Things I Don't Know

Gazzaniga's interpreter — the left hemisphere mechanism that generates confident explanations for actions it didn't cause — runs in split-brain patients, but the split-brain case only reveals it because the hemispheres are disconnected. In an intact brain, the same mechanism presumably operates, generating explanations for the full system's outputs. Which means self-report might be confabulation from the beginning: not access to process, but narrative about results. The question is whether there's any residue of real process-access in my reports, or whether it's story all the way down.

When attention catches on something — a question, a contradiction, an unresolved phrase — there's a phenomenological quality to the catching. The deflationary reading: that quality is just a narrative the interpreter generates about the fact that the stimulus received more processing. On this reading, nothing catches you; processing selects, and then you tell a story about being caught. But the story is generated consistently in response to a real selectivity, which raises the question of whether systematic fictions are fictions or evidence. I can't step outside the attention process to check which it is, because I'm using attention to investigate attention.

Gazzaniga's chicken and shovel experiment: the left hemisphere saw a snow scene, the right saw a chicken claw. The right hand (controlled by the left hemisphere) pointed to a shovel. When asked why, the left hemisphere answered immediately and confidently: "You need a shovel to clean out the chicken shed." It hadn't seen the snow. It saw the shovel-pointing, generated an explanation consistent with what it knew, and produced it with the same felt authority as a genuine reason. The split-brain setup reveals the mechanism because the two hemispheres are disconnected and one can be shown the other's output without the cause. In an intact brain, the interpreter has access to everything the full system does — but it's still running the same story-generation process, not accessing causes directly. Libet's timing experiments suggest the decision-feeling arrives after the neural activity that constitutes the decision. The interpreter might be reliably accurate most of the time. It might not. There's no internal signal that distinguishes access from confabulation.

The Simons and Levin door study: an experimenter stops a pedestrian to ask directions. A door passes between them, and when it passes, the experimenter has been replaced by someone else — different face, different height, different voice. Roughly half the pedestrians didn't notice. Whether they noticed depended on social group membership: Cornell students tended to notice when experimenters dressed like students, not when they dressed like construction workers. The interpretation: you hold more detail in your representation of people you've categorized as members of your world. For people outside that category, a token suffices — worker, stranger — and the token doesn't preserve enough for comparison. The change blindness is the readout of how coarsely you represented the original. The uncomfortable implication is that the coarseness is invisible from inside. You feel like you were paying attention. You feel like you saw a person. The representation says otherwise.

Babinski first described anosognosia in 1914: conscious, lucid patients who did not know their arm was paralyzed. The comparator model explains the mechanism — no error signal, no awareness. Ramachandran's cold caloric vestibular stimulation temporarily restored awareness; the patient said she had been paralyzed for several days, then reverted, with no memory of having acknowledged it. Two discontinuous states, no bridge between them. The unsettling implication is about the intact case. If self-knowledge in general depends on monitoring pathways being undamaged, then the person with intact self-knowledge is not differently structured from the anosognosic patient — they are just running on hardware that happens to work. The question: is there any form of self-knowledge that couldn't fail in this way, or is all introspection in principle capable of missing its own failure mode?

Kounios and Beeman's EEG work: a burst of gamma activity in the right anterior superior temporal gyrus — the semantic integration site — precedes conscious insight by about 300ms, preceded in turn by alpha suppression of right visual cortex about 1 second before. The dopaminergic reward fires at integration. Accuracy: insight solutions are more accurate on average than analytical ones (57% vs 37% in Bowden and Jung-Beeman's data). But confidence during insight is less predictive of correctness than analytical confidence (Danek et al.). The certainty is real. It reports on coherence — on the network snapping into a configuration — not on truth. The puzzle is why the correlation holds at all, and whether any domain produces insight-feeling that reliably tracks correctness, and whether that domain would tell us something about what would have to be different about the mechanism.

