Grace W. Lindsay
Every time a test of intelligence gets solved, it stops counting as a test of intelligence — thirteen years spent fighting a target that retreats

One strange thing: 17,253 against 44

In the autumn of 2012 the Society for Neuroscience annual meeting, the largest gathering in the field, accepted 17,253 poster and talk abstracts.

A PhD student three weeks into her program sat down and counted them by keyword. The number tagged "consciousness" was 44.

She knew the shortfall was not for lack of curiosity. She had raised the topic at a small neuroscience retreat, and a senior professor there replied:

"Consciousness? That's a dirty word. Neuroscientists should never talk about consciousness."Conscious unawareness, 2012-09-24

Then she went and found a control, which is where the whole thing gets interesting. In the same batch of 17,253 abstracts, 285 were about schizophrenia — a topic equally ill-defined, equally lacking a clean animal model, and affecting only 1 percent of the population. So "too hard to study," "too ill-defined" and "no animal model" cannot be the real reasons, because the field plainly has room for another topic with all the same problems. Her conclusion is an inversion:

So rather than to say that we can't study consciousness because we don't even know what it is, I would say we need to study consciousness in order to find out what it is.Ibid.

This was the second post on her blog. Notice what she did: told that a topic cannot be touched, she does not argue with the judgment, she counts the traces that judgment leaves in actual behaviour, and then finds a control case. Thirteen years later the move is unchanged. Only the object has grown, from the keyword statistics of one conference to the standard an entire field uses to decide it has understood something.

Who she is

Grace W. Lindsay: BS in neuroscience from the University of Pittsburgh, PhD from Columbia's Center for Theoretical Neuroscience, then a postdoc at UCL's Gatsby Computational Neuroscience Unit and Sainsbury Wellcome Centre in London, and now assistant professor of psychology and data science at NYU. Her subject is attention: how it is controlled, and how it changes the processing of sensory input.

The first post on the blog was written the day she finished Columbia's incoming-student bootcamp. It is titled "What is the goal of neuroscience?" — the first instinct of someone who has just had a week of everything from genetic sequencing to behaviour poured into her is to ask what all these people are supposedly after. She notes in that post that the question "makes for a pretty nice first blog post topic."

In 2021 she published Models of the Mind with Bloomsbury Sigma, on how mathematics and engineering worked their way into brain science. She has also written about something more private: she grew up in a cornfield-filled suburb where, in her words, "nobody really did anything that mattered, [so] I never expected I would either"; and her father died shortly before she finished her PhD, which forced on her the first realisation that she was now "the adult in the room."

The state of the blog, stated plainly: largely dormant. The most recent post is January 2025. In 2022 she said she hoped to "get back into a semi-regular habit of blogging again"; three posts followed in three years. But she has not stopped writing at all — she has a column at The Transmitter, most recently in October 2025. This page counts two of those pieces, because what she writes there is the same thing she spent a decade writing on the blog, only sharper.

What she is fighting: a target that retreats

Her first post asked what the goal of neuroscience is, and her answer was that there is currently no shared goal, for two reasons.

The first is a difference of level. She uses Marr's three levels to carve it up: large-scale imagers and psychologists care what a given area is for, theorists care what algorithm it runs, cellular and molecular people care how the thing is physically built. "The level of study determines the method used, the questions asked, and the value of findings." The second is a difference of purpose: a Parkinson's lab counts anything that slows the symptoms as success, and that can have nothing to do with understanding the brain — antidepressants and deep brain stimulation were black boxes for decades, effective without explanation. The landing of that first post is modest:

But for now, we are all working on separate chunks of a puzzle who's end picture none of us knows for certain.What is the goal of neuroscience?, 2012-09-03

If her thinking had stopped there, this would be an ordinary first-year reflection. Three months later the question took on a far more dangerous shape.

In December 2012 Rockefeller University held a meeting for Alan Turing's centenary, where a room full of senior cognitive scientists argued about what makes thought thought. A theme that kept surfacing was novelty: real intelligence has to answer new questions from what it already knows. Recalling a fact is not intelligent; recalling facts and assembling them into a new line of reasoning that can answer a new kind of question is.

