than the chat assistant, and less anthropomorphic. Hassabis's framing does not require a system that converses like a person; it requires a system that proposes and verifies in the way a working laboratory proposes and verifies. *"In the past, we trained answerers; now, we need to train researchers"* is the operational definition of the destination. ## Twenty-one months of artifacts DeepMind shipped a sequence of artifacts on the same line. AlphaFold 3 arrived in *Nature* on May 8, 2024, a diffusion-based architecture that pushed the AlphaFold lineage past proteins to the nucleic acids, ligands, and ions that pharmacology actually cares about. AlphaProof and AlphaGeometry 2 reached silver-medal standard on the International Mathematical Olympiad on July 25, 2024, and the full AlphaProof paper appeared online in *Nature* on November 12, 2025, in the issue dated March 19, 2026. The Nobel followed in October 2024; the Stockholm lecture in December. AlphaEvolve, the LLM-driven evolutionary search system disclosed in May 2025, produced a faster matrix-multiplication algorithm and improved best-known solutions on roughly a fifth of fifty open mathematics problems. Genie 3, in August 2025, rendered playable *"world models"* for a few minutes at 720p; DeepMind framed it, in its own blog, as a stepping stone because it gives agents the curriculum that a closed feedback loop needs. On December 10, 2025 Google DeepMind announced, as part of its UK government partnership, an automated materials laboratory for 2026, with researchers overseeing the research and directing its robotics. On February 11, 2026, Google DeepMind reported that the January 2026 version of Gemini Deep Think scored up to 90 percent on IMO-ProofBench Advanced, an Olympiad-level proof benchmark, as inference-time compute scaled, and about 38 percent on the PhD-level exercises of FutureMath Basic; the same report introduced Aletheia, a math research agent built on Deep Think. And on February 10, 2026, Isomorphic Labs announced its drug-design engine, IsoDDE, in a technical report claiming more than double AlphaFold 3's accuracy on the hardest protein-ligand generalization cases, with the model and weights kept proprietary; *Nature*'s news coverage on February 19 ran under the headline *['An AlphaFold 4' — scientists marvel at DeepMind drug spin-off's exclusive new AI](https://www.nature.com/articles/d41586-026-00365-7)*. Each artifact is a piece of the closed loop, and the loop is what is being assembled in public. ## Mathematics, where the closed loop closes tightest If the research-closed-loop framing is paying out anywhere, it is in mathematics. A proof either holds in Lean — the dependently typed formal proof assistant Microsoft Research developed and the mathematical community has adopted for machine-checkable proofs — or it does not, an Olympiad answer is either right or wrong, and the verifier is the same shape as the experiment. Mathematics is the science where the closed loop closes most tightly, which is why the nineteen months from July 2024 to February 2026 are the cleanest empirical record the program has produced. In July 2024, AlphaProof and AlphaGeometry 2 scored 28 of 42 — silver — on the IMO problems, with human formalization of the natural-language statements and roughly two to three days of compute. A year later, on July 21, 2025, Gemini Deep Think reached 35 of 42, end-to-end in natural language, within the four-and-a-half-hour competition window: gold-medal standard, certified by IMO President Gregor Dolinar at the same cutoff that gave 72 of 630 human contestants their gold medals. In February 2026, DeepMind reported that the January 2026 version of Deep Think reached up to 90 percent on the Olympiad-level IMO-ProofBench Advanced as inference-time compute scaled, and about 38 percent on the PhD-level exercises of FutureMath Basic. None of this resolves the harder question of whether a system can generate a genuinely novel conjecture, which is the bar Hassabis has set himself. It does mean that on the regime where the closed loop is cleanest, the curve is rising at the speed the framing predicts. ## The commercial counter The AlphaFold lineage has carried, since DeepMind's 2014 acquisition by Google, a structural tension between the science-first framing and the commercial obligations of the parent company. AlphaFold 2 and AlphaFold 3 were released free for the academic and non-commercial world, with the AlphaFold Protein Structure Database now holding over 200 million structures (per Hassabis's December 2024 Nobel lecture). That posture does not scale as a long-term commercial position, and Hassabis has never claimed it would. Isomorphic Labs, the DeepMind spin-out he chairs, is the proprietary line. The February 10, 2026 announcement of IsoDDE, kept proprietary, is the cleanest dated marker of where the science-first framing meets the revenue model. The sharpest version of the critique does not say Hassabis is insincere. It says the destination he names is