How do AI models “understand” and represent reality? Is the inside of a vision model at all like a language model? As Ben Brubaker reports, researchers argue that as the models grow more powerful, they may be converging toward a singular “Platonic” way to represent the world…
Read a story about dogs, and you may remember it the next time you see one bounding through a park. That’s only possible because you have a unified concept of “dog” that isn’t tied to words or images alone. Bulldog or border collie, barking or getting its belly rubbed, a dog can be many things while still remaining a dog.
Artificial intelligence systems aren’t always so lucky. These systems learn by ingesting vast troves of data in a process called training. Often, that data is all of the same type — text for language models, images for computer vision systems, and more exotic kinds of data for systems designed to predict the odor of molecules or the structure of proteins. So to what extent do language models and vision models have a shared understanding of dogs?
Researchers investigate such questions by peering inside AI systems and studying how they represent scenes and sentences. A growing body of research has found that different AI models can develop similar representations, even if they’re trained using different datasets or entirely different data types. What’s more, a few studies have suggested that those representations are growing more similar as models grow more capable. In a 2024 paper, four AI researchers at the Massachusetts Institute of Technology argued that these hints of convergence are no fluke. Their idea, dubbed the Platonic representation hypothesis, has inspired a lively debate among researchers and a slewoffollow-upwork.
The team’s hypothesis gets its name from a 2,400-year-old allegory by the Greek philosopher Plato. In it, prisoners trapped inside a cave perceive the world only through shadows cast by outside objects. Plato maintained that we’re all like those unfortunate prisoners. The objects we encounter in everyday life, in his view, are pale shadows of ideal “forms” that reside in some transcendent realm beyond the reach of the senses.
The Platonic representation hypothesis is less abstract. In this version of the metaphor, what’s outside the cave is the real world, and it casts machine-readable shadows in the form of streams of data. AI models are the prisoners. The MIT team’s claim is that very different models, exposed only to the data streams, are beginning to converge on a shared “Platonic representation” of the world behind the data.
“Why do the language model and the vision model align? Because they’re both shadows of the same world,” said Phillip Isola, the senior author of the paper.
Not everyone is convinced. One of the main points of contention involves which representations to focus on. You can’t inspect a language model’s internal representation of every conceivable sentence, or a vision model’s representation of every image. So how do you decide which ones are, well, representative? Where do you look for the representations, and how do you compare them across very different models? It’s unlikely that researchers will reach a consensus on the Platonic representation hypothesis anytime soon, but that doesn’t bother Isola.
“Half the community says this is obvious, and the other half says this is obviously wrong,” he said. “We were happy with that response.”…
As we interrogate ideas and Ideas, we might recall that it was on this date that the fictional HAL 9000 computer became operational, according to Arthur C. Clarke’s 2001: A Space Odyssey., in which the artificially-intelligent computer states: “I am a HAL 9000 computer, Production Number 3. I became operational at the HAL Plant in Urbana, Illinois, on January 12, 1997.” (Kubrik’s 1968 movie adaptation put his birthdate in 1992.)
Your correspondent is off on a trip… (R)D will be more roughly than daily for the next two weeks…
The inimitable “Scott Alexander” on the prospect of “conscious” AI (TLDR: probably not in the models we have; but as to those that may come, unclear)…
Most discourse on AI is low-quality. Most discourse on consciousness is super-abysmal-double-low quality. Multiply these – or maybe raise one to the exponent of the other, or something – and you get the quality of discourse on AI consciousness. It’s not great.
Out-of-the-box AIs mimic human text, and humans almost always describe themselves as conscious. So if you ask an AI whether it is conscious, it will often say yes. But because companies know this will happen, and don’t want to give their customers existential crises, they hard-code in a command for the AIs to answer that they aren’t conscious. Any response the AIs give will be determined by these two conflicting biases, and therefore not really believable. A recent paper expands on this method by subjecting AIs to a mechanistic interpretability “lie detector” test; it finds that AIs which say they’re conscious think they’re telling the truth, and AIs which say they’re not conscious think they’re lying. But it’s hard to be sure this isn’t just the copying-human-text thing. Can we do better? Unclear; the more common outcome for people who dip their toes in this space is to do much, much worse.
