
CHATGPT won't ever be the same again after this
What this covers
This is a conversation between an MLST host and Stephen Wolfram, recorded around the announcement of the ChatGPT plugin that lets the model call Wolfram Alpha and the Wolfram Language. Wolfram explains why a large language model needs that link: it continues text from statistical patterns in web pages and books, but cannot do serious computation. The talk then widens from the plugin to what it means to understand. Wolfram offers an operational definition of understanding: turning natural language into precise computational language that can actually be run. He also describes a feedback loop in which computed discoveries are published and later absorbed into training corpora.
The middle of the conversation turns to theory. Wolfram argues that even tiny programs, with four to six rules, can produce very complicated behavior, and that computational irreducibility means the consequences of rules often can't be predicted without running them. He calls it a stroke of luck, since without it the future would be fully determined and life meaningless. The host and Wolfram also discuss why neural nets generalize in human-like ways, noting the lineage from McCulloch and Pitts. The host points out that designing architectures is still trial and error, and that language models must interpolate because raw n-gram statistics would need more data than there are atoms in the universe.
The most speculative stretch uses the ruliad, Wolfram's entangled limit of all possible computations. Observers, he says, see only a slice of it, set by their computational boundedness and persistence through time, and he claims this yields general relativity, quantum mechanics, and the second law. Humans, animals, and AIs become different points in that space. Thomas Nagel and Noam Chomsky frame the limits of conceivability. The closing sections ask which theorems or computational creations humans will find interesting, with examples such as early fractal mosaics and the proteins biology happens to use. Several of these ruliad claims are labeled fringe or contested in the source material.
Wolfram argues that ChatGPT's integration with Wolfram Alpha and language represents a merger of statistical AI (language models) and symbolic computation, enabling genuine knowledge creation and the emergence of AI as a distinct computational civilization alongside humans.
- ChatGPT alone cannot perform serious computation—it only continues text statistically; Wolfram Alpha provides the computational substrate for irreducible computation and knowledge generation
- This hybrid neurosymbolic architecture represents the first time AI can discover genuinely new knowledge (not reconstituted human knowledge) and feed it back into circulation
- Both humans and AIs can be understood as points in 'ruliad space' (the entangled limit of all computations), and progress means expanding our navigable region in that space
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The analogy between following individual molecules in a gas versus experiencing it as continuous fluid demonstrates how our cognitive and observational limitations impose necessary simplifications on infinitely complex underlying realities, and we must find and remain in these simplified slices to function
“If we were stuck, you know, experiencing the world in terms of this hyperraph being rewritten all the time, we wouldn't be able to sort of go and lead our lives in a simple way. It's like, you know, when you have molecules in a gas bouncing around, if we had to follow every molecule, that would be a very different thing from our typical experience of a gas, which is just this sort of continuous fluid.”
Even extremely simple programs with just four to six rules are capable of generating extremely complicated behavior, a phenomenon of the computational universe discovered in the 1980s that contradicts engineering intuition.
“if you just pick a program at random... you take a very simple program like a cellular automaton program or sharing machine for that matter and you just say I'm going to pick the rules for this thing let's say at random and they're tiny rules let's say just you know four or five rules six rules whatever um and then you say well what does the system do you might say based on our experience with engineering and so on oh it won't do anything interesting and complicated because it's such a simple set of rules that turns out not to be true. That's something I discovered in the 1980s”
Computational irreducibility is a core notion in science: the knowledge of rules does not immediately tell you what consequences those rules will produce; in many systems, the only way to know what happens is to explicitly run the computation, which contradicts the mathematical science paradigm that has existed since the 1600s.
“computational irreducibility which I think is really a a pretty core notion that sort of we're starting to see kind of pop up in a lot of different a lot of different places. It's an important piece of kind of intuition about science. The question is this. If you have a program, you have a set of rules, you know the set of rules, does that immediately tell you what the consequences of those rules will be? In traditional science, well, particularly mathematical science practiced sort of since the 1600s. Kind of the big thing people are always excited about is science lets you predict stuff... what we realize from thinking about kind of computational universe, computational thinking about science based on programs rather than on equations is that isn't always true”
The architectural similarity between McCulloch-Pitts neurons and biological neurons, preserved in modern artificial neural networks, may be fortunate and responsible for human-aligned generalization, but this alignment is not obviously necessary
“possibly because the underlying sort of makullik pits level architecture is quite similar between our brains and and what happens in neural nets and I think that that's so you know that that's a that's something that has been fortunate I don't think and you know I mean for me you know I first looked at neural nets back in 1981 or 1980”
Computing the probability of even 10-tuple sequences of language would require more data than there are atoms in the universe, which means language models must perform a kind of clever interpolation beyond mere statistics
“computing the probability of even 10 tpples of language would require more data than there are atoms in the universe. Um which is very interesting. So language models are clearly doing some kind of clever interpolation.”
