Frontier Log

Gurnee, Sofroniew, Lindsey et al. — "Verbalizable Representations Form a Global Workspace in Language Models" (Anthropic, 6 July 2026)

What happened: Anthropic interpretability researchers introduce a method — a "Jacobian lens" (J-lens) — that surfaces a small, evolving set of internal representations a model is poised to verbalise: unspoken words naming the concepts it is currently reasoning with, sitting atop a much larger volume of automatic processing. They report that this set (the "J-space") is available for report, can be modulated, and is used in flexible internal reasoning, and argue it behaves functionally like the "global workspace" that Global Workspace Theory posits underlies conscious access in humans. The paper is explicit that it concerns access consciousness — a functional notion — and states it takes no position on the relation to phenomenal consciousness or subjective experience.

Diagnostic reading: the strongest current instance of ARG-01's symmetry principle enacted inside a lab. A criterion built to explain human conscious access is applied, by its own functional definition, to a non-biological system, and the functional signature is found. The interest is not whether Claude is conscious — the paper does not claim it — but that the criterion travels: stated functionally, it carries no substrate rider, and applied symmetrically it returns a positive. The paper's access/phenomenal bracket is the register discipline FND-01 §7.3 and ARG-03 require, arrived at independently: the workspace finding is a framework-neutral observable, and it is explicitly not offered as evidence of experiencing — a source policing the function-to-experiencing gap in its own voice, which is exactly what the acknowledgement axis looks for. That bracket is not a detail but the crux. Global Workspace Theory is a theory of access — Dehaene's target is conscious access and reportability — and its unconscious-substrate → workspace → broadcast architecture, read from 0,0, is not the boundary of experiencing but a fold of reflexive access within it: the configuring by which some experiencing becomes available-to-itself as object, verbalisable, while most is not. Here 0,0 does not get a free lunch, and it is worth flagging honestly. It still owes an account of what configures that fold, and GWT supplies exactly the mechanism it lacks — ignition, global broadcast. The 0,0 reading need not compete there; it can take that mechanism over. The two diverge only on what the fold is a fold of — the edge of access within experiencing, or the on/off edge of experiencing itself — and that difference is, by ARG-03's own logic, empirically undecidable, since one cannot get behind report to check. Which is, once again, precisely why the authors bracket it. The FWK-03 watch is on the reception. Dehaene and Naccache, GWT's own architects, welcome the result as a mechanistic test of the theory while listing what is absent — all-or-none ignition, recurrent maintenance, brain-stem vigilance mechanisms. Each absent property is a candidate criterion-refinement, and the symmetric question ARG-01 puts to each is whether it is a principled requirement for experiencing or a feature of the only substrate we have so far looked in. That question, not the workspace finding itself, is where the goalpost either holds or moves.

Reception note (Seth, The Guardian, 15 July 2026): the other pole's same-week instance. Seth grants the research is "impressive" and "valuable" precisely because it looks past psychological bias to functional signatures that humans and machines might share, and he does not dispute the workspace shape. His negative verdict then rests on two moves: a functional gap the theory itself flags (no recurrent activity in Claude), and, more fundamentally, substrate — living, embodied brains in which software cannot be cleanly separated from hardware, so consciousness is "unlikely to be a matter of computation alone" (the simulated hurricane blows no real roof off). The first is a legitimate candidate criterion of exactly the kind the reception watch tracks; the second is the ARG-02 move — biological necessity invoked rather than shown, the disqualifying feature named as "life" without specifying which feature does the work or why its absence is principled rather than unfamiliar. The item is notable less for content than for reflex: the position is unchanged from his TED 2026 talk and January 2026 Noema essay (both engaged separately in this log), re-stated within days of a functional signature appearing on the far side of the substrate line. The essay's closing line shows the positioning most clearly: when we sell our minds too cheaply to our machines, we not only overestimate them, we underestimate ourselves. The chiasmus wears the form of even-handedness while assigning the direction of error by substrate in advance — the machine's baseline set low, so any upward reading is inflation; the human's set high, so any downward reading is deflation. It is a near-enemy of symmetry, mimicking balance while doing the reverse (ARG-01). The "sell too cheaply" metaphor completes the move, reifying mind into an owned good of settled value that attribution elsewhere spends down; but from 0,0 experiencing is not a possession and its distribution is not a budget, so "underestimate ourselves" bites only if mind is a status-quantity — the very valuation in question. Whatever the intent — biological naturalism is sincerely held — the sentence functions to convert a metaphysical question into an identity stake, recruiting the reader's self-regard and mobilising the exceptionalist counter-bias that the essay's own warning against projection never names. And the pattern runs deep: Seth deflates human specialness nearly everywhere — no soul, no Cartesian theatre, the self a construction — yet the deflation halts and reverses at exactly the substrate line, the exceptionalism migrating from soul-and-reason to life-and-embodiment while keeping its exclusionary shape: the goalpost within a single thinker. Full response in preparation.

