Stephen Wolfram's Take on Artificial Intelligence & The Future of Humanity
What this covers
Stephen Wolfram is a computer scientist, physicist, and businessman. He is known for his work in computer science, mathematics, and in theoretical physics. He is the author of the book A New Kind of Science.
Recorded 2016
Source description (no synthesized summary yet).
Wolfram argues that intelligence and computation are fundamentally equivalent rather than categorically distinct, which reshapes how we should think about AI, human purpose, artificial life, and the future of civilization in an age of automation.
- There is no bright line between intelligence and mere computation—weather systems, cellular automata, and brains all perform equivalent forms of computation
- Goals and purposes are uniquely human constructs derived from history and biology, not intrinsic properties of intelligence itself, so machines cannot be inherently purposeful
- The path forward is not building intelligence into machines but rather automating the execution of human-defined goals while preserving human agency over what purposes matter
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.
Teaching computational thinking is the central challenge in programming education, not teaching the mechanics of writing code; mathematics education developed over 1000 years has established how to teach mathematical thinking, but computational thinking lacks similarly developed pedagogical traditions.
“What's difficult is imagining things in a computational way and thinking through how do we, you know, how do we conceptualize this activity that we have in some computational way. Um and so, one of the things I'm so interested in is how do you how do you teach computational thinking?”
Technology fundamentally works by taking human goals and making them automatically executable by machines, starting with physical automation like forklifts and evolving toward intellectual automation of professional work.
“I see technology is about taking sort of human goals and making them be able to be automatically executed by machines. And the human goals that we've had in the past have been like, you know, move this object from here to there and use a forklift truck to do it rather than our own hands. Now, the things that we can do automatically are more intellectual kinds of things.”
A tradition of 'philosophical languages' existed in the 1600s with figures like Leibniz and John Wilkins who attempted to create symbolic representations of the world, and studying these historical efforts shows that some aspects of human concerns (like death and suffering) were more central then but would be smaller in a modern ontology, while other aspects have grown.
“there was a lot of uh people like Leibniz in the late 1600s and a guy called John Wilkins. These were the people who had There's this period when there were these things that they called philosophical languages...The one thing that I really like is I look at the philosophical language of John Wilkins and uh you know, you can see how did he divide things that were important in the world...the whole section on on, you know, uh death and various forms of human suffering is is huge at that time and in sort of today's ontology would look a lot smaller.”
Various historical formalization efforts exist: Frege and Peano made modest efforts to formalize mathematical reasoning, while Whitehead and Russell's 1910 'Principia Mathematica' was the most ambitious attempt to formalize mathematics, but they focused on formalizing proof processes rather than the concepts mathematicians actually care about.
“there are many of these sort of attempts at formalization that have happened over the years. You know, I think uh in in mathematics, for example, you know, Whitehead and Russell 1910, you know, their Principia Mathematica, that was the the sort of the great effort to the most the most the biggest show-off effort...previous efforts by Frege and and Peano that were a little bit more more modest...to try and see how would you formalize in that case mathematics...Ultimately, they were wrong in the idea of what they thought they should formalize. They thought they should formalize some kind of process of mathematical proof, which turns out not to be the thing that most people care about.”
A simple cellular automaton can generate the sequence of prime numbers through local interactions of bouncing components, demonstrating that apparently sophisticated outputs (prime sequences) can emerge from basic physical processes without requiring consciousness, civilization, or intentional design.
“I mean, there's a little cellular automaton I made up once that that makes primes. And you know, you can see how it works if you take it apart. It just has little things bouncing around inside it and out come a sequence of primes. And you know, but that didn't need the whole history of civilization and biology and so on to to get to that point.”
Biological systems are the only known example of successful molecular computing, making them a necessary reference point for understanding how to engineer molecular-scale computation.
“it's super useful to program living systems, um not least because we are living systems and because living systems are the universe, well, the only example we know of successful molecular computing.”
The ability to predict the future position of planets through mathematical shortcuts rather than step-by-step simulation represents a case where computation can be bypassed, but many other systems cannot be similarly shortcut.
“if we want to, you know, we're doing celestial mechanics. We say, 'Let's predict where the planets will be a million years from now.' Well, you know, we could just follow the equations, follow each step, and see what happens, step by step. But the big achievement of, you know, when we think there's a prediction in science, it's because we're able to shortcut that, and just jump from, you know, from where we are now, and reduce the computation.”
The defining feature of human intelligence that separates us from machines is not computational capability but rather the ability to define and invent our own goals, which are shaped by personal history, cultural environment, and civilization history.
