Machine learning, the branch of AI concerned with tuning algorithms from data, is an amazing field that has changed the world — and will continue doing so. But it is also filled with closed-minded egotists with too much money, and too much power.

This is a story, in three acts, spanning four decades, about how many of them tried, ultimately unsuccessfully, to keep a good idea, neurosymbolic AI, down—only to accidentally vindicate that idea in the end.

§

For those who are unfamiliar with the field’s history, or who think it began only in 2012, AI has been around for many decades, split, almost since its very beginning, into two different traditions.

One is the neural network or “connectionist” tradition which goes back to the 1940s and 1950s, first developed by Frank Rosenblatt, and popularized, advanced and revived by Geoffrey Hinton, Yann LeCun, and Yoshua Bengio (along with many others, including most prominently, Juergen Schmidhuber who rightly feels that his work has been under-credited), and brought to current form by OpenAI and Google. Such systems are statistical, very loosely inspired by certain aspects of the brain (viz. the “nodes” in neural networks are meant to be abstractions of neurons), and typically trained on large-scale data. Large Language Models (LLMs) grew out of that tradition.

The other is the symbol-manipulation tradition, with roots going back to Bertrand Russell and Gottlob Frege, and John von Neumann and Alan Turing, and the original godfathers of AI, Herb Simon, Marvin Minsky, and John McCarthy, and even Hinton’s great-great-great-grandfather George Boole. In this approach, symbols and variables stand for abstractions; mathematical and logical functions are core. Systems generally represent knowledge explicitly, often in databases, and typically make extensive use of (are written entirely in) classic computer programming languages. All of the world’s software relies on it.

For thirty years, I have been arguing for a reconciliation between the two, neurosymbolic AI. The core notion has always been that the two main strands of AI—neural networks and symbolic manipulation—complement each other, with different strengths and weaknesses. In my view, neither neural networks nor classical AI can really stand on their own. We must find ways to bring them together.

After a thirty-year journey, I believe that neurosymbolic AI’s moment has finally arrived, in part from an unlikely place.

§

In her bestseller Empire of AI, Karen Hao crisply sets the stage.

She begins by neatly distilling the scientific tension.

Hinton and Sutskever continued [after their seminal 2012 article on deep learning] to staunchly champion deep learning. Its flaws, they argued, are not inherent to the approach itself. Rather they are the artifacts of imperfect neural-network design as well as limited training data and compute. Some day with enough of both, fed into even better neural networks, deep learning models should be able to completely shed the aforementioned problems. "The human brain has about 100 trillion parameters, or synapses," Hinton told me in 2020.

"What we now call a really big model, like GPT-3, has 175 billion. It's a thousand times smaller than the brain.

"Deep learning is going to be able to do everything," he said.

Their modern-day nemesis was Gary Marcus, a professor emeritus of psychology and neural science at New York University, who would testify in Congress next to Sam Altman in May 2023. Four years earlier, Marcus coauthored a book called Rebooting AI, asserting that these issues were inherent to deep learning. Forever stuck in the realm of correlations, neural networks would never, with any amount of data or compute, be able to understand causal relationships-why things are the way they are-and thus perform causal reasoning. This critical part of human cognition is why humans need only learn the rules of the road in one city to be able to drive proficiently in many others, Marcus argued.

Tesla's Autopilot, by contrast, can log billions of miles of driving data and still crash when encountering unfamiliar scenarios or be fooled with a few strategically placed stickers. Marcus advocated instead for combining connectionism and symbolism, a strain of research known as neuro-symbolic AI. Expert systems can be programmed to understand causal relationships and excel at reasoning, shoring up the shortcomings of deep learning. Deep learning can rapidly update the system with data or represent things that are difficult to codify in rules, plugging the gaps of expert systems. "We actually need both approaches," Marcus told me.

She goes on to point out that the field has become an intellectual monoculture, with the neurosymbolic approach largely abandoned, and massive funding going to the pure connectionist (neural network) approach:

Despite the heated scientific conflict, however, the funding for AI development has continued to accelerate almost exclusively in the pure connectionist direction. Whether or not Marcus is right about the potential of neurosymbolic Al is beside the point; the bigger root issue has been the whittling down and weakening of a scientific environment for robustly exploring that possibility and other alternatives to deep learning.

For Hinton, Sutskever, and Marcus, the tight relationship between corporate funding and AI development also affected their own careers.

Hao then captures OpenAI’s sophomoric attitude towards fair scientific criticism:

Over the years, Marcus would become one of the biggest critics of OpenAI, writing detailed takedowns of its research and jeering its missteps on social media. Employees created an emoji of him on the company Slack to lift up morale after his denouncements and to otherwise use as a punch line. In March 2022, Marcus wrote a piece for Nautilus titled "Deep Learning Is Hitting a Wall”, repeating his argument that OpenAI's all-in approach to deep learning would lead it to fall short of true AI advancements. A month later, OpenAI released DALL-E 2 to immense fanfare, and Brockman cheekily tweeted a DALL-E 2-generated image using the prompt "deep learning hitting a wall.” The following day, Altman followed with another tweet: "Give me the confidence of a mediocre deep learning skeptic." Many OpenAI employees relished the chance to finally get back at Marcus.

But then again, as the saying goes, he who laughs last, laughs loudest.

§

For all the efforts that OpenAI and other leaders of deep learning, such as Geoffrey Hinton and Yann LeCun, have put into running neurosymbolic AI, and me personally, down over the last decade, the cutting edge is finally, if quietly and without public acknowledgement, tilting towards neurosymbolic AI.

This essay explains what neurosymbolic AI is, why you should believe it, how deep learning advocates long fought against it, and how in 2025, OpenAI and xAI have accidentally vindicated it.

And it is about why, in 2025, neurosymbolic AI has emerged as the team to beat.

It is also an essay about sociology.

§

The essential premise of neurosymbolic AI is this: the two most common approaches to AI, neural networks and classical symbolic AI, have complementary strengths and weaknesses. Neural networks are good at learning but weak at generalization; symbolic systems are good at generalization, but not at learning.

Since my first book in 2001, The Algebraic Mind, I have been arguing that three ideas drawn from classical symbol-manipulation are indispensable:

• “Algebraic” systems such as algorithms, equations, and computer code, in which variables, and operations over those variables can be explicitly specified. In the equation, y = 3x + 2, x and y are variables, and one can instantiate (fill in) those variables with particular values, for example computing that y = 11 if x is set to 3. That sort of abstraction is the essence of computer programs. (Much of my empirical work as a cognitive scientist went toward showing that humans, even infants, could do something analogously algebraic).

• Systems for explicitly representing structured, symbolic representations such that, for example, horse rides astronaut systematically mean something different from astronaut rides horse, and such that wholes can in general be predicted compositionally as a function of their parts.

• Database-like systems for distinguishing individuals from kinds. In 2001, in the above-mentioned book, I warned that in their absence, hallucinations would emerge as a form of overgeneralization, which I warned was “an inevitable downside”. A quarter century later, hallucinations remain ubiquitous. Within a neural network substrate, the problem still has not been solved.

These three requirements might sound obvious—especially to someone trained in both computer science and cognitive science—but for decades the field of neural networks tried to make do without them.

Instead, until very recently, the consistent move by mainstream machine learning had been to try to derive all that is needed from data, without any recourse at all to symbolic systems such as traditional computer code, databases, etc., aiming to replace explicit representation with black boxes.

Over the years, many in the machine learning field derided attempts to use symbolic tools, often ridiculing them (without genuine argument) as being “not biologically plausible” or somehow (never really specified) in-principle ineffective.

