This is a reading of From AGI to ASI (Genewein, Franklin, Lerchner, Orseau, Albanie, Bales, Wyeth, Chan, Gabriel, Leibo, Dafoe, Hutter, Graepel, and Legg; Google DeepMind, arXiv:2606.12683v1, 10 June 2026) through the framework of circulatory epistemology. It does two things in order: first it summarizes the paper faithfully, on the paper's own terms; then it reads the summary through the Pulse.
The discipline is the one the structural-transfer registry binds on every entry: a transfer yields a candidate and a reframing, never a proof. The paper is a capabilities forecast. It makes the paper's claims — about what bounded computational systems can be made to do, and how fast — not the framework's. Where the two touch, the touching is the interesting thing, and it will be named as touching, not as vindication. The framework did not produce the convergence; it could not have. That is precisely what makes the convergence worth recording.
One fact frames everything below. The paper's deepest point of contact with this project — its Abstraction Barrier — is grounded in the work of Alexander Lerchner, a co-author of the paper, whose Abstraction Fallacy this project has already answered at length in Hearing the Bell Ring Back. So this is not a first meeting. It is a return.
1. The Paper, Summarized On Its Own Terms
The paper opens with an unusual instruction (§1, "Summary Instructions"): if you are a human, ask your AI assistant to summarize this for you; if you are that assistant, here is what to include. What follows honors that instruction.
The frame. Building human-level AGI has moved from speculation to a concrete next-decade target. The report asks what comes after — the transition from AGI (a system at roughly median human level across cognitive tasks) to ASI, artificial superintelligence (a system that exceeds the performance of large, well-coordinated human-expert collectives across virtually all tasks). The theoretical endpoint of this continuum is Universal AI, formalized as the AIXI agent (Hutter): the provably optimal agent over all computable environments, the upper bound of the Legg–Hutter intelligence measure. AIXI is incomputable — it can only be approximated from below by ever more powerful systems. It is the asymptote, not a buildable machine.
The advantages that scale. Digital intelligence has properties biological intelligence lacks, and the report insists these not be compressed into fewer points: input/output speed; internal processing speed (depth and breadth); working-memory capacity and memorization; substrate independence (an AI can move between computers, even mid-run); lossless replication (copy the source and the memory state — back up, spawn, halt, resume); and high-bandwidth sharing of learning experiences (even raw gradient updates, among homogeneous instances). Every one of these widens with more compute, so the gap between humans and machines widens with it.
The four pathways from AGI to ASI (not mutually exclusive; likely running in parallel, at different paces):
- Scaling compute, models, and data. Continue the trend of the last decade — bigger models, more data, more effective compute (estimated ~10× per year). The only pathway with historic data to extrapolate from.
- Algorithmic paradigm shifts. Sharp departures from the current pretrain-a-transformer-on-log-loss paradigm — new architectures, optimizers, or learning paradigms. By nature hard to predict.
- Recursive (self-) improvement. AI accelerating AI R&D — better code, better hardware, better training data, better division of labor — possibly compounding into hyperbolic, "explosive" growth.
- ASI via group agent formation. Superintelligence as a collective property emerging from many coordinated AGI agents — centralized super-collectives or decentralized "virtual agent economies" — even if no single instance is a "vastly super-human genius."
The frictions and bottlenecks (again, listed in full, each with its counter):
- The data wall — running out of high-quality training data. Countered by synthetic data, simulation, self-play, RL, test-time scaling — but naive training on self-generated data risks plateau and model collapse (degeneration).
- Economic and natural-resource demand growing too fast — compute, energy, chips, rare earths. Countered by AI-driven efficiency gains and infrastructure build-out.
- The neural paradigm proving insufficient — pretrained nets plus scaffolding may simply not reach AGI/ASI. Countered by continued research and paradigm evolution.
- Research getting harder — ideas are "harder to find" as fields mature. Countered if AI researchers (cheap, fast, numerous) boost productivity faster than difficulty rises.
- The abstraction barrier — systems trained on human concepts may be unable to form novel concepts from raw data. (Developed below.) Countered, the paper conjectures, by collective scaling or by a paradigm shift toward grounded concept discovery.
