The DeepMind architectural argument

On July 21, 2025, an advanced version of Google DeepMind’s Gemini scored 35 of 42 at the International Mathematical Olympiad — gold-medal range, certified by IMO President Gregor Dolinar, achieved end-to-end in natural language without a human translator formalizing the problems first. Five of six problems were solved cleanly inside the 4.5-hour competition limit. The cutoff for human gold that year was 35; only 72 of 630 contestants cleared it. The result drew the obvious headlines about machines doing competition mathematics. The less-obvious story is what the system underneath was: a pretrained Gemini, with a parallel-thinking inference layer bolted on top. A scaled language model with a search procedure stapled to it. A familiar move from this lab, repeated.

Between the May 2024 publication of AlphaFold 3 and the launch of Genie 3 in August 2025, Google DeepMind shipped five distinct frontier systems, and they did not share one architecture. AlphaEvolve and Gemini Deep Think build directly on Gemini; AlphaProof pairs a pretrained language model with AlphaZero-style tree search inside Lean; AlphaFold 3 predicts molecular structures with a diffusion-based model that has no language model underneath; Genie 3 is a generative “world model.” Each cleared a capability bar that scaling a general-purpose model, on the public evidence, had not cleared at the time.

The capability arc moves through five well-defined steps, each adding something new. AlphaFold 3, with Demis Hassabis as senior author, replaced AlphaFold 2’s protein-only architecture with a diffusion-based system that handles nucleic acids, ligands, and ions — the structural prediction problem reframed as a generative one. The Nobel Prize in Chemistry that October went to Hassabis and his DeepMind colleague John Jumper for AlphaFold 2’s protein-structure prediction, shared with the University of Washington’s David Baker. Gary Marcus, the cognitive scientist and longtime LLM critic, called it “the first Nobel Prize for Neurosymbolic AI.” Marcus’s reading is partial; AlphaFold itself is end-to-end neural. The framing matters anyway because Marcus is not a friendly reader of frontier-lab claims. When a critic whose published position is that pure scaling will not reach AGI endorses the architectural-diversity line as the right reading of the recipe, the line has corroboration from outside Hassabis’s camp.

The Nobel-anchored AlphaFold line told a reader who did not already follow this lab a clean fact. Pure-scale Transformers had not solved protein folding. A specialized architecture, trained on the same hardware substrate the rest of the field uses, did.

AlphaProof and AlphaGeometry 2, announced in a DeepMind blog in July 2024, went a step further into territory Marcus’s framing fits cleanly. AlphaGeometry 2, described in full in a February 2025 arXiv paper, solves 84 percent of the 2000–2024 IMO geometry problems. AlphaProof, the more architecturally novel system and the subject of a full Nature paper published online in November 2025, runs an AlphaZero-style reinforcement-learning agent inside Lean — a formal mathematical proof language — seeded by a pretrained language model. The system reached IMO silver-medal standard at 28 of 42 in 2024, a year before the all-natural-language Gemini Deep Think result cleared gold. Two ways into the same problem from the same lab: one through formal language and explicit search; one through pretraining plus parallel reasoning at inference. Both worked. Different architectures, different specialized components, same underlying bet that the language model alone would not have got there.

AlphaEvolve, announced in May 2025, is the most economically vivid of the systems. It pairs a Gemini-driven coding agent with an evolutionary search loop and an automatic evaluator: the language model proposes a candidate solution, the search procedure selects among generations of candidates, the evaluator scores each one against the target metric, and the loop iterates. The published results include a new matrix-multiplication algorithm and improvements on best-known solutions to 20 percent of fifty open mathematics problems, the kind of headline that travels in computer-science circles and not much further. The number that matters more, and the one DeepMind put at the top of the launch post, is the production one. By the time DeepMind disclosed the system in May 2025, AlphaEvolve had been recovering, on average, 0.7 percent of Google’s worldwide compute resources, continuously and in production, for over a year.

