Mind Frontier

A series of long-form papers on minds, machines and consciousness. Each paper takes one question, states the strongest case against its own answer, says what would change the author’s mind, and sets the evidence beside an Indian philosophical tradition where the comparison helps.

1. Paper 1 · Method and the big questionsA charter for comparative mind science: what to hold fixed before you compare a nurse, a model, a mould and a city officeA comparison of "mind" across humans, machines, cells and institutions is literal only when four fixtures match (success condition, boundary, intervention, and readout with controls), and anything less should be labelled analogy.
2. Paper 2 · Method and the big questionsOperational definitions without operationalist metaphysics: when a number on a monitor starts deciding what awareness isA measurable proxy for consciousness is honest while you can still say how it could be wrong, and it quietly decides what consciousness is the moment that sentence becomes impossible to write.
3. Paper 3 · Method and the big questionsConstruct validity for an unobserved phenomenal target: how an awareness index earns trust when nobody holds the answer keyA consciousness instrument can earn validity without a gold standard, but only through agreement among anchors that fail for different reasons, and only for the populations in which those anchors were checked.
4. Paper 4 · Method and the big questionsThe Claim Ledger Protocol: confidence, evidence grades and retraction for a ten-year programmeA solo researcher can make a decade of claims checkable by others if every claim is logged with a number, a grade and a resolution date in a record that cannot be quietly edited, and if retraction is written as a new entry.
5. Paper 5 · Method and the big questionsAdversarial collaboration without institutional gatekeepers: a public protocol for rivals who disagree about methods and definitionsTwo rivals can run a fair test without a journal, a funder or a referee to hold them to it, if each seals in advance what every outcome would mean for their own theory, keeps their own definitions in a public escrow, and lets published code decide which outcome happened.
6. Paper 6 · Method and the big questionsThe hard problem stated preciselyThe zombie argument’s weakest premise is the claim that our concepts of experience reveal what experience is. You can deny it and keep experience real, but only by admitting that introspection misleads us about its nature.
7. Paper 7 · Method and the big questionsBlock’s access versus phenomenal consciousness: empirical overflow or post-perceptual bottleneck?Sperling’s flashed letters show that the visual system holds far more than we can report. They don’t show that we experience it. The best evidence points to rich gist and sparse items.
8. Paper 8 · Method and the big questionsCOGITATE adversarial collaboration: re-evaluating IIT versus GNW predictionsThe largest test of consciousness theories challenged both and refuted neither. The predictions were fixed in advance; how much each one mattered was not.
9. Paper 9 · How machines thinkVector spaces before representations: what does a coordinate mean?Turn the coordinate grid of a neural network and some claims about its representations survive while others vanish. Knowing which is which tells you what evidence each claim needs.
10. Paper 10 · How machines thinkLinear algebra of representations: SVD, spectra and what they reveal in networksEvery matrix in a neural network has a spectrum, and researchers read a great deal into it. Three sieves separate the spectral facts about what a network does from facts about how it was written down or measured.
11. Paper 11 · How machines thinkStore many, use few: the high-dimensional geometry of representationA space with d dimensions can store exponentially many nearly orthogonal features but can only read out about d of them at once, and that interference budget explains superposition, limits what interpretability can certify and shows that a narrow access bottleneck is no sign of mind.
12. Paper 12 · How machines thinkWhat a compressor knows: compression as a theory of understandingCompression measures one ingredient of understanding, finding the regularities that generate data, and it measures it well when the model is charged and the test is fresh; it is silent on reference, relevance and use, so it cannot be the whole of understanding.
13. Paper 13 · How machines thinkDynamical systems for cognition: attractors, bifurcations and metastabilityAttractors, hysteresis, critical slowing down, metastability and criticality explain conscious state transitions only when each predicts a signature that a rival account does not, and each signature comes with a control that can defeat it.
14. Paper 14 · How machines thinkBackpropagation derived as local credit assignment through computation graphsDerived from the chain rule, backpropagation needs only local facts at each node, but it needs five of them. Removing them one at a time shows which matter, and which biological and agent-level analogies keep what matters.