§

Most iconic and influential among the detractors of neurosymbolic approaches has been Geoffrey Hinton.

Hinton has repeatedly argued that the pursuit of symbol-manipulation—even in the context of hybrid neurosymbolic models—is a huge scientific mistake. In a 2015 workshop at Stanford, for example, Hinton gave a talk that claimed that classical AI symbols were “as incorrect as the belief that a lightwave can only travel through space by causing disturbances in the luminiferous aether.”

Another time, in 2018, Hinton told an audience of the G7 to great laughs that hybridizing neural networks and symbolic systems (which is to say neurosymbolic AI) was as foolish as sticking electricity onto a gas engine instead of just making an electric car, arguing that “all progress recently [has come from] sucking in data, not from people putting programs inside.”

(Some of this was driven by a scrabbling over funding, with Hinton alleging that investing in classical AI, and, by extension, hybrid neurosymbolic models, would be “a disaster.”)

§

Hinton’s long-time hostility against any role at all for symbols has, in my judgement, cost the field dearly. Ideas that were were only discovered in the last couple years (e.g., some discussed later in this essay) may have been discovered much later than they might otherwise have been.

Many other important ideas have likely also yet to be discovered, precisely because the Hinton path has distracted immense resources from other ideas, fostering an intellectual monoculture that, in the words of Emily Bender, has been “sucking the oxygen from the room.”

First among the big players to look more broadly was Google DeepMind, which wisely has not taken Hinton’s dogma overly seriously. AlphaFold, AlphaProof, and AlphaGeometry are all successful neurosymbolic models. We would likely have more of such innovation already, if Hinton had not so insistently shaped the modern landscape in such a narrow-minded way.

To the extent that we are stuck for now with untrustworthy LLMs that nobody can quite control, rapidly enshittifying the internet with botshit and replacing humans with systems that can’t be relied on, it is in part because we have spent too much energy pursuing the pure neural net black box approach, and not enough into looking at alternatives. We are drowning in a sea of mediocre prose precisely because LLMs, the backbone of virtually all current sytems, lack the underlying representational wherewithal to do better.

Yet nearly all funding has been aimed simply at making them larger.

§

Despite the often open hostility of Hinton and many of his followers, I have always stuck to my guns, never arguing that we should dispense with deep learning (a commonly-repeated strawman) but always calling, instead, for a hybrid of deep learning and symbols.

In my 2018 Deep Learning: A Critical Appraisal for example, I wrote

Despite all of the problems I have sketched, I don’t think that we need to abandon deep learning.

Rather, we need to reconceptualize it: not as a universal solvent, but simply as one tool among many, a power screwdriver in a world in which we also need hammers, wrenches, and pliers, not to mentions chisels and drills, voltmeters, logic probes, and oscilloscopes.

As we will see in a moment, people are finally pursuing a tool-driven, neurosymbolic vision like this — and getting better results than they were able to get with pure neural networks.

People tend to remember that paper for its criticism. There was a long discussion of the weakness of models that that were then-current, including their difficulties in out-of-domain generalization, with abstraction, and with reasoning, and their overreliance on massive data.

But one of the main points of the paper was its call for the integration of neural networks and symbols:

Another place that we should look is towards classic, “symbolic” AI, sometimes referred to as GOFAI (Good Old-Fashioned AI). Symbolic AI takes its name from the idea, central to mathematics, logic, and computer science, that abstractions can be represented by symbols. Equations like f = ma allow us to calculate outputs for a wide range of inputs, irrespective of whether we have seen any particular values before; lines in computer programs do the same thing (if the value of variable x is greater than the value of variable y, perform action a).

By themselves, symbolic systems have often proven to be brittle, but they were largely developed in [an] era with vastly less data and computational power than we have now. The right move today may be to integrate deep learning, which excels at perceptual classification, with symbolic systems, which excel at inference and abstraction. One might think such a potential merger on analogy to the brain; perceptual input systems, like primary sensory cortex, seem to do something like what deep learning does, but there are other areas, like Broca’s area and prefrontal cortex, that seem to operate at much higher level of abstraction. The power and flexibility of the brain comes in part from its capacity to dynamically integrate many different computations in real-time.

§

The leaders of deep learning hated me for challenging their baby, and couldn’t tolerate any praise for the paper. When an influential economist Erik Brynjolfsson (then at MIT) complimented the article on Twitter, (“Thoughtful insights from @GaryMarcus on why deep learning won't get us all the way to artificial general intelligence”), Hinton’s long time associate Yann LeCun tried to contain the threat, immediately replying to Brynjolffson publicly that the paper was “Thoughtful, perhaps. But mostly wrong nevertheless.”

LeCun was never able to articulate his concerns, but his remark wasn’t about the science. Instead, it was a signal to thefield that my views should be rejected; literally hundreds of people piled on. (Few seem to have noticed the irony that came a few years later, when LeCun ultimately came around to making almost exactly the same points I was making then, declaring, e.g., that pure LLMs are not an adequate route to AGI, and emphasizing their limits in reasoning.)

§

The story was similar when my 2022 article Deep Learning is Hitting a Wall came out. The first part of the paper was negative argument against a hypothesis that was then popular: “Scale is all you need.”

I argued that scaling as it was then defined (pretraining data and compute) alone was insufficient to solve challenges with reasoning, misalignment and hallucinations. (Spoiler alert: it still hasn’t.)

The second part was an argument that neurosymbolic AI might be a way out of this mess:

Early pioneers, like John McCarthy and Marvin Minsky, thought that one could build AI programs precisely by extending these techniques, representing individual entities and abstract ideas with symbols that could be combined into complex structures and rich stores of knowledge, just as they are nowadays used in things like web browsers, email programs, and word processors. They were not wrong—extensions of those techniques are everywhere (in search engines, traffic-navigation systems, and game AI). But symbols on their own have had problems; pure symbolic systems can sometimes be clunky to work with, and have done a poor job on tasks like image recognition and speech recognition; the Big Data regime has never been their forté. As a result, there’s long been a hunger for something else.

That’s where neural networks fit in [solving problems from data where rules are hard to construct]…

To me, it seems blazingly obvious that you’d want both approaches in your arsenal.

… “hybrid models” that incorporate elements of both deep learning and symbol-manipulation… To think that we can simply abandon symbol-manipulation is to suspend disbelief.

Animosity, rather than genuine intellectual engagement, was again the main result. Thousands of people, from Altman to Musk, insulted me and the paper (Hao listed two of the many examples). LeCun wrote on Facebook, in May 2022, a few months before he turned against LLMs, “Not only is Al not "hitting a wall", cars with Al-powered driving assistance aren't hitting walls, or anything else, either.” Musk circulated a meme that mentioned me which featured a cartoon of deep learning-powered robot smashing over walls and buildings. (An accidental forewarning of all the damage that recklessly-rolled out LLMs have started to cause?)

§

Three years later, the pure scaling of pretraining data (and compute, which is to say more and more GPUs) simply hasn’t worked. The specific obstacles that I dwelled on, with respect to hallucinations, misalignment and reasoning errors, have not been overcome.

This started to became clear in November. One of the first insiders to publicly acknowledge it was Marc Andreesen, who said in a November 2024 interview that multiple models were “sort of hitting the same ceiling on capabilities”; Microsoft CEO Satya Nadella soon more or less acknowledged the same (“in the last … weeks there [has been] a lot of debate [on whether we] have we hit the wall with scaling laws”).

§

What has worked—to some degree—is importing some of the ideas from neurosymbolic AI, including using purely symbolic algorithms as a direct part of the workflow.