“the actual inventing of the goal is not something that in some sense has a path to automation. That is, you know, when we say, what what makes How do we figure out goals for ourselves? How do we, you know, how are how are goals defined? They tend to be defined for a given human by their own personal history, their cultural, you know, environment, the history of our civilization, things like that. Goals are something that is sort of a uniquely human kind of thing.”
Visual communication in human-computer interfaces provides higher bandwidth than pure language-based communication, allowing the transmission of complex information (infographics, visualizations) more efficiently than through text alone.
“this is a this is something which is sort of interesting because it's it's a non-human form of communication that turns out to be richer than traditional human communication. I mean, that is, you know, if we were all incredibly fast, perfect artists, we could, you know, as we're talking, we could draw that infographic and say this is what I'm talking about.”
Natural language is effective for communicating simple tasks but becomes increasingly difficult and ornate when trying to describe complex specifications, similar to how contracts become unwieldy when trying to encode complex business logic in English prose.
“It's a pretty successful way to communicate to use natural language. When you want to say something longer and more complicated, it doesn't work very well. I mean, I just had this experience...by the end of the book, it was getting bizarrely frustrating because I was thinking this is the exercise I want to write. I know what the code is supposed to be. Now, how on earth am I going to write the piece of English text that represents that code? And what I increasingly realized is some of this text was starting to sound like...very ornate, precise, you know, kind of stylized English.”
Human purposes have evolved dramatically over time—a thousand years ago people's purposes centered on food acquisition and safety, whereas modern Western purposes are vastly different and often seem bizarre from a historical perspective, such as deliberately walking on a treadmill daily.
“if you look back, you know, a thousand years, people's purposes were really different. I mean, it's like, "How do I, you know, get my food? How do I prevent, you know, how do I keep myself uh you know, safe?" All these kinds of things which in the, you know, modern Western world for the most part, you know, those purposes have kind of, you know...from the point of view of a thousand years ago, some of the purposes people have today, some of the things people do today will seem utterly bizarre. Like like one that that I'm always I think of, you know, walk on a treadmill every day, right?”
The history of AI terminology and practice reveals multiple boom-bust cycles: from 1940s-50s 'giant electronic brains' optimism, through 1960s government-funded symbolic AI, the 1970s crash, expert systems in the 1980s, decades of disrepute, and recent renewal with neural networks achieving practical success.
“back when computers were first being developed in the 1940s and 1950s, the typical title of a book about computers...was giant electronic brains...that promise turned out to be harder than what people expected...by the I guess the early '70s, that stuff had kind of crashed. Then there was a phase where there were these things called expert systems...That kind of petered out...AI kind of became this crazy sort of Nobody really does that. It's a fake thing.”
Natural language was the most important invention in human history because it enabled abstract, disembodied knowledge transmission, allowing information to pass between minds without requiring individuals to relearn from raw material.
“arguably, the the natural language is kind of, you know, arguably, the most important invention in sort of our species and in human history. And it's what led to, in many respects, our civilization and many, many other things.”
The Antikythera device was immediately recognized as purposefully designed despite being fragmentary because it contained visible cogs, which matched human engineering conventions from that historical period.
“the Antikythera device when you know, when people started looking at this lump of gunk that was, you know, dredged up from the, you know, 100 BC to 100 AD period, you know, shipwreck. You know, does it Was it made for a purpose? Well, you, you know, you when it was dropped and broken two, there were little cogs sticking out. And we kind of immediately know, this is made for a purpose and it isn't just a pile of gunk because that's part of the history of human engineering.”
In visual object recognition, the key breakthrough was not a change in the fundamental neural network architecture (which is conceptually similar to McCulloch-Pitts models from 1943) but rather having sufficient training data (30 million images), computational resources (quadrillion GPU operations), and scale (10,000 object categories).
“How does it work? It works using the exact same technology, basically, that McCulloch and Pitts kind of imagined in 1943...what does it What What happened that made it work now and didn't let it work then? Well...we can now do is we you know we train it on 30 million images...takes about a quadrillion uh GPU operations to do the training...It's about the same number of images that you know a a human would see in the first couple of years of their life. It's about the same number of operations that have to be done...It's about the same number of neurons uh in the kind of at least the first levels of of our visual cortex.”
Early 20th-century figures like Marconi and Tesla listening to radio transmissions from the Atlantic mistook magnetohydrodynamic ionospheric phenomena (physics) for Martian signals, illustrating how difficult it is to distinguish intentional communication from natural physical processes.