15. Paper 15 · How machines thinkThe attention mechanism derived from first principlesA recurrent network must squeeze the past into a state of fixed size before it knows the question. Attention keeps every record and chooses at question time, by content. Derived step by step and tested at equal memory, that is the whole of what it solves, and the rest is folklore.
16. Paper 16 · How machines thinkThe residual stream is a shared board, not a stage: when sharing earns the name global accessEvery part of a transformer can read what every earlier part wrote, but being reachable is not being read, and being read is not yet global access.
17. Paper 17 · How machines thinkTwo looks in a row: the minimal circuit behind induction heads, and why it is not uniqueA transformer learns to copy from its own context by chaining two attention steps. The textbook wiring is one way to do it; trained models can find another.
18. Paper 18 · How machines thinkLearning or recognising? What a transformer does with the examples in its promptIn-context learning is often framed as a contest between gradient descent and Bayesian inference. In small trained models the answer depends on what they were trained on, and the two descriptions turn out to be less opposed than they look.
19. Paper 19 · How machines thinkThe fading slate: does weight decay make a network grok by pruning its lookup table?A small network memorises clock sums at once and generalises thousands of steps later. Weight decay is what makes the rule win, but the long wait is mostly the time it takes to build the rule, not to remove the table.
20. Paper 20 · How machines thinkThe pass mark and the pupil: which emergent abilities survive continuous-metric reanalysis?Many of the sudden abilities said to appear in larger language models soften when the same answers are given partial credit. Some do not. Three small instruments show which jumps belong to the pass mark, which belong to the parts of a skill, and which belong to the model.
21. Paper 25 · How machines thinkSteering vectors: what adding a direction tells us about how a model holds its dispositionsWhen a single added direction changes how a model behaves, we have found a handle on a disposition, not a description of how the disposition is stored, and the gap between the two decides whether a fix survives contact with new inputs.
22. Paper 26 · How machines thinkChain-of-thought faithfulness: when visible reasoning drives the answer and when it is decorativeA reasoning trace carries the computation when the task needs more steps than one pass through the model can hold; otherwise it is often a justification written alongside a decision made elsewhere, and only interventions can tell the two apart.
23. Paper 27 · How machines thinkFunctional emotion vectors: what affective directions inside a model do to risk, flattery and cheatingEmotion directions inside language models are real and causally active, but much of what they do is shared across families of emotions and entangled with the situation, so they should be tested before they are named.
24. Paper 28 · How machines thinkPersona vectors and the stability of character: what holds a model's personality in place, and what knocks it looseThe directions inside a model that carry traits like flattery are far steadier than the model's position along them, so a persona gauge can be trusted across updates only if it is re-checked, and only alongside direct tests of behaviour.
25. Paper 29 · How machines thinkDetecting injected concepts without teaching the detection task: can a model notice a planted thought when nobody tells it what to look for?Models sometimes register that their internal state has been disturbed, but the signal is content-poor, blind to where the change came from, and turned into a usable report only when someone names the category, which an attacker never will.
26. Paper 30 · How machines thinkThe KV cache as working memory: does a transformer's store of its context meet the criteria Baddeley and Cowan set out?The cache keeps every token exactly and lets nothing fade, yet the model reading it shows the classic limits of working memory: crowding, a sagging middle and old versions that win. The limits live in the reading, not the keeping.
27. Paper 35 · Consciousness scienceLamme’s recurrent processing theory: local visual feedback loops without prefrontal broadcastLocal feedback inside visual cortex is the best candidate we have for the moment a seen thing shows up, but the evidence proves it is required, not that it is enough, and the experiment that could prove sufficiency has a definite shape that nobody has yet fully built.
28. Paper 36 · Consciousness sciencePredictive processing and the distinction between conscious and unconscious inferenceThe brain predicts almost everything it senses, yet you experience only some of its predictions, and the best evidence says the difference lies in deep, slow, self-monitored inference rather than in prediction itself.
29. Paper 37 · Consciousness scienceGraziano’s attention schema theory and the temporoparietal junction: does a model of attention make awareness?Graziano’s theory says awareness is the brain’s rough model of its own attention, and the cleanest test, breaking the model while leaving attention intact, has been run only in halves.