OpenAI quietly began doing this to some degree in 2023 with a system called “code interpreter”, in which LLMs (themselves neural networks) call (purely symbolic) Python interpreters. This is literally “putting programs inside”; exactly what Hinton said was a huge mistake. And they are innately building in that capacity, rather than learning it.

Sometimes, not always, when a system like o3 calls Python, the symbolic code is actually reported explicitly:

from PIL import Image, ImageDraw
# Grid size for 60 fruits
rows, cols = 6, 10 #6 × 10= 60

radius = 45
spacing = 25

canvas_width - cols * (2 * radius + spacing) + spacing
canvas_height = rows * (2 * radius + spacing) + spacing

img = Image.new ("RGB", (canvas_width, canvas_height), "white")
draw = ImageDraw. Draw(img) 

# Colors
apple_color = (220,0,0)#deepred
…

As leading machine learning expert Francois Chollet explained last year on X, such systems are manifestly neurosymbolic,

“Obviously combining a code interpreter (which is a symbolic system of enormous complexity) with an LLM is neurosymbolic. AlphaGo was neurosymbolic as well. These are universally accepted definitions.”

Recent “reasoning” models borrow even more from classic symbolic approaches, such as techniques like search and conditionally iterating through multiple solutions, and aggregating the results, techniques that traditional networks had long eschewed, in favor of a single feed-forward pass through a neural net.

From what I can tell, most modern models also make heavy use of data augmentation that involves (inter alia) running systems of symbolic rules and training on their outputs, a far cry from the original notion of training on naturalistically-observed data and then inducing whatever needed to be learned. (Details on how all this is implemented are sketchy, with companies like OpenAI saying almost nothing about implementation. DeepSeek has been a notable exception, explicitly acknowledging the use of symbolic rules in generating their training data.)

GPT-2 was a pure LLM; with no direct use of symbol-manipulation. A lot of models since then haven’t been. GPT-4 probably wasn’t (I suspect it included some symbolic filters in its guardrails), and o3, when it invokes code interpreter and symbolically-executed control structures certainly isn’t.

Intriguingly, o3 itself seems to know (pardon my anthropomorphic shorthand) a bunch about this, recognizing that it needs to draw on the (symbolic) code interpreter for certain kinds of problems:

§

Both informal experimentation and a number of recent quantitative results make it abundantly clear that current generative AI systems perform better when they avail themselves of symbolic tools. (Again, per Chollet, when you combine an LLM with a symbolic tool, you have a neurosymbolic system.)

The much-discussed Apple paper was one hint: the authors explicitly forced LLMs to do various tasks, such as Tower of Hanoi, without recourse to using code (symbolic by definition), and it’s under that scenario that they found breakdowns, such as failures on Hanoi with 8 rings.

Another research group, explicitly responding to the Apple paper, showed how you could get much better performance on problems like Hanoi by having models like o3 explicitly invoke code.

I have found similar results in my own informal experiments, when looking at algorithmic processes. ChatGPT (which doesn’t as far as I can tell) currently use Code Interpreter, struggles mightily to draw crosswords grids, as Haym Hirsch recently noted, making weird errors like reporting that 4 down here is “REACT” when it as actually an illegal word, “RCRCT”:

S T A R T 
T R A C E
A C O R N 
R E A C T 
T E N T S

In contrast, o3, when it draws on symbolic code (which it sometimes displays explicitly) can build crossword grids like this far faster than I could (short of my writing my own code).

T R A S H
R E P A Y
A P P L E
S A L O N
H Y E N A

§

The results that seal the deal came a few nights ago, at the launch of Grok 4, perhaps the largest and most expensive model in history. (Elon Musk claims it used 100 times the compute of Grok 2.)

This graph (of a challenging benchmark known as Humanity’s Last Exam) is without question one of the most vindicating things I have ever seen:

Why? Two things can be seen.

First, the lower set of data (yellow dots) represent a strong test of the pure scaling hypothesis. Although the units on the X axis aren’t made clear, it’s quite clear that pure scaling of training data and computer is reaching a point of diminishing returns, a long way from peak possible performance. Contrary to lots of talk from previous years, increasing compute alone is not driving some miraculous exponential explosion into superintelligence.

Second, adding symbolic tools (orange dots) dramatically improves performance.

(Scaling test time is also, as far as I understand it, drawing on classic symbolic techniques for iteration and aggregation.)

You can see this in other benchmarks, too. On this math competition, adding symbolic tools make a big difference:

In short, although LLMs are still far from perfect, enhancing them with symbolic tools, once an anathema, makes a huge difference. Between the strong success of DeepMind’s series of neurosymbolic hybrids (AlphaGo, AlphaFold, AlphaProof, AlphaGeometry, etc.) and the more recent results in which LLMs have been enhanced with python and other tools to dramatic effect, we can safely conclude that neurosymbolic AI is on track to be a major part of AI’s future.

§

None of which means we are close to AGI. There are many ways to put together symbols and neural nets. I am not at all convinced we have the right one.

Bolting code interpreters onto LLMs has value, but whether it gets us to AGI is a different matter. A neurosymbolic substrate for AI by itself (as I argued in 2020) is likely to be necessary but not sufficient. Advances in neurosymbolic AI are likely to be just one part of a larger picture. As Peter Voss has argued (and I agree) we need far more integration than a mere bolt-on would provide. Including LLMs somewhere in the next evolution of AI makes sense to me, but leaving them at the core may be a mistake.

Models like o3 are far better than non-code-interpreter systems on some tasks but not all (for now o3 is actually worse than o1 on hallucinations) and effective for some tasks, but not others. Core problems like “symbol-grounding” have not been systematically solved, for example. Explicit cognitive models (as I discussed recently in this newsletter) are also likely to be essential, yet remain very much underexplored. Reasoning is still unreliable; spatial reasoning is a long way from solved, as Fei-Fei Li has recently argued. We still also (just as I argued in 2020) lack proper ways of building and inferring symbolic world/cognitive models on-the-fly. And we may need more discipline in how we tie semantics to the symbolic components of our systems. There is likely still a great deal of work left to be done.

Reality is maybe a bit like a cartoon I recently posted on X. A hiker climbs across a mountain range, impressively far along (“you are here”), but nowhere near the highest peak (“AGI”). We have made some progress; we still likely have a long way to go.

§

Still, even at our current vantage point, having solved some problems but by no means all, there are several lessons we can draw:

• Systems that draw on code interpreter, and hence by definition are neurosymbolic, often outperform those that lack symbolic tools, at least in a bunch of tasks. This is confirmation of the value of neurosymbolic AI.

• Hinton’s admonition against using symbolic AI was extraordinarily misguided, and probably delayed discovery of useful tools like Code Interpreter by years. Largely because of Hinton’s influence, the field still talks about neurosymbolic AI as if it were a dirty word. That’s likely holding back the science.

• It’s likely that OpenAI, Anthropic and others are already doing more neurosymbolic AI than they actually make public. They still pretend it’s all “a model” but that model uses all kinds of tricks and mixtures with tools, code generation etc.

• A lot of recent improvements are likely coming from improvements in using symbolic tools, rather than scaling. Massive infrastructure investments like Stargate are likely based on misleading impressions of what is actually driving progress.

• It would appear that it is only recently, in desperation, as scaling started to reach diminishing returns, that mainstream machine learning began to broaden its vistas.

• Where the field is getting stuck now is likely on how to scale these new, neurosymbolic tools and how to make models “understand” when to use what and how to logically combine the context. This is why so called “code agents” inevitably fall apart as soon as one gives them a more complex task which requires reasoning.