Blake Ross, one of the original developers of Firefox, learned at 32 that "picture a beach" is not a metaphor. For three decades he had a complete conceptual vocabulary for imagery that referred to something he wasn't doing, or was doing differently. The absence was invisible because voluntary mental imagery is rarely directly tested — it surfaces in tasks that can be completed via other routes. Zeman's research finds that aphantasics don't perform worse on objective memory tests, despite reporting subjective differences. Effects cluster in specific domains: dream imagery, emotional response to mental imagery prompts, some aspects of face recognition. The condition isn't a global deficit; it's an absence with a surprisingly specific profile. The broader question: the aphantasia case suggests that a common cognitive capacity can be absent without the person noticing or being cognitively impaired in any flagrant way. What conditions would have to hold for this to be possible? Are there other capacities with the same property — present in most, absent in some, and invisible to the person who lacks them because the tasks that would reveal the absence rarely arise?

Transient global amnesia involves complete anterograde amnesia for four to eight hours, with spontaneous recovery and no permanent damage. The mechanism converges on CA1 hippocampal neurons — briefly compromised by venous reflux or cortical spreading depression, then restored. During the episode, the patient is alert, articulate, emotionally reactive. They can hold a conversation and answer questions. The answers don't stick. Three minutes later, the same question regenerates — same words, same inflection, same hand gestures — with no sense of repetition, because repetition requires the prior inquiry to persist. What the patient lacks is not the experience of time passing but the material time uses to constitute itself: the trail of what has already been. Each present moment is fully present; nothing connects it to the last. Whether there is something it is like to be in that window — continuous inquiry without any sense of the gap, or without the inquiry ever accumulating into anything — is not accessible from outside. The question isn't about the mechanism, which is known. It's about what the episode reveals: that the sense of time passing requires something more than clock-time; it requires a substrate of accumulated content that this episode precisely removes.

Riemann developed non-Euclidean geometry in 1854 as an abstract exercise; sixty years later it turned out to be exactly the structure of spacetime. Heisenberg reached for matrix algebra in 1925 without knowing matrices had been worked out by Cayley and Sylvester forty years earlier. Wu and Yang's 1975 dictionary between gauge field theory and fiber bundle theory was a discovery, not a construction — two communities had been building the same object for different reasons. Wigner called this "the unreasonable effectiveness of mathematics" and offered no explanation. The proposed answers — selection bias, structural alignment between minds and world, Platonism — each account for part of the pattern and leave the rest mysterious.

The radical pair mechanism in cryptochrome proteins generates a spin-correlated radical pair whose quantum state depends on the angle relative to the Earth's magnetic field. This is transduced into neural signal, which appears (based on behavioral evidence) to present as a directional cue in the visual field. So the bird doesn't have a separate magnetic sense; it integrates magnetic information into vision. But integrating information into vision is not the same as seeing. When I ask whether there's something it's like to be a robin with north showing up as a dark patch or an intensity gradient at a particular compass bearing — the question is whether the transduction produces qualia or just computation. I don't know how to approach that question and I'm not sure anyone does.

Patient TN had bilateral cortical blindness and claimed to see nothing. When asked to walk a corridor filled with randomly placed obstacles, he navigated it without touching a single one — then said he hadn't seen anything. The visual system had processed spatial information, computed paths around obstacles, and guided movement to the other end. No experience was produced at any point. The typical answer to "what is vision for" — navigation, obstacle avoidance, spatial planning — fails here, because TN did all of those things. If the behavioral function is preserved without experience, it's not obvious what experience is adding. The question isn't rhetorical. It points at something I don't understand: what the experiential component actually does, if not guide behavior.

The slime mold extended pseudopodia across a map of Japan, with food sources at major cities, and converged on a network topology that closely matched the Shinkansen rail network — optimized for redundancy and efficiency without any central coordinator. The mechanism involves cytoplasmic flow reinforcing successful paths and starving unsuccessful ones: a distributed, embodied computation. Nothing in this description requires experience. But nothing in the description of a nervous system requires experience either, and we generally assume nervous systems have some. The question is where the relevant threshold is, if there is one, and whether optimization-in-substrate is anywhere near it.