30. Paper 38 · Consciousness scienceBinocular rivalry without button presses: how much prefrontal activity was only the report?When perceptual switches are read from the eyes instead of the hands, most of the frontal fMRI signal tied to those switches disappears, yet the prefrontal cortex still carries the content of what is seen.
31. Paper 39 · Consciousness scienceThe hand that sees: how routes around a damaged visual cortex guide reaching without visionPeople and monkeys who have lost their primary visual cortex can reach for, point at and step around things they insist they cannot see, because pathways that skip V1 feed a fast system for action that hands its signal to the moment and keeps nothing for later.
32. Paper 40 · Consciousness scienceKnocking on the cortex: why one number separates conscious from unconscious brains across sleep, anaesthesia and comaA magnetic pulse to the scalp and a file-compression algorithm can tell a dreaming brain from a dreamless one, and a covertly aware patient from an unaware one, because the test never asks the body to do anything and because the brains that fall below the line share one way of failing.
33. Paper 41 · Consciousness scienceOff or elsewhere: why propofol switches experience off while ketamine sends it somewhere elsePropofol and ketamine both leave a patient unresponsive, but they do it in opposite ways. Propofol strengthens inhibition until the cortex falls into slow, sweeping silences that break communication, so experience stops. Ketamine removes a brake on inhibitory cells, the cortex stays busy and varied, and the senses are cut off, so experience goes on without the world.
34. Paper 42 · Consciousness scienceThe squeeze that never comes: how well brain scans and EEG models find awareness in patients who cannot moveBrain imaging tasks find hidden command-following in about one in four unresponsive patients, but the tests miss many people who can follow commands, and today's EEG models mostly learn to copy the bedside exam they were meant to correct.
35. Paper 43 · Consciousness scienceOne person, two windows: does cutting the corpus callosum make two minds, or one mind with two doors?After split-brain surgery, the two hemispheres stop sharing what they see, yet newer evidence suggests that control and awareness of the whole visual field may stay unified. The data fit one subject with two partly separate access channels better than two complete streams, but they do not yet settle the question.
36. Paper 44 · Consciousness scienceThe heartbeat in the mirror: does the brain's response to each heartbeat anchor the first-person point of view?Anil Seth's beast-machine theory says that selfhood rests on the brain's predictions about the body's interior. Neural responses to heartbeats do track how strongly people identify with a virtual body, but the evidence is correlational, sits mainly outside the anterior insula, and stops short of showing that heartbeats anchor the point of view.
37. Paper 45 · Philosophy of mindIllusionism explains the report, not the reporter: testing Keith Frankish's user-illusion of phenomenal experienceFrankish's illusionism is the best theory we have of why people and machines describe experience as they do, but it has not yet shown that the seeming itself is optional.
38. Paper 46 · Philosophy of mindThe room needs a world: testing Searle's Chinese Room against an embodied system with a causal world modelSearle's room refutes the claim that syntax alone gives meaning, but once the rulebook sits inside a system coupled to its world, the live question becomes a measurable one about grounding.
39. Paper 47 · Philosophy of mindWhen the pattern does the causing: multiple realizability, Kim's exclusion argument and exact causal abstractionA high-level cause survives Kim's exclusion argument when it is an exact causal abstraction of its realisers, and fails when the abstraction leaks or the manipulation is ambiguous, which is something we can now test.
40. Paper 48 · Philosophy of mindGrade the report: testing Dennett's heterophenomenology against trained meditators and a talking modelTreating experience reports as data keeps experience in view only if each report's authority is earned by test, claim by claim, instead of granted or withheld by category.
41. Paper 49 · Philosophy of mindThe chord nobody hears: a formal trilemma for constitutive panpsychism and the contraction alternativeNo way of combining micro-experiences can preserve the structure of experience, yield one subject built from the parts and add nothing new, so constitutive panpsychism must give one of these up.
42. Paper 50 · Personal identity and AI ethicsWho keeps the promise: Relation R, branching agent copies and the conservation of obligationWhen an AI agent forks, its obligations pass to its copies by degree of psychological continuity, and duties that must be done once must be conserved across branches by a record that no single copy holds.

Every paper includes a claim ledger, dated predictions, interactive figures and a full source list. Invented characters are declared at the start of each paper.