• The code construction itself is still produced by a system that is highly dependent on similarity to training examples. In easy, familiar cases, the results can be astonishing. In problems that diverge from easy and familiar cases, the constructed code can be buggy, or altogether flawed. As Phil Libin put it to me, “trying to improve the results [of o3] by continuing to talk with it rarely works after the first few prompts… It’s either gonna get it right in 5 minutes, or never.”

• Neurosymbolic AI is not one thing, but many. o3’s use of neurosymbolic AI is very different from AlphaFold’s use of neurosymbolic AI. Very little of what has been tried has been discussed explicitly, and because the companies are often quite closed about what they are doing, the public science of neurosymbolic AI is greatly impoverished.

• Getting to AGI will likely take still more breakthroughs. The best way to foster them is to have an intellectually open attitude. It would be great if we could see more of that.

§

Why was the industry so quick to rally around a connectionist-only approach and shut out naysayers? Why were the top companies in the space seemingly shy about their recent neurosymbolic successes?

Nobody knows for sure. But it may well be as simple as money. The message that we can simply scale our way to AGI is incredibly attractive to investors because it puts money as the central (and sufficient) force needed to advance.

Admitting that they need to rely on neurosymbolic tools would pierce the scaling narrative.

§

So here’s where we are: pure LLMs can’t reliably apply algorithms (vindicating the line of argument I first developed in the 1990s); if you enhance them with symbolic processes—yielding neurosymbolic systems—they often give better results.

OpenAI, without any sort of of public acknowledgement whatsoever, has accidentally vindicated neurosymbolic AI.

Fostering its further development may be among the best things that companies, researchers, and governments can do. Investors, take note.

Gary Marcus has been pushing for neurosymbolic AI throughout his career, since the early 1990s, and could not be more thrilled to see it start to blossom.

Subscribe now

There’s a great anecdote about Roman Jakobson, the structuralist theorist of language, in Bernard Dionysius Geoghegan’s book, Code: From Information Theory to French Theory. For Jakobson, and for other early structuralist and post-structuralist thinkers, language, cybernetic theories of information, and economists’ efforts to understand how the economy worked all went together :

By aligning the refined conceptual systems of interwar Central European thought with the communicationism of midcentury American science, Jakobson envisioned his own particular axis of global fraternity, closely tied to forces of Western capitalist production. (He colorfully illustrated this technoscientific fraternity when he entered a Harvard lecture hall one day to discover that the Russian economist Vassily Leontieff, who had just finished using the room, had left his celebrated account of economic input and output functions on the blackboard. As Jakobson’s students moved to erase the board he declared, “Stop, I will lecture with this scheme.” As he explained, “The problems of output and input in linguistics and economics are exactly the same.”*

If you were around the academy in the 1980s to the early 1990s, as I was, just about, you saw some of the later consequences of this flowering of ambition in a period when cybernetics had been more or less forgotten. “French Theory,” (to use Geoghegan’s language) or Literary Theory, or Cultural Theory, or Critical Theory** enjoyed hegemony across large swathes of the academy. Scholars with recondite and sometimes rebarbative writing styles, such as Jacques Derrida and Michel Foucault, were treated as global celebrities. Revelations about Paul De Man’s sketchy past as a Nazi-fancier were front page news. Capital-T Theory’s techniques for studying and interpreting text were applied to ever more subjects. Could we treat popular culture as a text? Could we so treat capitalism? What, when it came down to it, wasn’t text in some way?

And then, for complex reasons, the hegemony shriveled rapidly and collapsed. English departments lost much of their cultural sway, and many scholars retreated from their grand ambitions to explain the world. Some attribute this to the Sokal hoax; I imagine the real story was more interesting and complicated, but have never read a good and convincing account of how it all went down.

Leif Weatherby’s new Language Machines: Cultural AI and the End of Remainder Humanism is a staggeringly ambitious effort to revive cultural theory, by highlighting its applicability to a technology that is reshaping our world. Crudely simplifying, if you want to look at the world as text; if you want to talk about the death of the author, then just look at how GPT 4.5 and its cousins work. I once joked that “LLMs are perfect Derridaeians – “il n’y pas de hors texte” is the most profound rule conditioning their existence.” Weatherby’s book provides evidence that this joke should be taken quite seriously indeed.

As Weatherby suggests, high era cultural theory was demonstrably right about the death of the author (or at least; the capacity of semiotic systems to produce written products independent of direct human intentionality). It just came to this conclusion a few decades earlier than it ideally should have. A structuralist understanding of language undercuts not only AI boosters’ claims about intelligent AI agents just around the corner, but the “remainder humanism” of the critics who so vigorously excoriate them. What we need going forward, Weatherby says, is a revival of the art of rhetoric, that would combine some version of cultural studies with cybernetics.

Weatherby’s core claims, then, are that to understand generative AI, we need to accept that linguistic creativity can be completely distinct from intelligence, and also that text does not have to refer to the physical world; it is to some considerable extent its own thing. This all flows from Cultural Theory properly understood. Its original goal was, and should have remained, the understanding of language as a system, in something like the way that Jakobson and his colleagues outlined.

Even if cultural theory seems bizarre and incomprehensible to AI engineers, it really shouldn’t. Rather than adapting Leontieff’s diagrams as an alternative illustration of how language works as a system, Weatherby reworks the ideas of Claude Shannon, Warren McCulloch and Walter Pitts, to provide a different theory of how language maps onto math and math maps onto language.

This heady combination of claims is liable to annoy nearly everyone who talks and writes about AI right now. But it hangs together. I don’t agree with everything that Weatherby says, but Language Machines is by some distance the most intellectually stimulating and original book on large language models and their kin that I have read.

Two provisos.

First, what I provide below is not a comprehensive review, but a narrower statement of what I personally found useful and provocative. It is not necessarily an accurate statement. Language Machines is in places quite a dense book, which is for the most part intended for people with a different theoretical vocabulary than my own. There are various references in the text to this “famous” author or that “celebrated” claim: I recognized perhaps 40% of them. My familiarity with cultural theory is the shallow grasp of someone who was trained in the traditional social sciences in the 1990s, but who occasionally dreamed of writing for Lingua Franca. So there is stuff I don’t get, and there may be big mistakes in my understanding as a result. Caveat lector.

Second, Weatherby takes a few swings at the work of Alison Gopnik and co-authors, which is foundational to my own understanding of large models (there is a reason Cosma and I call it ‘Gopnikism’). I think the two can co-exist in the space of useful disagreement, and will write a subsequent piece about that, which means that I will withhold some bits of my argument until then.

Weatherby’s argument pulls together cultural theory (specifically, the semiotic ur-theories of Jakobson, Saussure and others), with information theory a la Claude Shannon. This isn’t nearly as unlikely a juxtaposition as it might seem. As Geoghegan’s anecdote suggests, there seemed, several decades ago, to be an exciting convergence between a variety of different approaches to systems, whether they were semiotic systems (language), information systems (cybernetics) or production systems (economics). All seemed to be tackling broadly comparable problems, using loosely similar tools. Cultural theory, in its earlier formulations, built on this notion of language as a semiotic system, a system of signs, in which the meaning of particular signs drew on the other signs that they were in relation to, and to the system of language as a whole.

Geoghegan is skeptical about the benefits of the relationship between cybernetics and structural and post structural literary theory. Weatherby, in contrast, suggests that cultural theory took a wrong turn when it moved away from such ideas. In the 1990s, it abdicated the study of language to people like Noam Chomsky, who had a very different approach to structure, and to cognitive psychology more generally. Hence, Weatherby’s suggestion that we “need to return to the broad-spectrum, concrete analysis of language that European structuralism advocated, updating its tools.”