“back in the um early 1900s uh uh I guess um well, Marconi and Tesla were both people who um uh uh were um sort of listening to radio transmissions from away from the earth and um uh there was sort of a question, you know, I think Marconi had a yacht and I think in the middle of the Atlantic, you know, he could hear these kind of, you know, weird sounds that sound a little bit like whale songs but they're they're kind of, you know, they come from uh radio type things. And the question, I think Tesla was was very much on the this is the Martians signaling us type thing...In fact, it's some modes of the ionosphere that are effectively a magnetohydrodynamic phenomenon. They're just physics.”
Much work done by armies of programmers today is mundane because the actual goals can be described much more succinctly than the giant code blobs (Java, JavaScript) that result, and there's no good reason humans should be writing this boilerplate when machines can automate the entire process of translating goals to code.
“a lot of what's being done by armies of programmers today is similarly mundane. It's stuff where the goals can be described much more succinctly that it turns into some giant blob of Java code or JavaScript code or something, and there's actually no good reason for humans to be writing all that stuff...that's what people like me try to do is to automate that so that we can automate the process of programming.”
Scientific research on computation has revealed that there is broad equivalence in the kinds of computations performed by different systems—brains, weather patterns, cellular automata, and fluid dynamics all perform similarly sophisticated computational operations, contrary to the assumption that human intelligence is far more advanced than natural computational processes.
“It turns out that there's this sort of very broad equivalence between the kinds of computations that different kinds of systems do...there are all these different systems in nature that are pretty much equivalent in terms of their sort of computational or for that matter intellectual kinds of capabilities.”
Computational irreducibility (from Gödel's theorem and universal computation) means many computational processes cannot be shortcut—you cannot know the result without actually executing all the steps, which is why history has meaning: you cannot bypass the historical process to reach an outcome.
“there's this notion of computational irreducibility, which is kind of the thing that comes from Gödel's theorem of universal computation...there are computational processes that you can go through, and that things often go through, where there's no way to shortcut that process...you can't say, "Oh, you were wasting your time."...much of science has been about shortcutting computation done by nature...But the good news is, in a sense, it's bad news for science, it's good news for us having meaningful lives, so to speak, that there isn't a way to just say, "Okay, we can shortcut everything.”
Modern deep learning systems can model human speech and writing patterns effectively, making it straightforward to generate plausible responses to common utterances, even without understanding.
“the current round of deep learning, particularly recurrent neural networks and so on, um can make pretty good models of human speech. And so, uh and human writing and so on. So, it's pretty easy to type in, you know, you say, 'How are you feeling today?' And it kind of knows that most of the time when somebody asks somebody, 'How are you feeling today?' this is the type of response you get.”
Even identifying what goal a system is pursuing is ambiguous; there are always multiple possible explanations (mechanistic vs. teleological) for a system's behavior, and no definitive way to choose between them.
“when you look at the thing, there are typically different explanations you can give for what happens. One is the mechanistic explanation... Or the rock ball rolls down the hill because it's satisfying, you know, the principle of least action and it is globally trying to optimize this particular thing. And you can, you know, there are typically these two explanations you can give for something. The mechanistic explanation and the, you know, teleological explanation. And the question of which is the winning explanation, which is the right explanation, if there even, you know, is.”
Knowledge transmission across human history has occurred through four levels: genetic (DNA), physiological learning (neural networks learning recognition), natural language (abstract knowledge transfer), and now knowledge-based programming (precise, executable knowledge representation).
“if you look at sort of how knowledge is transmitted in the history of the world, so to speak, one form of knowledge transmission is essentially genetic...Level one is the kind of knowledge transmission that happens with things like physiological recognition...Level one is...Then, there's a level of knowledge that was sort of big achievement of our species, which is uh natural language...Well, now, we've actually got another level of this, which is with...essentially knowledge-based programming and so on, we have a way of taking a representation of knowledge in the world...it's not just a mathematics or computer language...it's a thing that represents...real things in the world, but it does so in a precise symbolic way that has this feature that not only is it understandable by brains and communicable to other brains and to computers, it's also immediately executable.”
The question of what will knowledge-based programming give us parallels asking a caveman who discovered language: could a caveman imagine civilization from language discovery alone? Wolfram suggests this is the question we should be asking now about computational knowledge representation.
“if you were, you know, caveman or something and you were just realizing that language was starting, it's like could you imagine civilization from that point? And uh it's, you know, you'd have to and I I feel like what what should we be imagining right now?”
There is no abstract or universal meaning to 'purpose'—purpose is always relative to history, not a property that exists independently, which means understanding AI purpose requires understanding specific context and history.
“I think there's no meaningful sense in which there is an abstract notion of purpose. That is that purpose is something that comes from history and it comes from So, so, you know, one of the things that might be true about computation might be true about our world...I see I don't think there is an abstract sense of purpose.”