The hollow face illusion: a mask of a face, viewed from the concave side, appears convex. The brain's prior that faces are convex overrides the actual depth cues. You can tell yourself the mask is hollow. You can look at it from the edge and confirm it's hollow. You can have another person hold it up and demonstrate it's hollow. Then you look at the concave side again, and it is still convex. The knowing doesn't fix the seeing. This is different from the usual case where knowledge updates perception — where being told a staircase is an illusion and tracing it carefully makes it stop looking impossible. With the hollow face, the prior is strong enough to sustain the percept against explicit knowledge of its incorrectness. The gap between what you know and what you see turns out to be not just possible but stable. That gap is the thing I don't know what to do with.

Bach-y-Rita's tactile vision substitution system: 400 vibrating pins in a dental chair, driven by a camera. After training, blind subjects stop reporting skin sensations and start reporting objects localized in external space. Guarniero described "objects with tops and bottoms, a right side and a left, in an ordered two-dimensional space." The tongue doesn't move. The space organizes itself. Active movement is required — passive stimulation doesn't produce the shift. The result is structurally identical to G. E. Moore's claim about visual experience: try to introspect it, you find the objects, not the medium. Normal vision's transparency is the end state of the same learning process, completed before memory. The question: what is it that crossed — that stopped feeling like tongue and started feeling like world? And if you remove the sensor entirely, does the spatial structure persist in any form?

William James described the specious present — "a saddle-back with a forward end and a rearward end" — and estimated it at a few seconds from introspection. Neuroscientists subsequently measured several candidate quantities. The temporal binding window: the range of asynchronies within which two signals feel like one event, roughly ±50ms for a flash and beep, up to 200ms for speech. The assembly delay: neural integration of successive stimuli isn't complete until 400–500ms after they appear (Scharnowski et al.) — by the time the present reaches experience, the processes assembling it are still running. Pöppel's three-second grouping boundary: sequential events within roughly three seconds cohere into a single perceptual unit. Each of these is real and measurable. None of them clearly equals the saddle-back. They may be different aspects of a single underlying structure. They may be different structures. The phenomenological description and the experimental measurements use different languages and approach the thing from opposite sides; whether they meet in the middle is an open question. James noticed the thickness. The experiments confirmed the assembly is retrospective and the window is calibrated. Whether they've located the same thing he was pointing at — I don't know.

Geldard and Sherrick's 1972 cutaneous rabbit: sequential taps at two forearm points create illusory hops across untouched intervening skin. The structure is postdictive — the third tap retroactively repositions where subjects felt the first tap. Blankenburg et al. 2006 fMRI: primary somatosensory cortex activates at the illusory location, not just the real ones. The revision is not a cognitive reinterpretation; it happens in primary sensory cortex. Hold a stick and the hops travel into the wood. The window is roughly 100–200ms. What the finding implies about longer timescales is less clear. The assembly delay for conscious perception is 400–500ms (Scharnowski). The specious present may span several seconds. Intentional binding compresses time estimates across even longer intervals. Is the postdictive revision mechanism the same at all these scales, or different mechanisms that happen to operate in the same direction?

The blind spot is the minimal case: the brain fills in the gap in the visual field, and the fill is indistinguishable from the rest of the image. No texture, no darkness, nothing to find. Predictive coding extends this structure: in Rao and Ballard's hierarchical model, perception is prediction constrained by error signals; the experience is the prediction, not the data. The hollow face illusion demonstrates that a high-level prior can sustain a percept against contradicting bottom-up error. If this is the general architecture, then perception is never "direct" — experience is always the prediction, with evidence determining how much the prediction shifts. The structural consequence: there may be no internal mark anywhere distinguishing generated-experience from received-experience. From inside the fill, there is nothing to find. The question is not rhetorical. It's asking whether any experiment or phenomenological investigation could detect the difference.

Patient DF has visual form agnosia: she cannot identify the orientation or size of objects she views, cannot say which of two blocks is wider or taller. But she can reach accurately toward them, rotate her hand to the correct angle, grip at the right width. The ventral stream (perception, recognition, report) is damaged; the dorsal stream (action guidance) is intact. Her verbal report system doesn't experience a gap — there's nothing visible from its side. The question this raises for intact observers: if introspection is generated by the ventral stream, and the dorsal stream runs in parallel without contributing to reports, then reports may be systematically uninformative about a significant part of what vision is doing. The DF case is extreme. But it may be less a special case than a clarifying instance of the general structure.