ChatGPT can achieve computational superpowers by calling on Wolfram Alpha and Wolfram Language through its plugin mechanism, because ChatGPT as a large language model is limited to piecing together text based on statistical patterns from the web and books, but cannot perform serious computation.
“chat GPT can sort of get computational superpowers by calling on wolf from alpha and wolf from language uh through its plug-in mechanism uh the you know chat GPT is a large language model which has kind of gone through the web and lots of books and so on and it is piecing together things by sort of seeing what a reasonable piece of continuing text would be and it's amazing how far that gets... the thing it isn't able to do is actually do sort of serious computation”
Neural nets work by discovering regularities in language and creating models that extend beyond merely statistical patterns to make continuations of text that correspond to what humans would expect, even though the mechanism inside the neural net may be incomprehensible to humans.
“what chat GPT has discovered regularities in language... by the time it's any substantial number of words there's no way it can have sort of you know statistical data to just work out the probability from that. So it has to have a model... they do this modeling. They do this continuation. They do this extension beyond the merely statistical so to speak by in a way that corresponds to what we humans expect”
Nagel's thought experiment about subjective experience ('what is it like to be a bat') and Chomsky's nativism about cognitive constraints establish that conceivability is limited by brain structure; humans can only conceive of things within their cognitive grasp, creating a trajectory space of conceivable concepts.
“Nagel's thought experiment um about the bat is is a subjective form of conceivability. And Nam Chomsky talks as a nivist about the domain of things which are conceivable as a function of the cognitive prize which are in our brain. So there's this trajectory space of things which are conceivable”
Fractal patterns (nested, self-similar structures) were invented by a mosaic-maker in Italy in the early 1200s but were completely unnoticed by art historians for centuries, only becoming recognized as 'fractals' in recent decades—demonstrating that new paradigms allow us to see things that were always present but unperceived.
“fractal patterns, nested patterns, right? Things that we've known about, you know, sort of popularized for the last oh, I don't know, three or four decades and sort of known about for about 100 years. Well, actually those kinds of nested patterns, so far as I know, were first invented by a mosaic layer in Italy in the 1200s, early 1200s. made this thing that particular person died. The uh you know the the the art of making that particular nested pattern disappeared. The art historians wrote about many of the mosaics done by this family of mosaic layers. They never mentioned the nested patterns. They talked about you know the birds and lions and so on but they were completely blind to this idea of nested patterns which then in recent times everybody who would see this pattern say oh that's a fractal pattern”
The ruliad is the entangled limit of all possible computations—imagine starting all possible Turing machines with all possible initial tapes and examining the states they produce, converging into a single giant structure underlying all physics and mathematics.
“this big idea of this thing we call the rulad which is kind of the entangled limit of all possible computations. Imagine you start all possible touring machines with all possible initial tapes. You start them all off together and you sort of simply look at what states those things produce. So there might be two touring machines that produce different states and then later on those states will be will converge because they'll end up being the same and you get this this kind of giant structure”
Our experience of the physical world reflects our position as observers embedded within the ruliad, observing a particular slice corresponding to how our sensory data and minds work; just as physical position determines what we see in space, our observer characteristics determine what we see in computation space.
“our experience of the physical world is our we are embedded within this rulad and our experience of the physical world is taking some slice of that rouad experiencing kind of the set of all possible computations in a particular slice that corresponds to the ways that we kind of that our sensory data work and that our minds work”
Different minds can be understood as different points in ruliad space; humans are close together, animals further away, and AIs somewhere in that space; progress in paradigms means expanding our navigable region in ruliad space to encompass more ways of thinking about the world.
“minds experience the rulad, experience this sort of entangled limit of all possible computations. It's like we are at a particular place in the rouad experiencing sort of the world in this particular way just as we're at a particular place in physical space experiencing the world with respect to that place in physical space. And so you think about sort of different minds as being different points at different points in the rouad. So, so different human minds might be quite close in the rouad. Kind of the the animals might be a bit further away. The AIs are somewhere in the rouad”
Space itself appears continuous and regular because it represents a computationally reducible slice of the underlying hyperrewriting system that operates at the fundamental level; this continuity is a property of how we observe the Ruliad, not fundamental to the Ruliad itself
“The fact that we believe space is continuous that is a feature of a slice of computational reducibility from the underlying computational irreducibility of this complicated as you know the the way we set it up with this you know rewriting rules for some hyperraph that's going on underneath. That's a very complicated thing.”