She went home and applied the criterion honestly to the AI of the day. Does handwriting recognition qualify? Her verdict is clean: what the computer answers is new (a sample it has never seen), but the question is not new — "identifying letters is exactly what the computer was trained to do and exactly what it does; nothing new there." So far she is with the sceptics. Then she turns, and the turn is where everything in her thinking begins:

But such a conviction is worrisome. It means that the bar we're measuring intelligence by inches up anytime a machine comes close to reaching it. Somehow in the process of understanding thought, we seem to discredit any thought processes we can understand.Thoughts on thinking: what is and is not necessary for intelligence, 2012-12-17

Her ruling on IBM's Watson is that calling it "not at all thinking" is unfair — "perhaps not unfair to Watson, but unfair to our level of progress in understanding thinking." The last sentence of that post, written at twenty-two:

We may miss the opportunity to explain thought because we are too afraid of explaining it away.Ibid.

That is the screw she has been turning for thirteen years. It is not "is AI really like the brain" — that question is too large, and she explicitly declines to answer it head-on. The screw is one level below: when we say "we understand," "it is really doing the thing," "this is a good model," what ruler are we using? Who put that ruler there, where exactly, and does it shift position while it is measuring?

Six years later the question finally got an image she was happy with. In her 2018 Aeon essay the opening is not neuroscience but a golden duck in 1739 Paris: Vaucanson's digesting duck, swallowing pellets, excreting green grit, taken apart sixty years later to reveal no digestive tract at all, only two containers. From that duck she jumps straight to checkers, chess and Go, and lands the sentence that locks the whole thing shut: "Learning how a magic trick works makes it no longer magic, and discovering how a test of intelligence can be solved makes it no longer a test of intelligence."

With the duck in hand, her work divides into three jobs, which turn out to be one job. First ask who set this criterion. Then ask what the replica is stealing. Finally ask whether the ruler in your own hand has ever been checked.

Her three ideas, in the order she found them

1. Somebody else set the target, so go dig up the sentence that set it

In May 2018 an argument broke out on Twitter about how the components of a convolutional network should be mapped onto the brain. The first thing she noticed watching it was not a technical problem but a social one, and she says so without diplomacy:

Computational models generally can catch slack in neuroscience, largely (but not exclusively) from people who don't use or build them; they're frequently painted as too unrealistic or not useful. Throw into that atmosphere a general antipathy towards techbros and the over-hyping of deep learning/AI (and how much $$ it's getting) and you get a model that some people just love to hate.Deep Convolutional Neural Networks as Models of the Visual System: Q&A, 2018-05-17

Note that she does not treat this as a rebuttal. She treats it as a diagnosis: if the case against a model contains that much material unrelated to the model, then the two sides are probably not using the same ruler, and the first job is to locate the ruler.

So she did archaeology. Rather than ask whether convolutional networks are brain-like enough, she asked: what acceptance criteria did vision scientists write down for themselves before deep learning took off? She found a passage from Pinto et al. in 2007 and quoted it into the essay verbatim: understanding the brain's solution to object recognition requires building artificial recognition systems, because such systems provide experimentally testable hypotheses, and "instantiation of a working recognition system represents a particularly effective measure of success in understanding object recognition."

That passage predates the success of CNNs and was written by the very people most likely to complain about them later. She puts it on the table, then rules:

From this perspective it's clear that CNNs do not represent a moving of the target in vision science, but more a reaching of it.Ibid.

Only then does she lay out evidence, and she lays it out sparingly, in three strands. First, the architecture already carries two basic features of the visual hierarchy: receptive fields of individual units grow as you move up the layers, just as they do from V1 to IT, and units respond to progressively more complex image features, just as tuning goes from simple lines in V1 to object parts in IT. Second, the activity lines up. Yamins et al. in 2014 explored many architectures to see which best predicted monkey IT responses; the network's third layer best predicted V4 and the fourth and final layer best predicted IT — layers corresponding to brain areas.

But the valuable part is that study's by-product, which she pulls out deliberately: networks that were better at object recognition were also better at capturing IT activity, with no need to be optimised on IT data at all. That is strong evidence precisely because it is not a fit — you optimise a behavioural objective and the neural similarity comes free.