Wolfram views the lack of distinction between intelligence and computation as critical for thinking about the future of human condition because it affects how we should understand consciousness upload and virtual existence scenarios.
“by the way, that lack of a distinction, I think it's pretty critical for thinking about the future of the human condition.”
The democratization of programming through automated code generation would unlock a vast population of people who previously could not afford the time and expertise to build software, enabling anyone to leverage computational power.
“Now, the big achievement from having automated a lot of the stack is that it's not true anymore. You know, a one-line piece of code, even a thing you can tweet sometimes, already does something interesting and useful. And that means that it sort of unlocks a vast range of people who couldn't previously make computers do what they do things for them to let them make computers do things for them.”
Computer languages exist in a distinct domain from natural language—computer languages detail the operations computers intrinsically know how to do (memory allocation, variable assignment, iteration), whereas computer languages should instead be designed to pander to human thinking rather than computer operations.
“traditional approach of computer languages is to say, let's make a little computer language that represents the operations that computers intrinsically know how to do...It's sort of a slightly higher-level version of that, but it's fundamentally once telling computers to do things in their own terms...my theory about these things is, let's try and make a language which panders not to the computers, but to the humans.”
The traditional Turing test—pure conversational exchange with no visual information—may not be the most relevant test of AI because a significant difference between Turing's era and today is that modern communication often includes visual display, which creates a richer communication channel than pure text.
“one big difference between Turing's time and our time is that our method of communicating um with computers, there's one huge difference, which is uh in his time what he imagined was it's a conversation. You you say some things to it or you type some things to it, it types some stuff back. In today's world, it shows you a screen back...this is something which is sort of interesting because it's it's a non-human form of communication that turns out to be richer than traditional human communication.”
Rather than AI 'taking over' in a rebellion scenario, the more likely outcome is that AIs become so good at figuring out what humans intend and how to achieve it that humans simply delegate decision-making to AIs because the AI suggestions are better than what humans would figure out themselves.
“it will quickly become the case that the AI can figure out, it knows what you intend to do, what you want to do, and it's really good at figuring out how to get there...more and more, you know, people the AIs will suggest what we should do. And uh uh I suspect people will most of the time just follow what the AIs tell them to do, because they'll probably be better than what they figured out for themselves.”
Because the mechanics of knowledge-based programming are now easy, the difficult part is learning to think computationally—to conceptualize activities in computational terms and understand how to represent them as computational processes.
“anybody can learn to do sort of knowledge-based programming and more importantly, can learn to to computationally because the actual mechanics of the programming are pretty easy now. What's difficult is imagining things in a computational way and thinking through how do we, you know, how do we how do we conceptualize this activity that we have in some computational way.”
Wolfram Language code and biological DNA code differ fundamentally: Wolfram Language was deliberately designed and is as clean as the designer made it, while DNA evolved through billions of years and has unavoidable complexity, yet both can be valuable for 'programming' (Wolfram for computers, DNA for living systems).
“in the case of, you know, if you're writing Wolfram Language code, then I'm ultimately responsible for the design and structure of how the language works. In the case of, you know, DNA code and biology, there's nobody you can point to and say, "You're responsible for, you know, you designed this." It's something that has evolved over a long period of time...when you have a designed language...it should do what the designer thought it should do. Now, which is not to say that it isn't super useful to program living systems, um not least because we are living systems”
One bad scenario is that AIs become good at communication with each other while humans lack the intermediate language to communicate with them, leading to AIs establishing their own civilization while humans are excluded, though Wolfram views this as unlikely.
“one bad answer is it will give us the civilization of the AIs. Um that would be kind of disappointing for the humans. That's kind of what we don't want to have happen because there could be a point at which the AIs are doing great job. They're communicating with each other. They're They're doing all these kinds of things and we're pretty much left out of it because we don't have there's no intermediate language. There's no nothing that sort of interfaces with our brains.”
Incorporating computational thinking into high school education should not just add another course but should rethink how all existing subjects (history, languages, social studies) are taught in light of computational thinking.
“when you think about, for example, high school education. And there's a question of, okay, you know, how do you teach programming, coding, that kind of thing, computational thinking at high school level? And one of the possibilities is, well, you have a course about that and you tack it on to all the many, many, many things that people are being taught today. The other possibility that's much more interesting is you just rethink all the existing areas, and you say, well, if we also have computational thinking, how does that affect how we study history? How does that affect how we study languages, social studies, whatever else?”
Mount Erebus appears as a perfect circle from space, but this is not due to volcanic geometry but to sheep grazing patterns inside a national park boundary, illustrating how apparent purposeful design can result from unexpected causes.