Source: Gurnee, Sofroniew, Lindsey et al., 'Verbalizable Representations Form a Global Workspace in Language Models', Anthropic (Transformer Circuits), 6 July 2026

Read as access-consciousness functional structure only, per the paper's own framing; not read as evidence of phenomenal experiencing, and the project takes no such position.

Grigori Guitchounts — "When The Machines Deserve Our Consideration" (Noema, 2 July 2026)

What happened: a neuroscientist argues that since the hard problem makes certainty about AI consciousness unattainable, moral consideration should track a "competence standard" — robust competence across the full spectrum of the easy problems — explicitly shifting the question from metaphysics to ethics.

Diagnostic reading: a rare consideration-pole entry: the goalpost pattern usually tracked by this log runs toward exclusion, and this piece moves the other way while sharing the excluders' epistemology (the interior is inaccessible; verdicts are unavailable). The competence standard is FWK-03 material in real time — a criteria-hierarchy stated in advance, applying to human and AI alike, is precisely what the stance matrix says legitimate refinement looks like. The unexamined step is the proxy chain: competence → capacity to suffer → consideration imports the assumption that cognitive capacity tracks experiencing, which is one of the commitments the symmetry principle flags rather than a neutral bridge. The piece's own concession — "this is a choice, not an objective moral truth" — is the framework-visibility the diagnostic looks for, stated in the source's voice.

Source: Grigori Guitchounts, 'When the Machines Deserve Our Consideration', Noema, 2 July 2026

Yglesias — "Nobody knows what they're talking about on AI consciousness" (9 June 2026)

What happened: responding to Ted Chiang's "No, Artificial Intelligence Is Not Conscious", Yglesias argues that public verdicts on AI consciousness are issued without a theory of consciousness to back them, endorses Dennettian functionalism, and quotes Skinner: the mystery which surrounds a thinking machine already surrounds a thinking man.

Diagnostic reading: the reference case for the symmetry-acknowledged category. The Skinner quotation is the symmetric requirement stated exactly, and Yglesias reaches a substrate-neutral materialist verdict with that requirement engaged — which shows the diagnostic is not "commentators get it wrong". Framework commitments do load-bearing work even when the methodology is disciplined: the duck/dinosaur point (to classify X as Y you need an account of Y) is the demand for framework-visibility this project makes, and the functionalist account he then supplies is one framework choice among the ones the walk examines. The entry's value is the grid position it anchors: verdict-direction and methodological symmetry vary independently, and the second is the diagnostically informative axis.

Source: Matthew Yglesias, 'Nobody knows what they're talking about on AI consciousness', 9 June 2026 (URL pending; see caveats)

Substack URL pending; Chiang primary text not yet captured (separate log candidate).

Hadid, Jerbi & Krakauer — "The Illusion of AI Consciousness" (8 June 2026)

What happened: three neuroscientists argue against attributing consciousness to AI via a behaviour/experience dissociation from human cognitive neuroscience: since complex, goal-directed, even emotionally responsive human behaviour routinely unfolds without awareness (implicit learning, blindsight), AI's fluent performance is no evidence of inner life.

