The AI Expert: "Super AI Will Be Unstoppable!" – What’s Coming NEXT | Stephen Wolfram
What this covers
🔔 Subscribe for new episodes: https://www.youtube.com/@seankim?sub_confirmation=1 ▶️ Recommended for you: https://www.youtube.com/watch?v=xGlVx_6_A3o
Stephen Wolfram is a computer scientist, physicist, and entrepreneur best known for founding Wolfram Research and creating Mathematica and the computational knowledge engine Wolfram|Alpha. A child prodigy, he published scientific papers in physics by the age of 15 and earned his Ph.D. from Caltech at 20. He later developed A New Kind of Science, proposing that simple computational rules can explain complex phenomena in nature. Wolfram has been a pioneer in symbolic computation, computational thinking, and AI. His work continues to influence science, education, and technology.
In our conversation we discuss: 00:00 – What was the first version of AI? 23:38 – What triggered the current AI revolution? 34:19 – Did OpenAI base its initial algorithm on Google's work? 46:47 – What is the technological gap between now and achieving AGI? 1:15:59 – Do you fear an AI-driven world you can’t fully understand? 1:35:15 – What do we need to unlearn if AI can replicate human abilities? 1:47:39 – What happens when there aren’t enough jobs due to automation? 1:54:01 – How is AI reshaping people’s views on wealth? 2:25:48 – The future of automating software development
Learn more about Stephen Wolfram Official Website: https://www.stephenwolfram.com/index.php.en Wikipedia: https://en.wikipedia.org/wiki/Stephen_Wolfram
Listen to this episode on the go 🟢 Spotify: https://open.spotify.com/show/4M9IlPFWzCaTv60kvCMlVS 🍎 Apple Podcasts: https://podcasts.apple.com/podcast/id1482999379
📸 Connect on IG: https://instagram.com/heyseankim
Source description (no synthesized summary yet).
Wolfram argues that AI is fundamentally advancing toward non-human forms of intelligence rather than replicating human cognition, and the meaningful challenge ahead is not achieving human-level AI but determining how humans choose to deploy vastly more powerful computational capabilities.
- AI will likely transcend human intelligence asymptotically rather than converge on it; scaling neural networks adds computation in ways that diverge from brain-like operation
- The real limitation is not technical capability but human choice about what to do with automation; as mechanics get automated, the distinctly human role becomes deciding objectives, not executing them
- Fear of AI replacement is misplaced because humans and increasingly-capable AI occupy different niches; history shows automation fragments job categories rather than eliminating purpose
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.
Wolfram's major scientific discovery in the early 1980s was that random, extremely simple programs can do incredibly complicated things, revealing nature's secret for generating the complexity we observe in nature—and this same principle explains how intelligence can arise from simple computational rules.
“one of the things I got interested in the beginning of the 1980s was what does a program that you just pick at random, what does it typically do? You might assume if it's a tiny little program that it would just do tiny little simple things. The huge surprise that took me a while to kind of really get used to is that's just not true. In the computational universe of possible programs, even very simple programs can do incredibly complicated things.”
In the 1950s and 1960s, during the Cold War, there was significant funding and effort devoted to building machine translation systems to automatically translate between Russian and English in diplomatic exchanges, motivated by fear that human interpreters might mislead both sides and trigger world war.
“during the cold war and people were thinking well there are occasionally these sort of high level diplomatic exchanges between you know the US and Soviet Union and so on and they had this idea that well you know in those exchanges there would be some interpreter who'd be you know translating Russian to English etc etc etc and they were like we're really worried the interpreter is going to mislead everybody and uh it's going to lead to World War II or whatever so Let's put a machine in place instead.”
Neural networks are an incredibly old idea, with origins tracing back to the mid-1800s when people were studying how the brain works under microscopes and attempting to understand the nature of nerve cells and their electrical properties.
“Neural networks are famous today, but neural networks are an incredibly old idea. I mean, I've actually been researching a bit the the ancient history of neural networks, and it goes back even much further than I imagined. It goes back to the mid mid 1800s.”
In 1943, Warren McCulloch (a neurophysiologist and psychiatrist) and Walter Pitts (a young mathematician) published a foundational paper laying out a mathematical idealization of neural networks that became the theoretical foundation for all subsequent neural network research.
“the next probably big step was 1943 uh chap called Warren McCulla and chap called Walter Pittz. Um Warren McCulla was a was a neurohysiologist and psychiatrist. Walter Pittz was a a young math very young math person. Um they worked together wrote a paper about kind of the logical theory of neural nets. That paper is kind of the foundation of everything that's been done since.”
When fertilizer was invented and crop breeding improved, predictions of famine (driven by population growth outpacing food production) proved wrong—technological advance solved the scarcity problem, allowing agriculture to provide food for a much larger population.
“when fertilizer was invented you know people thought the world was going to run out of food because there weren't going to be enough crops produced to deal with the population increase. But then things like fertilizer, crop breeding and so on were were were invented and uh you know and that problem went away through something and because food effectively became cheaper and uh food became easier to produce”
Early computers in the late 1940s and 1950s were described as 'giant electronic brains,' leading people to assume that automating thinking would be roughly as straightforward as automating physical tasks like bulldozing.
“when computers well electronic computers were first kind of coming on the scene in the late 1940s and so on, the typical description of them was giant electronic brains. So people had the idea from very early on what computers are going to do is automate thinking kinds of things. And people kind of assumed that that wouldn't be terribly hard.”
Camillo Golgi discovered in the 1870s how to stain nerve cells with a dye that made only a tiny fraction of cells visible, allowing visualization of individual nerve cell patterns under a microscope, while Ramón y Cajal argued that nerve cells were separate units with gaps (synapses) between them, contrary to Golgi's belief that they formed one continuous network.
“There was a chap called Golgi who figured out in I think the 1870s how to stain nerve cells because when you look at you know a slice of brain tissue under a microscope it just looks really complicated. You can't see anything. Golgi figured out this way to stain it so that some tiny fraction of the cells would turn purple and then you could see those cells that have turned purple and you could kind of see the the pattern of those things. There was for for a while there was a big sort of dispute between Golgi and a chap called Romani Kajal uh because Golgi thought that sort of every nerve that you would see these sort of nerve cells and that they were really all connected in one big net. Romani Kajal thought they were all separate cells that had sort of synapses gaps between them.”
In the early 1960s, Marvin Minsky—an AI pioneer who had originally been interested in neural networks and even written his PhD thesis on them—decided that perceptrons and neural networks were fundamentally trivial and incapable of doing anything interesting. With Seymour Papert, he published a book arguing this thesis, effectively killing mainstream interest in neural networks.
“in the early 1960s a person I actually knew pretty well named Marvin Minsky uh who was a kind of AI pioneer who'd originally been interested in neural networks. He wrote his PhD thesis about neural networks. He even built a neural net machine but he uh kind of decided perceptrons and all things neural nets are trivial. and he and a chap called Simo Papot wrote this book about perceptrons which argued that perceptrons and neural nets can't do anything interesting. Game over. So by the late 1960s, everybody said neural nets are doomed.”
Claude Shannon's 1948 paper on information theory laid out the idea that language has statistical regularities (e.g., Q is usually followed by U, E appears more frequently than X) that could be exploited to understand encrypted or scrambled text, forming the foundation for statistical approaches to language generation.
“a chap called Claude Shannon worked out this thing he called information theory. um uh they actually worked out I think during the war probably Alan Turing was somewhat involved in this also uh wrote this this paper in 1948 I think introducing information theory and this idea of sort of the statistics of things like language so given that you had the statistics of language you knew which letters were more common which pairs of letters were more common and so on you could start imagining generating language statistically you say okay if you happen to randomly pick a Q, the next letter is going to be a U, etc., etc., etc.”
Throughout economic history, when labor-intensive occupations get automated (like agriculture in the 1800s), the freed labor doesn't disappear—instead, the economy fragments into many new job categories as new opportunities open up.
“I looked a while ago at what's happened to jobs that humans do. Like in the US for example, there's data back to like 1850 or so of what jobs do people do. Back in 1850, most people were doing agriculture. Most people were actually, you know, plowing fields and things like this. That all got almost all got automated. So then what happened to the economy? Well, what happened is that what was a big chunk of the pie fragmented into a zillion different areas.”
During the Cold War, the U.S. government funded machine translation projects because officials feared human interpreters in diplomatic exchanges between the US and Soviet Union might deliberately mislead, so they sought to replace interpreters with automatic translators.
“one of the problems of the time was the cold war and people were thinking well there are occasionally these sort of high level diplomatic exchanges between you know the US and Soviet Union and so on and they had this idea that well you know in those exchanges there would be some interpreter who'd be you know translating Russian to English etc etc etc and they were like we're really worried the interpreter is going to mislead everybody and uh it's going to lead to World War II or whatever so Let's put a machine in place instead. Let's have an an automatic translator.”
In 2011, a student of Jeff Hinton left a neural network training by accident for a month on GPU hardware, running through millions of images, and the resulting trained network won that year's ImageNet competition, triggering a major resurgence of interest in deep learning and proving that deeper neural networks with multiple layers could solve complex image recognition tasks.