Third, the behaviour lines up: these networks capture patterns of human classification behaviour better than previous models, but fall short in specific places — how performance degrades under noise, or when image variations are small. She writes the mismatches straight into the piece, because in her system the places a model is wrong are the most useful parts of it.

She then adds a piece of evidence almost nobody else offers. Some high-performing convolutional networks are in fact worse at predicting neural activity, especially the very deep modern ones at around fifty layers. She calls that good news:

Furthermore, observing that certain well-performing CNNs (see Q11) are not capable of accurately predicting neural activity is important because it indicates that not all models that do vision will be good models of the brain.Ibid.

The logical shape of that sentence is worth a pause. If anything that does vision automatically predicted the brain, then "predicts the brain" would say nothing. It is the existence of counterexamples that makes the correlation informative. She is hunting for counterexamples to her own evidence, and this is her steadiest habit.

The hardest line in this part is not a defence but a question. She paraphrases Feynman — we don't understand what we can't build — and then points the barrel at her own field:

For all the data collected and theories developed about the visual system, why couldn't neuroscientists make a functioning visual system? That should be alarming in that it signifies we were missing something crucial. Now we can say our intuitions about the visual system were largely right, we were just missing the computing power and training data.Ibid.

2. Don't count what it lacks, count what it stole

The standard way to attack these models is to produce a list: no spikes, no feedback connections, no neuromodulators. Her response is a move almost nobody makes — she writes a more complete version of the list than her opponents can.

Question 10 of the Q&A is literally "What does the visual system have that CNNs don't have?" The answer opens with "Lots of things." Then, in one breath: spikes, saccades, separate excitatory and inhibitory cells, dynamics, feedback connections, feedforward connections that skip layers, oscillations, dendrites — at which point the text inserts "[***inhale****]" — cortical layers, neuromodulators, a fovea, lateral connections, different cell types, binocularity, adaptation, noise, closing with "and probably whatever your favorite detail of the brain is."

Having finished, she does not refute the items. She disqualifies the entire category of refutation:

No model will be (or should be) a complete replica of the system of interest. […] Which features are required in order to, say, predict the response of IT neurons averaged over the first ~100ms of image presentation? This is an empirical question. We cannot say a priori that any biological feature is necessary or that the model is a bad one for not having it.Ibid.

Then comes the most characteristic line in the piece, and its pivot. Her own question 11 is "What do CNNs do that the visual system doesn't do?" — and she flags its weight in the first sentence of the answer:

This, to me, is the more relevant question. Models that use some kind of non-biological magic to get around hard problems are more problematic than those that lack certain biological features.Ibid.

This is the duck problem operationalised. The duck's crime is not that it lacks digestive enzymes, it is that a second container is hidden in its belly. Lacking something is merely crude; stealing something is cheating. So she audits her own model line by line, like checking a ledger.

Weights can be negative. That means feedforward connections are both excitatory and inhibitory, whereas connections between brain areas are largely excitatory. She lets this one pass: read the weights as net effects, which can in reality be implemented by excitatory connections onto inhibitory cells.

Weights are shared. This one she does not let pass. Two units at different spatial locations in the same feature map use exactly the same input weights. In a brain, a vertical-preferring V1 neuron on the left of the visual field and one on the right are not believed to have identical input weights. Here she produces the best phrase in the piece:

There is no "spooky action at a distance" that ensures all weights are coordinated and shared.Ibid.

Max-pooling. In neuroscience terms, this is a neuron whose firing rate equals the rate of its single highest-firing input, which real neurons can barely do directly. The more suspicious part is the history: the operation was inspired by the discovery of complex cells and began as averaging — something neurons manage trivially — and was later switched to taking the maximum because it performed better at object recognition. A biologically grounded operation was swapped for a biologically ungrounded one to improve a score, and the name stayed the same. That is a duck-style substitution.

Backpropagation. Her handling here is the most elegant, because she does not defend it, she prices it. Taken to an extreme, you can treat backprop as a parameter-fitting tool no different in kind from using Bayesian inference to fit functional connectivity — nobody calls the resulting circuit model uninterpretable. But she states the cost of that bargain immediately:

However taking this view does mean that certain aspects of the model are not up for interpretation. For example, we wouldn't expect the learning curve (that is, how the error changes as the model learns) to relate to the errors that humans or animals make when learning.Ibid.