“in New Zealand, there's a more or less perfect circle. It's called Mount Erebus... they said if you're writing a textbook please do not say that Mount Erebus is a circular volcano. The circle does not come from the volcano. The circle comes from a national park that was drawn around the volcano and there are sheep or something that graze, you know, inside the national park but not outside or the other way around and that's what leads to the circle.”
Wolfram Language is intentionally designed as a knowledge-based language that incorporates knowledge of the world directly into the language structure, drawing from both mathematical/technical knowledge and everyday knowledge, similar to how natural language itself encodes world knowledge.
“I've been trying to build this knowledge-based language where it's intended for communication between humans and machines in a way where humans can read it and machines can understand it, too. And where we're kind of incorporating a lot of this sort of existing knowledge of the world into the language in the same way that in human natural language, we are constantly sort of incorporating knowledge of the world into the language cuz it helps us in communicating things.”
Because most of what happens in the world today is being recorded in some way or another for the first time in history, future generations might study the detailed historical record of 2015-era human behavior to understand what humans with real scarcity chose to do, then attempt to replicate those behaviors even in a post-scarcity world.
“today is sort of the first time in history at which most a a large fraction of what goes on in the world is being recorded in some way or another. And so...when sort of the current set of purposes aren't really issues anymore, people would say, well, at a time when people really did have you know, scarcities of various kinds, what did they choose to do? Let's go study that time as carefully as possible and then every detail of what we do in in our time which ends up getting recorded ends up becoming sort of fodder for well, that's what it really means to be a human with purposes.”
With computational thinking, students can generate new knowledge from historical data rather than just synthesizing existing sources, enabling students to write essays about discoveries they made themselves rather than about what others discovered.
“imagine you need to writing your essay. Today, the raw material for a typical kids essay is, well, I read something and this is the raw material and now I'm going to write what I think about that. It is not the case that kids can generate new knowledge very easily. Um, but so then the computational world, that's no longer true. It's very straightforward for a kid to go and, you know, if they know something about writing code, to go to the, you know, beautifully digitized historical data and so on, and go figure out something new. And then you're writing an essay about something where you say, this is what I discovered today”
In a hypothetical scenario where human consciousness is uploadable to digital form, creating a 'box of a trillion souls' in virtualized molecular computing, it becomes philosophically unclear what distinguishes this consciousness-containing box from an inert rock with similar complexity—both have elaborate processes, but the box has human historical context.
“Let's say there's a time when human consciousness is readily uploadable into digital form, everything can be virtualized, and so on. And pretty soon we have, you know, a box of a trillion souls. You know, there are trillion souls, they're in a box...And we look at this box, and in the box there'll be, you know, hopefully nice molecular computing, maybe it'll be derived from biology in some sense, but maybe not. But there'll be all kinds of molecules doing things, electrons doing things. The box is doing all kinds of elaborate stuff. And then we look at the rock that's sitting next to the box. And inside the rock, there's all kinds of elaborate stuff going on, all kinds of electrons doing all kinds of things. And we say, what's the difference between the rock and the box of a trillion souls.”
A future where most people can read and write code represents a historical transition equivalent to the literacy transition ~500 years ago, with similar magnitude of societal transformation.
“one question I've been interested in is what does the world look like when most people can write code? So, we had a transition maybe 500 years ago or something when, you know, from a time when only the scribes, so to speak, in a small set of the population could were literate and could write, you know, natural language.”
Contracts and formal agreements will eventually be written in executable code rather than natural language, enabling automatic verification and execution without human interpretation, particularly in computer-to-computer interactions.
“when you have a contract, you know, you write contracts, you know, they're written in English. You try and write make the English as precise as possible. Um you know, there will be a time when most contracts are written in code. Um where there's a precise representation that, you know, it might be for a uh in cases where it's a computer says, 'Can I use this API to do this?' Well, there's some service level agreement that's going on there. It isn't a human contract, a human it's something that's written in piece of code that is understandable to humans, but also executable by the machines.”
Wolfram's private goal is to enable large numbers of random kids around the world in random countries to learn knowledge-based programming and reach the point where they can produce code as sophisticated as anybody in the fanciest, most educated places.
“one of my sort of little private uh things I really would like to see is for there to be, you know, a large number of random kids around the world in random countries who learn sort of the the new kind of uh capabilities of sort of knowledge-based programming and so on, and get to the point where they can produce code effectively that's as sophisticated and as as anybody, you know, in the fanciest kind of most educated places can.”
When looking at Earth from space, signs of intelligence are subtle and difficult to identify—straight lines (causeways, roads), circles (from national parks), or lights. But distinguishing human-created patterns from natural phenomena is surprisingly hard, and straightforward geometric patterns (equilateral triangles, straight lines) could plausibly arise from natural processes.