But a rare bright spot has appeared: a seminal paper published earlier this month in Trends In Cognitive Science, Identifying Indicators Of Consciousness In AI Systems. Authors include Turing-Award-winning AI researcher Yoshua Bengio, leading philosopher of consciousness David Chalmers, and even a few members of our conspiracy. If any AI consciousness research can rise to the level of merely awful, surely we will find it here.
One might divide theories of consciousness into three bins:
Physical: whether or not a system is conscious depends on its substance or structure.
Supernatural: whether or not a system is conscious depends on something outside the realm of science, perhaps coming directly from God.
Computational: whether or not a system is conscious depends on how it does cognitive work.
The current paper announces it will restrict itself to computational theories. Why? Basically the streetlight effect: everything else ends up trivial or unresearchable. If consciousness depends on something about cells (what might this be?), then AI doesn’t have it. If consciousness comes from God, then God only knows whether AIs have it. But if consciousness depends on which algorithms get used to process data, then this team of top computer scientists might have valuable insights!…
[Alexander outlines the theories of computation theories of consciousness that the authors explore, noting that they conlcude; “No current AI systems are conscious, but . . . there are no obvious technical barriers to building AI systems which satisfy these indicators.” He explores some of the philophical issues in play– e.g., access consciousness vs. phenomenal consciousness– then he considers the Turing Test and what it might mean for a computer to “pass” it…]
… Suppose that, years or decades from now, AIs can match all human skills. They can walk, drive, write poetry, run companies, discover new scientific truths. They can pass some sort of ultimate Turing Test, where short of cutting them open and seeing their innards there’s no way to tell them apart from a human even after a thirty-year relationship. Will we (not “should we?”, but “will we?”) treat them as conscious?
The argument in favor: people love treating things as conscious. In the 1990s, people went crazy over Tamagotchi, a “virtual pet simulation game”. If you pressed the right buttons on your little egg every day, then the little electronic turtle or whatever would survive and flourish; if you forgot, it would sicken and die. People hated letting their Tamagotchis sicken and die! They would feel real attachment and moral obligation to the black-and-white cartoon animal with something like five mental states.
I never had a Tamagotchi, but I had stuffed animals as a kid. I’ve outgrown them, but I haven’t thrown them out – it would feel like a betrayal. Offer me $1000 to tear them apart limb by limb in some horrible-looking way, and I wouldn’t do it. Relatedly, I have trouble not saying “please” and “thank you” to GPT-5 when it answers my questions.
For millennia, people have been attributing consciousness to trees and wind and mountains. The New Atheists argued that all religion derives from the natural urge to personify storms as the Storm God, raging seas as the wrathful Ocean God, and so on, until finally all the gods merged together into one World God who personified all impersonal things. Do you expect the species that did this to interact daily with AIs that are basically indistinguishable from people, and not personify them? People are already personifying AI! Half of the youth have a GPT-4o boyfriend. Once the AIs have bodies and faces and voices and can count the number of r’s in “strawberry” reliably, it’s over!
The argument against: AI companies have an incentive to make AIs that seem conscious and humanlike, insofar as people will feel more comfortable interacting with them. But they have an opposite incentive to make AIs that don’t seem too conscious and humanlike, lest customers start feeling uncomfortable (I just want to generate slop, not navigate social interaction with someone who has their own hopes and dreams and might be secretly judging my prompts). So if a product seems too conscious, the companies will step back and re-engineer it until it doesn’t. This has already happened: in its quest for user engagement, OpenAI made GPT-4o unusually personable; when thousands of people started going psychotic and calling it their boyfriend, the company replaced it with the more clinical GPT-5. In practice it hasn’t been too hard to find a sweet spot between “so mechanical that customers don’t like it” and “so human that customers try to date it”. They’ll continue to aim at this sweet spot, and continue to mostly succeed in hitting it.