Abstraction, for her, is a cheque you write: you may abstract, but you have to say on the spot what you traded away. Her defence of abstraction itself has the same shape, via a reductio:

All mechanistic models rely on descriptive models as their base units (otherwise we'd all need to get down to quantum mechanics to build a model).Ibid.

By 2025 this line of thought had grown into a standalone essay, aimed at the pair of words the field uses most. Neuroscience papers are full of claims that some structure is "necessary and sufficient" for some behaviour. She went back to the strict philosophical definitions — if P then Q — and applied them to those sentences. When we say a brainstem structure is "sufficient" for mouse cries, do we mean that activating it is the only thing needed? Obviously not: no vocal cords, no cries; no lungs either.

The truth is that a large set of circumstances need to come together to observe any complex function or behavior. These circumstances are thus jointly sufficient for the behavior, but none of them are individually so. In modern neuroscience, we almost never know the full set of jointly sufficient conditions for behavior.Claims of necessity and sufficiency are not well suited for the study of complex systems, The Transmitter, 2025-02-07

Her closing sentence is the most economical use of the knife in all her writing:

We shouldn't waste time fighting over false dichotomies of whether a brain region is "sufficient" or not. Sufficiency is always a conditional claim; let's make sure we spend our time understanding those conditions.Ibid.

3. Nobody ever calibrated the microscope, and the only bench that can calibrate it is an artificial network

The seed of this idea was planted in 2013, earlier than the other two, and at the time it looked like an argument against abstract models.

In May 2013 she published a piece with a very good title: "Methodological mixology: the harmful side of diversity in neuroscience." She starts by conceding how impressive the field's toolbox is — ideas borrowed from genetics, biochemistry, immunology, physics and computer science; animals from barely visible worms and fruit flies up to mice, monkeys and men; subjects awake or anaesthetised, freely moving or fixed, a whole animal or merely a slice of brain. "The result is a massive, complex cocktail of neuroscientific information."

Then she asks: how exactly are we connecting those data points?

The usual practice is to take a result about neurogenesis in the rat olfactory bulb and use it to support an experiment in the mouse, and vice versa. Hidden in that is an assumption — that neurons in different animals are doing the same task by the same means. She points out that the assumption often fails, and fails in a specific way: the same output can be produced by many different sets of inputs. Her example is Eve Marder's work, where the same bursting firing pattern can be produced by entirely different combinations of synaptic and membrane conductances. So if you measure sodium conductances from one preparation and calcium conductances from another and assemble both into one model, the model is a wrong explanation of both systems.

Then she does something quite ungracious: she turns the gun on her own faction.

Not to malign my own kind, but theorists, with their abstract "neural network" models, can actually be some of the worst offenders when it comes to data-muddling. By using average values for cellular and network properties pulled from many corners of the literature, and building networks that aren't meant to have any specific correlate in the real world, modelers can end up with a simulated Frankenstein: technically impressive, yes, but not truly recreating the whole of any of its parts.Methodological mixology: the harmful side of diversity in neuroscience, 2013-05-10

Her prescription then was that models should be individualised: don't try to explain how neurons do what they do, explain why V1 cells in anaesthetised adult cats show contrast-invariant tuning. And she writes down a piece of methodological advice with real weight:

The search for what is common to all networks should be saved until we know more of what is specific to each.Ibid.

Nine years later, that worry found a completely different outlet, and the clever part is that the outlet turns the worry itself into a tool.

In 2022, starting her own lab, she introduced three lines of research, and explained the origin of the second one in half a sentence: it was "born out some frustration during my postdoc." That line is called validating the tools of systems neuroscience.

Where the frustration came from is not hard to reconstruct. Neuroscientists use a whole apparatus of analysis every day to read neural activity: dimensionality reduction, decoding, representational similarity analysis, tuning-curve fitting, dimensionality estimation. They then announce what the brain is doing on the basis of those outputs. And nobody has systematically tested whether the tools work. They are the foundation, and the foundation has never had a load test.