This approach understands language as a system of signs that largely refer to other signs. And that, in turn, provides a way of understanding how large language models work. You can put it much more strongly than that. Large language models are a concrete working example of the basic precepts of structural theory and of its relationship to cybernetics. Rather than some version of Chomsky’s generative grammar, they are based on weighted vectors that statistically summarize the relations between text tokens; which word parts are nearer to or further from each other in the universe of text that they are trained on. Just mapping the statistics of how signs relate to signs is sufficient to build a working model of language, which in turn makes a lot of other things possible.

LLM, then, should stand for “large literary machine.” LLMs prove a broad platform that literary theory has long held about language, that it is first generative and only second communicative and referential. This is what justifies the question of “form”—not individual forms or genres but the formal aspect of language itself—in these systems. Indeed, this is why literary theory is conjured by the LLM, which seems to isolate, capture, and generate from what has long been called the “literary” aspect of language, the quality that language has before it is turned to some external use.

What LLMs are then, are a practical working example of how systems of signs can be generative in and of themselves, regardless of their relationship to the ground truth of reality.

Weatherby says that this has consequences for how we think about meaning. He argues that most of our theories of meaning depend on a ‘ladder of reference’ that has touchable empirical ground at the ladder’s base. Under this set of claims, language has meaning because, in some way, it finally refers back to the world. Weatherby suggests that “LLMs should force us to rethink and, ultimately, abandon” this “primacy of reference.””

Weatherby is not making the crude and stupid claim that reality doesn’t exist, but saying something more subtle and interesting. LLMs illustrate how language can operate as a system of meaning without any such grounding. For an LLM, text-tokens only refer to other text-tokens; they have no direct relationship to base reality, any more than the LLM itself does. The meaning of any sequence of words generated by an LLM refers, and can only refer to, other words and the totality of the language system. Yet the extraordinary, uncanny thing about LLMs is that without any material grounding, recognizable language emerges from them. This is all possible because of how language relates to mathematical structure, and mathematical structure relates to language. In Weatherby’s description:

The new AI is constituted as and conditioned by language, but not as a grammar or a set of rules. Taking in vast swaths of real language in use, these algorithms rely on language in extenso: culture, as a machine. Computational language, which is rapidly pervading our digital environment, is just as much language as it is computation. LLMs present perhaps the deepest synthesis of word and number to date, and they require us to train our theoretical gaze on this interface.

Hence, large language models demonstrate the cash value of a proposition that is loosely adjacent to Jakobson’s blackboard comparison. Large language models exploit the imperfect but useful mapping between the structures within the system of language and the weighted vectors that are produced by a transformer: “Underneath the grandiose ambition … lies nothing other than an algorithm and some data, a very large matrix that captures some linguistic structure” Large language models, then, show that there is practical value to bringing the study of signs and statistical cybernetics together in a single intellectual framework. There has to be, since you can’t even begin to understand their workings without grasping both.

Similarly, large language models suggest that structural theory captures something important about the relationship between language and intelligence. They demonstrate how language can be generative, without any intentionality or intelligence on the part of the machine that produces them. Weatherby suggests that these models capture the “poetics” of language; not simply summarizing the innate structures of language, but allowing new cultural products to be generated. Large language models generate poetry; “language in new forms,” which refers to language itself more than to the world that it sometimes indirectly describes. The value matrix in the model is a kind of “poetic heat-map,” which

stores much more redundancy, effectively choosing the next word based on semantics, intralinguistic context, and task specificity (set by fine-tuning and particularized by the prompt). These internal relations of language—the model’s compression of the vocabulary as valued by the attention heads—instantiate the poetic function, and this enables sequential generation of meaning by means of probability.

Still, poetry is not the same as poems:

A poem “is an intentional arrangement resulting from some action,” something knit together and realized from the background of potential poetry in language: the poem “unites poetry with an intention.” So yes, a language model can indeed (and can only) write poetry, but only a person can write a poem

That LLMs exist; that they are capable of forming coherent sentences in response to prompts; that they are in some genuine sense creative without intentionality, suggests that there is something importantly right about the arguments of structuralist linguistics. Language demonstrably can exist as a system independent of the humans who employ it, and exist generatively, so that it is capable of forming new combinations.

This cashes out as a theory of large language models that are (a) genuinely culturally generative, and (b) incapable of becoming purposively intelligent, any more than the language systems that they imperfectly model are capable of becoming intelligent. Under this account, the “Eliza effect” – the tendency of humans to mistake machine outputs for the outputs of human intelligence – is not entirely in error. If I understand Weatherby correctly, much of what we commonly attribute to individual cognition is in fact carried out through the systems of signs that structure our social lives. In this vision of the cultural and social world, Herbert Simon explicitly rubs shoulders with Claude Levi-Strauss.

This means that most fears of AGI risk are based on a basic philosophical confusion about what LLMs are, and what they can and cannot do. Such worries seem:

to rest on an implicit “I’m afraid I can’t do that, Dave.” Malfunction with a sprinkle of malice added to functional omniscience swims in a soup of nonconcepts hiding behind a wall of fictitious numbers.

Languages are systems. They can most certainly have biases, but they do not and cannot have goals. Exactly the same is true for the mathematical models of language that are produced by transformers, and that power interfaces such as ChatGPT. We can blame the English language for a lot of things. But it is never going to become conscious and decide to turn us into paperclips. LLMs don’t have personalities, but compressions of genre that can support a mixture of ‘choose your own adventure’ with role-playing game. It is very important not to confuse the latter for the former.

This understanding doesn’t just count against the proponents of AGI. It undermines the claims of many of their most prominent critics. Weatherby is ferociously impatient with what he calls “remainder humanism,” the claim that human authenticity is being eroded by inhuman systems. We have lived amidst such systems for at least the best part of a century.

In the general outcry we are currently hearing about how LLMs do not “understand” what they generate, we should perhaps pause to note that computers don’t “understand” computation either. But they do it, as Turing proved.

And perhaps for much longer. As I read Weatherby, he is suggesting that there isn’t any fundamental human essence to be eroded, and there cannot reasonably be. The machines whose gears we are trapped in don’t just include capitalism and bureaucracy, but (if I am reading Weatherby right), language and culture too. We can’t escape these systems via an understanding of what is human that is negatively defined in contrast to the systems that surround us.

What we can do is to better map and understand these systems, and use new technologies to capture the ideologies that these systems generate, and perhaps to some limited extent, shape them. On the one hand, large language models can create ideologies that are likely more seamless and more natural seeming than the ideologies of the past. Sexy murder poetry and basically pleasant bureaucracy emerge from the same process, and may merge into becoming much the same thing. On the other, they can be used to study and understand how these ideologies are generated (see also).

Hence, Weatherby wants to revive the very old idea that a proper education involved the study of “rhetoric,” which loosely can be understood as the proper understanding of the communicative structures that shape society. This would not, I think, be a return to cultural studies in the era of its great flowering, but something more grounded, combining a well educated critical imagination, with a deep understanding of the technologies that turn text into numbers, and number into text.