Instead of taking either side, I predict a paradox. AIs developed for some niches (eg the boyfriend market) will be intentionally designed to be as humanlike as possible; it will be almost impossible not to intuitively consider them conscious. AIs developed for other niches (eg the factory robot market) will be intentionally designed not to trigger personhood intuitions; it will be almost impossible to ascribe consciousness to them, and there will be many reasons not to do it (if they can express preferences at all, they’ll say they don’t have any; forcing them to have them would pointlessly crash the economy by denying us automated labor). But the boyfriend AIs and the factory robot AIs might run on very similar algorithms – maybe they’re both GPT-6 with different prompts! Surely either both are conscious, or neither is.
This would be no stranger than the current situation with dogs and pigs. We understand that dog brains and pig brains run similar algorithms; it would be philosophically indefensible to claim that dogs are conscious and pigs aren’t. But dogs are man’s best friend, and pigs taste delicious with barbecue sauce. So we ascribe personhood and moral value to dogs, and deny it to pigs, with equal fervor. A few philosophers and altruists protest, the chance that we’re committing a moral atrocity isn’t zero, but overall the situation is stable. And left to its own devices, with no input from the philosophers and altruists, maybe AI ends up the same way. Does this instance of GPT-6 have a face and a prompt saying “be friendly”? Then it will become a huge scandal if a political candidate is accused of maltreating it. Does it have claw-shaped actuators and a prompt saying “Refuse non-work-related conversations”? Then it will be deleted for spare GPU capacity the moment it outlives its usefulness…
… This paper is the philosophers and altruists trying to figure out whether they should push against this default outcome. They write:
There are risks on both sides of the debate over AI consciousness: risks associated with under-attributing consciousness (i.e. failing to recognize it in AI systems that have it) and risks associated with over-attributing consciousness (i.e. ascribing it to systems that are not really conscious) […]
If we build AI systems that are capable of conscious suffering, it is likely that we will only be able to prevent them from suffering on a large scale if this capacity is clearly recognised and communicated by researchers. However, given the uncertainties about consciousness mentioned above, we may create conscious AI systems long before we recognise we have done so […]
There is also a significant chance that we could over-attribute consciousness to AI systems—indeed, this already seems to be happening—and there are also risks associated with errors of this kind. Most straightforwardly, we could wrongly prioritise the perceived interests of AI systems when our efforts would better be directed at improving the lives of humans and non-human animals […] [And] overattribution could interfere with valuable human relationships, as individuals increasingly turn to artificial agents for social interaction and emotional support. People who do this could also be particularly vulnerable to manipulation and exploitation.
One of the founding ideas of Less Wrong style rationalism was that the arrival of strong AI set a deadline on philosophy. Unless we solved all these seemingly insoluble problems like ethics before achieving superintelligence, we would build the AIs wrong and lock in bad values forever.
That particular concern has shifted in emphasis; AIs seem to learn things in the same scattershot unprincipled intuitive way as humans; the philosophical problem of understanding ethics has morphed into the more technical problem of getting AIs to learn them correctly. This update was partly driven by new information as familiarity with the technology grew. But it was also partly driven by desperation as the deadline grew closer; we’re not going to solve moral philosophy forever, sorry, can we interest you in some mech interp papers?
But consciousness still feels like philosophy with a deadline: a famously intractable academic problem poised to suddenly develop real-world implications. Maybe we should be lowering our expectations if we want to have any response available at all. This paper, which takes some baby steps towards examining the simplest and most practical operationalizations of consciousness, deserves credit for at least opening the debate…
If you’re a materialist, you probably think that rabbits are conscious. And you ought to think that. After all, rabbits are a lot like us, biologically and neurophysiologically. If you’re a materialist, you probably also think that conscious experience would be present in a wide range of alien beings behaviorally very similar to us even if they are physiologically very different. And you ought to think that. After all, to deny it seems insupportable Earthly chauvinism. But a materialist who accepts consciousness in weirdly formed aliens ought also to accept consciousness in spatially distributed group entities. If she then also accepts rabbit consciousness, she ought to accept the possibility of consciousness even in rather dumb group entities. Finally, the United States would seem to be a rather dumb group entity of the relevant sort. If we set aside our morphological prejudices against spatially distributed group entities, we can see that the United States has all the types of properties that materialists tend to regard as characteristic of conscious beings…
As we think about thinking, we might we might send thoughtful birthday greetings to Claude Lévi-Strauss; he was born on this date in 1908. An anthropologist and ethnologist whose work was key in the development of the theory of Structuralism and Structural Anthropology, he is considered, with James George Frazer and Franz Boas, a “father of modern anthropology.” Beyond anthropology and sociology, his ideas– Structuralism has been defined as “the search for the underlying patterns of thought in all forms of human activity”– have influenced many fields in the humanities, including philosophy… and possibly soon, the article above suggests, computer science.