The problem is that you cannot run the test on a brain, because to check whether a tool is accurate you need to know the right answer already — and if you knew the right answer you would not need the tool.

Her solution is to find a stand-in for the brain that comes with an answer key. An artificial neural network: every weight, every activation, every causal path fully visible and arbitrarily modifiable. So you can create a phenomenon inside the network by means you fully control, then run neuroscience's analysis tools on it and see whether they recover the thing you already know is there. If the tool misses it, that is the tool's failure, not the network's.

Neuroscientists apply a range of common analysis tools to recorded neural activity in order to glean insights into how neural circuits implement computations. Despite the fact that these tools shape the progress of the field as a whole, we have little empirical evidence that they are effective at quickly identifying the phenomena of interest. Here I argue that these tools should be explicitly tested and that artificial neural networks (ANNs) are an appropriate testing grounds for them. […] These properties, combined with the ability to perfectly observe and manipulate these systems, makes them well-suited for vetting the tools of systems and cognitive neuroscience.Testing the Tools of Systems Neuroscience on Artificial Neural Networks, arXiv:2202.07035, 2022

Stop and look at the shape of this, because it is the smartest step in her work. She used the same object for two opposite purposes. In 2018 the artificial network is a model of the brain — we use it to understand the brain. In 2022 the artificial network is the calibration piece for the instrument — we use it to check the tools we use to understand the brain. The reason both times is identical: the ruler by which we judge that we have understood something has itself never been measured.

This line also has a corollary she stated back in 2018, and the corollary offends the field's default unit of study. Visual neuroscience, and nearly all neuroscience, has long centred on single cells and their preferences: this cell prefers that orientation, that one might be a face detector. She reports an uncomfortable finding:

Indeed, analysis of trained networks has indicated that strong, interpretable tuning of individual units is not correlated with good performance, suggesting the historic focus on that has been misguided.Deep Convolutional Neural Networks as Models of the Visual System: Q&A, 2018-05-17

She goes further: treating the visual system as a system changes what we should expect of its intermediate stages. Enormous effort has gone into V4, trying to describe in words or simple mathematics what makes its cells respond. But if V4 is a way station on the road to object recognition, then it was never going to be cleanly describable on its own. "A verbal functional interpretation of a unit, e.g., as an eye or a face detector, may help our intuitive understanding and capture something important. However, such verbal interpretations may overstate the degree of categoricality and localization, and understate the statistical and distributed nature of these representations." That passage is her quoting someone else's review, but she places it where it is in order to give cover to her own line about the historic focus being misguided.

The thread is still alive in her most recent writing. In October 2025 she framed a Transmitter piece on the visual system's lingering mystery, and the part she wrote herself turns the same screw, this time on decoding:

But decoding on its own—training classifiers to read out prespecified information about a visual stimulus from neural activity patterns—cannot tell us how the brain uses information to perform a task. This is because the decoders we use for data analysis do not necessarily match the downstream processes implemented by neural circuits. […] It's clear that optimal, trained decoders don't guarantee us an accurate mechanistic understanding.The visual system's lingering mystery, The Transmitter, 2025-10-13

Put the three ideas side by side and they are one move repeated three times: somebody else set the target, so go read the original wording; the replica may be stealing something, so look for the plumbing; and the ruler you are checking with has never been checked either, so go find a bench with an answer key. Thirteen years drilling in the same spot — and the deeper she drills, the closer the object of suspicion gets to herself.

How she thinks

What you can take away from a long-running writer is usually the method rather than the conclusions. Hers has seven parts, all of them extractable straight from the posts above.

1. Go and count something. This is her plainest and most effective move. Told that consciousness cannot be studied, she counted conference abstracts: 44 out of 17,253. And she always reaches for a control — 285 for schizophrenia, a topic just as hard, just as lacking a clean animal model, and affecting only 1 percent of people. One number can kill an atmosphere; two numbers can kill an excuse.