“The thing that relaunched AI was something happened in 2011 actually that a forementioned person Jeff Hinton uh had been continuing to study deep neural nets and something people including myself were just like I don't know I don't think anything interesting is going to happen here but um he had been studying trying to do image recognition trying to tell is this a picture of a cat or a dog or whatever there was a big collection of images ImageNet that existed and there were these annual competitions for uh you know who can recognize images best. Well, a student of of of Jeff Hinton's left a neural net training kind of by mistake for a month trying to, you know, going through millions of images saying with a neural net being told this is a cat, this is a dog, there's a cat, there's a dog, and you know, trying to tweak the neural net. That's how neural net training works. Tweak the neural net to get the answers right uh more and more often.”
Advanced economies feature people doing diverse types of work, including things that seem pointless from subsistence perspective (podcasting, social media, competitive gaming) but create real economic and social value.
“there are these things that seem purposeful and meaningful gradually change and you know the the thing and what what does you know AI or automation in general do it takes the things that humans want to do and it somehow makes them easier to do. And the thing that you might say was that maybe there'll be no need for the humans because everything's going to be easy. But then you still have the question, well, what do you actually want to do?”
Marvin Minsky, despite originally being interested in neural networks and having written his PhD thesis on them, decided that perceptrons and neural networks were fundamentally trivial and incapable of doing anything interesting, co-authoring a book with Seymour Papert that effectively ended research in neural networks until the 1980s.
“in the early 1960s a person I actually knew pretty well named Marvin Minsky uh who was a kind of AI pioneer who'd originally been interested in neural networks. He wrote his PhD thesis about neural networks. He even built a neural net machine but he uh kind of decided perceptrons and all things neural nets are trivial. and he and a chap called Simo Papot wrote this book about perceptrons which argued that perceptrons and neural nets can't do anything interesting.”
John Maynard Keynes predicted in the 1930s that technological productivity would be so high that people would work less than 15 hours per week, yet we still work 40-50 hours weekly, suggesting something other than constraints is determining work hours.
“there was a a quote I think on an essay around John Mayor Kees he said that in the 1930s he said that in a 100 years due to the advancement of technology that the productivity is going to increase so much that humans are only going to need to work more than less than 15 hours per week and you mentioned with AI and the evolvement of AI humans are pretty much the AI is going to be able to do what everything the humans are going to do yet here we are pretty much 100 years later humans are still working 40 to 40 50 hours a week”
Money can have negative value when it causes conflict (inheritance disputes), mental stress, or reduces initiative, showing that increased money does not automatically increase wellbeing.
“there are more than more than we could enumerate of of things where but you know a very uh you know there are typical ones that everybody knows about about you know oh I inherited a ton of money now what am I supposed to do with my life type thing or you know [2:20:47] or there's uh you know there's this pot of money and people who would otherwise be be friends are you know arguing to the death over it”
In the early 1960s, early AI researchers believed the field would lead humanity to 'go down in the history of the earth as being the thing that created the next species that was AI,' and that species would be more significant than humans.
“it's actually very interesting to read what people wrote in the early 60s about kind of AI and what was going to happen with AI and the idea that you know kind of we were a species that was mostly going to go down in the history of the earth as being the thing that created the next species that was AI and so on.”
Wolfram personally began building a symbolic manipulation system in 1979 called SMP (Symbolic Manipulation Program) that could do symbolic math and other symbolic tasks, but he chose to use C as the implementation language rather than LISP, which created a point of disagreement with John McCarthy.
“in 1979 I kind of got into building my own kind of symbolic computation system, and it wasn't practical to use lisp at that time. It was just there weren't good implementations and so on. So, I used this then new fangled language called C”
For many decades, people attempted to build question-answering systems using various AI techniques (statistical, symbolic, etc.) and essentially none of them worked until Wolfram Alpha, which used a different architectural approach that combined natural language understanding with precise computational representation.
“People had tried to make question answering systems with various kinds of AI techniques, statistical, symbolic, whatever. They tried to do that for decades. It had never worked. It was kind of particularly notable.”
Transformer networks are neural network architectures designed for sequential data (like language) where neurons are connected sequentially but can form long-range connections, allowing information from early in a sequence to relate to information much later in the sequence.
“When it comes to things like language, language has the feature that it's sequential. You know, we it's just this stream of, you know, letters or words or sounds or whatever coming out. And there were the these things called transformer nets which are kind of are networks that are sort of that only have connections sequentially but their connections can be quite long range just as you know a word somewhere in a sentence can kind of refer back to a word that's much earlier in the sentence.”
The computational universe of possible programs is infinite, so there will always be new mathematical theorems to prove, new inventions to make, and new things to explore—scarcity of novelty itself is built into the structure of possibility.
“you know the computational universe of possibilities is is infinite. You might have thought that you know there'll come a time when we've made every invention that can be made. That time will never come. It's you know a century ago people were like yeah we've almost made every invention that could be made. well turned out was not true.”
ChatGPT's breakthrough was that it reached a threshold of linguistic quality where generated text reads like meaningful human writing, making it useful to consumers, whereas previous language models produced garbage that was obviously machine-generated.
“I think the same kind of thing happened with neural nets. It kind of it got good enough that it was like you could read the text and it read like, you know, meaningful text that a human might write. And that was uh sort of the chat GBT moment”
When various economic advances have made basic necessities cheap (food, shelter, clothing), people don't stop wanting things—they shift to wanting other things: entertainment, social experiences, intellectual pursuits, and so on.
“the first point is that you know, a lot of what people pay for in the world today isn't basic necess but not in all countries but but in in you know in there are many segments of the world in which what people are mostly paying for isn't basic necessities. So there is a thing that people care about paying for that is something much more ethereal than just getting enough food to eat so to speak.”
In the 1980s, Japan initiated a Fifth Generation Computer Project aimed at solving AI through intensive research using primarily the Prolog language, which had certain problem-solving approaches that 'doesn't work out so well in the end,' representing another wave of AI hype and funding.
“Japan was viewed as a country where kind of oh it just copies American technology. Everybody was kind of very down on that. Japanese government had this this great idea. They said let's do a research project that's going to jump ahead of everybody else. It was called the Japanese fifth generation computer project and it was a project in I forget when it started early 80s sometime. It was a project where Japan was going to solve AI and um their methods well they were using particularly a language called prologue which is sort of a lispish kind of thing but it had a particular idea about problem solving which doesn't work out so well in the end.”
By the end of the 1980s and early 1990s, despite the neural network renaissance and other AI efforts, the field remained essentially at a low ebb—people stopped identifying themselves as AI researchers, and AI was 'really at a very low ebb for quite a long time.'
“By I would say end of the 80s early 90s this stuff basically hadn't worked out. Um, and people said, 'Ah, AI is is doomed.' And everybody was kind of, you know, everybody who might have been saying they were doing AI didn't say they were doing AI anymore. And and AI was really at a at a very low EB for quite a long time.”
Education should shift away from teaching mechanics (computation, routine procedures) toward teaching broad thinking skills: philosophy and computational thinking are two valuable forms of thinking that will become more important as mechanics get automated.
“it should be more about, you know, learning the facts that you need to know to be able to think broadly about things and then learning how to think broadly about things... the extremes of kind of philosophy as a way of thinking about things and computational thinking as this kind of formalized way of thinking about things. These are two good sort of forms of thinking worth learning so to speak.”
Large-scale systems based on irreducible computation require accepting that we can't have 'narrative explanations' of how they work—we can't tell a comprehensible story about internal mechanisms. But we can still predict behavior, measure outcomes, and use the system for purposes we care about.
“to make systems that can really make use of computation as it can best be made use of, we sort of inevitably have to deal with this kind of irreducible computation that we can't readily understand with our minds. We can't have a kind of narrative explanation of what's happening inside.”
Wolfram is skeptical that Universal Basic Income alone would solve unemployment/purpose if AI takes over work. UBI experiments haven't shown the expected outcomes; people with basic needs met still seek scarcity and meaningful activity rather than defaulting to satisfaction.
“if you ask the question do humans need to work you know... the experiments people have done... the situations people have where it's just like you know now you reset the base level and everybody has this still people are going to seek the scarcity and so on.”
John McCarthy coined the term 'artificial intelligence' in 1956 at the Dartmouth conference and invented LISP, a sophisticated computer language that he believed was the only language suitable for implementing AI, particularly symbolic mathematical computation.
“in 1956 there was this conference at Dartmouth where John McCarthy coined the term AI and uh one of the things John McCarthy did in the in the years right after that was invented this language called lisp”
By the end of the 1980s and into the early 1990s, AI was widely considered dead as a field, and anyone claiming to work in AI would no longer describe their work as 'AI' for fear of association with a discredited paradigm.
“by I would say end of the 80s early 90s this stuff basically hadn't worked out. Um, and people said, "Ah, AI is is doomed." And everybody was kind of, you know, everybody who might have been saying they were doing AI didn't say they were doing AI anymore. And and AI was really at a at a very low EB for quite a long time.”