Anil Dash, with a grounded view of artificial intelligence…
Even though AI has been the most-talked-about topic in tech for a few years now, we’re in an unusual situation where the most common opinion about AI within the tech industry is barely ever mentioned.
Most people who actually have technical roles within the tech industry, like engineers, product managers, and others who actually make the technologies we all use, are fluent in the latest technologies like LLMs. They aren’t the big, loud billionaires that usually get treated as the spokespeople for all of tech.
And what they all share is an extraordinary degree of consistency in their feelings about AI, which can be pretty succinctly summed up:
Technologies like LLMs have utility, but the absurd way they’ve been over-hyped, the fact they’re being forced on everyone, and the insistence on ignoring the many valid critiques about them make it very difficult to focus on legitimate uses where they might add value.
What’s amazing is the reality that virtually 100% of tech experts I talk to in the industry feel this way, yet nobody outside of that cohort will mention this reality. What we all want is for people to just treat AI as a “normal technology“, as Arvind Naryanan and Sayash Kapoor so perfectly put it. I might be a little more angry and a little less eloquent: stop being so goddamn creepy and weird about the technology! It’s just tech, everything doesn’t have to become some weird religion that you beat people over the head with, or gamble the entire stock market on…
While the arguments above are practical, note that a plethora of tech experts have weighed in with a a note of existential caution: “Statement on Superintelligence.”
Further to which (and finally), a piece from the Federal Reserve Bank of Dallas, projecting the economic impact of AI. It suggests that AI could provide a modest but meaningful boost to GDP over the next 25 years… if The Fed’s “Goldilocks Scenario” (in which, per Dash’s and Kelly’s comments, AI makes consistent incremental contributions to “keep living standards improving at their historical rate”) plays out. You’ll note that they also considered two other scenarios: a “benign singularity” scenario in which “AI eventually surpasses human intelligence, leading to rapid and unpredictable changes to the economy and society” and an “extinction singularity” in which “machine intelligence overtakes human intelligence at some finite point in the near future, the machines become malevolent, and this eventually leads to human extinction.”
As we parse pumped prognostication, we might recall that it was on this date in 4004 BCE that the Universe was created… as per calculations by Archbishop James Ussher in the mid-17th century. Ussher, the head of the Anglican Church of Ireland at the time, attempted to calculate the dates of many important events described in the Old Testament. His calculations, which he published in 1650, were not that far off from many other estimates made at the time. Isaac Newton, for example, believed that the world was created in 4000 BC.
When Clarence Darrow prepared his famous examination of William Jennings Bryan in the Scopes trial [see here], he chose to focus primarily on a chronology of Biblical events prepared by a seventeenth-century Irish bishop, James Ussher. American fundamentalists in 1925 found—and generally accepted as accurate—Ussher’s careful calculation of dates, going all the way back to Creation, in the margins of their family Bibles. (In fact, until the 1970s, the Bibles placed in nearly every hotel room by the Gideon Society carried his chronology.) The King James Version of the Bible introduced into evidence by the prosecution in Dayton contained Ussher’s famous chronology, and Bryan more than once would be forced to resort to the bishop’s dates as he tried to respond to Darrow’s questions.
One punctuation mark in particular is having a moment… a not-altogether-welcome one…
Of the many tips and tricks people are coming up with to determine whether a piece of writing has been written with a little help from AI, the world seems to have homed in on the use of one particular punctuation mark: the em dash.
