A polemical counter-reading of the paper. Where 19 - Insights and Takeaways catalogued what worked, this lights the rest on fire.
1. The 10×/yr is laundered uncertainty. Three independently uncertain trends — hardware (1.5×), investment (2.5×), algo efficiency (3×) — multiplied together and called "conservative." Compounding errors does the opposite of conservatism. And "algorithmic efficiency" is measured against past benchmarks, which is partly circular: of course the cost to match AlexNet drops when nobody's optimizing for AlexNet anymore. This headline number (03 - Effective Compute — 10x per year) is doing structural work in every downstream argument, and it's a much softer claim than presented.
2. The continuum framing is sleight-of-hand. They invoke the Legg-Hutter score to make AGI/ASI/UAI a smooth continuum and dodge every definitional question (02 - The Intelligence Continuum). But Legg-Hutter is an average over all computable environments weighted by Kolmogorov complexity — incomputable, dependent on an arbitrary choice of universal Turing machine, and with zero empirical connection to "is this AI useful." They borrow the formalism's weight to launder an informal claim.
3. AIXI is window dressing. Treating AIXI as the "theoretical ceiling" (05 - Universal AI (AIXI), informally) is like treating perfect-chess minimax as the ceiling of chess AI: true, useless, ornamental. The bridge they want — "modern pretraining is amortized Solomonoff induction" — is hand-wave. They cite Catt 2023 / Grau-Moya 2024 / their own Genewein 2026 like these settle the claim. They don't. The bridge is "this could be argued" with a citation glow-up.
4. The multi-agent ASI argument is question-begging. "If individual models plateau at human-level, running millions in parallel is ASI" (09 - Pathway 4 — Multi-Agent Collectives). But human institutions of millions of coordinated experts already exist — are they ASI? The paper waves "AI collectives would coordinate better" without showing any reason that's true. Better than what? Better than the most coordinated humans, who are also stipulated to be the AGI baseline? The argument needs a multi-agent scaling law it explicitly says doesn't exist yet. They're cashing a check the paper itself confesses is unwritten.
5. The "research-gets-harder inverts" move cheats. Bloom et al. is about ideas getting genuinely harder to find — exponentially more researcher-years needed for equal output. The paper (12 - Bottleneck — Research Gets Harder) reduces this to "researcher count is the bottleneck, and AI makes researcher count elastic, so the friction reverses." But if ideas are intrinsically harder, then a million artificial researchers hit the same diminishing returns faster, not later. The conclusion only works if the bottleneck is people, not difficulty. The paper assumes its own conclusion.
6. The physical-limits section is a fig leaf. Speed of light, Landauer, Bekenstein, Bremermann (17 - What ASI Cannot Do) — none of these were ever in dispute. Nobody serious claimed ASI violates thermodynamics. Listing well-understood bounds creates an aura of formal rigor while admitting (in the very next paragraph) that the bounds are practically vacuous. It's an alibi: "we considered the limits."
7. The abstraction barrier kills three pathways, not one. Lerchner's hypothesis (13 - Bottleneck — Abstraction Barrier) is that models trained on human concepts can't discover novel concepts from raw data. If true, this caps:
The paper quarantines this critique inside the Pathway-1 section and pretends Pathway 4 is the escape hatch. It isn't. The barrier is structural and it applies everywhere. They don't notice — or they noticed and chose not to write it down.
8. The "AGI day is the wrong frame" reveal is a strawman victory. Who, specifically, was claiming AGI is a discrete event? Yudkowsky-style FOOM scenarios already model continuous takeoff curves. Bostrom already distinguished slow / fast / moderate. The continuous-cascade framing has been mainstream for at least a decade. The paper presents an unobjectionable reframe as a major insight (01 - The Big Picture).
9. The Hassabis Einstein quote is unfalsifiable theatre. "Could a 1900-vintage model derive GR?" There is no way to construct or evaluate this counterfactual. Using a CEO sound bite to ground a claim about a fundamental limit (15 - Is Superintelligence Super-Creative) is appeal-to-authority dressed as humility. Remove the quote and the abstraction-barrier section visibly thins out.
10. "Assume alignment is solved" is the foundation, not a caveat. The paper opens with this assumption (16 - Goals, Agency, Alignment) and proceeds for 50 pages as if it were a side note. It is the entire frame. Imagine From Fission to Fusion: Pathways opening with "we assume confinement is solved." Without alignment, the pathway analysis is moot (none of these systems get deployed in the assumed form). With alignment unsolved, the bottleneck taxonomy is academic. Pretending this is a minor scope choice while making it the precondition for every claim is intellectual cowardice.
11. The open-research-questions section is a recruitment pitch. A substantial fraction of cited "future work" is already-published DeepMind in-house papers (18 - Open Research Questions: Tomašev × 3, Chan, Morris × 2, Trivedi, Meulemans, Davidson). The paper presents as field-mapping; it's field-shaping. That's not a sin — DeepMind has earned a perspective — but the document hides its agenda behind survey rhetoric.
12. The conservatism is the tell. Every claim is hedged into evacuation: "cannot be ruled out", "might", "potentially", "possibly", "to some degree." A document that won't commit to a prediction isn't forecasting — it's posturing for plausible deniability in both directions. If they think ASI in two decades is plausible, write that. If they don't, say so. The hedging is a corporate-comms reflex, not epistemic humility.
13. The Borg comparison is the indictment. When your most concrete description of post-AGI social form is Star Trek's authoritarian hive-mind (09 - Pathway 4 — Multi-Agent Collectives), you have not thought hard enough about social form. That's not a research direction; it's an admission that the "what does an ASI society look like?" question has been entirely outsourced to mid-century science fiction. The most consequential question in the paper, and the answer is "Borg or markets, we'll get back to you."
14. The author list is a structural critique. Fourteen authors, almost all DeepMind, including two of the field's most cited theorists (Hutter, Legg), the head of multi-agent (Leibo), and the head of governance (Dafoe). This is a corporate position paper from the world's most heavily-funded AI lab, dressed as academic synthesis. They thank Tomašev and Krier for "review" — the same people whose papers are then cited as future work. The peer review is in-house. The agenda-setting is in-house. The "this is what the field should research next" is in-house. None of this is fake; it's all true and the work is real. But it is not the neutral landscape map it presents as. It is a 50-page argument that DeepMind's current research bets are the bets that will matter. Of course it concludes that.
The paper writes:
"Upward of 90% of this document are human authored with no direct involvement of a language model."
Read this again in context. They are publishing a report about how AI will transform cognitive labor. They feel the need to certify that AI did not write it. Either (a) they don't trust their own technology to write a position paper, or (b) they think the audience won't trust an AI-written report on AI's future. Either reading is devastating. The lab building toward ASI does not yet trust its tools to author the document about the tools' future.
That sentence is the actual epistemic state of frontier AI in June 2026, accidentally confessed in a footnote.