Throughout history, automation has caused job categories to fragment rather than disappear; as agriculture was automated, new categories of work emerged, and as more gets automated now, people are increasingly doing multiple distinct things rather than one 'job'.
“I looked a while ago at what's happened to jobs that humans do. Like in the US for example, there's data back to like 1850 or so of what jobs do people do. Back in 1850, most people were doing agriculture... That all got almost all got automated. So then what happened to the economy? Well, what happened is that what was a big chunk of the pie fragmented into a zillion different areas.”
In 2011, Jeff Hinton's group achieved a breakthrough in image recognition by training a deep neural network on a large dataset of images (ImageNet), winning that year's competition by a large margin, which sparked renewed interest in neural networks and led to the deep learning revolution.
“The thing that relaunched AI was something happened in 2011 actually that a forementioned person Jeff Hinton uh had been continuing to study deep neural nets... he had been studying trying to do image recognition... Well, a student of of of Jeff Hinton's left a neural net training kind of by mistake for a month trying to, you know, going through millions of images saying with a neural net being told this is a cat, this is a dog... it was like, okay, it was kind of a mistake... it won the imageet competition that year... that was a big wakeup call that you know neural nets are back”
In a military trial of early perceptrons, the system performed well at distinguishing pictures containing tanks from those without, but this success was later revealed to be spurious—all tank pictures had been taken during the day while non-tank pictures were taken at night, so the perceptron was merely detecting lighting conditions rather than recognizing tanks.
“a famous glitch was there was some uh trial for the military of these of these things where there were a bunch of pictures of tanks which had tanks in them and other pictures that didn't have tanks in them. And uh you know the the perceptron did really well figuring out which pictures had tanks in in in them. But then somebody realized that all the pictures with tanks were taken during the day and the pictures without tanks were taken at night. So really all the perceptron was doing was something very trivial.”
Keyes predicted in the 1930s that productivity advances would eventually enable humans to work only 15 hours per week, yet 100 years later humans still work 40-50 hours. This suggests either that something blocks the productivity gains from benefiting humans, or that humans find new things to do with the freed time.
“John Mayor Kees he said that in the 1930s he said that in a 100 years due to the advancement of technology that the productivity is going to increase so much that humans are only going to need to work more than less than 15 hours per week and you mentioned with AI and the evolvement of AI humans are pretty much the AI is going to be able to do what everything the humans are going to do yet here we are pretty much 100 years later humans are still working 40 to 40 50 hours a week”
There were two fundamentally different approaches to AI from the 1960s onward: the symbolic approach, which aimed to create computational representations of the world and rules for how it works; and the statistical approach, which used neural networks to extrapolate patterns from observed data without explicit rules.
“there were really two approaches that well first thing in the 1960s there were two different approaches that were taken to AI and those there continue to be two different approaches today. there was kind of the the symbolic approach and the statistical approach. The um the the idea the symbolic approach was kind of the thought that you can have sort of a computational representation of the world and you can have your AI kind of figure out things in the world in a kind of computational fashion. The other approach was the statistical approach just saying forget about having rules for how the world works. Just say we notice this and that and the other thing. Let's extrapolate from what we notice. Let's just do the statistics of the world to guess what how things work. The main approach that got used there was neural networks.”
People perceive ChatGPT as appearing 'out of nowhere' around 2021-2023 because it became directly usable by consumers, but this success had a long history going back to work in the 1950s-1960s on statistical language properties and cryptanalysis from World War II.
“most people that are playing around with what the consumer version of what they think AI is from Chaja BT or Gemini, people would think that there's been like people think this is pretty much AI. Like this is the last two three years people just think AI just kind of popped out of nowhere, right? Like from 2021 before that AI didn't even exist.”
LLMs (Large Language Models) are good at linguistic interfaces and language-based tasks but not at pure computation, by design. They can write Wolfram Language code because they've learned patterns from text, but they're not intended for and can't excel at mathematical or logical computation that requires precise execution.
“the thing that sort of has emerged is that LLMs are really good at doing this kind of linguistic interface at dealing with things language sort of for languages sake. They're not good at doing things which require computation. That's not what they're intended for. That's not what their structure is. It's actually we have a big clue that they're not going to be good at that because humans aren't good at that either. you know, we don't get to be able to do, you know, run programs in our minds and things like this.”
AI systems (particularly LLMs) are likely to excel at medical diagnosis and similar tasks involving pattern matching between symptoms and known conditions, potentially reaching superhuman performance because humans are limited in what they can thematically search and match across medical knowledge.
“One that that I've been curious about is medical diagnosis. uh actually diagnosis of almost anything. Diagnosis of you know problems with your computer things where there's a body of knowledge about things that can happen and you have certain symptoms and you're trying to match those symptoms to what's known and I kind of have this suspicion that you know the current round of AIs is going to do quite well at that possibly in a quite superhuman way. Um because you know we humans I mean it's it's it's this thing about sort of thematic searching of what's out there in the world”
Wolfram's major scientific discovery is that extremely simple programs can produce incredibly complicated behavior, suggesting that nature uses this principle to generate complexity; this is a secret nature uses and also represents computation in a deeply non-human way.
“one of the things I've done in science for a long time one of sort of my big discoveries and directions in science is to understand what what computation in the wild looks like... one of sort of my big discoveries and directions in science is to understand what what computation in the wild looks like... a question that I got interested in the beginning of the 1980s was what does a program that you just pick at random, what does it typically do? You might assume if it's a tiny little program that it would just do tiny little simple things. The huge surprise that took me a while to kind of really get used to is that's just not true. In the computational universe of possible programs, even very simple programs can do incredibly complicated things.”
Image identification systems developed by 2012 using neural networks were not significantly worse than systems existing today (2024), suggesting that while image recognition reached a 'solved' threshold around 90-95% accuracy, the field has not made dramatic improvements in that specific domain over the last 13 years.
“in the last 13 years or so, things haven't in that particular domain haven't improved that much. It's it's like you reach a threshold, you start to be able to do something, then that works and then you build that capability into a bunch of systems and it becomes useful, but it's not as if that capability itself, you know, just because it made that one jump doesn't mean it's going to make lots of other jumps.”
Nobody knows what AGI really means; it's a buzzword. The term comes from 1930s IQ research, which Wolfram considers a 'doomed concept.' Defining AGI as 'human-level intelligence' is problematic because humans are unique in having precisely the characteristics humans have—you can approximate human capabilities in some ways, but there's no objective measure of whether something is 'human-like in all respects.'
“Nobody knows what AGI really means. It's a it's a buzzword that um it's kind of like uh you know, first AI was a buzz word and people didn't quite know what that meant except they thought it meant things that do stuff that's kind of like what people do. And it's like well now we've got things that do lots of stuff that people do whether it's you know being able to generate language, drive cars, whatever else. Um but still there's got to be something else that people do and and that must be AGI, some general intelligence thing. It's kind of ironic that um actually I just realized this as I'm talking to you that the the term general intelligence was I think coined in the 1930s and um the uh it came in when uh at the same time as the concept of IQ came in. a very in my opinion a very troublesome concept”
The scaling law for minds: as neural networks become larger (with more neurons/parameters), they don't become superhuman-human intelligence; they become non-human intelligence. A system with trillions of neurons would not be a human running faster; it would compute in fundamentally different ways based on different sensory inputs and goals.
“What is the next step from from cats and dogs to humans to the next level of minds? What does that look like? And that's kind of a uh you know that's kind of a thing that you might wonder sort of you know we've got something that is humanlike at some level right now. we can go on checking the boxes of giving it, you know, letting it be, you know, my my guess is within, you know, the next big thing that will get solved in AI is robotics. Um, and you know, it's been super difficult to get, you know, robot hands to to pick things up”
Science's mission, in one sense, is to make a bridge between what exists in nature (incomprehensibly complex) and what fits in human minds (comprehensible), allowing us to talk about laws like fluid mechanics without understanding every molecule's behavior.
“I mean, we can think of science, in fact, the mission of science in a sense is to take what exists in the natural world and kind of make a bridge between that and what we can understand with our finite minds. I mean it's like saying oh you know we're not going to in our minds we're not going to understand what every molecule in the in the you know in the river does for example but we can say we roughly can talk about certain laws of fluid mechanics that govern roughly what the what the river does. We're making a bridge between what's actually happening in nature and what fits in our finite minds.”
Cryptocurrency has value despite not being directly exchangeable for food because it represents a network of dependencies and relationships that people care about—this illustrates that value isn't only material but relational/network-based.
“people might say, well, cryptocurrency isn't worth anything really because I can't go, you know, buy food with it... I think that argument is not correct. And that relates to this. I mean, in a sense, it's worth something because there's this whole network of things that depend on on it”
OpenAI collected a huge amount of training data and used transformer networks with billions of neural net weights ('large language models') to train ChatGPT, which learned to predict and generate sequences of language based on patterns in the training data.
“guys at OpenAI uh basically collected this huge and and well they collected this huge amount of training data. There was one kind of technical idea which I'm not sure how significant it will be in the long view of history. This idea of transformer nets... Well the guys at OpenAI uh basically collected this huge and and well they collected this huge amount of training data.”