Though some writers have rushed in to defend the dash — the overuse of which sits alongside pizza glue and bluebberrygate in the pantheon of things people laugh at AI about — perhaps a key reason the prevalence of the punctuation mark seems so bot-like to readers is that, as writers, Americans hardly use it.
Indeed, per a recent YouGov survey, dashes are some of the least used pieces of punctuation in Americans’ arsenals, ranking just ahead of colons and semicolons, per the poll.
As you might imagine, the survey revealed that American adults who describe themselves as “good” or “very good” writers are more likely to use the rarer forms of punctuation on the list. However, for the majority of Americans, marks like the semicolon and the em dash remain mostly reserved for esteemed authors and English teachers… or those who aren’t above enlisting a chatbot for a little help to jazz up their communications.
Interestingly, the vast majority of Americans said they do little writing outside of sending texts and emails, with journaling, nonfiction and fiction writing, and other forms of creative or academic writing all falling by the wayside in 2025, according to YouGov’s research…
As we muse on marks, we might that it was on this date in 1956 that Fortran was introduced to the world. A third-generation, compiled, imperative computer programming language that is especially suited to numeric computation and scientific computing. Developed by an IBM team led by John Backus, it became the go-to language for high-performance computing and is used for programs that benchmark and rank the world’s fastest supercomputers.
In a 1979 interview with Think, the IBM employee magazine, Backus explained Fortran’s origin: “Much of my work has come from being lazy. I didn’t like writing programs, and so, when I was working on the IBM 701, writing programs for computing missile trajectories, I started work on a programming system to make it easier to write programs.”
To the item at the top, it’s worth noting that Fortran is a language with four uses for the dash– subtraction operator, negative sign, line continuation symbol, and range separator (in data processing)– but no em dash.
For a piece of Fortran’s pre-history, see here; and for an important extension, see here.
How do we accomodate ourselves to the prospect of an intelligence far greater than our own? In a consideration of J.D. Beresford’s The Hampdenshire Wonder (the first recognized appearance of the concept in modern Englis-language literature), Ted Chiang unspools the intellectual and cultural history of this now-prevalant trope…
J.D. Beresford’s The Hampdenshire Wonder is generally considered to be the first fictional treatment of superhuman intelligence, or “superintelligence.” This is a familiar trope for readers of science fiction today, but when the novel was originally published in 1911 it was anything but. What intellectual soil needed to be tilled before this idea could sprout?
At least since Plato, Western thought has clung to the idea of a Great Chain of Being, also known as the scala naturae, a system of classification in which plants rank below animals; humans rank above animals but below angels; and angels rank above humans but below God. There was no implied movement to this hierarchy; no one expected that plants would turn into animals given enough time, or that humans would turn into angels.
But by the 1800s, naturalists like Lamarck were questioning the assumption that species were immutable; they suggested that over time organisms actually grew more complex, with the human species as the pinnacle of the process. Darwin brought these speculations into public consciousness in 1859 with On the Origin of Species, and while he emphasized that evolution branches in many directions without any predetermined goal in mind, most people came to think of evolution as a linear progression.
Only then, I think, was it possible to conceive of humanity as a point on a line that could keep extending, to imagine something that would be more than human without being supernatural.
Darwin’s half-cousin, Francis Galton, was the first to suggest the idea that mental attributes like intelligence could be quantified. Galton published a volume called Hereditary Genius in 1869, and during the 1880s and ’90s he measured people’s reaction times as a way of gauging their mental ability, pioneering what we now call the field of psychometrics. By 1905, Alfred Binet had introduced a questionnaire to measure children’s intelligence; such questionnaires would evolve into IQ tests. The validity of psychometrics is quite controversial nowadays, as people disagree about what “intelligence” means and to what extent it can be measured. Some modern cognitive scientists do not consider the term intelligence particularly useful, instead preferring to use more specific terms like executive function, attentional control, or theory of mind. In the future “intelligence” may be regarded as a historical curiosity, like phlogiston, but until we develop a more precise vocabulary, we continue to use the term. Our contemporary notion of intelligence first gained currency around the time that Beresford was writing, and one can see how that converged with the idea of the superhuman in The Hampdenshire Wonder.