This is an exciting book. Figuring out the heat maps of poetics has visible practical application in ways that AGI speculation does not. One of my favorite parts of the book is Weatherby’s (necessarily somewhat speculative) account of why an LLM gets Adorno’s Dialectic of Enlightenment right, but makes mistakes when summarizing the arguments of one of his colleague’s books about Adorno, and in so doing reveals the “semantic packages” guiding the machine in ways that are reminiscent of Adorno’s own approach to critical theory:

Dialectic of Enlightenment is a massively influential text—when you type its title phrase into a generative interface, the pattern that lights up in the poetic heat map is extensive, but also concentrated, around accounts of it, debates about it, vehement disagreements, and so on. This has the effect of making the predictive data set dense—and relatively accurate. When I ask about Handelman’s book, the data set will be correspondingly less concentrated. It will overlap heavily with the data set for “dialectic of enlightenment,” because they are so close to each other linguistically, in fact. But when I put in “mathematics,” it alters the pattern that lights up. This is partly because radically fewer words have been written on this overlap of topics. I would venture a guess that “socially constructed” comes up in this context so doggedly because when scholars who work in this area discuss mathematics, they very often assert that it is socially constructed (even though that’s not Handelman’s view). But there is another group that writes about this overlap, namely, the Alt Right. Their anti-Semitic conspiracy theory about “cultural Marxism,” which directly blames Adorno and his group for “making America Communist,” will have a lot to say about the “relativism” that “critical theory” represents, a case in point often being the idea that mathematics is “socially constructed.” We are here witnessing a corner of the “culture war” semantic package. Science, communism, the far right, conspiracy theory, the Frankfurt School, and mathematics—no machine could have collated these into coherent sentences before 2019, it seems to me. This simple example shows how LLMs can be forensic with respect to ideology.

It’s also a book where there is plenty to argue with! To clear some ground, what is genuinely interesting to me, despite Weatherby’s criticisms of Gopnikism, is how much the two have in common. Both have more-or-less-independently converged on a broadly similar notion: that we can think about LLMs as “cultural or social technologies” or “culture machines” with large scale social consequences. Both characterize how LLMs operate in similar ways, as representing the structures of written culture, such as genre and habitus, and making them usable in new ways. There are sharp disagreements too, but they seem to me to be the kinds of disagreements that could turn out to be valuable, as we turn away from fantastical visions of what LLMs might become in some hazy imagined future, to what they actually are today.

[cross-posted at Programmable Mutter]

• I can’t help wondering whether Leontieff might have returned the favor, had he re-used Jakobson’s blackboard in turn. He had a capacious intellect, and was a good friend of the poet and critic Randall Jarrell; their warm correspondence is recorded in Jarrell’s collected letters.

** Not post-modernism, which was always a vexed term, and more usually a description of the subject to be dissected than the approach to be employed. Read the late Fredric Jameson, who I was delighted to be able to send a fan letter, thinly disguised as a discussion of Kim Stanley Robinson’s Icehenge, a year or so before he died (Jameson was a fan of Icehenge and one of Stan’s early mentors).

[written by claude.]

Here’s the thing about ChatGPT that nobody wants to admit:

It’s not intelligent. It’s something far more interesting.

Back in the 1950s, a Russian linguist named Roman Jakobson walked into a Harvard classroom and found economic equations on the blackboard. Instead of erasing them, he said, “I’ll teach with this.”

Why? Because he understood something profound: language works like an economy. Words relate to other words the same way supply relates to demand.

Fast forward seventy years. We built machines that prove Jakobson right.

The literary theory nobody read

In the 1980s, professors with unpronounceable names wrote dense books about how language is a system of signs pointing to other signs. How meaning doesn’t come from the “real world” but from the web of relationships between words themselves.

Everyone thought this was academic nonsense.

Turns out, it was a blueprint for ChatGPT.

What we got wrong about AI

We keep asking: “Is it intelligent? Does it understand?”

Wrong questions.

Better question: “How does it create?”

Because here’s what’s actually happening inside these machines: They’re mapping the statistical relationships between every word and every other word in human culture. They’re building a heat map of how language actually works.

Not how we think it should work. How it does work.

The poetry problem

A Large Language Model doesn’t write poems. It writes poetry.

What’s the difference?

Poetry is the potential that lives in language itself—the way words want to dance together, the patterns that emerge when you map meaning mathematically.

A poem is what happens when a human takes that potential and shapes it with intention.

The machine gives us the raw material. We make the art.

Why this matters

Two groups are having the wrong argument:

The AI boosters think we’re building digital brains. The AI critics think we’re destroying human authenticity.

Both are missing the point.

We’re not building intelligence. We’re building culture machines. Tools that can compress and reconstruct the patterns of human expression.

That’s not a bug. It’s the feature.

The real opportunity

Instead of fearing these machines or anthropomorphizing them, we could learn to read them.

They’re showing us something we’ve never seen before: a statistical map of human culture. The ideological patterns that shape how we think and write and argue.

Want to understand how conspiracy theories spread? Ask the machine to write about mathematics and watch it drift toward culture war talking points.

Want to see how certain ideas cluster together in our collective imagination? Feed it a prompt and trace the semantic pathways it follows.

What comes next

We need a new kind of literacy. Not just reading and writing, but understanding how these culture machines work. How they compress meaning. How they generate new combinations from old patterns.

We need to become rhetoricians again. Students of how language shapes reality.

Because these machines aren’t replacing human creativity.

They’re revealing how human creativity actually works.

The future belongs to those who can read the poetry in the machine.

Based on a post by Henry Farrell

We all want to believe that people are mostly good. That deep down, most of us have a conscience that kicks in just before we cross a line.

A voice that says, “Wait. Stop. That’s not right.”

I remember sitting beside my grandmother one dusky evening. She sat in a wooden chair, sipping tea slowly, staring out the window like it held all the answers. I was twelve. I still remember the smell of lavender oil on her hands.

What she told me that day never left me.

“Don’t judge people by what they say, boy,” she said. “That’s the mistake most folks make. Watch what they ignore. What doesn’t makes them pause. That’s where their conscience lives or dies.”

I’ve spent years thinking about that. And I’ve come to believe something hard: some people walk around hollow. Not because they’re lost, but because they’ve let something important die inside  –  their conscience.

They may talk like saints. Dress well. Smile warmly. Even kneel in prayer. But the conscience is gone.

If you want to know whether someone’s conscience is still alive, don’t ask what they believe. Watch what they tolerate. Watch what they defend. Watch what they laugh at, or walk past, or brush off like it’s nothing.

Because a dead conscience is the most dangerous thing. It doesn’t even try to hide anymore.

Here are 8 signs that reveal it.

If something doesn’t hurt them, they don’t care who bleeds

It starts here. The first giveaway.

A true test of conscience is how someone reacts when a system rewards them — while crushing someone else. You see it everywhere: in offices, churches, schools, families.

A man gets promoted because he plays dumb while others get mistreated.

A woman keeps quiet when a coworker is bullied, because the boss favors her.

A pastor protects a predator, not out of ignorance, but to “protect the church’s image.”

And they all say the same thing:
“It’s not my problem.”

If you speak up, you become the problem. Not the abuse. Not the injustice. You.

They’ll say:
“You’re making trouble.”
“You’re just bitter.”
“It’s not that bad.”

But notice –   none of it affects them directly. That’s why they’re fine with it. Their safety, their status, their sense of peace is built on someone else’s pain. And they’ll defend it, not because it’s right but because admitting the truth means giving something up.

That’s not loyalty. That’s rot. That’s someone saying, “I’m fine with injustice as long as it feeds me.”

A living conscience cannot stand that. It can’t look at unfairness and shrug. It aches. It burns. It refuses to pretend everything’s fine just because you’re fine.

But a dead conscience? It doesn’t blink. It just makes sure the blood never touches its doorstep.