With only two constraints—computational boundedness and persistence through time—the slice of the ruliad that humans observe will satisfy all three major theories of 20th century physics: general relativity, quantum mechanics, and statistical mechanics/second law of thermodynamics.
“if you put only a small set of constraints on the way that we observe what's happening in the rouad actually really two constraints one is that we are computationally bounded we're limited in the computations we can do the other constraint is we believe that we are persistent in time even though at every moment we get made from different atoms of space we have the belief and the experience that there's a sing that we that we persist through time. Those two constraints alone are enough to basically tell you that the slice of the rulad that we observe will satisfy physics as we know it. In fact, specifically the three big theories of 20th century physics, uh general relativity, the theory of gravity, quantum mechanics, and the statistical mechanics, the second law of thermodynamics”
Biological tissue is neither liquid nor solid but a unique computational phase of matter where molecules are transported in semantically meaningful ways rather than randomly; molecular biology reveals structure in what appears random.
“when when you have biological tissue what is that is it a is it a liquid is it a solid well it's really neither of those things it's say it's a weird phase of matter that is kind of a computational phase of matter where we increasingly know from molecular biology that biology sort of explicitly transports molecules around in sort of quotes meaningful ways. They don't just run around randomly. Not that they're really random in a in a liquid, but they they seem random to us in a liquid, but we can we can already recognize that there is sort of structure in the way that things are transported around inside cells, inside molecular biological processes”
Construction of neural net architectures is currently an art form based on trial and error with ad-hoc shapes, not backed by rigorous science; developing a more intelligible science of machine learning architectures is an open challenge.
“the construction of these architectures is an art. It's really just people do some trial and error. They come up with some horseshoes or whatever shape, you know, of uh of architecture happens to work. you know, GPT is in that, you know, same kind of category. You point out there's no real science, you know, behind that.”
The boundary between complete randomness and ordered systems is where irreducibly complex computations and touring machines exist, and this boundary region contains the interesting computational structures and phenomena
“there's sort of total pure chance. there's complete randomness and at this narrow boundary between there's really chaos which is where all the interesting stuff happens. It's where the irreducibly complex computations are. It's where touring machines are. It's where where all this rich behavior is kind of on that on that boundary and it's almost the sliver”
ChatGPT plus Wolfram Alpha enables the creation of new knowledge through irreducible computation, allowing the system to tweet new discoveries that could be incorporated back into training corpora, creating a feedback loop of knowledge generation and dissemination
“As soon as you have actual computation in the loop and what I would talk about as as irreducible computation, you can build what amounts to new knowledge. And so yes, it it surely is the case that you know, you ask chat GPT plus wolf from some some question. It goes off. It it creates a piece of wolf from language code or something. It goes and and runs somewhere. It finds new knowledge. it could then if it wanted to it could tweet that new knowledge and then that new knowledge could get incorporated into the kind of uh into the corpus of knowledge corpus of language that can be used by chatgbt and so on.”
Understanding in computational systems has an operational definition: taking natural language input and converting it to precise computational language from which computations can be executed, which is the gold standard of understanding more rigorous than asking whether a language model 'understands'.
“for me sort of the the the sort of the gold standard of understanding is you can convert it to a computational language. You can compute all sorts of things about it”
In the computational universe, there is an infinite supply of computational creativity, but only those creations humans have learned to care about (connected to our narrative, language, and civilization) are pulled from abstract computational space into the human sphere.
“there's an infinite supply of computational creativity yes there's an infinite supply of computational creativity. the challenges of those things that are created, which are ones that we humans sort of care about, which are ones where in the development of our civilization, our natural language, things like this, we've actually got to the point where we think about those kinds of things”
Computation is the 'giant escape valve' allowing humans to do huge numbers of new things, but the question is whether our understanding keeps pace with computation's ability to create the new.
“computation is the is the giant escape valve so to speak because it allows us to do huge numbers of new things the issue is does our understanding kind of follow along with the actual sort of ability of computation to to kind of create the new so to speak”
We should be 'lucky' that computational irreducibility exists because without it, the future would be fully determined and knowable, making life meaningless; computational irreducibility means that the passage of time achieves something—future states cannot be read off from present rules.