The architectural-diversity record otherwise sits in a register a reader can dismiss as benchmark theater — competition mathematics, narrow scientific puzzles, things hyperscalers can subsidize as branding. AlphaEvolve does not sit in that register. It is a system whose architectural premise (language model plus evolutionary search plus automatic evaluator) was already, in May 2025, paying for itself at a scale only Google can measure. The 0.7 percent is not “0.7 percent of an academic compute cluster.” It is 0.7 percent of the worldwide compute of the company that runs YouTube, Search, Gmail, Maps, and the largest single AI training fleet in the world. The architectural move did not just clear a benchmark. It cleared a P&L review.

Genie 3, in August 2025, took the line into “world model” territory. The system retains “consistency for a few minutes at a resolution of 720p,” which sounds modest until one places it against what came before — Genie 2, in December 2024, held consistency for up to a minute, with most published examples lasting 10 to 20 seconds — and against what it enables, which is unlimited synthetic training environments for agents. DeepMind’s own framing in the launch post called Genie 3 “a key stepping stone on the path to AGI.” Hassabis on the December 2025 Hannah Fry podcast made the architectural claim explicit: “world models” plus automated experimentation, joined into a research closed loop, is the pair of prerequisites he reads as necessary for the next jump. Genie 3 is the “world model” half, shipped.

Gemini Deep Think’s IMO gold in July 2025 closes the arc the IMO result above opened. Same lab, same time window, a fifth specialized component — parallel-thinking search at inference time — bolted onto a scaled language model. End-to-end natural language. Gold-medal range. Five problems out of six. Seen after the other four, it reads less as a standalone announcement than as the predictable next move from a lab that has spent two years matching specialized machinery to the problem in front of it.

Hassabis’s framing of the recipe has been stable enough across 2024–2026 that a reader can pull it from any of half a dozen interviews and get the same answer. On Dwarkesh Patel in February 2024 he called large language models “a necessary, but probably insufficient component of an AGI system” and described “AlphaZero-like planning mechanisms on top that make use of that model.” On the Big Technology Podcast in May 2025, on stage with Sergey Brin, the Google co-founder, at Google I/O, he said the lab had “always been big believers in what we’re now calling the thinking paradigm” and that the path to AGI “may require one or two more new breakthroughs.” The lab does not run a single architecture into the wall. It runs a portfolio of architectures, some built on its base language model and some not, picks the combination the problem rewards, and ships. Scaling matters; scaling without the auxiliary components has not, in this lab’s hands, produced these results.

The skeptical reading deserves its strongest form. At its November 2025 launch, Gemini 3 Pro sat at the layer Gemini-app users touch — Google CEO Sundar Pichai confirmed the standalone-app count at over 650 million on the Q3 2025 earnings call (Oct 29, 2025), rising to 750 million per TechCrunch coverage of the Q4 2025 earnings (Feb 4, 2026), and past one billion monthly users by Google’s August 11, 2026 count. It was itself a mixture-of-experts Transformer, a design in which only a subset of parameters fires per token, with reasoning post-training. At the chat-window layer, it was not categorically different from OpenAI’s o-series or Anthropic’s Claude 4 family of the time. The architectural diversity that powers the bull case lives largely in the auxiliary Alpha systems. Those systems serve narrow tasks — protein structure, competition mathematics, “world model” rollouts, hyperscaler compute optimization — rather than the general-purpose deployment surface a billion users meet through a chat box.

By that reading, what DeepMind has published is impressive narrow AI. The architectural move that the bull case credits, specialized machinery matched to the problem, is the same move the field has been making since AlphaGo, executed at higher scale on more substrates. Whether the lessons from any of these systems transfer to the general deployment surface is the open question. Gemini 3’s Deep Think mode imports the parallel-thinking-on-top-of-LLM lesson from the IMO result. AlphaEvolve runs inside Google’s compute fleet, not inside a user-facing API. Genie 3’s “world model” substrate is, by DeepMind’s own framing, a stepping stone — meaning it is not yet the destination.

