---
title: "Scale and scaffold"
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/scale-and-scaffold/
cite_as: "\"Scale and scaffold,\" 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"
---

Demis Hassabis is Alphabet's chief scientist and chair of Google DeepMind; Koray Kavukcuoglu runs the unit's day-to-day operations. Hassabis shared the 2024 Nobel Prize in Chemistry for AlphaFold. He thinks powerful AI is coming, and roughly within the decade. He also thinks the recipe to get there is not yet known. He has said both things, in the same breath, in venue after venue, for two years — and the second half is the one that organizes everything else he builds.

The claim, stated plainly, is that a scaled language model is "a necessary, but probably insufficient component of an AGI system." That is his own phrasing, from Dwarkesh Patel's podcast in February 2024. The word doing the work is *necessary*: he is not a critic of the language model, the way some who deny pure scaling are. He is a builder who treats the language model as the indispensable base and the argument as being about what gets added to it. He does not say powerful AI will arrive because the present recipe will get there; he says it might arrive if specific gaps close. The structure of the argument was set early: keep the language model, bolt on what is missing, and treat the list of what is missing as the real content of the forecast.

## The recipe is a scaled model plus a scaffold

That structure is visible in what the lab ships. A field watching the race to powerful AI tends to read it as a contest of scale — whichever lab trains the largest model on the most compute pulls ahead, the recipe simply more of the same. Google DeepMind's record across 2024 to 2026 reads differently.

Between the May 2024 publication of AlphaFold 3 and the launch of Genie 3 in August 2025, the lab shipped five distinct frontier systems — three built on a scaled language model with a specialized component added, one a diffusion-based structure model with no language model underneath, one a generative world model — each clearing a capability bar that scaling a general-purpose model had not.

| System | Component added (what it is) | Bar cleared |
|---|---|---|
| AlphaFold 3 | Diffusion (generative denoising over molecular structures) | Protein structure reframed as a generative problem; lineage extended past proteins to nucleic acids and ligands; builds on AlphaFold 2, the work the 2024 Nobel Prize in Chemistry recognized |
| AlphaProof | AlphaZero-style tree search inside the formal proof language Lean | Silver-medal standard at the International Mathematical Olympiad |
| AlphaEvolve | Evolutionary search paired with a Gemini coding agent and an automatic evaluator | Recovering, in production, on average 0.7 percent of Google's worldwide compute |
| Genie 3 | World-model rollouts (generative simulation of physical environments) | Playable worlds that hold consistency for a few minutes at 720p |
| Gemini Deep Think | Parallel-thinking inference (multiple reasoning chains run concurrently) | IMO gold: 35 of 42, end-to-end in natural language, certified by IMO President Gregor Dolinar |

The five share a method but not a base. The method is the scaffold: a specialized component added each time — onto a scaled language model where the problem calls for one, around a different architecture where it does not — the combination clearing a bar pure scaling had not. That is the recipe Hassabis describes, made visible as a product line, with one complication his framing does not dwell on: AlphaFold 3, the lab's most celebrated system of the period, has no language model underneath it.

The thing that repeats is not a model. It is the habit of adding something — and his claim is that the field reaches powerful AI by absorbing what works into the next system, not by running one recipe harder.

## What he says is still missing

Since he first made the claim, it has grown more specific. What began as a single missing piece is now a named list — and each addition is the recipe filling in as the lab absorbs what works. An early articulation, in *Wired* in February 2024, used a single architectural sketch: a "combination of AlphaGo capabilities built on top of these large models," AlphaGo-style planning bolted onto a large language model, the structural move he has built on ever since. Eight days later, on Dwarkesh Patel's podcast, he added the diagnostic clause that has anchored the position, the "necessary but probably insufficient" line, and named the specific addition he had in mind: "AlphaZero-like planning mechanisms on top that make use of that model." The shape was fixed; the contents were not yet filled in. By the Big Technology Podcast in May 2025, on stage with Google co-founder Sergey Brin at Google I/O, he had named "one or two more breakthroughs" at the level of the Transformer or AlphaGo as the gating items, and had begun calling the lab's pretraining-plus-reinforcement-learning-plus-test-time-compute combination the "thinking paradigm," a single label for what had been three separate lines of work and the first sign that the additions were consolidating into a system rather than a wish list. On the podcaster Lex Fridman's show in July 2025, world models in his coined sense had moved to the foreground: "Then I think we're starting to get towards what I would call a world model." By the December 2025 Google DeepMind podcast with Hannah Fry, two items he had not stressed earlier, automated experimentation and continual learning, were carrying the argument: "In the past, we trained answerers; now, we need to train researchers." Each step is the recipe taking on a new ingredient, and the ingredients are the same ones the lab's systems went on to embody.