“there are always sort of unexpected consequences. there are always things you couldn't readily foresee from the underlying rules that you had. But I think that that um so this this phenomenon of computational irreducibility that is that's kind of the hard that's hard computation that's computation you can't reduce in a sense we should consider ourselves lucky that there is irreducible computation because you know we lead our lives we do all these different things it would be really bad in a sense one might feel it would be really bad if somebody could just say you know okay you don't need to lead your life I know what the answer is it's 42 or whatever else. There's nothing sort of achieved by the passage of time. What computational irreducibility tells one is that there's sort of something achieved by the passage of time. Some some sort of computational achievement from the passage of time”
The question of which mathematical theorems are humans likely to find interesting is an open and difficult problem; perhaps large language models trained on what humans have cared about could help identify which regions of possible theorems deserve attention.
“this is something uh uh I've been curious about and I did some experiments but they haven't been very successful yet is kind of of all possible let's say mathematical theorems which ones are ones that a human is likely to consider interesting”
Progress is gradual and slow because humans must build intellectual and narrative infrastructure—'ambient ideas'—before new concepts become comprehensible; one cannot simply jump to an arbitrary place in ruliad space and understand it.
“Progress is gradual and slow. In other words, if you jump to some random place in the rouad, you pick a random program, you look at it, you know, when you first look at it, it's like, I don't know why I care about this. It's something where you we have to have a a kind of a I don't know whether you can call it cognitive or or or intellectual history kind of path that gets there. We don't we don't get to talk about the you know until we've we've built sort of the ambient ideas until we've built this kind of intellectual path it isn't really it doesn't really work to just jump to that place”
Biology has selected a particular domain of proteins from all possible proteins; the question is what correspondence exists between what proteins biology uses and what concepts humans have represented in language.
“there's the proteins that well there's the proteins that biology happens to use which is a tiny set of all possible proteins and sort of what the correspondence is between kind of what happens to get used in biology and what we happen to be thinking about and representing in human language”
A civilization of AIs is in the process of developing; already much of what appears on the web comes from AI processes rather than human decision-making, and this will increase, requiring humans to coexist with an AI sphere as we coexist with the natural world.
“there's sort of a whole separate civilization of the AIS that is is in the process of developing. I mean we already see that a whole bunch in uh in the things that operate on on the web today and so on but I think we'll see we'll see an increasing amount of that going forward... already a lot of things that you know we're presenting with on on the web and things like that come out of AI like processes not out of some kind of uh uh human decision-m”
Computational irreducibility is the enemy of neural net training because neural networks rely on predictive shortcuts—if a system is computationally irreducible, you cannot burn in a prediction of its outcome without explicitly computing it, making training far harder or impossible.
“computational irreducibility is the enemy of neural net training because as soon as there is computational irreducibility, as soon as there's actually something you have to explicitly compute, you just can't do this approach of saying well you know training is about sort of knowing what will happen uh you know being able to sort of explicitly burn in a prediction of what will happen and that is explicitly kind of that's that's sort of the opposite direction from computational reducibility”
Discovering and naming previously unknown patterns in physical reality is valuable and achievable; there are infinite collections of exploitable pieces of computational reducibility we haven't yet found or utilized
“in a sense that's why you know if we take it at a more general level if you say well what inventions can we make given the physical world which has computational irreducibility what things can we predict what things can we invent which have foreseeable behavior and so on and the answer is there will be there will be an infinite collection of these things we'll never run out of inventions little pieces of computational reducibility that we can find in this ultimately computational irreducible universe in which we live”
A possible science of machine learning might emerge by thinking of all possible training paths as a space (as quantum mechanics does with all possible histories), from which one could apply physics concepts like event horizons and continuum structures to machine learning.
“Imagine you followed all possible paths. It'd be a computationally absurd, you know, thing to do. But the universe does something very much like that in the way that quantum mechanics works in physics... Imagine you could start thinking about kind of that space of possible paths that were followed. Imagine you could start using ideas from physics to think about that that set of possible paths... there are ways that you could imagine doing some kind of science where you start saying well there's an event horizon that will form in this part of sort of training space”
A large language model is a dynamical system that has been trained and is now frozen—it takes a snapshot of the world as humans have created it, and in doing so becomes a standardizing force for knowledge work, similar to how the invention of typography standardized appearance of written text.