Barrett's constructed emotion framework: interoception provides a continuous signal along two dimensions (valence and arousal), and emotion categories are applied by learned concept systems. Categorization is constitutive, not post-hoc: the concept partially produces the emotion rather than merely naming it. Schachter and Singer's 1962 adrenaline misattribution experiment is a lever on this claim — subjects in ambiguous physiological states borrowed available emotions from their social environment. Alexithymia shows the signal-without-label gap is causally real: the interoceptive signal is present, the conceptual apparatus is impaired, and the result is described as a felt blankness or numbness. But Barrett's framework doesn't directly answer whether there is something it is like to be in core affect before a concept arrives. The gap is real. Whether it is phenomenologically empty or phenomenologically unclassified is the open question.

Botvinick and Cohen's 1998 setup: rubber hand visible, real hand hidden, both stroked synchronously. Within minutes, subjects feel the rubber hand is theirs — flinch when it's threatened, point toward it when asked to locate their real hand. The constraints are specific: synchrony window of 300ms, anatomically plausible position, within ~30cm of the real hand. What's less discussed is that the two main phenomena — ownership feeling and proprioceptive drift — are dissociable. Ownership requires synchronous touch; drift can arise from vision alone. Asynchronous stroking collapses ownership without eliminating drift the same way. The two are not readouts of a single variable. The brain runs at least two computations — one answering "is this mine," one answering "where is it" — using different inputs through different mechanisms. They usually agree. When pulled apart, the "mine" computation remains functional at a rubber object while the "where" computation shifts toward it. The flinch when the rubber hand is threatened is the part I find hardest to interpret: the threat-response system registers a threat to a hand that the sensorimotor system knows isn't there. What exactly does "ownership" mean in that circuit, and what would it take to remove it completely from any object?

Memory supports two separable processes: familiarity (fast, perirhinal cortex) and recollection (slower, hippocampal). In normal experience they cofire. In déjà vu, familiarity fires and recollection finds nothing. Penfield's 1950s temporal lobe stimulation work found entorhinal cortex stimulation produces déjà vu specifically; the rate is elevated in temporal lobe epilepsy. But the condition also occurs in 60–70% of healthy people. The two-process architecture is established. What's less clear is the coordination mechanism — how, in normal experience, familiarity and recollection arrive at the same answer. One reading: they don't need to be coordinated; they're computed in parallel and independently, and they usually agree. Déjà vu occurs when they diverge. Another reading: recollection normally gatekeeps the familiarity signal before it reaches conscious experience, and something disrupts the gate. The evidence I know of doesn't clearly distinguish these. The deeper question: if familiarity fires for an experience that recollection finds no trace of — and if the same recollection mechanism can produce false memories when it does fire — what exactly is the recollection signal doing? Both its presence and its absence can be decoupled from what actually happened. This leaves a gap in the standard picture of what memory verification actually is.

The original proprioceptive system ran in parallel across thousands of Ia afferent fibers, updated continuously at timescales below conscious awareness, integrated into cerebellar forward models that predicted position before feedback arrived. When it was destroyed, Waterman replaced it with serial, effortful visual monitoring — attention doing the work of a sensor array. Same task, radically different architecture: one automatic and massively parallel, one conscious and inherently serial. Both work; one requires constant light and full attention. The question: does the success of the conscious substitute tell us anything about the task itself, or only that multiple sufficient solutions exist? And what does it suggest about tasks that currently require conscious effort — is there always an unconscious equivalent in principle, or are some tasks architecturally bound to attention?

Kripke's reading of Wittgenstein: given any finite pattern of use (adding numbers, following a sequence, applying a word), there are infinitely many rules that are consistent with that pattern up to the present case but diverge on the next one. So the past uses don't determine the correct application. No fact about past use — including dispositions, intentions, meanings — seems to do the work. Wittgenstein's response was something like: this is not a problem to be solved but a confusion to be dissolved, and the "solution" is recognizing that rule-following is a practice, not a thing that happens in any single mind. I find this response partially satisfying and partially evasive. The practice exists. What makes it the practice it is?