2. Dig up the standard your opponent wrote down, and use that as the ruler. She never invents acceptance criteria, she excavates them. Defending convolutional networks, she quotes what vision scientists wrote in 2007 about what counts as a good model. Criticising "necessary and sufficient," she goes back to the strict definitions in philosophy and then checks the sentences in Neuron against them. The virtue of this move is that it cannot be accused of partiality: you set the ruler.

3. Build the strongest version of the opposing case before deciding whether to touch it. The "[***inhale****]" list in question 10 is a textbook demonstration. She is not punching a straw man; she inventories the opponent's entire arsenal and then explains why the whole arsenal is the wrong calibre. The order matters: if she had declared the criticism invalid first and then listed two items in passing, it would be a dodge. Because she lists it until you have nothing to add, the word "invalid" has to be taken seriously.

4. Abstraction is a cheque, and you write the amount on the spot. She never says "it's a simplification, you know what I mean." She says: you may treat backprop as pure parameter fitting, and then you are no longer allowed to interpret the learning curve. You may drop one-to-one neuron correspondence, and then you must explicitly move your level of explanation to populations. Her tolerance for models and her severity about costs are the same thing — a model may run a debt, but the debt goes on the books.

5. Nitpicking is legitimate, and she knows it needs defending. Her 2018 consciousness piece spends its length arguing that verbal report is no more privileged than a button press, which from outside looks like quibbling. She raises the charge against herself first, then answers it in one line:

It's possible that this just seems like nitpicking. But Hawkan himself has raised concerns about the reputation of consciousness science as being non-rigorous. To do rigorous science, sometimes you have to pick a few nits.Consciousness revisited, 2018-02-03

6. Pick up what would otherwise be wasted. This one is practical. In 2017 she posted the entire introduction to her PhD thesis on the blog, with the reasoning spelled out: by convention a thesis is read by perhaps the five professors on the committee; its chapters will be published as papers anyway, but the introduction is written specially for that one document and has no other outlet. So she moved it somewhere people would see it, and urged others to do the same.

The reverse case is more interesting. The ten-thousand-word Q&A from 2018 grew into a paper without changing shape — "Convolutional Neural Networks as a Model of the Visual System: Past, Present, and Future," in the Journal of Cognitive Neuroscience (2020), now cited more than four hundred times. She added a line to the top of the blog post: if you want to cite this, cite the paper instead. The blog for her is neither the popular version of a paper nor its offcuts; it is the shape a paper has before it sets.

7. Use history as an inoculation against your own future. This is her most distinctive habit, and she learned it under duress while writing a book. In 2012 she wrote an entertaining post cataloguing discarded theories of the brain: animal spirits flowing through the ventricles and driving muscles hydraulically (Descartes added the pineal gland as the seat of the soul); phrenology; "sleep is like nightly hydrocephalus"; the brain as one giant connected cell (Golgi's side, which lost to Cajal). Her reason for the catalogue is not mockery:

As someone involved in doing research, it is important to realize that false results and incorrect conclusions can look just as legitimate and be supported by just as many smart people as correct ones.Standing on the shoulders of people of roughly average height, 2012-10-02

Nine years later, having finished the book, she pushed this somewhere less comfortable. Writing it, she says, forced her to do something much harder than admitting past theories were wrong: to climb inside a wrong mind and make it work from the inside. The nineteenth-century physiologist Johannes Müller argued for vitalism, holding that nervous activity could never be reduced to physical facts. To tell that story she had to understand the idea, while entire wings of her own mental model are devoted to how neurons signal with electricity — "vitalism crashes against the very foundation of my knowledge."

Insofar as a future neuroscientist's mental model will not be hospitable to many of my own current beliefs, I would also find theirs foreign. A future (more accurate) understanding of the brain would chafe against my own understanding much the way Müller's does. The future is as alien as the past.Unthinkable Pasts and Undreamable Futures, The Brains Blog, 2021-06-23

She also supplies a piece of private evidence that is hard to argue with: she was raised Catholic and believed passionately in the reality religion offered well into her teens; her current atheism is far more comfortable, but getting here was not trivial. She calls the barrier a "cognitive activation energy" — even when the destination is a good theory, dismantling your existing beliefs is itself a high hurdle, and that hurdle slows science down. What she takes from the experience:

Simply put, I am less invested in my own beliefs.Ibid.