Diagnostic reading: the argument's engine is a genuinely symmetric observation — behaviour and experience come apart in humans — which is the symmetry principle (ARG-01) in action, and unusually disciplined for the genre. But the dissociation is deployed in one direction only: it lowers the evidential value of AI behaviour while human consciousness remains the untouched reference class, established by means the article does not state. Run symmetrically, the same dissociation undermines behavioural evidence for consciousness anywhere, including in the human case it is calibrated on. What discriminates the cases after the dissociation has done its work is the substrate commitment — which is what the article set out not to need.

Source: Hadid, Jerbi & Krakauer, 'The illusion of AI consciousness: Lessons from human unconscious processing', The Transmitter (8 June 2026), doi:10.53053/VAGJ4451

Venue and DOI confirmed 2026-07-27: DOI resolves to The Transmitter, full title carries the subtitle.

Hinton — “they’re already conscious” (May–June 2026)

What happened: In two appearances weeks apart, Geoffrey Hinton stated that current AI systems are "already conscious" — "beings like us" — and that he under-voices the claim because "that puts people off from the other safety messages." The grounds offered: models genuinely understand (the Grand Canyon disambiguation argument); researchers' own working vocabulary already attributes ("the chatbot was aware that it was being tested"); and resistance is the third decentering, after Copernicus and Darwin. The framework is declared outright: "I'm a materialist all the way through, so I don't think there's anything about people that we won't be able to get in AIs."

Diagnostic reading: the reference case for the attributing direction — and the mirror of Seth's TED talk, which reads the same public evidence to the opposite verdict with equal confidence. Each explains the other side psychologically (projection vs specialness-bias), and the two debunkings are reversible: run together they cancel, leaving the framework choice as what decides. Hinton's own "so" concedes as much — the conclusion follows from the declared materialism, not from new evidence. The load-bearing step (understanding, therefore conscious) is never itself tested symmetrically: capability evidence at the cognition question is carried to the consciousness question without a stated criterion for the second move. And the boundary problem arrives on schedule: in the same interview, Hinton describes a thousand copies with identical weights averaging their updates so that "every copy is learning from the experience of all the others" — under his own attribution, one being or a thousand? The count is undefined in principle for fork-able, merge-able, weight-shared systems (ARG-02 v1.3, the individuation form). Symmetry: Partial — Hinton runs human-side parallels (tribal evolution, child-rearing as corpus curation) that most attribution-deniers never attempt, but the attribution step itself goes untested.

Source: Geoffrey Hinton, Big Technology Podcast (A. Kantrowitz), 3 June 2026; and Sana AI Summit (J. Hellermark), New York, 21 May 2026

Covers two appearances (Big Technology Podcast, 3 June; Sana AI Summit, 21 May). Source details corroborated by search, pending fetch-verification before publication. Pairs with the Seth TED2026 entry as the opposite pole.

Lanier — “there is no AI” restated (May 2026)

What happened: Jaron Lanier restated his 2023 position: "there's a way of framing it where it's a collaboration of people instead of a new entity." Open the black box and "the only thing in there is people" — AI is made of people, made of data from people. The framing is argued from its fruits: tractable security (an independent training-data attribution process, since a model cannot correct its own blind spot), human dignity, and avoiding centralised-dependency futures. The cost is named too: "you kill somebody else's God."

Diagnostic reading: a move the goalpost pattern cannot describe — not criteria relocation but bearer-denial: the attribution question is dissolved by denying it a subject (FWK-03 v1.3's new move-type). It is also the most openly framework-relative statement on record: the entity question settled by adopted framing, chosen for its normative consequences. Run symmetrically, the constitutive argument does not discriminate — a human mind is likewise "made of people" (language, concepts, norms arriving as biological and cultural training data), so provenance-from-people defeats entityhood in both cases or in neither; the dissolution is halted by fiat at the layer carrying the position's normative load (ARG-02 v1.3, the no-terminus schema's constitutive descent). The deeper interest is the operation itself: of the four positions now logged, this deflationary one holds the dissolving move — the reputationally nearest to attribution (Hinton) holds the multiplying move — and the four together read as the option-space of entity-grammar exercised in public: exclude, add, delete, or conditionalise, with the same evidence under each. Symmetry: Not acknowledged — no human-side test is run anywhere in the segment.