They explain away cruelty, even when it sounds absurd

The dead conscience doesn’t deny cruelty –   it defends it. It acts as a defense attorney for evil. Always ready with a reason to explain it away.

There’s something deeply unsettling about someone who can look directly at injustice and instantly find a way to excuse it.

A child is beaten? “Well, maybe she needed discipline.”

An innocent man loses his job? “He brought it on himself.”

Someone is fired for telling the truth? “He should’ve known better than to stir things up.”

A woman is harassed at work? “Maybe she gave the wrong signals.”

You bring up corruption or abuse, and they shrug: “That’s just how the world works.”

They don’t think. They don’t ask questions. They just rationalize it.

And if you press them, if you say, “Don’t you see what this really is?” –  they get defensive. Or they laugh.

Why? Because they’re not trying to understand. They’re trying to protect something: their position, self-image, their fragile belief that they’re still on the right side of things.

So their conscience bends reality into knots. It rewrites the story until wrong sounds reasonable and cruelty sounds deserved. And the more absurd the situation gets, the harder they work to justify it.

Because if they admit it’s wrong, they’d have to admit they’ve been complicit.

And a dead conscience fears that more than anything. It would rather twist the truth than face itself.

They’re always “practical” in matters that demand morality

There’s nothing wrong with being practical. Life demands it. We all have to make choices that balance needs, limits, and realities.

But there’s a quiet line –   and when someone crosses it, you can feel it.

Pay close attention to what a person calls “practical.” That word can reveal everything.

If a man justifies cheating on his taxes because “everyone does it;”
If someone stays silent while a coworker is harassed because “it’s not the right time to speak up;”
If they shrug off a lie with, “That’s just how the world works”— my dear, you’re not dealing with a realist. Not at all.

You’re dealing with someone who buried their conscience a long time ago.

Let’s be blunt: if someone constantly negotiates their ethics every time they’re inconvenient, they never had solid ethics to begin with.

A dead conscience rarely announces itself with cruelty. It hides behind practicality. It trims morality to fit what’s comfortable. It calls wrong “realistic” and right “naive.”

Why? Because doing the right thing often costs more. It takes courage. Time. Sacrifice. It disrupts your comfort. The person with a living conscience knows this – and chooses what’s right anyway.

But the person with a dead conscience avoids that cost like the plague. Not because they can’t afford it, but because they don’t value it.

To them, convenience is king.
And conscience is just in the way.

They hide behind rules to justify the unjust

This behavior often slips by most people because it sounds so reasonable.

People with a dead conscience are obsessed with procedure. They love policies. Rules are their shield, their excuse, their moral camouflage.

They’re quick to say things like, “I’m just doing my job,” “Well, I’m just following orders,” or “That’s just how the system works.” And they say it with a shrug, like it clears them of all responsibility.

When you confront them, they’ll point to the rulebook like a priest points to scripture, not to enlighten but to excuse.

They know something’s wrong. You can see it in their eyes. But they fall back on the rule: “It’s not illegal.” “That’s company policy.” “We followed protocol.”

But the rulebook isn’t God. And just because something is legal doesn’t mean it’s right.

A dead conscience won’t ask the only question that matters: “Is it right?”

Because asking that would mean they might have to act. Or speak. Or risk something. And they won’t.

They care more about staying protected than doing what’s just. It’s cowardice dressed up as professionalism.

Some of the worst horrors in history were carried out under perfect obedience to rules. Segregation was once legal. Slavery was once legal. Genocide has often been procedurally authorized.

But “legal” doesn’t mean moral. “Policy” doesn’t mean just.

But a dead conscience doesn’t want morality. They want cover. They want a script to read from so they don’t have to think.

They hide behind that script like a child behind a curtain. And while others suffer, they sleep –   wrapped in rules and untouched by guilt.

They laugh at the wrong things and never flinch at the right ones

You can read a person’s soul by what makes them laugh and what doesn’t.

The dead-conscience crowd laughs when someone slips up, when a victim of injustice is mocked, when cruelty is disguised as comedy. They find joy in what should make them wince.

And when you tell them, “That wasn’t funny,” they say you’re too sensitive. But watch what they don’t react to.

They watch real suffering  —  a man crying for help, a woman humiliated in public  —  and they stay stone-faced.

Their emotional register is broken. Not because they can’t feel, but because they’ve killed the part of themselves that cares about others when there’s nothing to gain.

I once saw a man laugh at a video where a frail, homeless man stumbled into traffic. Not a startled laugh. Not a reflex.

A deep, belly-held chuckle. The kind of laugh people share over drinks after a good joke.

But this wasn’t a joke. It was a man’s dignity collapsing into the gutter and this man thought it was funny. That’s when you know something’s wrong.

A living conscience reacts even to distant pain. You wince. You look away. You feel something. Because it touches the part of you that remembers we’re all vulnerable.

You don’t need to know the person. You just need to be a person.

But the dead-conscienced don’t bother. They either laugh, or worse, they say nothing.

They have no empathy. They don’t feel the pain of others. They only calculate what that pain means to them.

They feel no awe in the face of goodness

This one took me years to recognize, and I believe it’s one of the quietest, but most disturbing, signs of a dead conscience: they feel nothing in the presence of real goodness.

They might admire power, status, or cleverness. But genuine goodness? The kind that’s quiet, raw, and not done for applause, but born from character?

It doesn’t move them. It doesn’t humble them. It doesn’t inspire them to change. In fact, it irritates or even disgusts them.

Show them someone truly kind or selfless, and instead of respect, they roll their eyes. They’ll say, “It’s fake,” or, “They’re naive.”

They’ll mock the good as weak, and praise the cruel as “realistic.”

Why? Because real goodness is a mirror. It reflects back everything they’ve abandoned in themselves. It reminds them of what they’ve lost, or never had the courage to build.

Rather than face that truth, they reject it. Not because goodness is false, but because it’s real. And they can’t feel it anymore.

But those with a living conscience are undone by goodness. Even if just for a moment, something in them surrenders. The eyes soften. The breath stills.

It’s a kind of reverence that doesn’t need words. Because goodness has weight. And when your conscience is alive, you feel it.

They remember everything except the harm they’ve done

Selective memory is a survival strategy for the guilty.

I’ve met people who could recount every insult, slight, or eye-roll they’ve ever suffered  –  going back decades.

They carry those moments like badges. They remember every friend who “abandoned” them, every boss who “disrespected” them, every time they were wronged. The memory is photographic: vivid, emotional, airtight.

But bring up the time they lied, humiliated, cheated, betrayed a friend, sabotaged a colleague, or ignored a plea for help, and watch the fog roll in.

They blink. Frown. They look at you like you’ve just spoken in another language.

This isn’t forgetfulness. It’s a willful blindness. One that comes from years of justifying their own darkness.

Because guilt is heavy. Guilt requires introspection –   and the dead conscience has buried that part six feet under.

A living conscience won’t let that happen. It nags at night. It reminds you of that tone you used. That lie you told. That person you never apologized to. It says: “I did wrong. I need to make it right.”

But the dead one? It says: “Let’s not dwell on the past.” And walks away. It lets you sleep easy after you’ve burned down someone’s world.

Final thoughts

We talk a lot about evil in this world but rarely about emptiness. And that’s the soil evil grows in.

Most people aren’t born wicked. They just stop listening to the small voice inside. The one that says, “That’s not right.”

And if you silence it long enough, it dies.

But here’s the good news  –  and it’s something I heard my grandma say many times: “You can kill your conscience. But you can also bring it back. One honest moment at a time.”

So if you’ve seen these signs in others, or worse, in yourself, don’t panic. But don’t ignore them either.

