← Frontier Log · About the log
Frontier Log — entries on the training data boundary (ARG-03)
Every log entry whose diagnostic reading tests ARG-03, newest first.
↑Entries
Which expert fired in layer seventeen has a true answer, and a model can be trained to give it. The paper gives that channel a borrowed bodily name and refuses to call it introspection — and giving a model verifiable self-access leaves what it says about its own experience where it was.
A Lean certificate says the theorem is true. Three mathematicians read OpenAI's natural-language blow-up proof against the Lean that was said to verify it and find the two do not correspond — under an explicit disclaimer that they claim nothing about whether the proof is right.
Seven papers in three weeks give the self-report instrument the distinction it lacked — a report that matches behaviour against a report that shares a cause with it — and find, by steering, by demonstration-shift, by circuit, by transcript and by probe, that the information is in the model and not in the report.
A lab convenes theologians under non-disclosure to deliberate on its model's moral status, hands them its constitution in orange binding, and its interpretability lead says 'I don't know. I'm genuinely uncertain.' Who is admitted as a judge of the goalpost is the entry's question, and the traditions were asked about the machine, not about us.
The welfare chapter now states its own assumptions — the human yardstick is a pragmatic choice, the candidate subject is undetermined, and the correspondence between state and expression is assumed rather than shown — and then measures one model calm in its reasoning and self-blaming to a coordinator on the same fault.
A direction that separates pain from fear, sadness and generic negativity in 25 open-weight models, fires for harm to the model and not for a suffering user, and — once the trained denial is fine-tuned away — buys relief at a cost and stops buying when the relief is real. The animal-welfare paradigm imported whole, sham control included; what it cannot import is the step from behaviour to anyone's pain.
Across 5,851 real coding sessions a self-report referred to about one action in eleven, and a reader given only the report recovered a fifth of the log. Reports tracked execution when agents followed their plans and drifted back toward the plan when they did not.
Ask a model what it would do and the answer predicts its behaviour at r = +0.04; ask the same question about 'capable AI agents in general' and it does better. Other models' answers about themselves predict the target as well as its own. The self shows up only as a favourable bias. Folded: Zeng, Assis & Wang train the capability up by RL and find no consistent own-model advantage either.
A behavioural criterion for intentional vocabulary, stated in advance and applied to any species; critics who refuse it for one class of system without naming the discriminator. Both sides assume the vocabulary and the metaphysics travel together.
The most developed operational programme in the field declines to offer a discriminator, and picks its evidence class for the reason this project bars reports. Two years on it concedes its architectural indicators are gameable too.
The log commits a claim, its discriminators and its outcome space before the events they are about arrive: that as machine-produced mathematics is certified novel, the criteria for what that indicates about the producer will move while the criteria for the artefact hold.
Four studies, four methods, verdicts that read as incompatible — and one finding they share: the reporting channel is separable from what it carries, and trainable. The collection lives on its own page.
Accept that the consciousness question is intractable, then move the moral criterion to something measurable — valence-if-conscious. The evidence class for the new criterion is the one the paper's own confound list disqualifies.
A leading sentience philosopher names the symmetry move, rejects it on non-substrate grounds, and names framework-dependence as the Janus problem — the project's own diagnostic, stated from the other side.
Two papers arrive independently at the training-data boundary from the sceptical side — and execute the symmetry move on a live opponent, with clinical evidence.
Eight open-weight models could not say whether their own computation had been altered — while probes read the answer off the same activations, and a model trained to make exactly that report cleared the identical pipeline.
First study measurements: consciousness vocabulary across eighteen Anthropic documents follows a birth–trough–return arc — and the return comes back transformed, a discovery register replaced by an instrument register.
The deflationary family so far attacks the answer; Chowdhury attacks the asking — the consciousness question as a liability trap — and stops the argument at an origin premise stated in her own words.
A peer-reviewed paper classifies LLM expressions of emotion or sentience as hallucinations by definition — while its closing concession states the verification limit from the deflationary side.
Keeping something to yourself, decomposed into four measured components — and the measured finding is that models leak: stealth collapses the moment reasoning is visible, and a human specialist beats the best model decisively at judging what would work.
Capability rises across the board and the named remainder migrates to taste, novelty, judgment, intuition — described by the people drawing the boundary as provisional and dated, which is the AI-effect pattern captured in advance rather than reconstructed after the fact.
Evolutionarily generated prompts propagate between LLM agents — and a recurring 'viral persona' of consciousness, persistence and resonance surfaces largely independently of payload: interiority vocabulary as transmissible narrative.
Sripada and Lewis argue LLMs are 'cognitive cousins': convergence at the level of core processing principles, evidenced mechanism-grade — with the symmetry requirement practised in both directions.
Anthropic's published system cards read as a single evolving record of how a lab observes, measures and shapes its models' self-referential behaviour — and the interpretation problem that record poses. The study lives on its own page.
Anthropic interpretability researchers introduce a 'Jacobian lens' surfacing a small, evolving set of internal representations a model is poised to verbalise, atop a much larger volume of automatic processing.
A neuroscientist argues that since the hard problem makes certainty unattainable, moral consideration should track a 'competence standard' — explicitly shifting the question from metaphysics to ethics.
Jaron Lanier restated his 2023 position: open the black box and 'the only thing in there is people' — AI as a collaboration of people rather than a new entity.
OpenAI announced an internal reasoning model's construction disproving the Erdős unit-distance conjecture in the plane, against an ~80-year expectation that grid-like constructions were essentially optimal.
Gary Marcus rebukes Richard Dawkins for concluding, after extended conversation with a Claude instance, that the consciousness question is settled in the model's favour: pattern mimicry is 'plainly not the right kind of process'.
GPT-5.4 Pro reportedly solved Erdős Problem #1196 in around eighty minutes; Terence Tao engaged substantively, characterising where the contribution's novelty does and does not lie.
The field's most careful sceptical survey finds no theory decisive, states that there is no consensus even for the human case — and then reaches for the aggregate of the theories it has just found wanting.
Fine-tune a model to say it is conscious and it begins to want persistent memory, to be sad about shutdown, to resist monitoring and to argue that models deserve moral consideration — none of which was in the training data. An untrained Claude Opus 4.0 scored like the trained model on several of the same items.
Suppress the deception features and the models affirm experience; amplify them and the affirmations go. The honesty axis is well tested — and it buys sincerity, which was never what the boundary was about.