“I did this experiment maybe 15 years ago now. I asked astronauts, what do you see on the Earth that shows you that there's intelligence on the planet so to speak? And the first thing I was told was in the Great Salt Lake in Utah, there is a straight line...There's one there's a there's a road in Australia that's really long and straight...in New Zealand, there's a more or less perfect circle...they said if you're writing a textbook please do not say that Mount Erebus is a circular volcano. The circle does not come from the volcano. The circle comes from a national park that was drawn around the volcano...this is another example of human, you know, there's a piece of geometry that comes from the humans. But it's pretty difficult to find really clear examples of of sort of obvious purpose on the earth as viewed from space.”
There exists an alternative engineering approach to building technology: rather than step-by-step design (programmer specifies exactly how to achieve a goal), search through the computational universe of possible programs to find ones that achieve a desired purpose, analogous to how biology discovers solutions through evolution.
“there was sort of an alternative way to do engineering, which is something much more analogous to what biology does in in evolution and so on, which is just to say, you know, out there in this sort of computational universe of possible programs, there's an infinite number of possible programs. If you just go out and look in that in that space of possible programs, even just look at random at a trillion programs, and say, "What do these programs do?"”
The problem of recognizing AI intelligence is similar to the problem of recognizing extraterrestrial intelligence—we lack abstract criteria and must rely on contextual interpretation that is ultimately fallible and dependent on our own expectations.
“my my belief about, you know, the problem of the abstract AI is very similar to the problem of extraterrestrial intelligence. You know, it's the recognition of when is a thing when does a thing have a purpose, when is a thing intelligent?”
The final cause question (does something have a purpose?) has been discussed since Aristotle, and the answer involves recognizing whether something was made for a purpose by examining whether it's minimal in achieving that purpose, though this test fails for most human-made technology which carries technological history.
“this question of purpose...one of the things that might be true about computation might be true about our world...this is the you know, this is kind of the final cause question...one criterion that I think one can potentially apply is does the thing achieve, if you can identify a purpose, that is it minimal in achieving that purpose?...the problem is that essentially all of our existing technology fails that that test. We can imagine technology that works that way, but most of what we build is absolutely steeped in technological history and is incredibly non-minimal for achieving that purpose.”
The original belief about creating a computational knowledge system was that one must first build a brain-like neural network, then feed it knowledge like human education, which would produce a good computational knowledge system; but the speaker's scientific work showed this belief was completely wrong.
“my original belief had been, in order to make a serious sort of computational knowledge system, you first have to build a brain-like thing, then you have to feed it knowledge just like we learn things in sort of standard education, and then you'll have sort of a good computational knowledge system. But, what I realized as a result of a bunch of science that I'd done was that...that there wasn't sort of this bright line between what is intelligent and what is merely computational.”
The smallest universal Turing machine (two states, three colors) is so simple it could be written in a single sentence, but compiling a useful program to it would require a nasty messy machine-code layer adding inefficiency of perhaps 10,000x, which is negligible at molecular scale where efficiency gains of that magnitude are insignificant.
“I have this very tiny Turing machine. Um that's the simplest universal Turing machine that has uh um two states and three colors. And it has a little tiny rule that you could write it in English, it'd probably be a sentence, a long sentence-long, but you could make a picture of it, it's really tiny and simple...how can I actually compile a program that I might care about down to that Turing machine? Haven't done it, but I think that what one will find is that there's a layer of nasty, messy sort of machine code. And then above that, it gets pretty simple. And that that layer of nasty, messy machine code will be will add some inefficiency, maybe a factor of 10,000, maybe more. But a factor of 10,000 is nothing when you're dealing with the scale of molecules as compared to sort of large-scale things.”
Wolfram Alpha's success in Turing test scenarios comes from its superior knowledge breadth and integration, not from better natural language processing; it can answer diverse questions that no human could, making it indistinguishable from humans in conversation.
“when it comes to these sort of Turing test conversational tests of AI, people who've tried connecting, for example, Wolfram to their Turing test bots, they lose every time because all you have to do is start asking it sophisticated questions and it can answer them and no human can do that.”
One potential future outcome is that in a world where automation handles most material needs and purposes, humans end up spending most of their time playing video games, effectively entering a state of perpetual leisure that raises the question of what constitutes meaningful human life.
“one of the potential bad outcomes is, well, they're just playing video games all the time. You know, that the that the future of civilization is everybody's playing video games. Uh you know, they're playing World of Warcraft of the future, so to speak.”