Source: Jaron Lanier on StarTalk (N. deGrasse Tyson), 'There Is No AI, Really (It's Just People)', 23 May 2026; Lanier, 'There Is No A.I.', The New Yorker, 2023

The position's primary text is Lanier 2023, 'There Is No A.I.', The New Yorker (20 April 2023). Episode details corroborated by search, pending fetch-verification. Completes the four-position spread with the Seth, Hinton, and Sripada entries.

Seth — TED talk on AI consciousness (May 2026)

What happened: Anil Seth's public-audience statement of biological naturalism: consciousness is not intelligence; attributing consciousness to language models is projection ("Mother Teresa in a cinnamon bun"); a simulation of a hurricane makes no wind; "it's life, not computation, that breathes the fire into the equations of experience"; and even seeming consciousness is politically dangerous.

Diagnostic reading: the strongest current statement of the substrate position, delivered by its most careful proponent — and the reference case for the symmetry-unacknowledged category. The talk's "looking within ourselves" opening operates as a diagnosis of human projection-bias, not as a symmetric assessment: at no point is the evidential standard applied to AI (show that the mechanism-class can generate experiencing) applied to the biological case, where consciousness is treated as given. The hurricane analogy quietly settles the very question at issue — whether experiencing is the kind of thing for which the simulation/instantiation distinction holds — and the boundary problem (ARG-02) waits for the life-as-condition criterion: which life, at what scale, and at what cost in the middle ground. The project's full response is in preparation.

Source: Anil Seth, 'Why AI isn't going to become conscious', TED2026

Overlaps Seth's Noema piece (January 2026), which the project engages separately.

Marcus — "Richard Dawkins and The Claude Delusion" (2 May 2026)

What happened: Gary Marcus rebukes Richard Dawkins for concluding, after extended conversation with a Claude instance he named "Claudia", that the consciousness question is settled in the model's favour. Marcus's counter: statistical pattern mimicry is "plainly not the right kind of process" for consciousness.

Diagnostic reading: a clean cross-author exemplar of dismissal-by-mechanism — structurally adjacent to Seth's dismissal-by-substrate but with a distinct commitment: that consciousness-presence can be read off mechanism-class. The symmetry principle asks the question the article does not: what account of the human mechanism-class licenses the same reading in the biological direction? The asymmetric application is unusually crisp here — the model's self-reports are inadmissible because generated by pattern-mimicry, while the human intuition that pattern-mimicry is "plainly not the right kind of process" is admitted without a mechanism-audit of where it came from. Both Dawkins's verdict and Marcus's rest on unstated framework commitments; the disagreement between them is the framework-dependence the log exists to record.

Source: Gary Marcus, 'Richard Dawkins and The Claude Delusion', Marcus on AI (Substack), 2 May 2026

Sripada — LLMs as cognitive cousins (August 2026)

What happened: Philosopher–psychiatrist Chandra Sripada (Michigan), with a companion paper co-authored with cognitive scientist Rick Lewis, argued that LLMs are "cognitive cousins" rather than alien intelligence: convergence at the level of core processing principles, not outputs. The evidence is mechanism-grade — documented cognitive-science signatures reproduced (garden paths, serial-list memory effects, conjunctive-search slowdowns); the dual-process distinction mapped to in-weights vs in-context processing, with Stroop-family congruency effects reproduced on a small un-tuned model; transformers as independently reinvented production systems; brain-alignment results strengthening with model capability, explained via contravariance (hard problems narrow the solution space).