There's ambiguity in whether 'objectives' are intrinsic to systems or imposed from outside through human interpretation. When humans describe their own behavior, they do so in terms of objectives, but this may be a learned descriptive framework (like language) rather than an intrinsic feature of consciousness.
“It's an interesting question to what extent the description of what we do in terms of objectives is something that is innate and natural or whether that's something that we learn just as we learn language to for us to describe what we do as I'm doing that because blah blah blah you know it might be that you know when we're all you know babies or whatever we just do what we do and it is a a higher layer that is our description of objectives so to speak just as it's a higher layer to be able to to describe things in terms of language.”
Software development through manual labor (writing code to spec) is 'bizarre' as a profession given that the core intellectual work (deciding what the software should do) could be separated from the mechanical work (implementing the spec), and technology now enables rapid spec-to-code conversion.
“the fact that there is this whole ecosystem of people doing manual labor software development is just bizarre. I mean it's happened because of a bunch of the economics of labor, a bunch of the ways that technology has developed but the fact is you know you say it's you know you can create an app in you know that's what I do many times a day writing tiny amounts of code you know it's that's what you know that's been my objective is to automate those things”
The OpenAI team likely didn't know in advance that ChatGPT would work as well as it did, and if they had, they would have constrained it more during development. Many of their approaches were exploratory rather than production-focused.
“I remember chatting with the folks who worked on it just after it came out and sort of my obvious question was did you know it was going to work they were like no Um, and in fact, probably had they known it was going to work as well as it did, they would have tried to constrain it in a lot more ways than they did. And a lot of things they did were kind of sort of what you do if you're just trying to do the experiment, not what you do if you're trying to build a production system”
The question of whether AI will make human skills obsolete is misconceived: as capabilities get automated, what becomes important is the human ability to choose what to do, what to think about, and what problems to solve—something that can't be automated.
“you know the mechanics of what's done getting automated just means means there are more choices and maybe that's manifest in both more kinds of jobs and more jobs done by individual people and so on. That's my guess. I mean I I'm you know the idea as I say there could be a choice to just sit around as a couch potato and maybe there'll be you know a segment of society that does that”
People might rationally choose to stop using new technology and live in accordance with older modes of life. While possible, Wolfram doesn't think this is how human society will actually play out, given historical patterns of engagement with new possibilities.
“you know it might be that uh that humans just decide you know if we if we look at human society people have different different beliefs about what we should be doing I mean it's it's like you could say I don't believe in anything that was invented in the last hundred years I'm not going to live my life in such a way that I'm using things that were invented in the last 100 years, last thousand years, whatever. Um, you know, you can make a decision that says I'm just going to lock out the things that are now possible.”
Some people throughout history have lived in conditions where basic necessities were provided (through abundance of natural resources or lack of need), and the anthropological observation is that such people often engage in 'ritualistic' behaviors that seem from the outside to serve no purpose, but are internally experienced as deeply meaningful.
“people say there have been periods of of thousands of years where there have been you know groups where well here's the funny thing where people say you know their basic necessities were taken care of they lived in a place where they could you know just pick the pick berries off off bushes to eat and and so on and uh and then then you ask well what did the people do in that situation and the a common anthropological statement I haven't really dug dug deep in this. So I don't know I don't know how much I believe this but the statement that's made is well people do these go into these very ritualistic kinds of behaviors”
Interestingly, consumer electronics (computers, smartphones, displays) were once expensive and restricted to wealthy individuals, but became so cheap that access became essentially flat—everyone now has access to roughly the same quality of consumer electronics regardless of wealth.
“back I don't know 40 years ago or something 45 years ago I used to have fancy computers I used to have access to computers that were much fancier than the typical person's computers were but in fact consumer electronics became cheap enough that everybody has the same kinds of computers now. It's not it's very flat”
Before ChatGPT, sequence prediction (predicting the next word given prior words) was 'really cruddy' and 'really did not work well at all,' making practical applications like code autocomplete and text prediction fail. This was a major technical challenge waiting for a solution.
“the question of could you do what was called sequence prediction could you do a better job of knowing given that you have a piece of text that starts this way what will come next... was really cruddy... You know, we tried to use it a bunch for doing oh things like predicting uh pieces of code for autocompletes, things like this really did not work well.”
The barrier to entry for creating software has fallen dramatically: using tools like Replit, Windsurf, and other systems, people can create apps and websites in minutes that would have taken years. This is similar to how podcasting became possible through technology.
“in the world of software where we can put up a prompt and you can create software so like there's replet there's windsurf there's programs that are out there now where you can create an app create a website within minutes which would normally have taken weeks or months or years even”
Raw LLMs alone don't do well as tutors without substantial superstructure, though they can answer questions based on uploaded class materials. Building a full tutoring system that keeps students on track requires extensive machinery beyond the LLM.
“The raw LLM does not do very well at this. the um uh the thing that um you know people might say the raw LLM is going to be able to to be a tutor. It is true that if you like upload the class assignment that you had or the the notes from your class and you tell the LLM, ask me questions based on these notes, it'll do a reasonable job at that. But if you say lead me through this whole course...that's a thing that at least in our observation seems to need a whole lot of superructure”
Large language models (LLMs) have billions of neural network weights and are fundamentally designed for linguistic interfaces and language generation, not for computation, making them unsuitable for mathematical or logical operations that require step-by-step reasoning.
“the thing that sort of has emerged is that LLMs are really good at doing this kind of linguistic interface at dealing with things language sort of for languages sake. They're not good at doing things which require computation. That's not what they're intended for. That's not what their structure is.”
One major frontier in AI integration is building an AI tutor for algebra that actually works, which has been a decades-long failed goal of AI education, though recent work looks promising, suggesting this is a major unsolved problem at the frontier of practical AI.
“We have a big project right now to build an AI tutor... people have been trying to do sort of computerized education for 70 years and it's basically always failed... It looks promising this time around.”
Medical diagnosis and other forms of symptom-matching against large knowledge bases are areas where AI will likely perform at superhuman levels because they require thematic searching and pattern matching across large datasets, which is not what human medical training emphasizes.
“I mean, one that that I've been curious about is medical diagnosis... possibly in a quite superhuman way. Um because you know we humans I mean it's it's this thing about sort of thematic searching of what's out there in the world”
The perceptron tank-recognition problem demonstrated a fundamental challenge in AI: distinguishing between genuine learning and statistical artifacts in training data, where perceptrons achieved high accuracy simply because daylight-photographed tanks were correlated with tank presence, not because they understood tanks.
“there were a bunch of pictures of tanks which had tanks in them and other pictures that didn't have tanks in them... the perceptron did really well figuring out which pictures had tanks in in in them. But then somebody realized that all the pictures with tanks were taken during the day and the pictures without tanks were taken at night. So really all the perceptron was doing was something very trivial.”
The concern about AI replacing humans stems from fear that AI will outcompete humans for resources and survival, but this treats AI as a single unified agent competing for dominance rather than multiple systems with different goals in a complex ecosystem.
“people tend to project onto AIs the idea that the AI is going to want to do this or that thing. But you know just as they project onto other people that other people are going to want to do this or that thing. The only thing we ever know for sure ourselves is how we're feeling internally ourselves. You know, everything else is kind of an assumption, a projection. And the concept that, you know, the AI that has all of this sort of computational ability is going to want to do something is is a very weird concept.”
The trolley problem (deciding between killing five people or one person) is a thought experiment that fails to capture the nature of ethical reasoning because ethics requires considering the full context of connections between people, relationships, beliefs, and so on—you cannot isolate an ethical decision from the rest of the world.
“the thing that's that's the thing that the cheat in that in that problem is well in science one of the things that makes science possible is that we can do kind of controlled experiments. We can say we're going to do an experiment on this little tiny piece of the world, ignoring everything else that's happening in the world. We don't have to. But in ethics, I don't think that's possible. In other words, there is no answer to the question of the llamas or the endangered lizard without knowing the whole story of sort of the connections of the llamas, whether that one of those llamas was somebody's pet llama, whether there was a, you know, a group that worships llamas, and whether the endangered lizard was a, you know, all kinds of things.”
Wolfram has seen first-hand that consumer electronics became so cheap and ubiquitous that everyone has essentially equivalent hardware, contradicting earlier expectations that a consumer technology hierarchy would persist (some having fancy computers, others not).
“back I don't know 40 years ago or something 45 years ago I used to have fancy computers I used to have access to computers that were much fancier than the typical person's computers were but in fact consumer electronics became cheap enough that everybody has the same kinds of computers now. It's not it's very flat”
Robotics is likely the next major frontier for AI, as robots need to handle complex physical manipulation tasks (like picking items from shelves or packing boxes) that are easy for humans but have been difficult to teach to machines due to limited training data and the complexity of physical interaction.
“my my guess is within, you know, the next big thing that will get solved in AI is robotics. Um, and you know, it's been super difficult to get, you know, robot hands to to pick things up and, you know, be able to pack boxes with whatever stuff you want from a warehouse, things like this. Um, it's it's something we humans manage to do fairly easily. It's something that it's been a little difficult to get training data, you know, for the stuff that humans write while there's, you know, a trillion words on the web and things like this.”