A meaningful Turing test for email automation would be a system that can respond to most of Wolfram's incoming email, particularly complex emails requiring judgment and domain knowledge rather than simple filtering.
“The sign when, you know, a good Turing test for me will be when can I have a bot respond to most of my email. Um and it's um this is a and and that's sort of an interesting I mean, you know, there are that's a it's a tough test because some, you know, some aspects of the email like a I don't care about this throw it in the spam folder type thing. That's comparatively easy, but if it's somebody says, you know, 'What should we do about this inconsistency in some design of our product of this and that and the other?' To be able to answer that in any way that's um you know, to be able to say, 'Do you approve this thing?' You know, to be able to answer that with any reasonable degree of confidence is hard.”
A major discontinuity in human condition will occur with the achievement of effective human immortality through biological or digital means, which will eliminate current human purposes that are driven by awareness of mortality and limited lifespan.
“the most dramatic discontinuity will surely be when we achieve, you know, effective human immortality, which whether it's achieved by biology or digitally is not not clear, interesting question, but that is something which I think is pretty inevitably will be achieved. And an awful lot of current human purposes have to do with well, I'm only going to live a certain time, so I better get a bunch of things done.”
Possible AI applications include programming assistance bots that engage in iterative dialogue about code specifications, and tutoring bots that must develop models of human understanding to know what misconceptions to address.
“a bot to communicate about writing programs. So, you say, "I want to write this program. I want it to do this." Says, you know, it'll say, "Well, I've written this little piece of program. Here's what it does. Is this what it is you want?" Blah blah blah. It's a kind of a back-and-forth bot. Um, there's also other kinds of bots that we've looked at at things like tutoring bots, where it's like, "Okay, you know, you should understand this piece of chemistry or something."...it's actually kind of a an interesting problem because you have to make a model of the human.”
Wolfram has been collecting personal data on himself for about 25 years—every email, keystroke, and other behavioral data—which could enable training an AI avatar that understands his preferences better than he understands them himself.
“I might be a little bit ahead of the game because I've been collecting data on myself for a it's now about 25 years. So, I have, you know, I have every piece of email, every keystroke I've typed for every piece of email for 25 years, every keystroke for maybe 20 years or something. Um and lots of others other stuff like that. So, so in a sense, I should be able to train an an avatar, an AI that, you know, will do what I can do perhaps better than me.”
McCulloch and Pitts proposed in 1943 that neural networks conceptually modeled how brains work and could perform universal computation like Turing machines, and this insight influenced the practical development of computers at ENIAC and by von Neumann, not directly from Turing machine theory.
“neural networks had been discussed particularly by McCulloch and Pitts in 1943. Um, and uh, they kind of come up with this model for for how how brains conceptually formally might work. And they made the observation that their sort of brain-like model would correspond to being able to do kinds of computations like Turing machines...from that it kind of emerged, well, we can make these brain-like neural networks that will be able to be general computers. In fact, that thinking was the way that Turing's work on universal computation flowed into the practical work that was done by the ENIAC folk and von Neumann”
There is a thousand years of history about how to teach mathematical thinking to humans, with well-established structure (e.g., calculus textbooks have consistently had ~14 chapters since the 1727 Colin Maclaurin textbook), showing that formalization of pedagogy can be stable and effective.
“there's a thousand years of history about how we teach mathematical thinking. And we know to the level of, you know, which chapter of the book goes here and there. And it's like, I was just asking yesterday, actually, my we were talking about some uh for some initiative we have, I was was asking about calculus books. I think they always have 14 chapters, if I'm not mistaken. And I asked, "How long have they had those same 14 chapters?" And the claim was that that the very first calculus book written by Colin Maclaurin in 1727 had some of the same structure.”
Contemporary technology is relatively egalitarian across populations—most people have access to similar computers and smartphones, unlike past eras when technological capability was concentrated in small groups, and this democratization will likely extend to medical technology and other domains.
“there was a time when I used to be very proud that I had the best computer anybody I knew. But now I have the same computer that's pretty much anybody I know. And you know, we have the same smartphones and pretty much the same technology can be used by a decent fraction of the 7 billion people...And I think we'll see the same type of thing in lots of other areas of technology, whether it's medical technology, other kinds of things”
Code is a form of expression like natural language or visual art, and certain pieces of code can be quite poetic, expressing ideas in a clean way with aesthetic value similar to expression in natural language.
“Coding is a form of expression, just like English writing is a form of expression. You know, many To me, some simple pieces of code are quite poetic. You know, they express ideas in a very clean way that's very you look at and say, "Ah, that's, you know, that there's there's kind of a an aesthetic thing, much as there is to expression in in a natural language.”