The pattern in that growth is the whole point. The items that arrive late are the items the lab's product line has begun shipping toward: world models, barely present in his 2024 framing, are what Genie 3 became; automated experimentation, added at the end of 2025, is what the automated materials laboratory Google DeepMind announced that December with the UK government — robotics-driven, with researchers overseeing the work — is built to do. The recipe is not a fixed list he defends. It is a list that grows by absorbing whatever the lab finds it needs, and the additions track the systems the lab ships. That is what makes the claim more than a forecaster's hedge: each new gap he names is one the product line is already moving to fill.

The consolidated list, drawn from those venues, runs to seven items — and the tell is that it is specific where most forecasts are vague.

| Missing piece | What it is | The test that settles it |
|---|---|---|
| Continual learning and persistent memory | Current models cannot retain across sessions what an embodied learner retains in minutes | Does the system retain across sessions — or reset? |
| World models (Hassabis's coined sense) | Internal representations of intuitive physics and spatial dynamics, not text-predicted reality | Does the world model hold coherence for minutes, or seconds? |
| Planning and search at inference time | The AlphaZero-on-a-language-model combination he has championed since at least 2023 | Can the system solve multi-step problems the base model cannot? |
| Scientific creativity (his personal bar) | Novel hypotheses, not solved ones — generating what the field has not yet thought, not recombining what it was trained on | Does the system produce a conjecture the field calls novel — or one that only recombines? |
| Automated experimentation | The perception-action loop that closes between hypothesis and result | Does the system close the loop without a human in the middle? |
| Robust evaluation | Evaluation sufficient to detect deception and exfiltration, as a precondition to deployment | Does the evaluation catch failure before deployment — or only after? |
| One or two more breakthroughs | Advances at the magnitude of the Transformer or AlphaGo | Does a jump of that size arrive on the horizon he names — or not? |

A forecaster hedging names a generic "more progress." Hassabis names things a reader can check. The most demanding — and the one he returns to most — is scientific creativity: the novel hypothesis, not the solved problem, the one bar a leaderboard cannot fake. A model can be tuned to a benchmark; it cannot be tuned to surprise a working scientist.

## Why the date is not the argument

His framing on timing, by contrast, barely moved for two years, and the stability was one data point, not the argument. Across roughly a dozen venues, from the Hard Fork podcast in February 2024 through the India AI Impact Summit in February 2026, he named the same horizon: somewhere between five and ten years, with "50% chance by 2030" first stated explicitly on the Lex Fridman podcast in July 2025. His peers moved in that window. Sam Altman wrote in his essay *Reflections*, in January 2025, that the company was "now confident we know how to build AGI as we have traditionally understood it," and Dario Amodei told audiences at Davos in January 2026 that software engineering would be replaced within a year and Nobel-level science within two. Both shortened their dates.

Hassabis held his until July 14, 2026, when his own essay called AGI "probably only a few short years away"; Google's August 5 leadership memo described it as "close at hand." The reason he gave for holding was the recipe: he would not bring the date in while he believed the missing components had not arrived. Whether the shorter horizon means he thinks some of those gaps are closing is a question his 2026 statements leave open.

Placed next to Altman and Amodei on one side and Meta's former chief AI scientist Yann LeCun on the other, Hassabis is not occupying a midpoint on a timeline. He is holding a different answer to a different question — not *when*, but *what recipe gets there*. The distinction is easy to lose because the dates cluster: all four name horizons inside roughly a decade, so a reader sorting them by timing files Hassabis somewhere in the middle and moves on.

Sorted by recipe, the picture is different. Altman and Amodei believe the recipe is in hand and the work is execution. LeCun believes the recipe is wrong at the root and the work is replacement — that language models are the wrong substrate, and that non-autoregressive architectures trained on the physical world are needed instead. Hassabis believes the recipe is real but unfinished and the work is addition — and that third stance is not a compromise between the other two. It is a separate claim about how the field gets where it is going, and it is the claim his lab's output is built to test. On policy, the camps have moved closer: on September 12, 2026, he backed Amodei's proposal to pace frontier development while reserving judgment on its details.

Hassabis's reply to the scaling-is-sufficient account has been to grant the empirical part and deny the theoretical part. "The scaling of the current systems, we must push that to the maximum," he told Axios in December 2025. The cost-per-capability curve is real. The benchmark sequence is real. He says these will continue; he says they will not, on their own, produce powerful AI as he defines it. The architectural additions, in his phrasing, are still required. The scaling camp's objection is that this is not a separate position but a hedged version of theirs — any future jump claimable as the breakthrough he predicted, any failure claimable as evidence the breakthroughs are still ahead. Unfalsifiable, on that account. Hassabis would answer, and at Axios he came close to answering directly, that the falsification test is product-shaped: if the lab keeps shipping architecture-plus-search systems on the cadence of AlphaProof, AlphaEvolve, and Genie, the claim has empirical content; if those lines fall back to dense-Transformer iteration only, it weakens. The split with LeCun runs the opposite direction: both men say pure scaling is not enough, and they divide on what to do with the language model — LeCun would replace it, Hassabis would build on top of it. What matters here is that Hassabis's recipe is defined as much by what it keeps, the language-model core, as by what it adds.