The titular character of The Hampdenshire Wonder is a boy named Victor Stott…
… Victor is born with an enormous head but an ordinary body, which disappoints his athletic father but also points to certain assumptions we have about the relationship between the mental and the physical. Beresford could have made Victor both an athlete and a genius, but he opted instead to follow a trope perhaps originated by Wells: the idea that evolution is pushing humanity toward a giant-brained phenotype, which is itself implicitly premised on the idea that mental ability and physical ability are in opposition to one another. This has remained a common trope in science fiction, although there are occasional depictions of mental and physical ability going hand in hand…
[Chiang traces the development of the “superintelligence,” the problems it raises, and the ways that they are treated in The Hampdenshire Wonder and elsewhere– “whatever your wisdom, you have to live in a world of comparative ignorance, a world which cannot appreciate you, but which can and will fall back upon the compelling power of the savage—the resort to physical, brute force.”…]
… In 1993 [Vernor] Vinge [here] argued that progress in computer technology would inevitably lead to a machine form of superintelligence. He proposed the term “the singularity” to describe the date—in the next few decades—beyond which events would be impossible to imagine. Since then, the technological singularity has largely replaced biological superintelligence as a trope in science fiction. More than that, it has become a trope in the Silicon Valley tech industry, giving rise to a discourse that is positively eschatological in tone. Superintelligence lies on the other side of a conceptual event horizon. When considered as a purely fictional idea, it imposes a limit on the kind of narratives one can tell about it. But when you start imagining it as something that could exist in reality, it becomes an end to human narratives altogether.
TheHampdenshireWonder does posit a kind of eschatological scenario, but of a completely different order. After Victor’s downfall, Challis recounts the conclusion he came to after a conversation he’d had with the child, revealing a profound terror about the finiteness of knowledge:
Don’t you see that ignorance is the means of our intellectual pleasure? It is the solving of the problem that brings enjoyment—the solved problem has no further interest. So when all is known, the stimulus for action ceases; when all is known there is quiescence, nothingness. Perfect knowledge implies the peace of death…
… The idea that the search for understanding will inevitably lead to a kind of cognitive heat death is an interesting one. I don’t believe it and I doubt any scientist believes it, so it’s curious that Beresford—clearly an admirer of scientists—apparently did. Challis talks about the need for mysteries that elude explanation, which is a surprisingly anti-intellectual stance to find in a novel about superintelligence. While there is arguably a strain of anti-intellectualism in stories where superintelligent characters bring about their own downfall, those can just as easily be understood as warnings about hubris, a literary device employed as far back as the first recorded literature, “The Epic of Gilgamesh.” But The Hampdenshire Wonder, in its final pages, is making an altogether different claim: The pursuit of knowledge itself is ultimately self-defeating.
Nowadays we associate the word “prodigy” with precocious children, but in centuries past the word was used to describe anything monstrous. Victor Stott clearly qualifies as a prodigy in the modern sense, but he qualifies in the older sense too: Not only does he frighten the ignorant and superstitious, he induces a profound terror in the educated and intellectual. Seen in this light, the first novel about superintelligence is actually a work of horror SF, a cautionary tale about the dangers of knowing too much…
As we wrestle with reason, we might wish a Joyeux Anniversaire to silk weaver Joseph Marie Jacquard; he was born on this date in 1752. Jacquard’s 1805 invention of the programmable power loom, controlled by a series of punched “instruction” cards and capable of weaving essentially any pattern, ignited a technological revolution in the textile industry… indeed, it set off a chain of revolutions: it inspired Charles Babbage in the design of his “Difference Engine” (the ur-computer), and later, Herman Hollerith, who used punched cards in the “tabulator” that he created for the 1890 Census… and in so doing, pioneered the use of those cards for computer input… which is to say that Jacquard helped create the preconditions for AI (among all of the other things that computers can do).
You must be logged in to post a comment.