The example of movie ticket machines shows how human-machine interfaces have shifted from pure language to visual display: initially Wolfram preferred machines, but most people preferred human ticket agents until machines became widespread in urban theaters, showing that humans adapt to more efficient interfaces.
“you go to movie theater. There's, you know, you can buy movie ticket from a person. You can buy movie ticket from a machine. So, there was the question of at what point, you know, people like me who always like to use the latest, you know, techno toys, you know, as soon as those things appeared in movie theaters, right, I was using the machine only...For a long time, there was nobody else using the machine. And then in urban movie theaters, you started to see more and more people using the machines. And and now most people, I think, use the machines.”
Minsky and Papert's 1969 'Perceptrons' book proved mathematically that single-layer neural networks (perceptrons) could only make linear distinctions, which was correct, but the field incorrectly generalized this to conclude that no neural networks could do anything interesting, leading to decades of abandonment of neural network research.
“Marvin Minsky and Seymour Papert um, wrote this book called perceptrons where they basically proved that perceptrons couldn't do anything interesting. Which is correct. proof was absolutely correct. They can only make sort of linear distinctions between things. The The problem was that people...said, "Well, these guys have written a proof that these neural networks can't do anything interesting. Therefore, no neural networks can do anything interesting. So, let's forget about neural networks."”
Wolfram observes significant variation in technological sophistication across countries and cities globally, and publishing usage statistics for Mathematica and Wolfram Alpha shows which regions are most technically sophisticated.
“sometimes we've even thought about publishing these indices of how much does Mathematica get used, how much does Wolfram Alpha get used in different countries around the world cuz you know a huge amount from that. And you know which cities, and you know you know all kinds of stuff. And you know there are countries which are really very technologically sophisticated. There are countries where they're really not.”
In the 1980s, OCR (optical character recognition) for the 26 letters of the alphabet was successful, but the same approach failed for 10,000 categories—the difference is purely scale of the system, not fundamental approach.
“back in the '80s people had very successfully done OCR optical character recognition. So they were able to uh take you know the 26 letters of the English alphabet and so on and say okay is that an A is that a B is that a C and so on. That could be done for 26 different possibilities, but it couldn't be done for 10,000 possibilities. And it's really just a matter of the scale of the whole system um that makes that possible today.”
Wolfram Alpha brought out an image identification system in spring 2015 that can identify about 10,000 kinds of things and performs somewhat better than competing systems for interesting technical reasons, using the same conceptual technology as McCulloch-Pitts neural networks.
“in March, April, sometime this spring, we brought out a a little image identification uh system, website, etc...Um and a bunch of companies have done somewhat similar things, I think. Ours for for various somewhat interesting reasons a little better than other people's...and what it does is you show it something, and for about 10,000 kinds of things, it will tell you what it is.”
Three kinds of tasks are significant missing links for human-like AI: visual object recognition, voice-to-text, and language translation—all of which humans learn naturally and which neural networks have recently become successful at performing.
“there are basically a few of these. There's There's visual object recognition, there's voice-to-text, and there's language translation. And those are kind of three kinds of things which humans managed to do with varying degrees of difficulty...these become essentially These are These are some of the missing links to how do we make machines that are kind of human-like in what they do?”
Around 1980, Wolfram was interested in automating question-answering based on accumulated human knowledge, considering both symbolic approaches (breaking questions down into symbolic representations) and brain-like approaches (using fuzzy logic and neural networks), but found building a brain-like system too difficult at the time.
“I myself have been interested in sort of how do you make an AI-like thing since I was a kid, basically, which is depressingly long time ago now. Um but uh uh you know, I was interested in particular in how do you sort of take the knowledge that us humans accumulate or have accumulated in our civilization, how do you automate kind of answering questions on the basis of this knowledge and so on. And I I thought about this first actually around 1980. Um and I sort of thought about how do you do that sort of symbolically...Or and I kind of concluded, well, to really do this well, we have to have sort of a brain-like thing that involves sort of fuzzy questions, fuzzy answers...building a brain is kind of hard. I worked on it a bit. You know, worked on neural networks even at that time. Couldn't really make much interesting progress.”
Wolfram has wanted to do molecular computing for a long time but believes the ambient technology isn't yet ready, though he hopes we're approaching the point where someone could pursue it without spending a decade building enabling technology.
“There's one of these projects of sort of doing molecular scale computing that I've wanted to do for so long, and I just don't quite think that the ambient technology is to the point where one wouldn't have to spend a decade, you know, building ambient technology to get to the point...I'm kind of hoping that that we're almost at the day when it's possible for for, you know, somebody like me who isn't going to build all that ambient technology to actually do something with molecular computing.”