The counter-argument has empirical legs. Gary Marcus’s “first Nobel Prize for Neurosymbolic AI” framing reads the AlphaFold result as evidence that neural-symbolic combinations matter. Whether the broader Alpha line embodies that thesis depends on which of the five systems one looks at. AlphaProof is unambiguously a neural-symbolic combination. AlphaFold 3 is end-to-end neural with a diffusion head. AlphaEvolve is more cleanly described as LLM-plus-search-plus-evaluator than as neurosymbolic. Genie 3 is a generative “world model.” Gemini Deep Think is a parallel-thinking search procedure. The five systems do not embody one architectural thesis — they embody five different ones, and they do not share a base: AlphaFold 3 has no language model underneath it at all.

Five independent architectural moves all cleared bars that general-purpose scaling had not, which suggests the recipe is matching architecture to problem rather than any one specific component. The pure-scale camp at OpenAI and Anthropic, on the published record, has shipped fewer of these specialized auxiliary systems and converged faster on the general-purpose product surface. Noam Brown, OpenAI’s lead researcher on inference-time scaling, has restated the scaling thesis around major OpenAI releases — including the September 2024 o1 launch, the April 2025 GPT-4.1 launch, and the April 2026 GPT-5.x launch — though not across every release in between. Dario Amodei, Anthropic’s chief executive, in his February 2026 Dwarkesh Patel interview, said his basic model of scaling had not changed since 2017 — same seven factors, same bet. Neither lab has published an Alpha-line equivalent at DeepMind’s cadence. The deployment surfaces converge. The published architecture record does not.

The publication asymmetry is itself part of what the record shows. Both interpretations are available. One: DeepMind publishes more architectural detail because it has more architectural variety to publish, and the variety is the recipe. Two: DeepMind publishes more architectural detail because Google DeepMind operates on a longer scientific-publication cadence than its peers, and the variety is presentational. Hassabis at Davos in January 2026 framed the asymmetry his way: no single competitor, on his reading, is pursuing the “world models” and agent and science line at the same breadth. Independent verification is hard — OpenAI does not publish architecture details for its frontier models, and Anthropic’s disclosures on Claude 4 and Sonnet 4.5 are sparse. The reader is left to weigh the asymmetry on the available evidence.

The follow-up signals to watch are dated and concrete. The April 2026 DeepMind blog announcing Decoupled DiLoCo reported training a 12-billion-parameter model across four US regions over a 2–5 Gbps wide-area network “more than 20 times faster than conventional synchronization methods” — distributed training across geographically separated data centers, which is one of the architecture-plus-scale recipe components the lab has been quietly betting on. The bet matters because the next jump in scaled language models is grid-bound: cross-region distributed training is one of the few paths to keep parameter counts climbing without single-site power and cooling becoming the binding constraint.

Isomorphic Labs’ IsoDDE, announced in a technical report on February 10, 2026, kept proprietary and covered in Nature’s news pages on February 19 under the headline “‘An AlphaFold 4’ — scientists marvel at DeepMind drug spin-off’s exclusive new AI,” is the commercial continuation of the AlphaFold lineage. Whether the architecture-plus-scale line keeps shipping at its recent cadence — five major systems in the 15 months from May 2024 to August 2025 — through 2027 is the test. If the cadence holds, the reading that DeepMind works the problem-specific quadrant more broadly than its rivals, whose entries there include OpenAI’s o-series and Sora and Meta’s V-JEPA 2, is the reading the empirical record supports. If it slows in favor of pure Gemini scaling, the reading weakens, and the chat-window-convergence skeptic gains ground.

Hassabis’s own phrasing, on the December 2025 Hannah Fry podcast, was that the lab has shifted from training “answerers” to training “researchers.” The Alpha-line is, on his telling, what training researchers looks like in 2024–2026: “world models” combined with automated experimentation, joined into a research closed loop.