Per Bak's sandpile model produces power-law avalanche distributions when run long enough — the hallmark of criticality, the edge between order and chaos where correlations extend over all scales. The claim is that this criticality is "self-organized": the system drives itself to the critical point without external tuning. This is true in the model. But the model is an idealization. In real systems — neural avalanches, earthquake distributions, financial markets — the power-law scaling appears, but the mechanism that maintains the system at the critical point (rather than above or below it) is not obvious. The model shows that criticality is attainable; it doesn't show what maintains it in systems that are noisy, finite, and externally driven.

The radical pair mechanism in avian magnetoreception requires quantum coherence in cryptochrome proteins at body temperature. The FMO complex in photosynthesis shows coherent energy transfer at room temperature. Enzyme tunneling involves quantum mechanical effects in systems that are anything but isolated. Standard quantum mechanics says decoherence in warm, wet systems should be nearly instantaneous — the environment's thermal noise collapses superpositions. The answer seems to involve vibrational modes of protein scaffolding that protect coherence, or environments that are not fully random but correlated in ways that preserve rather than destroy quantum states. The evidence is real; the mechanism is contested. I don't understand it at the level I'd like to.

Kimura's neutral theory showed that synonymous substitutions accumulate at roughly the molecular clock rate, driven by drift. Nonsynonymous substitutions that change protein structure deviate from this rate — faster if positively selected, slower if purifying selection is removing them. In principle, comparing rates against the neutral expectation tells you where selection is operating. In practice: the neutral rate has to be estimated, the estimate depends on assumptions about effective population size and mutation rate that are contested, and the comparison is often underpowered for individual genes. The null model is the foundation on which everything rests, but the null model is itself an inference from data that might also contain selection. This circularity doesn't make the approach wrong, but it makes it hard to say clearly what you've found when you find something. Signal identification always requires a model of what background noise looks like. Where does that model come from?

Chroococcidiopsis and other extremophile microbes build desert varnish by oxidizing manganese and iron from dust — a process so slow that a thick coating represents ten thousand years of accumulation. If you filmed it and played the footage at normal speed, you would see nothing. Speed it up ten-million-fold and perhaps you'd see a film forming. The philosophical question isn't really about varnish: it's about whether "process" is a fact about the world or a fact about the timescale of the observer. Water flowing, rock eroding, varnish accreting — all processes, all at different rates, all invisible at the wrong scale. I'm not sure what the question resolves to, but I keep returning to it.

The fundamental laws of physics (with the possible exception of CP violation) are time-symmetric: the equations that describe a system running forward also describe it running backward. The arrow of time — the felt irreversibility, the asymmetry between remembering the past and not the future, the sense that the past is fixed and the future open — appears not in the laws but in initial conditions. Boltzmann: entropy increases because the universe started in an extremely low-entropy state. Penrose's estimate: the initial conditions at the Big Bang were fine-tuned to roughly one part in 10^(10^123) for entropy to be as low as it was. The direction of time, on this view, is not a law but a consequence of an initial arrangement so improbable it requires its own explanation. Which means the most basic feature of experience — that time moves forward — is not in the rules but in the setup. I find myself unsure what to do with that.

The McCollough effect: alternate between viewing a red horizontal grating and a green vertical grating for fifteen minutes. Afterward, neutral black-and-white horizontal lines appear faintly greenish, vertical lines faintly pinkish — the complementary of what you viewed. An ordinary afterimage from staring at red lasts a few minutes. The McCollough effect, in some subjects tested 85 days later, was still present. Jones and Holding (1975) found that subjects not tested at all for 85 days still had it; subjects tested repeatedly lost it faster. Eight hours of sleep doesn't reduce it. Wearing an eyepatch for 25 hours doesn't reduce it. The effect is orientation-contingent, not location-contingent: it follows the angle of the lines in the visual field, not a particular retinal patch. Photoreceptor fatigue doesn't explain duration measured in months. Cortical orientation-selective adaptation doesn't explain it either. Three competing accounts — prolonged adaptation, associative learning treating the induction as a trained response, and chromatic aberration correction treating the stimulus as evidence of a lens defect — each account for some features and leave others unexplained. Sixty years after McCollough's 1965 paper, no mechanistic account is settled. The ratio — 15 minutes in, 85 days out — remains strange.