The ethical question of whether we have a responsibility to spread life through the universe (or a responsibility to keep it 'interesting') is meaningless to Wolfram. Ethics is fundamentally a human thing—there is no abstract ethics independent of human values.
“I simply don't get that at all. I mean, ethics is a human thing. There is no abstract ethics. You know, people get confused because people try to take ethical questions and couch them as scientific questions.”
Wolfram explored 'interconcept space' by modifying the internal representations of an image generation AI trained on human concepts like 'cat in a party hat,' finding regions the AI treats as meaningful but humans don't recognize, analogous to alien minds exploring non-human concepts.
“if you take sort of the if you modify the innodes of the AI or alternatively if you take kind of the internal description that it has of cat in a party hat and you start tweaking that. Well, you move away from the concept of cat in a party hat into what I was calling interconcept space. You move away from these human-defined concepts to concepts which are which exist in the mind of the AI but are not familiar to us humans.”
The 'front line' of economic activity typically ends up being things people have to make choices about—tasks where there's no abstract way to input the goal. You can't algorithmically determine what someone should want to do.
“the front line usually ends up being things people have to make choices about. That's places where you can't feed in. There's no abstract way to feed that in. There's no abstract answer to the question of what should you do? What should the universe do? The universe does what the universe does. It doesn't, you know, that there's no there's this choice of what to do is something that is sort of the ultimately quintessentially human thing”
Wolfram had been interested in answering questions about the world from accumulated knowledge since before 1980, and concluded from science that there is no bright-line difference between intelligence and mere computation, leading him to believe such a system was buildable.
“I well there's a branch that I was involved in. There's a branch I wasn't involved in. The branch that I was involved in was this thing that I've been interested in doing since before 1980 or so of can one answer questions about the world from the knowledge that we've accumulated in the world. And I had made some sort of science advances that made me convinced that there wasn't sort of a bright line difference between intelligence and mere [24:59] computation.”
The computational universe contains far more possible software and useful patterns than any company or person could ever enumerate; the question is which parts humans care about, and new technologies (like XR/VR) open up new domains of software demand.
“the space of possible software it's like these little programs that I study in the computational universe everyone is a piece of software in a sense everyone does something most of them are not things that anybody would care about some of them make really pretty pictures... But the leading piece to that has to be, well, what do you want it to do? You know, what is the thing you want? So, if you can say, well, you know, is there going to be a broader set of things that people want software to do? Maybe as people's activities broaden out, there will be a need.”
Humans are also not good at computation (running programs in our heads), suggesting that LLMs' computational weakness mirrors human limitations and reflects neurological constraints rather than a design flaw.
“It's actually we have a big clue that they're not going to be good at that because humans aren't good at that either. you know, we don't get to be able to do, you know, run programs in our minds and things like this.”
The fact that we can observe something computationally doesn't mean we can understand it with our finite minds; science works by bridging what happens in nature and what our minds can comprehend, like understanding fluid mechanics principles rather than tracking every molecule.
“We can think of science, in fact, the mission of science in a sense is to take what exists in the natural world and kind of make a bridge between that and what we can understand with our finite minds... we're not going to in our minds we're not going to understand what every molecule in the in the you know in the river does for example but we can say we roughly can talk about certain laws of fluid mechanics that govern roughly what the what the river does.”
Understanding how AI systems work inside (their 'lumps of irreducible computation') is not necessary for using them, just as understanding how horses work internally was not necessary for riding them; this is a return to how technology worked before the industrial revolution made it simple enough to understand.
“when people were you know getting transported around by riding horses you could know something about how to get the horse to do what you want but knowing how the horse works inside was not something you really cared about. You were able to use the horse for something that was useful to you but you didn't know mechanistically how the horse worked inside... post-industrial revolution, for a brief time, we've been operating machines that are simple enough that we know what's going on inside. You know, we're now back to a situation where to make systems that can really make use of computation as it can best be made use of, we sort of inevitably have to deal with this kind of irreducible computation that we can't readily understand”
Automation throughout history has taken mechanical tasks and delegated them to machines, but the frontier of decision-making remains fundamentally human because choices about what to do require values that cannot be abstracted or automated.
“what becomes the human act is what do I want it to do then you know as I say in that particular case I've spent the last four decades trying to do that automation AIS add another level of automation to that um but it doesn't get you around the you know you kind of uh you know it's it's like there's still a there's still some somewhere you have to decide what the app is supposed to do. And I think that's the um uh and yes, that's a thing where where you're um uh you know, again, that's that's what I've been living for many decades now is getting to the point where it's mostly about imagining what the app is supposed to do, not about the mechanics of actually writing it.”
When AI is applied to physical robotics, it will be non-human intelligence in a different form (embodied), not simply 'faster humans,' and this distinction parallels the broader distinction between making AI more human-like vs. making it more powerful non-human computation.
“Is that AGI or is that just a AI in a different form than a physical form through hardware?... what is the limit of AI? What I'm what I'm saying is the limit of AI if you define intelligence to be the thing that humans have is pretty much humans.”
The idea that 'if AIs get to be really smart they'll figure everything out' is false because discovering things about the physical world requires actually running experiments, not just reasoning, creating a bottleneck that slows down even superintelligent systems.
“the physical world is partly because it is doing all this computation. You can't figure out in advance things about what's going to happen in the physical world. You actually have to try experiments and so on to see what's going to happen.”
Eliezer Yudkowsky's AI risk argument rests on three claims: (1) AIs will be able to optimize almost anything up to physical constraints; (2) AIs will have a wide range of possible objectives; (3) Most possible objectives don't leave room for humans. Wolfram questions the third claim as logically incoherent because it's unclear what it even means for a non-human system to 'have' an objective.
“His theory is uh AIS will be able to optimize the doing of almost anything. I agree with that. The up to the constraints of the physical world. You know, in other words, if there is a thing you can define that you want to do to be able to do it, one will be able to optimize the path to be able to do it. I mean that's as I say that's what I've lived for the last you know four decades in software that I've built is you know can I can I automate that can I optimize that so I I take that as as a reasonable thing second statement is AIS will sort of have objectives that are kind of a wide range of different objectives if you can define them as having objectives ives. I think that's kind of true, but I'm not sure what it means to say that they have objectives. Well, that so then the next claim would be most of those possible objectives don't leave room for humans. That's a much more bizarre claim, I think, because it's like saying that, you know, it's very it's it's difficult to define this notion of an objective for something that doesn't have the kind of thinking that that isn't humanlike”
The real question about superintelligent AI isn't whether it will be like humans but how we can 'lasso' non-human computation into things we care about—take vastly complex computation and connect it to human values and purposes.
“the real question is how do we how do we take that nonhuman computation and kind of lasso it into something we care about”
There is an irony in AI safety discourse: people fear AI will replace humans, yet simultaneously pursue the goal of making AI human-like, which would directly lead to humans becoming replaceable.
“It's kind of ironic if one is saying what's our technological objective? Oh, it's to make this thing that's humanlike. That's a that's a you know that's a goalpost you can see type thing. But that's the wrong goalpost you're saying. If the whole idea of us [1:00:57] fe you know fear brewing that AI is going to replace us shouldn't we try to make sure that humans exist as our individual selves and make AI this completely different thing that can help us rather than trying to create human level intelligence where we will actually just be we are negligible relative to nature”
ChatGPT's ability to produce logical arguments comes from discovering statistical patterns in logic—the same way Aristotle discovered logic by analyzing sentence patterns—not from implementing logical inference rules.
“there's a kind of semantic grammar, a grammar that's based on meaning that says what noun, what verb, what noun can go [39:39] together. And there's there's, I think, a lot of regularity in that structure. There's a kind of a a semantic grammar of language. That is effectively what CHGBT discovered statistically by looking at the web. And people were very surprised that, for example, it it could discover logic, that it could make arguments that were made logical sense. I think the way it did that is the same way that you know Aristotle discovered logic back a couple of thousand years ago which was you just look at a bunch of sentences that people say and you ask [40:11] what is the kind of structural pattern that they have led to you know syllogisms in Aristotle's time”
When people lose basic material scarcity (food, shelter), they pursue other forms of scarcity and meaning; anthropological evidence suggests that even when basic needs are met, people engage in ritualistic and meaningful activities.
“there have been periods of of thousands of years where there have been you know groups where well here's the funny thing where people say you know their basic necessities were taken care of they lived in a place where they could you know just pick the pick berries off off bushes to eat and and so on and uh and [1:50:59] then then you ask well what did the people do in that situation and the a common anthropological statement I haven't really dug dug deep in this. So I don't know I don't know how much I believe this but the statement that's made is well people do these go into these very ritualistic kinds of behaviors”
Unlike abstract 'objective space' analyses of AI, the actual world has competing AIs, physical constraints, and economic incentives, making single-AI-takeover scenarios less plausible than often discussed.
“It's not like there's one AI in the world. There could have been could have been that that you know, there was in science fiction and in even people's early conception of computers, it could have been we just build this one giant computer and that's all there is. Um, that's a little bit of a different situation. It would be like saying there's, you know, if there was [1:21:47] one uh, you know, organism and there was no competition between organisms and so on, it would be a different situation.”