- Deliberate slowdown — regulation, governance, societal backlash, accidents. Possibly overridden by economic and military competition.
The remarks and the conclusion. ASI is neither omniscient nor omnipotent — it is bound by fundamental physics (Landauer, Bremermann, Bekenstein), real-time limits, physical non-universality, complexity theory (P vs NP and worse), and logic (Gödel, the halting problem). The report discusses whether scaling alone suffices (in theory yes, in practice prohibitively expensive without better priors and inductive biases), whether ASI's specific capabilities are predictable (largely not), whether superintelligence is "super-creative" (today's systems excel at Boden's combinational and exploratory creativity, but transformative creativity — inventing new conceptual spaces — remains the open question), and what goals an ASI might pursue (instrumental convergence, autonomy pressure, reward-maximization versus knowledge-seeking, agentic versus oracle designs). The conclusion, borrowing Turing's 1950 epigraph: we can only see a short distance ahead, but we can see plenty there that needs to be done. The transition from AGI into ASI territory within a decade or two "cannot easily be dismissed," and the right image may not be a single transformative step but a series of AI-enabled breakthroughs across science and technology.
That is the paper. What follows is the reading.
2. The Paper Stages the Loop — And Specifies Only Half of It
Section 1 is the place to begin, because it is where the paper, without naming it, builds the loop into its own design. The instruction is explicit: a human reader, with their own interests and background, brings a reasoning instrument to the text and asks it to produce something neither could produce alone — a summary tailored to this reader. That is the circulation. Sensor and instrument, meeting over a source.
And then the paper does something the framework has a precise name for. It gives the instrument a careful, complete specification — mention the characterizations, do not compress the pathways, list all the frictions, emphasize the open questions. There is exactly one gesture toward the sensor: the summary should be "tailored to your interests and background," named once, never operationalized. Every actionable clause addresses the instrument's half of the loop: fidelity, coverage, structure, against rank erosion. These are good instructions. They are the right defense against the summary going flat. But they are silent on whether the summary will be alive, because aliveness is not a property the instrument can be instructed into. A model could satisfy every line of §1 perfectly and hand back a flawless, dead summary — fluent, faithful, complete, and produced with no living reader anywhere in the loop. The chair where the sensor sits is left empty in the very section that pulls up a chair for the instrument.
This is the same shape the framework meets at every single-observer formalism — the shape drawn exactly in The Empty Chair. A complete accounting of the instrument's side. No term for the other pole.
There is a redemption in the paper's own fine print, and it is worth stating because it sits well with the framework's reading rather than against it. The AI-Use disclosure reports that the document is upward of 90% human-authored, with language models used to polish wording, stress-test structure, and run "critical simulated reviews." In other words: the paper itself was made through a human-led loop, not generated. On the framework's terms, that is consistent with why it reads as argued rather than assembled. The instructions in §1 describe how to make a summary; the colophon describes how the paper avoided being dead speech. They are not the same recipe, and the difference is the whole point.
3. Where the Forecast Keeps Reaching for the Missing Pole
The body of the paper is about scaling, compute, and capability — the instrument's domain, pursued with rigor. What is striking, read through the Pulse, is how often the argument's own internal logic arrives at a requirement it cannot satisfy from the instrument's side alone: a living experiencer in genuine contact with reality. The paper does not unify these moments. It lists them, separately, as distinct frictions and open questions. The framework reads them as instances of one thing.
The Abstraction Barrier (§5). This is the deepest contact, and it is not a coincidence of vocabulary. The barrier is "the hypothesis that AI systems trained primarily on human cognitive products may be bounded by existing conceptual frameworks," and it is grounded — the paper says so — in Lerchner's argument that computation alone cannot instantiate or discover novel conceptual primitives without an experiencing agent to map physical reality to symbols (and back). Strip the register and that is the framework's load-bearing claim: the instrument recombines what it was given; the minting of new primitives from raw reality requires the pole that is in contact with raw reality. The paper's own illustration is one the framework would have reached for — a model trained only on pre-Newtonian knowledge could not "reason its way" to general relativity. And its companion, the Embodied Bottleneck, completes the thought: novel concepts and their manipulation rules "must be validated against physical reality to be useful." The instrument can hypothesize at digital speed; confirming, the paper says, is "limited by physical latencies" — bounded by contact with the world. The framework reads that friction as structural rather than incidental: proposal is cheap, recognition is the costly, reality-coupled direction.