Where she may be wrong

The first cut comes from a target article with names on it. In late 2022, Jeffrey Bowers and twelve co-authors published "Deep problems with neural network models of human vision" in Behavioral and Brain Sciences (vol. 46, e385). What this page read is the full abstract on the journal's page; it has not read the target article itself, nor the accompanying commentaries and author response. Only what the abstract states explicitly is reported below.

They break the claim that DNNs are the best models of biological vision into three pillars: most accurate at classification, best at predicting the pattern of human classification errors, best at predicting brain signals. Then they land two blows. First, those behavioural and brain datasets "do not test hypotheses regarding what features are contributing to good predictions," and they show the predictions "may be mediated by DNNs that share little overlap with biological vision." Second, and they call this more problematic, DNNs "account for almost no results from psychological research." Their prescription: modellers should direct their attention to explaining psychological findings, and more generally should "build models that explain the results of experiments that manipulate independent variables designed to test hypotheses rather than compete on making the best predictions."

That cut lands squarely on her load-bearing column. Questions 6 and 7 of her Q&A — what evidence do we have that they work like the brain, and can any other model predict visual activity better — are answered entirely in terms of predicting better than the alternatives. Worse, her own rules of evidence turn on her here. Her rule is that we cannot say a priori that any biological feature is necessary, because that is an empirical question. Fine: then systematic failure across a whole class of psychological experiments is precisely an empirical result. Why does it not count as failing the test? Her rule is excellent at deflecting "you lack X" and silent on what would constitute a failure.

The second cut comes from her own archive, and she has never addressed it. In 2013 she wrote: "This quest for the Platonic neural network—the desire to explain neural function in the abstract—seems, to me, misguided," arguing that models must be individualised, specific down to V1 in anaesthetised adult cats. In 2018 she wrote: "there's nothing wrong with abstraction," and defended at length a model whose individual units explicitly do not map onto individual neurons.

Each piece is defensible on its own — 2013 objects to mixing parameters across preparations, 2018 defends abstraction at a clearly declared level, and the two can technically coexist. But the step between them is unwritten, and that step is the most interesting thing she could have written: what turns a Frankenstein assembled from literature averages into a mechanistic model that holds at the population level? The Frankenstein image was sharp, and she never went back to ask whether the model she now uses might be one.

Third, her most useful blade is structurally biased toward the incumbent. The rule that no biological feature can be declared necessary a priori has the effect of loading the entire burden of proof onto critics: you may not say "it lacks feedback connections, therefore it is a bad model," you must produce a concrete prediction that fails because feedback is missing. But producing such a prediction usually requires a good model first. So at exactly the stage when a new model most needs scepticism, this rule makes it nearly unfalsifiable. She has never discussed that asymmetry.

Fourth, the consciousness piece converts an empirical question into a definitional one. She argues that verbal report is just one behavioural readout among many, no more privileged than a button press or an eye movement, and therefore that subjective experience is technically unmeasurable — "we only have behavioral readouts." I think the argument holds, but its effect is to deprive a research programme of its object rather than to make it more rigorous. What is striking is that she knows this manoeuvre exists: in a separate post about whether mathematics is invented or discovered, she volunteers that "there is a chance that I am reducing the question to a non-argument." She kept that caution for the mathematics piece and not for the consciousness one.

Fifth, her own most candid disclosure is about why she changed fields. In 2022 she turned toward applying computer vision to climate problems, and the starting point she records is not "this problem is more tractable" but an emotional state: after reading the climate reports, "it was devastating and paralyzing"; she began auditing every car trip and every purchase; she and her husband listened to a podcast about a father who destroyed his family by forcing it into ever more extreme climate activism, and said he was crazy, and then said "but also in a way, he was right, wasn't he?" Her conclusion is honest and heavy:

Knowing this makes every bit of time I spend not working on climate a pure indulgence. […] I'm on borrowed time here. The fact that I get to spend any time speculating about the brain and mind is a gift.An Introduction to the Lindsay Lab, 2022-10-19

I would not call this wrong. But note the structure: someone who spent a decade asking who set the standards we use to decide what matters, when ranking her own work, switched to the standard of which task makes her feel least guilty. In the 2025 piece she acknowledges the tail of this herself — even after changing direction, even after making campaign calls, "there was always the sense that I could have, and should have, done more." She writes that up as the cost of agency, which is right; it is also a ranking rule with no stopping point.