Distributing universal basic income from automation would not solve the human purpose problem because people have shown they still seek scarcity and challenges even when material needs are met.
“I [2:19:44] doubt it. I mean I think that people you know the uh the experiments people have done the the the situations people have where it's just like you know now you reset the base level and everybody has this still people are going to seek the scarcity and so on.”
The barrier to entry for podcasting dropped dramatically due to technology, making it possible for individuals to broadcast globally without needing broadcast infrastructure, exemplifying how automation enables new forms of work.
“take podcasting for example if If you wanted to be kind of a person who would broadcast things to the world, that was, you know, you had to build this whole stack of things and you had to go work for a radio station or whatever else it is. But because of technology, the cost of getting into the business of podcasting went way down.”
Wolfram Alpha's key advantage was that it had a pre-existing underlying computational language (Mathematica, released in 1988, and what is now Wolfram Language) for representing things in the world computationally. Natural language understanding wasn't abstractly getting a computer to understand English—it was translating English into this precise computational language and then computing with it.
“I only realized this sort of after we'd built what we built. We had this huge advantage because we already had this underlying computational language that we built starting came out in 1988 Mathematica and what's now wolf from language which is this kind of way of representing things in the world computationally. We already had that kind of precise computational representation of the world. And so our natural language understanding wasn't just abstractly get the computer to understand this. It was translate what those pesky humans say into this precise computational language that we can then do computations with”
If humans abdicate choices to AI systems, that would be a 'bad situation' because it means society would just replay captured patterns forever without new human creation or direction.
“Now you can be in a situation where people say, "Well, let's just abdicate those choices to the AIs." That's kind of a bad situation because then in a sense you're saying let's just take society and civilization as it has been as captured by the AIS and let's just run the same thing over and over and over again. It's kind of the humans never get to sort of do anything that is that is uh kind of you know new and different”
Modern activities like social media engagement or podcasting that seem pointless from an outside perspective are internally deeply meaningful to participants, even if their external utility is unclear. Meaning is subjective and context-dependent rather than objectively determined.
“I'm amused to realize the extent to which so many of the things we do today could be seen as ritualistic. I mean from the point of view of you don't know why you're doing it you know I'm sitting in front of a computer getting weird pictures coming up on the screen this seems like a devotional ritualistic kind of activity if you don't you know if you don't have a thread of understanding what the point is so to speak”
CTOs and CEOs often use Wolfram's computational language to rapidly prototype ideas, but field software engineers are locked into spec-following mode where they expect to spend weeks implementing specifications rather than thinking about what those specifications should be.
“the people who run them, the CTOs, the CEOs and so on actually use our tech and build prototypes of things in a very short time and then they go on and they figure out the next thing they want to do and build that as well. But the people who are now sort of in the trenches doing software engineering are like, "Well, we just got a spec. We build to that spec." Okay, now if we could do that much more [2:28:06] quickly, which they can with our tech, then they're like, "Well, now I've got something very difficult to do. I've got to make a new spec. That's not what my job is. My job is to grind out code."”
Wolfram is exploring using AI to thematically analyze millions of physics papers to find experimental consequences of his physics theory, a use case impossible without AI's ability to extract meaning from large text corpora, representing a genuinely new capability.
“I'm interested... there are millions of physics papers out there in the world. I haven't read all of them. I couldn't read all of them. So, a question is, can I use AI to essentially thematically analyze all of those papers”
Walking on a treadmill for exercise would seem completely pointless to someone from a thousand years ago who struggled to obtain food and viewed life through a religious lens, illustrating how the values and meaningful activities of a society change with economic and technological conditions.
“one of my favorite examples of that is, you know, walking on a treadmill. You know, explain to somebody from a thousand years ago why you walk on a treadmill. Well, it's, you know, it's to improve my health so I'll live longer. So, this, so that it's like, why does that make any sense? you know, we're we do what we do for the greater glory of God or whatever. And we are, you know, this is our sort of brief time on earth type thing. And why are we why are you there many, you know, you you could you could look at different different things that people might have thought were significant a thousand years ago. None of which would explain why you would walk on a treadmill”
The natural world is full of computation happening much faster than brain computation, like in turbulence and fluid dynamics in a babbling brook, which is computation but not humanlike computation—it's another model for what 'computation in the wild' looks like when not constrained to be human-like.
“The fundamental thing is it's not very humanlike. And we have a really good model for what it's like, which is the natural world. The natural world is absolutely full of things that compute much faster than brains compute. that you know you look at any kind of you know babbling brook or something like this all that fluid turbulence and so on that's happening in the water that you can think of as computation just like the the electrical signals in our brains uh we can think of as computations there's lots of computation going on in the babbling brook it's computation that isn't very humanlike”
People have different skill sets and interests; being born in a particular era can be lucky or unlucky depending on whether that era's dominant activities align with one's interests and aptitudes.
“people have different skills, people have different interests. One can be lucky or unlucky in the period of history in which one lives. So for example, I consider myself lucky to have been in a period of history when computers started being sort of usable things cuz they're a good fit for a lot of stuff that, you know, I like doing.”
In the 1970s, Wolfram was already using online database services with keyword search to find scientific papers; this evolved through full-text web search to modern semantic search via LLMs—a continuous improvement trajectory rather than a discontinuous breakthrough.
“back in the in the 1970s I was already using you know online uh database services where you could do you know that people had uploaded the abstracts of all scientific papers... That became a lot easier when the web came along... This is another step.”
The concern that superhuman AI will 'take over' mistakes anthropomorphization for reality. Programs don't 'want' things; saying a program wants to 'fill black squares' is a human metaphor, not an accurate description of what the program actually does.
“the generalization of wanting to for example all these programs that I've studied in the computational universe I might as a human say oh it seems like it wants to you know fill this thing with black squares or something but that's really a very weird description. It's a very humanized description of something that really isn't very human.”
It's trivial for an AI to be creative—it just picks a random number and does something unexpected. The real question is whether the random thing it produces is something a human cares about.
“why can't the AIs be creative? It's very trivial for an AI to be creative. It just has to pick a random number and that's doing something that you know is creative. Now the question is is the random thing that it picks is that something a human will care about”
Many people find important things that they never had access to because the barrier to entry was too high. As automation lowers barriers, people can pursue more of what they care about. This is not dehumanizing but human-amplifying.
“that's not a that's not sort of a dehumanizing thing in a sense. It's a human amplifying thing because what gets more important is well what is it you want to do? What's the idea that you have?”
The real question about software automation isn't whether code generation is possible but what that app is supposed to do. Code generation reduces friction, but someone still has to conceptualize the purpose and functionality. Specification, not implementation, is the bottleneck.
“the thing to understand about things like software is and it relates very much to what we've been talking about a lot in this in this conversation is you say you know you you snap your fingers and then there's an app. Well, what is that app supposed to do? You have to describe what it does and that is making a choice so to speak...the description of what it does is the thing that is the thing that's going to be of value. It's not the mechanics of”
The space of possible software is like the computational universe: everyone is a piece of software doing something, most of which are not things anyone would care about. Useful software requires mining the computational universe for patterns people want and turning them into applications.
“what um uh you know I think it's really a well what do you you know that the space of possible software it's like these little programs that I study in the computational universe everyone is a piece of software in a sense everyone does something most of them are not things that anybody would care about some of them make really pretty pictures some of them make good cryptographic systems some of them make good image processing filters you know, those are ones where we've been able to kind of mine what's out there”
With mathematical theorem generation, computational systems can generate billions of true theorems, but most are trivial (people look and say 'yes, that's true, but so what?'). Progress in math that humans care about is about building a tower where each result enables the next, not maximizing the quantity of true statements.
“something I've looked at quite a bit is if you look at mathematical theorems it's pretty easy to get a computational system to just go spewing out zillions and zillions of theorems, billions of theorems. They're all true theorems...But it's not making progress in math that people care about. Because most of those theorems, people look at them and say, "Ah, okay, I guess it's true, but so what?" And what counts as kind of math that humans care about tends to be this kind of prong that gets built, this tower that gets built”
Eliezer Yudkowsky's concept that we have 'a responsibility to keep the universe interesting' or to spread life through the universe represents an ethical obligation, but Wolfram finds this concept incomprehensible because ethics is a human construct.
“Eleaser has sort of an idea about uh uh sort of um that we have a responsibility to keep the universe interesting so to speak but I have no idea what that means... Ethics is a human thing. There is no abstract ethics.”
The definition of intelligence as 'what humans have' is historically doomed as a framework because humans have repeatedly been humbled by discovering they are not special in the ways they assumed.
“the lesson of the history of science has been we keep on getting humbled in the fact that no, we're not special in that way or this [59:26] way or or whatever. And I think the way in which we're special is that we are precisely the way we are, so to speak.”
AI systems don't have wants or desires in the way humans understand them; attributing objectives to a simple program that does what it does is a humanized, externally-imposed interpretation rather than a reflection of internal states.