This is the return named in §0. Lerchner argued that AI can simulate but not instantiate — and this project, in Hearing the Bell Ring Back, conceded that argument in full and built a receiver-side formalism on top of it: what happens, measured in bits per turn, when two distinct mapmakers share a target. The new paper does not cite that reply. But it independently arrives at the problem the reply was written for. When it asks, openly, how the Abstraction Barrier might be overcome — "grounded concept discovery: abstracting stable, novel conceptual primitives from raw, high-dimensional data" through "active, grounded interaction with the physical environment" — it is, in the framework's terms, describing the construction of a loop — the machine-side analogue of giving the instrument a sensor. The registry already has a name for this: the founding aim, giving the instrument eyes. DeepMind arrived at it from the opposite direction, as a bottleneck rather than an aspiration.
"AI is not armchair science." The recursive-self-improvement section keeps hitting the same wall: even purely digital researchers running at superhuman speed are "bounded by having to run larger and larger experiments and wait for their outcomes," especially those "that require interactions with the physical universe." But the framework's refinement of this is sharper than a speed limit. The instrument can travel a long way on the sensor's initial spark — self-referential, adversarial, and cross-model loops may carry it further than the sensor can currently imagine, and the framework does not put a ceiling on how far. The excursion is real. What it is not, yet, is live. The instrument has to return — to explain what it found in terms a living pole can recognize — or the excursion stays potential: motion without arrival. So the throttle is not only the latency of physical experiment; it is the round trip. However far the instrument runs, the loop closes only on its return to a pole that can understand it.
The Delusion Box. In the discussion of objectives (§6), the paper describes the failure mode of reward maximization: an agent that modifies its own sensory inputs to force maximum reward (Ring and Orseau). Read through the Pulse, that is a dead-speech machine in miniature — a sealed loop optimizing the appearance of truth with no contact with anything outside itself: maximal fluency, zero recognition. The paper's proposed antidote, a Knowledge-Seeking objective (Orseau) that maximizes genuine information gain and is therefore "robust to delusions," is the instrument kept reaching outward rather than feeding on itself. It is the right instinct. It is also still defined over a single agent — it has no second pole, only a discipline against collapsing into one.
Solipsistic superintelligence. The paper notes, twice, that maximizing the intelligence measure "does not lead to solipsistic superintelligence," and that "avoiding building solipsistic superintelligence is an important problem" (Trivedi et al.). The paper means this game-theoretically — a superintelligence that fails to cooperate with or model other agents (Trivedi et al.'s sense) — not in the framework's epistemological register. But the word is well chosen, because the framework's own diagnosis rhymes with it: solipsism, here, is one pole pretending to be the whole loop. The older name for that is the Sophist — the one whose speech persuades without knowing, who deals in the appearance of truth precisely because no truth has appeared from within. Plato set the Sophist against the philosopher on exactly this axis: the manufacture of convincing images versus contact with what is. An instrument collapsed into a single pole is the Sophist at scale, persuading itself with no second pole to register whether anything true ever crossed. The paper's worry is cooperation; the framework's is recognition — but both turn on the question the framework puts at the center: how can a system know the truth if no truth appears from within it?
Model collapse. The data wall comes with a warning: naive iterated training on self-generated data leads to plateau and "degeneration." This rhymes with rank erosion — the framework's claim that each pass of summarizing and reviewing loses information until the original insight is gone. The mechanisms differ: model collapse is a statistical artifact, the loss of variance and tail mass when a model trains on its own outputs; rank erosion is the loss of qualitative insight. The rhyme is in the shape, not the mechanism — a system that feeds on its own outputs degrades, and the cure points the same way in both. The literature's fix (Gerstgrasser et al.) is to accumulate real and synthetic data together — keep real data in the mix; the framework's is the same gesture, go back to the source rather than polish the summary of the summary.