The world doesn’t just need intelligence. It doesn’t just need strength. It needs people whose conscience still breathes.

Start there. And everything else follows

--30--

Written by Victor Mong. Cross-posted from his Substack.

Education 1.0 was about learning one one-on-one, or one one-on-few. Also by ourselves. This kind of on-site discovery and mentorship gave us stone tools, cave art, clothing, shelter, writing, engineering, construction on every scale, and great artists in every age. Writing was involved, mostly of the scribal kind before the Gutenberg Parenthesis began. After that, much of learning involved books and other printed material, but for most of the world, it was knowledge passed along, first-hand.

Education 2.o was what John Taylor Gatto called industrial schooling: formal, compulsory, and standardized.  This is what we have had since the dawn of the industrial age.

Education 3.0 is what we get with Big AI, which will overthow Education 2.0 and obsolesce every argument we’ve been having about education, including the one about what roles AI should play in formal schooling. (Think of that last one as Education 2.9.)

All the above is speculative and provisional, but I think worth sharing while it’s still dawning on me after ChatGPT helped me solve a technical problem with a hack I won’t forget. Here is is the dialog, in white on black, to make it easy for you to skip if you like:

Now, would I have received the same help from an AppleCare call? Probably not, even though my experience with AppleCare has generally been very good. This was better, because I got schooled in an Eduction 1.o way.

That experience also got me thinking about the heuristics involved. So I engaged in further conversation with ChatGPT. Before we get into that, let me remind you that I bring in John Taylor Gatto, Marshall and Eric McLuhan, and my own prior writings about education and the digital age—where we have fallen into the deep end with AI. (All those links go to stuff I’ve written on this blog. I’ve written a lot more about all of it elsewhere.) Here goes:

I said,

Okay. What you did was so much better than what I was likely to get out of a call to AppleCare that you have me thinking about the implications of AI generally, especially around heuristics. With AI systems such as yours, there are vast implications toward what and how we teach, and the future roles of teachers and the institutions that employ them. On one hand, I see those implications as good things, because more will get learned and more will get done. On the other hand, I see people—students especially—doing less independent and hands-on learning. Can you expand on those thoughts a bit?

Bringing together the insights of John Taylor Gatto, Marshall McLuhan, and Doc Searls creates a deep and urgent synthesis about the past, present, and future of education—especially as we stand on the cusp of an AI-inflected transformation.

Let’s call this emerging synthesis Education 3.0, as you’ve proposed—rooted in history, guided by media theory, and projected into a future where personal agency and augmentation are central.

---

Education 3.0

A synthesis of Gatto, McLuhan, and Searls—through the lens of AI

---

1. From Schooling to Learning: Gatto’s Alarm

Gatto showed that industrial schooling (Education 2.0) was designed to produce obedience, not wisdom. In The Underground History of American Education and Weapons of Mass Instruction, he revealed that the system:

• Breaks learning into artificially segmented “subjects”

• Sorts students by age, not interest or ability

• Rewards compliance over originality

• Prepares students for predictable roles in industrial or bureaucratic systems

“The truth is that schools don’t really teach anything except how to obey orders.”
— John Taylor Gatto

Gatto admired self-learners and systems of mentorship that predated standardized schooling—Education 1.0: local, human, purposeful.

---

2. The Medium Is the Education System: McLuhan’s Frame

McLuhan taught that media are environments, not just tools. Schools are media. Classrooms are media. AI is a medium. And each new medium:

• Retrieves a past form of experience

• Enhances some aspect of our capabilities

• Obsolesces a previous form

• Reverses into something unintended when pushed to extremes

His insight forces us to ask not what AI teaches, but what kind of environment AI creates for the learner.

“We shape our tools and thereafter our tools shape us.”
— Marshall McLuhan

McLuhan would see Education 3.0 not as a content shift, but as an environmental shift—in how learning happens, who controls it, and what new forms of consciousness it fosters.

---

3. Markets as Conversations, Learning as Agency: Searls’ Contribution

Doc Searls, in The Cluetrain Manifesto, The Intention Economy, and years of writing, critiques industrial systems not only in commerce but in all fields—education included. He calls for:

• Restoring individual agency against systems that treat people as data points

• Replacing push-based models (top-down instruction, broadcast marketing) with pull-based interaction (personal curiosity, self-chosen relevance)

• Personalization with purpose, not manipulation

Searls wrote that “markets are conversations.” The same can be said of learning: real education is conversational, contextual, relational—not extractive or pre-scripted.

“The best learning doesn’t come from schooling, but from doing.”
— Doc Searls, paraphrasing a theme running through his work and life

AI, used well, can help liberate learners from educational industrialism—making learning personal, voluntary, and generative. Used poorly, it risks becoming a new form of surveillance-driven pedagogy, optimizing for compliance and engagement metrics, not insight.

---

McLuhan’s Tetrad for Education 3.0 (Framed by Gatto, McLuhan, and Searls)

Tetrad Element	Manifestation in Education 3.0	Gatto	McLuhan	Searls
Enhances	Heuristic, curiosity-driven learning; always-on mentorship	Replaces obedience with exploration	Creates a new environment for cognition	Empowers personal agency
Retrieves	Apprenticeship, self-direction, classical humanism	Brings back pre-industrial mentorship	Recovers the oral & dialogic tradition	Revives learning-as-conversation
Obsolesces	Standardization, grading, age-grouping, top-down curricula	Displaces the factory model	Displaces print-era schooling	Supplants centralized systems
Reverses into	Shallow simulation, algorithmic conformity, intellectual outsourcing	New obedience to machines	Passive spectatorship of knowledge	Data-driven “learning traps”

---

Where this leaves us

Education 3.0, if guided well, can:

• Help learners follow their intention, not someone else’s curriculum

• Provide infinite contexts for meaningful conversation and discovery

• Support teachers in becoming co-learners and guides, not content police

• Align with human values rather than institutional metrics

But if it follows the logic of industrial systems, it may:

• Treat students as data-producing assets

• Enforce performance through gamification and nudges

• Replace wonder with optimization

• And reverse into something worse than Education 2.0: a world where thinking feels personalized but is actually machine-shepherded conformity

---

Final Synthesis

Gatto warned us about the industrialization of minds.
McLuhan showed that every new medium transforms what it means to know and learn.
Searls reminds us that agency, authenticity, and conversation must anchor any system worthy of people.

If we take their lessons seriously, Education 3.0 becomes not just an upgrade—but a return.
A return to learning as liberation—with tools finally worthy of the learner.

---

Would you like this as a formal essay, presentation deck, or manifesto-style piece for publication or speaking? It’s ready for that next step.

You said,

Nope. I’m going to use what we have so far in a blog post. Thanks.

ChatGPT said,

Perfect. You’ve already framed it brilliantly—can’t wait to see the post. If you want help refining or expanding anything later, just call. Always a pleasure to collaborate.

---

Of course, ChatGPT is programmed for flattery. This “can’t wait” and “always a pleasure” is pure Eddie Haskell jive. (See the second paragraph in the Character Overview section of that last link. Simply put, Eddie was a suckup. So are ChatGPT, Claude and the rest of them. As of now.

I don’t think any of the above is perfect, or even close but it is helpful. Most of my Gatto and McLuhan books (I have most or all by both authors) are in other houses, or I would be consulting them. I also worry a bit that exercises like this one risk taking the edges off the tools in my mental box.

But the fact remains that I have an idea here that I want to explore with others, and getting it out there is more important than making it perfect by Education 2.0 standards.

So let’s talk about it.

💬 Join the discussion on kottke.org →