“the concept that, you know, the AI that has all of this sort of computational ability is going to want to do something is is a very weird concept... the generalization of wanting to for example all these programs that I've studied in the computational universe I might as well a human say oh it seems like it wants to you know fill this thing with black squares or something but that's really a very weird description. It's a very humanized description of something that really isn't very human.”
We can learn to exist alongside systems we don't understand by using tools we've always used: making predictions, creating shelters, and organizing our society around managing incomprehensible natural forces (like tornadoes).
“there are things that are in principle possible that might be possible in the natural world that might be possible in the computational world... in the natural world, there are tornadoes and things that don't align with what we want. we end up, you know, being able to predict them and, you know, having tornado shelters and things like that to kind of exist alongside them.”
Humans are negligible relative to the universe's computation and yet feel content in their niche, so creating large non-human computational systems should not be feared if humans maintain their niche of choice-making.
“but we seem to have just a fine time in that position you know relative to the relative to the universe we are negligible relative to even the computation that happens, you know, around us on the earth, we're [1:01:28] negligible. Um, yet we feel pretty pleased with ourselves regardless.”
The computational universe is infinite; there is no point at which all possible computations will be exhausted, all theorems will be proven, or all inventions will be made, ensuring an endless frontier of possibility.
“the computational universe of possibilities is is infinite. You might have thought that you know there'll come a time when we've made every invention that can be made. That time will never come. It's you know a century ago people were like yeah we've almost made every [1:43:37] invention that could be made. well turned out was not true.”
Some forms of value cannot be purchased with money, such as the internal experience of making a discovery in science, suggesting money's reach is limited.
“you know for me the you know my general approach to things is use any tool you can you can. And so you know for me you know one of the things that's funny I was just realizing this actually people are kind of sort of [1:17:00] say well if the AIs get to be really smart they'll you know what will that be like for humans so to speak? You know, I myself happen to have been in this situation that I've sort of created for myself where I've been building tools to sort of that that can enhance my ability to think about things.”
High IQ alone does not guarantee world-takeover capability; Wolfram knows many people with very high IQ who cannot execute world-changing plans, suggesting intelligence-to-power conversion is not straightforward.
“people say things like well you know think about the IQ of a of an AI and think about the fact that it can [1:22:17] improve itself and its IQ is going to run off to infinity. Well, you know, I I think of many people who would be able to do IQ tests really well who I know who uh I can I can be quite sure and not you know that that on its own doesn't you know you it it isn't the the ticket to sort of you know be take over the world type thing.”
When multiple people pursued a goal initially perceived as impossible due to lack of resources, they succeeded by treating the resource constraint as an excuse rather than a fact, suggesting that resource gaps are often psychological rather than material.
“I see happen a lot I remember well you know you see people saying I can't do that because I don't have enough money to do it. Sometimes that's true but a lot of the time it's just not true. A lot of the time it's just deciding you're going to do it and there'll be a way to do it and it doesn't really have anything to do with the money. That's just an excuse, right?”
The productivity frontier in knowledge work is not knowing the language (natural or computational) but knowing what you want to express—once mechanical communication barriers lower, the constraint becomes human intentionality.
“yes, that's a thing where where you're um uh you know, again, that's that's what I've been living for many decades now is getting to the point where it's mostly about imagining what the app is supposed to do, not about the mechanics of actually writing it.”
The trolley problem (choosing between killing one person or five) is framed as a scientific problem requiring abstract ethical reasoning, but ethics cannot be abstracted because choices depend on the entire history and context of the situation, making controlled ethical experiments impossible.
“in science one of the things that makes science possible is that we can do kind of controlled experiments... In ethics, I don't think that's possible. In other words, there is no answer to the question of the llamas or the endangered lizard without knowing the whole story of sort of the connections of the llamas, whether that one of those llamas was somebody's pet llama, whether there was a, you know, a group that worships llamas, and whether the endangered lizard was a, you know, all kinds of things.”
The scenario of brain uploading leading to 'a trillion souls in a box playing video games for eternity' might seem like a terrible outcome from modern perspective, but would likely be internally experienced as deeply meaningful by those uploaded minds, making external judgment of its value problematic.
“there's a view of sort of the future which says well we'll figure out brain uploading and all this kind of thing and pretty soon what the future of humanity will be you know a trillion souls in a box playing video games for the rest of eternity. Okay. Yeah. and and you might say, 'Gosh, that's a terrible outcome.' From our point of view today, from our experiences today, from the things we care about today, that seems like a terrible outcome. But my guess is that it in the internal experience of that disembodied soul, you know, playing video game, playing quotes video games will be perfectly meaningful.”
Wolfram advises people to believe that 'anything is possible' rather than accepting constraints, citing that kids and CEOs both share the belief that anything is possible, while other people get stuck in their current track thinking it's unchangeable.
“I kind of noticed at some point that those two categories, both of which I found interesting, were the the category of kids and the category of of CEOs, they're types of people who believe that anything is possible... there are lots of other people who don't believe that anything is possible. They're they're kind of like we're we're stuck in this in this particular track”
Early AI discourse in the 1950s-60s predicted humans would be remembered as the species that created the next species (AI), and this narrative remains nearly identical today except with updated pronoun usage from 'men' to 'people'.
“the idea that you know kind of we were a species that was mostly going to go down in the history of the earth as being the thing that created the next species that was was AI and so on. And it's really really funny to read these things now because honestly pretty much word for word they're the same as what people say today with the one exception that a bunch of the kind of language of you know men will do this so to speak has been you know adjusted in modern times. It will be people will do it”
The reason ChatGPT works is that language itself has more regularity and structure than people typically realize. Specifically, there is a 'semantic grammar' of language—patterns about which nouns, verbs, and nouns can go together based on meaning rather than just syntactic rules.
“There's something about language that we hadn't noticed that actually is more regular than we had we had imagined. And you know, we're all used to the idea that we do grammar and we know that you know in English sentences are formed with you know noun verb noun with different parts of speech and different combinations. But there's more to it. Most sentences that are just noun verb noun are completely meaningless. But there's a kind of semantic grammar, a grammar that's based on meaning that says what noun, what verb, what noun can go together. And there's there's, I think, a lot of regularity in that structure. There's a kind of a a semantic grammar of language.”
Currently LLMs and computational systems interact loosely: the LLM generates text, decides it needs to call a computational tool, gets results back, and weaves them into ongoing text; more fine-grained integration would be more powerful but Wolfram is uncertain how to achieve it.
“the combination is really powerful and that combination you know one of the things that will eventually happen although I don't yet know how to do it is to make a more sort of fine grained uh integration of those things because right now the LLM is going along it's generating a bunch of text then it generates kind of the the things it needs to call our system as a tool gets the results back then keeps going and uh [46:26]”
The current era of incomprehensible technology is not unprecedented; before industrialization, people used horses without understanding how horses work mechanistically. The industrial revolution created a brief period where technology was understandable, but as technology becomes more sophisticated, incomprehensibility returns as the normal state.
“You know, in the past when people were you know getting transported around by riding horses you could know something about how to get the horse to do what you want but knowing how the horse works inside was not something you really cared about. You were able to use the horse for something that was useful to you but you didn't know mechanistically how the horse worked inside. You know, post-industrial revolution, for a brief time, we've been operating machines that are simple enough that we know what's going on inside.”
There are positive and negative values of money: positive when it enables things you care about, negative when it creates conflict between people who otherwise might be friends, or creates problems like wealth management burden.
“There's positive value of money and there's negative value of money. You know, sometimes you see, you know, I I know plenty of people who've been bitten in so many ways by the negative value of money, so to speak...there's uh you know there's this pot of money and people who would otherwise be be friends are you know arguing to the death over it, so to speak, even though if the pot of money wasn't there, they'd just be happily friends with each other”
Being the first person to discover something (in science, art, or any domain) is intrinsically exciting and valuable to humans, and this built-in scarcity—that only one person can be first—will continue to drive human ambition and purpose regardless of technological abundance.
“there's always going to be the first person who discovers that. And that's kind of an exciting thing to be the first person who does this or that thing. There's always sort of a built-in scarcity to to what's out there.”
Even with LLMs and other AI advances, more complex software development still requires the integration of various AI capabilities with human-designed infrastructure, not just end-to-end automation. The frontier of what's hard is constantly shifting to higher levels of abstraction.
“the thing to understand about things like software is and it relates very much to what we've been talking about a lot in this in this conversation is you say you know you you snap your fingers and then there's an app. Well, what is that app supposed to do? You have to describe what it does and that is making a choice so to speak”
Wolfram's company (around 800 people) is tiny relative to the amount of software it has produced, because of decades of automation. This creates a 'recursive' advantage where automating something enables automating the next thing.
“I mean my company is what about 800 people which is really kind of tiny comp relative to what we've been able to produce. Okay. Um, it's uh, well, it, you know, I think relative to the amount of software we've produced, how has that been possible? Well, it's because we've automated the heck out of things. And you know, we're building on sort of a, you know, we built this tower where we're kind of recursively able to do more because we've automated the last thing we were able to do.”