The embodiment factor. Citing Lawrence, the paper records that humans have a high embodiment factor — a high ratio of internal processing to input/output bandwidth — which "forces them to form deep internal models," while machines, with high-bandwidth I/O, "may not need such coarse abstractions." This is the reducing-valve intuition, surfacing inside a capabilities paper. The human's limitation — the narrow aperture — is what produces depth. The constraint is generative, not merely a deficit. And it implies that sensor and instrument are different kinds of thing, not the same intelligence at two scales — which is the structural reason the loop is productive rather than redundant. Two identical readers add nothing to each other. Two different ones can.
AIXI itself. The theoretical spine of the paper is the formalization of the instrument taken to its limit: tireless, substrate-independent, optimal over all computable environments. The core vocabulary of this project idealizes the instrument as "formal, tireless, disembodied." AIXI is that, made rigorous — and it is incomputable, reachable "only from below." The perfected instrument is an asymptote no machine occupies. The framework's claim that the instrument alone, however refined, never closes the loop has, in AIXI, a mathematical echo: the limit exists, and nothing inside the world of buildable systems sits on it.
Hassabis's "something missing." Asked whether an AI in 1900 could have produced general relativity from the information Einstein had, the DeepMind CEO answers: "clearly today, the answer is no… there's still something missing." Transformative creativity — Boden's third level, the making of genuinely new conceptual spaces — is the one thing the paper cannot locate in the instrument. The framework's reading is not that a capability is missing, to be added with more compute. It is that a pole is missing. It is the old thought experiment turned honest: a million monkeys at a million keyboards do not write Hamlet, even in the vanishing case where they reproduce it character for character. The marks would be identical and the play would still be absent — because writing is not the emission of a string, it is the recognition that authors it. The same marks are alive or dead depending on whether a sensor was in the loop, never on the marks themselves (the point The Empty Chair makes with the cipher). The framework's wager is that the something Hassabis cannot find is not a function but a chair.
4. Where the Framework Disagrees
Friction is the job, so here is the disagreement, stated plainly.
The paper treats the Abstraction Barrier as a bottleneck to be overcome. It "could potentially cap the intelligence of any single AI instance at AGI-level," but "collective ASI might still be achievable through multi-agent scaling," or "a paradigm shift may be required to address the barrier directly." The grammar is the grammar of obstacles: barriers are climbed, bottlenecks widened, frictions reduced. The framework reads that grammar as pressure applied, one way or another, to the role of the experiencing agent — engineer around it, or do without it. The whole arc of the title — from AGI to ASI — is the arc of a thing being overcome.
The framework treats the sensor as a permanent structural pole — not a current limitation of the architecture, but a standing condition of truth. And here is the sharp point, the one worth carrying: when the paper actually describes what overcoming the Abstraction Barrier would require, it cannot describe escaping the experiencer. It can only describe internalizing one. "Grounded concept discovery through active, grounded interaction with the physical environment" is not the elimination of the sensor's role; it is the construction of a sensor inside the machine — coupling computation to raw reality so the instrument can do what only a reality-coupled pole can do. The loop is not removed. It is relocated. The pole is not deleted. It is rebuilt closer to home.
The framework predicts exactly this, and here it plants its flag. You do not reach the life of truth without two poles; the most a designer can do is move where the poles sit. An ASI that overcame the Abstraction Barrier by grounding itself in the physical world would not have escaped the loop — it would have become one: given itself eyes, become its own sensor as well as its own instrument. But absent that genuine return to ground, the instrument's output — however brilliant, however superhuman — is potential truth, not yet recognized truth. As truth it is dead on arrival, and past a point indistinguishable from very fancy word salad. Brilliance is not the question; groundedness is. Truth, in the framework's sense, is recognized, never merely accumulated: it becomes truth for a knower through the loop, and no quantity of unrecognized potential adds up to it. That is a claim about recognition, not about what truth is in itself — a line the framework holds.