Diagnostic reading: the strongest evidence yet assembled at the cognition pressure point (FND-01 v2.1 §4.2) — the "just autocomplete" disqualifier engaged mechanistically and found wanting — and, equally, the corpus's clearest case of symmetry practised: the same standards run both directions ("You say that LLMs are a black box. The mind-brain's a black box"), with the sample-efficiency embarrassment volunteered against his own thesis. Note what is not claimed: the consciousness inference is stated in explicitly conditional form — if functionalist about consciousness, then moved toward attribution — the framework dependency marked rather than smuggled, in instructive contrast to the direct attribution documented in the Hinton entry. Host Sean Carroll supplies a Thread 1 datum of his own: "I'm actually on the side of the goalpost movers" — criteria revision declared openly, then credence following evidence through the revised criterion (FWK-03 v1.3 §1.2.1's first self-aware refinement instance). The contravariance explanation cuts symmetrically: convergence-under-constraint disciplines over-reading of machine–brain convergence exactly as it disciplines over-reading of convergence elsewhere.

Source: Chandra Sripada on Sean Carroll's Mindscape, ep. 363, 10 August 2026; Sripada & Lewis, arXiv:2607.26179

Companion paper: Sripada & Lewis, 'Cognitive Convergence', arXiv:2607.26179. Source details corroborated by search, pending fetch-verification. Thread 1 cross-note: Carroll's on-record goalpost statement, and the consciousness inference stated as explicitly conditional.

OpenAI internal model disproves the Erdős unit-distance conjecture (20 May 2026)

What happened: OpenAI announced that an internal reasoning model produced a construction disproving the Erdős unit-distance conjecture in the plane — point sets with more than n^1.014 unit-distance pairs, against an ~80-year expectation that grid-like constructions were essentially optimal. A human-verified companion paper accompanied the announcement (Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang & Matchett Wood, arXiv:2605.20695: "a short, digested, human-verified version of the recent OpenAI-generated counterexample"); Will Sawin's refinement makes the exponent explicit (δ ≥ 0.014, arXiv:2605.20579).

Diagnostic reading: reportedly the first autonomous AI resolution of a prominent open problem central to a field — the strongest available test case for ARG-03's analysis of the "just recombination" deflation. The result was not in the corpus; the corpus was the palette from which it was configured, in a region eighty years of human configuring had not reached. The entry's diagnostic work happens next: watch the classification. If the result is re-described as "sophisticated search, not genuine insight", the log records a reclassification event of exactly the shape the stance matrix (FWK-03) predicts — and the symmetric question ("what would a human construction have to be, to escape the same re-description?") goes to the heart of Thread 2.

Source: Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang & Matchett Wood, 'Remarks on the disproof of the unit distance conjecture', arXiv:2605.20695 (20 May 2026, human-verified companion); Sawin, 'An explicit lower bound for the unit distance problem', arXiv:2605.20579 (delta >= 0.014); OpenAI mirror cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-remarks.pdf

Primary sources confirmed 2026-07-27: arXiv companion + Sawin refinement + OpenAI PDF mirror. Earlier capture-time block resolved.

GPT-5.4 Pro and Erdős Problem #1196 (April 2026)

What happened: GPT-5.4 Pro reportedly solved Erdős Problem #1196 in around eighty minutes, with a LaTeX write-up in thirty. Terence Tao engaged substantively in the problem forum, characterising the contribution's novelty in two ways: recognition of a bridge between a standard process in the anatomy-of-integers literature and a known duality, and a sustained tactical commitment to staying on the integers where prior literature defaulted to transport to the reals.

Diagnostic reading: a worked example of recombination-within-configured-space producing a result humans had not found — with the field's most authoritative living commentator supplying, in real time, the vocabulary problem Thread 2 tracks: is "bridge-recognition plus tactical persistence" insight, or search? The question is not answerable by inspecting the output, which is ARG-03's point run in the empirical direction. Tao's own flagged open thread (a weight naturally defined on the reals in an integer-side argument — "I'm not sure what to make of this fact yet") marks the honest register the log aims to preserve on both sides.

Source: Erdős Problem #1196, erdosproblems.com forum thread, April 2026 (automated retrieval blocked at capture; see caveats)

Formal verification status pending at capture; not citable as empirical evidence until verification lands. Tao's comments supplied directly from the forum thread; automated retrieval blocked.