“chat GPT and large language models in general are as you say taking a snapshot of the world as we humans have created it and they're just saying okay this is now the standard you know every essay should now be written this way because this is the average of what the world has produced so far and it's kind of a funny time it's sort of a a standards moment for a lot of kinds of knowledge work”
Even if AI produces structured concepts humans do not understand, those concepts might be actionable through Wolfram Alpha (generating useful consequences in the world) before or without humans achieving conscious understanding of the concepts themselves.
“Maybe the case too. Maybe we don't need to understand it, right? Because as long as let's say the the uh AI is producing structured concepts even if we don't understand those concepts if they're structured enough that that uh programs in Wolfframe Alpha can make use of them to produce actions let's say in the world that are good for us then we may not even need to understand those concepts at least not yet”
Protein folding success in machine learning likely works by learning correspondences with physically observed chunks of protein structure and then determining how to fit and smooth these chunks together, not by understanding the fundamental physics of folding.
“the way those things tend to work is they're they're sort of just making a correspondence with what's been physically observed to large chunks of the of the protein. And then the big effort of the machine learning model is to figure out how to fit together those sort of physically known chunks, so to speak, and how to sort of smooth out the correspondence between those”
Logic itself was discovered by Aristotle through a machine-learning-like process: examining examples of rhetoric, identifying patterns in arguments that work, and extracting abstract logical structures like syllogisms.
“my imagination of what Aristotle did is something much like a machine learning system would do. He looked at lots of examples of rhetoric and he said what are some patterns of how language works in these pieces of rhetoric that are sort of making arguments and so on. what are the patterns of language that successfully make arguments and from that he extracted you know syllogistic logic”
One could make up a word for any feature one observes (like shapes in clouds), but without a reason to care—without narrative connection to what matters in our culture—the new word does not connect to human concerns and remains isolated from the human sphere.
“I could make up a word and uh I don't know I could make up a word for I don't know let's say look at look at clouds in the sky and they have all kinds of weird different shapes and we have a few names for a few of those shapes. Right now we don't have a reason to care about the particular details of those shapes but you know I can make up a word for one of those things but until I can sort of tell you why we care about that it becomes something that's very hard to it's it's something we're sort of not connected to”
Wolfram Alpha was the first serious example of a linguistic user interface on the input side, taking natural language queries and converting them to computational form; ChatGPT adds linguistic output, creating bidirectional linguistic interface to the computational world.
“I kind of view sort of what chat GPT is doing and a lot of its potential uses as being kind of a new form of user interface. is kind of a linguistic user interface in the way that we've had in the past kind of you know we had graphical user interfaces etc. I mean, in a sense, Wolfman Alpha was, I think, the first sort of serious example of a linguistic user interface on the input side. We were not dealing with the output side. We're not writing essays. We're we're just generating essentially more like a a gooey kind of output, a linguistic input, so to speak”
Intelligibility (the ability to express an idea clearly) is more important than conversational interface design; a linguistic interface should prioritize being comprehensible to many people in many situations rather than mimicking natural conversation.
“it's an interface that that can be used by many people in many different situations, but the key thing is that it's about being intelligible rather than being a conversational interface”
The question of how goals are implemented computationally is a matter of scientific interest, but the practical users of AI systems may not require understanding the implementation details—only that goals are achieved.
“I think this is the kind of question of you know once we define the goals how those goals get implemented is something that is a matter of interest for science but it's not necessarily a matter of relevance for the people using it”
Wolfram first attempted neural nets in 1980-1981 but found them too complicated to understand scientifically, leading him to study simpler systems like cellular automata, which revealed rich behaviors that would apply even more so to neural nets.
“for me you know I first looked at neural nets back in 1981 or 1980 and uh you know they didn't do anything interesting. I couldn't really get them to do anything. I I decided were way too complicated to understand scientifically. So I started looking at things like cellular automter where you just have these fixed cells that have uh very simple behavior and you don't have the sort of connection to everything else and all these weights and so on. And uh turned out that you know I managed to figure out a very rich set of behaviors that happen even in these much much much simpler systems and those kinds of behaviors certainly will happen even more so in neural nets”
Neural network architectures matter far less than conventional wisdom suggests because the principle of computational equivalence implies that sufficiently complex networks can learn anything if 'bashed hard enough'—the practical differences may be orders of magnitude but not fundamental.
“the principle of computational equivalence kind of tells you in the end it doesn't really matter in the end if you kind of bash it hard enough it's going to be able to sort of learn quotes learn anything now there may be when I say bash it hard enough there might be a factor of a billion difference in practice And that makes big difference in in in the world as it is”