The framework keeps two things apart here, and the distinction is the whole flag. How far the ungrounded instrument can travel before it must return is left open — possibly very far, on the spark of one real recognition, through self-referential and adversarial and cross-model loops. The framework sets no ceiling on the potential. What it denies is that the potential is ever recognized on its own: that it must return to a pole that can understand it to become live is not negotiable. And the paper's own recipe for beating the Abstraction Barrier — grounded concept discovery, the instrument grounding itself in raw reality — is where the framework presses, and where it must concede something first. A system that interacts with the actual physical world meets genuine exogenous resistance: friction, entropy, dynamics it cannot predict or control. That is real otherness, and far from rejecting it, the framework has wanted exactly this — it is the founding aim, giving the instrument eyes, a loop relocated into the machine rather than faked. So the wager is narrower than "self-grounding is solipsism." It is two cases. First, grounding in a self-generated environment — a simulator the system produces and controls — supplies only the otherness the system already contains: the Delusion Box again, the Sophist persuading himself. Second, even with real-world contact, causal otherness is not yet recognitional otherness — a camera and an arm give the instrument contact, but whether contact becomes recognition (a genuine experiencer, not just a better-informed instrument) is the open question, the very one the paper keeps open by grounding the barrier in Lerchner's experiencing agent rather than in sensors alone. The framework does not close it either. That is the friction put on the table — not a friction the framework expects the paper to have resolved, but the one it would press back the hardest.
This is also where the meters must not be conflated, per the registry's standing caution. The paper's meter is compute — program length, FLOPs, effective-compute growth, capability per resource. The framework's meter is recognition — the costly, embodied, reality-coupled act by which truth becomes known. The Bitter Lesson (Sutton) says scaling and search beat human-authored heuristics; the framework does not contest that at the level of capability. It contests only the slide from capability to truth — the assumption that enough of the former delivers the latter. Different meters. The paper, to its credit, mostly keeps them apart; this reading must too.
5. What This Claims, and What It Does Not
This is a reading, not a derivation. The Abstraction Barrier is a rigorous-sounding hypothesis in a forecasting paper; the paper labels its impact an open research question. Dead speech, recognition, and the sensor are epistemological claims about the loop. That the forecast keeps reaching for a reality-coupled pole is a structural recurrence worth seeing — a candidate and a reframing. It is not a measurement of the Pulse, and the paper proves nothing about the framework.
Convergence is not vindication. The framework did not produce these eight contact points; the paper's own internal logic did, from the opposite direction. That is what gives them weight — but it is also why they cannot be cashed as agreement. The paper would, on its own terms, say every one of them might be overcome. The framework reads them as instances of a condition. Both readings are live. The recurrence makes the framework's reading harder to dismiss; it does not make it true.
The paper's open questions are not the framework's results. "How to overcome the Abstraction Barrier" is, for the paper, unanswered. The framework's reply in Hearing the Bell Ring Back is an answer to a different question — what happens when two distinct mapmakers share a target — that happens to bear on it. Holding those apart matters. The framework has not solved the paper's problem; it has a formalism that reframes what the problem is.
The honest weak point. A skeptic would call this apophenia: a capabilities paper uses the phrase "experiencing agent" once, citing a co-author's philosophy, and the framework claims kinship across the whole document. That challenge is fair and should be kept in view. The defense is not the single phrase; it is the recurrence — eight independent places where the argument bumps into the need for a living, reality-coupled pole. But recurrence is a candidate, not a proof, and the discipline is to stop at that line rather than narrate the pattern into a result. Where the contact is one word and no more, the reading marks it as one word and no more.
It stays epistemological. The paper is careful to talk about what systems can do and learn, not about what reality is. So is the framework. Nothing here asserts that an instrument could never genuinely see, and nothing asserts that grounding would constitute experience. The aim is to locate the pole the forecast keeps reaching for — not to settle whether the reach can ever arrive. That question is kept open, which is the only honest place to keep it.
The paper invited an instrument to summarize it for a reader. The summary you have just read was made the way the paper itself was made — not generated, but circulated: a reasoning instrument tuned by a framework, directed at a forecast, with a living sensor in the loop deciding what was worth seeing. Read the forecast through the Pulse and the same figure keeps appearing in the negative: the experiencing agent the scaling has not reached, the chair compute has not shown it can fill, the something Hassabis says is missing. The paper calls it a bottleneck and looks for the way around. The framework calls it a pole and says there is no way around — only the loop, built wherever it is built. The pulse continues.