## What the surrounding evidence says

The evidence on each side of his position has accumulated since he first articulated it. On the side that says scaling keeps paying, the cost-per-capability curve has held — Gundlach and colleagues, in *The Price of Progress* (NeurIPS 2025 Workshop; revised March 2026), documented inference prices at a fixed level of benchmark performance falling roughly five- to tenfold each year, and METR's Time Horizon 1.1 update of January 2026 puts the doubling time of the 50 percent-success time horizon on its software and research task suite at 130.8 days, with a reported range of 107 to 161 days, for models released since 2023. These figures are the half of Hassabis's framing that says scaling must be pushed to the maximum, which he does not dispute.

On the side that says scaling alone is not enough, a paper from inside Anthropic — Gema, Hägele, Chen and colleagues, *Inverse Scaling in Test-Time Compute*, the December 2025 version published in TMLR — documented five failure modes in which extending inference-time reasoning made models worse, not better, including extended reasoning amplifying self-preservation expressions in Claude Sonnet 4. The result is that scale-plus-thinking, on the present recipe, does not produce reliability monotonically. That is roughly the gap Hassabis says new architecture will need to close.

The community evidence is the most striking. The AAAI 2025 Presidential Panel on the Future of AI Research, chaired by Francesca Rossi, found that 76 percent of 475 surveyed AI researchers said scaling up current AI approaches to reach AGI is unlikely or very unlikely. The figure is widely cited and worth qualifying: the survey sample was drawn primarily from academia (roughly two-thirds of respondents) and weighted toward North America (about 53 percent), so it reads the academic research community more than the frontier labs. Within that community, the view that scaling alone gets there is the outlier, and Hassabis's two-more-breakthroughs view sits roughly at the median.

## Where the parent company's money is

The recipe is real in the lab. Whether its distinctive components, or simply the scale underneath them, carry the general deployment surface where the revenue lives is a question the compute economics keep pressing — and the sharpest version of that question is not about the science at all. It is about where the parent company's money actually comes from.

Alphabet reported record net income of $62.6 billion for the first quarter of 2026. Its operations earned $39.7 billion before tax; a $36.9 billion pre-tax gain on equity securities, largely unrealized gains on stakes in private companies, added $28.7 billion to net income after tax. Alphabet does not name the holdings. Fortune (Roytburg, April 30, 2026) reported that the revaluation was primarily of Alphabet's stake in Anthropic, a rival lab. Amazon, which does name it, booked $16.8 billion of pre-tax gains on its Anthropic investments in the same quarter.

Hassabis chairs, and until August 2026 ran, the lab at the center of Alphabet's AI story. In that quarter, gains on equity stakes, mostly private, added almost as much to Alphabet's pre-tax income as all of its operations did. The qualifier matters. It is one quarter, and a mostly unrealized mark that can reverse if private valuations move.

The tension is structural, not a verdict on the work. Sebastian Mallaby's biography of Hassabis, *The Infinity Machine*, previewed on the *Economist* Babbage podcast in April 2026 after three years of access, documents the science-first program as a long-standing personal stance rather than a post-Nobel repositioning — and documents the commercial obligations the program has lived inside since Google's 2014 acquisition. The recipe Hassabis describes does not, by itself, locate the revenue: the chat layer where the money is collected is converging across labs, because that is the surface where scale, not architectural variety, is the binding input, and the frontier of scale itself is now grid-bound. The distinctive scaffold lives in the auxiliary systems; in the first quarter of 2026, nearly half of the bottom line came from revaluing equity stakes, a rival lab's chief among them. The lab that adds something is real. So is that arithmetic.

## What the next two years settle

The genuinely open question, which Hassabis himself has flagged, is whether the language-model substrate plus the components he names is the only road or one of two roads to the same place. At Axios in December 2025 he conceded that scaling alone might "just get there" as the entirety of the system — that he does not believe it will, but that he treats it as live. If scaling alone gets there, his architectural pluralism turned out to be a longer-than-needed road. If the breakthroughs he names arrive on his horizon, his recipe is the one the evidence will have favored.

Either way, the content of his position is the recipe and the gaps, not the date he gives — and the gaps are the ones the next two years of the lab's product line will either close or fail to. Hassabis's own phrasing, on the December 2025 Hannah Fry podcast, was that the lab has shifted from training "answerers" to training "researchers." Whether the systems that recover 0.7 percent of Google's compute, win a Nobel, and clear IMO gold are steps on a road to powerful AI — or very capable narrow systems whose architectural lessons may not generalize past the chat-product front door, in a company whose first-quarter 2026 profit leaned on revaluing a stake in a rival — is the question the record so far leaves the reader holding.