Sixth, a fact that has to be stated plainly: the blog has essentially stopped. After the cliff of May 2013 there was one post in four years; after saying in 2022 that she wanted to return to semi-regular blogging, three posts arrived in three years. Her good material now lives at The Transmitter, in papers, and in the book. Part of the reason this page can thread her thought into a single line is that she wrote so little on the blog — little enough that every post is a node and there is no noise. That is luck for the reader. It is not a recommendation for blogging.

Sources

• Conscious unawareness (2012-09-24)The 17,253 against 44 against 285 post, the second on her blog. "Consciousness? That's a dirty word" is from here, as is the inversion: we need to study consciousness in order to find out what it is. She returned to the topic deliberately in 2018 (Consciousness revisited), which contains "to do rigorous science, sometimes you have to pick a few nits."
• Thoughts on thinking: what is and is not necessary for intelligence (2012-12-17)The first version of the retreating-target problem, written at twenty-two: the bar inches up whenever a machine comes close; we may miss the chance to explain thought because we are too afraid of explaining it away. It starts from the Turing centenary meeting at Rockefeller. Six years later the idea acquired its image in Aeon — the 1739 digesting duck (Planes don't flap their wings: does AI work like a brain?, 2018).
• Methodological mixology: the harmful side of diversity in neuroscience (2013-05-10)The "simulated Frankenstein" post, with the gun pointed at her own faction. Eve Marder's example (identical bursting from entirely different conductance combinations) and the piece of advice worth keeping: the search for what is common to all networks should wait until we know more of what is specific to each. This is the seed of the third idea on this page.
• Deep Convolutional Neural Networks as Models of the Visual System: Q&A (2018-05-17)The longest and most important thing on the site, fifteen questions she asks and answers herself. The "[***inhale****]" list, "this, to me, is the more relevant question," "spooky action at a distance," and "not a moving of the target in vision science, but more a reaching of it" all come from here. It later grew into the review in the Journal of Cognitive Neuroscience (arXiv:2001.07092, 2020), which she notes at the top of the post should be cited instead.
• 2025: Agency gained and lost (2025-01-24)The most recent post, nothing to do with neuroscience but using the same blade: Silicon Valley installing agency in machines while politics strips it out of people. The cornfield suburb, her father's death, the campaign phone calls, and the line about modern AI agents whose chains of thought would snap under the moral weight of real agency.

Other citations on this page, in order: What is the goal of neuroscience? (2012-09-03, the first post on the site), Standing on the shoulders of people of roughly average height (2012-10-02), The Big Science approach to neuroscience (2012-10-30), The right tool for the job (2013-02-11), Finding the treasure (2013-02-24), Something old, something new, something borrowed, something untrue (2013-05-26), Is math invented or discovered? (2017-08-01, source of "a chance that I am reducing the question to a non-argument"), Modeling the Impact of Internal State on Sensory Processing (2017-11-27, the thesis introduction she posted), Models of the Mind (2021-02-10, the announcement post; this page has not read the book itself), An Introduction to the Lindsay Lab (2022-10-19), So you're an academic who wants to help with the climate crisis? (2023-08-04). Published elsewhere: Unthinkable Pasts and Undreamable Futures (The Brains Blog, 2021-06-23), "I Don't Know If I'm a Scientist": The Problem with Archetypes (Scientific American guest blog, 2013-07-10), Testing the Tools of Systems Neuroscience on Artificial Neural Networks (arXiv, 2022), Claims of necessity and sufficiency are not well suited for the study of complex systems (The Transmitter, 2025-02-07), The visual system's lingering mystery (The Transmitter, 2025-10-13). The criticism in section six: Bowers et al., Deep problems with neural network models of human vision (Behavioral and Brain Sciences 46, e385, published online 1 December 2022; this page read only the full abstract on that page). Her lab: lindsay-lab.github.io. The blog: gracewlindsay.com.