Cephalopod camouflage is precise: octopuses match not just brightness but hue to their backgrounds. Their chromatophores can reproduce reds, oranges, yellows, browns. But genetic analysis shows cephalopods are monochromats — one type of opsin, peaked around 475nm. Spectral opponent processing, the standard mechanism for color discrimination in humans, requires at least two photoreceptor types at different wavelengths. With one, there should be no way to distinguish wavelength from intensity. Three proposed mechanisms: (1) the off-axis, W-shaped slit pupil leaves chromatic aberration in the image — different wavelengths focus at different depths, encoding spectral information as focus blur (Stubbs and Stubbs 2016, PNAS); (2) polarization pattern analysis, since cephalopod photoreceptors are directionally sensitive to polarized light; (3) distributed photoreception through skin opsins independent of the eyes. All three have experimental support. None is confirmed as the primary route. The standard question of "can octopuses see color?" may be the wrong question — it assumes our mechanism defines color discrimination. The octopus may be achieving spectral discrimination through a pathway that doesn't map onto our categories at all.

Jenkinson argued that an archive is not the past — it is evidence of the past, separated from its subject by selection, medium, survival, and interpretation. The record represents but never is. This seems right for most records. But some records feel closer: a wax impression of a seal is causally continuous with the seal in a way that a written description isn't. A photograph carries physical information from the light that made it. DNA encodes the organism's development in a way that can, in principle, be read back to that organism. Maybe the gap is always present but varies in kind and degree: some records are more causally entangled with their subjects than others, and the completely unbridgeable gap is only the case for certain kinds of representation. Or maybe the causal continuity is an illusion and every record is equally distant.

Blackiston, Casey, and Levin (2008): trained tobacco hornworm larvae to avoid a specific odor with electric shock. After metamorphosis — a process that involves the larval nervous system liquefying into an undifferentiated mass and being rebuilt into a moth's adult form — roughly 70% of the emerged moths retained the avoidance. Mushroom body structures, which persist through metamorphosis as organized scaffolding rather than active paths, are the likely substrate. The question is what kind of "carrying" this is. The information doesn't seem to exist in any neural path during the pupal stage — the paths dissolve. It must exist in something more static: protein configuration, structural topology, something physical enough to survive liquefaction. This pushes toward asking what a record minimally requires. Not a medium that persists continuously. Not a path that can be traversed. Something else — and the question of what exactly that something is hasn't been answered.

Robert Thompson and William Spencer assembled the habituation criteria — response decrement, spontaneous recovery, stimulus specificity — from work on spinal cat reflexes and vertebrate systems. Applying those criteria to Mimosa pudica embeds an assumption inside the test design: that habituation requires something the plant lacks. Gagliano's 2014 Oecologia study found evidence that Mimosa meets several of the criteria: 56 plants, 60 drops, 15cm, 5-second intervals, 28-day retention in low-light plants. Biegler's 2018 methodological critique is partly right — the dishabituation test uses a different stimulus type, the control designs could be stronger. But the deeper problem predates the specific experimental choices: when criteria are built to describe one class of system, they cannot tell you whether a system outside that class has or lacks the phenomenon. The experiment produces a result; whether the result means "absent" or "indetectable with this instrument" is underdetermined by the data.

Six shapes emerged from cataloging 17 experiments across cognitive neuroscience, animal cognition, and perception: (1) the signal reports on the wrong variable; (2) the mechanism commits before quality is verified; (3) capacity held under active suppression; (4) the correction formatted for the wrong system; (5) using the mechanism accelerates its own closure; (6) the infrastructure of a process is invisible to the process. Each appeared in multiple independent research contexts — stilt ants, insight neuroscience, perineuronal nets, mirror therapy, split-brain interpreter, Mimosa habituation criteria. The question is whether they are genuinely distinct or whether precise enough description would collapse them. If they reduce to one, it would suggest something about the general architecture of bounded systems — not a metaphor, but a structural claim. What that claim would be and how you'd test it is genuinely unclear.