the same destination Sam Altman and Dario Amodei, Anthropic's chief executive, are racing toward; what differs is who pays for the trip and who keeps the keys. ## Three leaders, three emphases Putting the three frontier-lab leaders side by side is the way to see why this matters. Amodei's *Machines of Loving Grace*, the *Adolescence of Technology* update of January 2026, and the February 13, 2026 Dwarkesh interview describe an AGI that is *"more capable than everyone in only 1–2 years,"* a destination he calls a *"compressed 21st century,"* and a near-term displacement of 50 percent of entry-level white-collar jobs over one to five years. *Machines of Loving Grace* also describes AI scientists that design and run experiments, so the destination is closer to Hassabis's than the products suggest; the difference, on this publication's reading, is that Anthropic sells the general-purpose model first and expects the science to follow. Altman's *Reflections* essay of January 2025 names confidence in the recipe — *"We are now confident we know how to build AGI as we have traditionally understood it"* — and frames the destination as superintelligence. OpenAI, too, leads with the product. Hassabis differs in emphasis. He defines AGI by the full range of the brain's cognitive capabilities, and the bar he sets himself is the research closed loop with novel-hypothesis generation; the named timing moved from the *"5 to 10 years"* and *"50% chance by 2030"* first stated on Lex Fridman in July 2025 to *"probably only a few short years away"* in his July 14, 2026 essay; the destination is scientific discovery with the commercial line carried by Isomorphic Labs as a separate vehicle; and the framing of the products — Gemini, Project Astra, Antigravity — is as the funding base and the practice ground for the research engine, not as the engine itself. Hassabis's emphasis differs from the other two more than his destination does. [Amodei's compressed 21st century](/feature/dario-amodei/compressed-21st-century) points at a similar capability level — Nobel-grade science within a year or two — but Anthropic's business, on this reading, treats the system as a paying product first and a science engine second. Hassabis points at the same capability level and treats the science engine as the point, with the products as scaffolding. A skeptic can fairly ask whether the distinction will survive contact with quarterly earnings. Google CEO Sundar Pichai confirmed the standalone Gemini app at over 650 million active users on the October 29, 2025 earnings call, rising to 750 million by early February 2026; on August 11, 2026, Google said it had passed one billion monthly users. The application layer is what funds the research line, and the application layer is the place where the destination Hassabis names is hardest to keep distinct from the destinations the other two name. ## What would test the claim The first test is the lab's own 2026 milestone. If an end-to-end automated experimental cycle ships in a real domain — a wet-lab biology loop, an autonomous materials-screening pipeline, anything that runs *propose, design, execute, observe, revise* without a human in the inner loop — the research-closed-loop framing has corroborated itself empirically. If the milestone slips, the framing was rhetoric ahead of the engineering, and the destination remains aspirational in the same way that Altman's *"superintelligence"* and Amodei's *"country of geniuses"* are aspirational. The Isomorphic Labs clinical readouts are the second test: whether the proprietary IsoDDE line produces decisions that pay out beyond the in-house pipeline. And the third, the hardest test, the one Hassabis set himself, is whether Gemini Deep Think for research mathematics, or its successor, generates a novel conjecture that the field accepts as real, distinct from solving the existing problem set. The claim has been sharpened since the Nobel lecture, with the December 2025 closed-loop articulation and the UK government partnership's plan for an automated DeepMind materials laboratory in 2026, and it has been put to work in the AlphaFold–AlphaProof–AlphaEvolve–Aletheia line. In September 2026 it also gained an institution: Hassabis became one of three directors, with Shane Legg and James Manyika, of the new DeepMind Institute, which studies AGI's implications for society. Half of the 2024 chemistry prize, shared by Hassabis and his DeepMind colleague John Jumper, rewarded AlphaFold, a system that solved an existing problem; the other half went to David Baker, the University of Washington biochemist, for protein design. Hassabis's December 2024 question to himself was whether the system that can solve existing problems will also come up with new explanations for the universe. He has not answered it. He has said, repeatedly, that he believes the answer is yes, that the means is the research closed loop, and that the automated laboratory planned for 2026 is the place where the question begins to settle.