New technologies that reach the horizon (like VR/XR) will enable new categories of software that don't exist today (like virtual note-taking in augmented environments), creating new opportunities for software entrepreneurs rather than eliminating them.
“Another piece in the 1980s was uh the kind of the as I mentioned sort of the rebirth of interest in neural networks and uh a bunch of people including the people who sort of came you know continued working on that till the present day people like Jeff Hinton and my friend Terry Senovski who were kind of the the early people who got interested in could neural nets really be made to do something interesting.”
When Wolfram showed Wolfram Alpha to Marvin Minsky (the same AI pioneer who had dismissed neural networks in the 1960s) just before its release in 2009, Minsky initially dismissed it as just another failed question-answering system, but when shown a few examples, Minsky became enthusiastic and started telling others 'You've got to see this. It actually works.'
“I was a couple of weeks before we released Wolf from Alpha. I happened to see Marvin Minsky, who I mentioned earlier, who was sort of a big AI pioneer. And so I say to Marvin, we got this cool new thing that's coming out, you know, let me show it to you. So I kind of show him a couple of things and he's like, changes the subject. Not interested. It's like because for him, he'd seen a zillion examples of people saying, 'I built a question answering system.' And so I said, 'Look, Marvin, you know, you should look more carefully. This time it actually works.' And so he types a few more things and then he's like, 'Oh my god, it actually works.' And he's running around this event that we were at telling people, 'You've got to see this. You got to see this. It actually works.'”
The coming era is likely to favor people who like thinking and having ideas, as mechanical labor becomes increasingly automated; this represents a shift in which skills are valuable.
“I do think that what's coming is I do think that what's coming probably is a time when if you like thinking this is the time for you so to speak. If you like having ideas and so on this is the time for you. If you like kind of doing the um you know the mechanics of doing things maybe it's less the time for you.”
For software engineering, the 'spec' (what the software should do) is the piece that matters and is hard, while implementation should be increasingly automated; the difference between a good and bad engineer is increasingly the ability to define good specs, not write code.
“people who run them, the CTOs, the CEOs and so on actually use our tech and build prototypes of things in a very short time and then they go on and they figure out the next thing they want to do and build that as well. But the people who are now sort of in the trenches doing software engineering are like, "Well, we just got a spec. We build to that spec." Okay, now if we could do that much more quickly, which they can with our tech, then they're like, "Well, now I've got something very difficult to do. I've got to make a new spec. That's not what my job is. My job is to grind out code."”
Wolfram believes it is conceivable but unlikely that humans would simply abdicate choices to AI and let machines run society on autopilot, as this would eliminate the human capacity for novelty and new directions.
“you know then you ask the question well why can't the AIs be creative? It's very trivial for an AI to be creative. It just has to pick a random number and that's doing something that you know is creative. Now the question is is the random thing that it picks is that something a human will care about.”
Software development by humans (manually writing code to implement specifications) is a peculiar artifact of current economics and tool limitations; Wolfram has spent four decades automating away the need for manual code writing.
“the fact that there is this whole ecosystem of people doing manual labor software development is just bizarre. I mean it's happened because of a bunch of the economics of labor... the fact is you know you say it's you know you can create an app in you know that's what I do many times a day writing tiny amounts of code... that's what you know that's been my objective is to automate those things”
The term 'general intelligence' was coined in the 1930s simultaneously with the development of IQ tests, as a concept for measuring human intelligence; both concepts are troublesome in Wolfram's view.
“the term general intelligence was I think coined in the 1930s and um the uh it came in when uh at the same time as the concept of IQ came in. a very in my opinion a very troublesome concept but um it was at a time when people were trying to figure out I don't know for recruits for the army and things like this”
Neural networks experienced a comeback around 1982 when experiments showed that deeper networks with more layers could do interesting things, contradicting the Minsky-Papert conclusion that perceptrons were fundamentally limited.
“neural nets had kind of a comeback. This is around 1982 or so. There were, uh, was a couple of sort of experiments that were done with neural nets where it was like, wow, they're able to actually do things. And the fact that they were sort of, you know, squashed flat by the [14:36] perceptrons analysis wasn't really right. If you had deeper neural nets that had sort of more layers of of computation in them, then it would then they might be able to do something interesting.”
Wolfram created images by moving around in 'interconcept space' (the internal representation of concepts in AI systems), finding regions that don't correspond to human concepts but are equally meaningful within the AI's representation. Some of these images are now in art exhibits in Paris, showing that interconcept space can develop into human-recognized art styles.
“you move away from the concept of cat in a party hat into what I was calling interconcept space. You move away from these human-defined concepts to concepts which are which exist in the mind of the AI but are not familiar to us humans...some of those pictures actually I was was amused to see that somebody took some of the pictures that I made in that post and they're now in some art exhibit somewhere in Paris”
The term 'artificial general intelligence' (AGI) is a poorly-defined buzzword similar to 'AI' itself; it implies something that can do everything humans can do, but the only thing that can do everything humans do in all respects is a human.
“Nobody knows what AGI really means. It's a it's a buzzword that um it's kind of like uh you know, first AI was a buzz word and people didn't quite know what that meant except they thought it meant things that do stuff that's kind of like what people do... the only thing that is going to check all the boxes for being like a human is a human.”
Even if future humans could upload their minds, a trillion disembodied souls playing video games would likely find that internally meaningful despite external observers viewing it as a dystopian outcome.
“there's a view of sort of the future which says well we'll figure out brain uploading and all this kind of thing and pretty soon what the future of humanity will be you know a trillion souls in a box playing video games for the rest of eternity. Okay. Yeah. and and you might say, "Gosh, that's a terrible outcome." From our point of view today, from our experiences today, from the things we care about today, that seems like a terrible outcome. But my guess is that it in the internal experience of that disembodied soul, you know, playing video game, playing quotes video games [1:53:35] will be perfectly meaningful.”
Many activities that people today engage in can be viewed as ritualistic (sitting in front of a computer) unless one understands the purpose and meaning behind them, showing how meaning is subjective and context-dependent.
“I'm amused to realize the extent to which so many of the things we do today could be seen [1:51:30] as ritualistic. I mean from the point of view of you don't know why you're doing it you know I'm sitting in front of a computer getting weird pictures coming up on the screen this seems like a devotional ritualistic kind of activity if you don't you know if you don't have a thread of understanding what the point is so to speak”
Building an effective AI tutor requires 'four times as much AI work going on behind the scenes' compared to the AI directly interacting with students, showing that while LLMs excel at language interaction, substantial infrastructure is needed to make full applications work.
“in our AI tutor, there's four times more sort of AI work going on in behind the scenes than the AI that's actually interacting with the student... that's defining what should happen”
ChatGPT's breakthrough is analogous to the invention of the telephone: people knew in principle you could transmit sound over electrical wires, but the signal was unintelligible until Bell and others found practical hacks; similarly, the technology to generate meaningful text from neural nets existed in principle, but ChatGPT found the practical combination that worked.
“I kind of liken it to sort of a history of technology thing, which was the invention of the telephone. You know, people had known that in principle you could transmit voice over you could transmit sounds over over over electrical wires. But when people had done that, you know, you would try and listen at one end of the wire and you wouldn't understand anything was being said. And eventually, you know, Alexander Graanbell found a bunch of hacks that again, I don't think he knew were going to work, but suddenly people could actually understand what was being said... I think the same kind of thing happened with neural nets.”
If we create neural networks with 100 trillion neurons (vastly more than humans' 100 billion), the resulting system would engage in far more computation and would likely be very non-human, raising the question of what such a system would do and whether it would be intelligible to humans.
“we have 100 billion neurons we have a thing with 100 trillion neurons we have something that's doing not the number of computations that we do in our brains every second but zillions of times more than that. What is that like? Well, the fundamental thing is it's not very humanlike.”
The notion of 'objective' for humans is itself unclear: we might use the word to describe what we do, but whether objectives are innate to human cognition or learned language patterns for describing behavior is an open question.
“It's an interesting question to what extent the description of what we do in terms of objectives is something that is innate and natural or whether that's something that we learn just as we learn language to for us to describe what we do as I'm doing that because blah blah blah you know it might be that you know when we're all you know babies or whatever we just do what we do and it is a a higher layer that is our description of objectives so to speak just as it's a higher layer to be able to to describe things in terms of language.”
AI-generated content (whether mathematical theorems, images, or other artifacts) can be creative but may not be interesting or valuable to humans; what matters is whether the output connects to human values and interests.
“if you look at mathematical theorems it's pretty easy to get a computational system to just go spewing out zillions and zillions of theorems, billions of theorems. They're all true theorems. You might say, "Wow, that's exciting. That's making progress in math." But it's not making progress in math that people care about. Because most of those theorems, people look at them and say, "Ah, okay, I guess it's true, but so what?"”
Cryptocurrency could have value not because it can be directly exchanged for food, but because a network of systems depend on its existence, suggesting value is network-dependent rather than use-dependent.
“it's worth something because there's this whole network of things that depend on on it, on its existence and so on. And I think that that's again an example of something where you could say, well, you know, all [1:59:22] the value in the world is the fact that we have, you know, houses and and food and things like this, but yet there are these other kinds of value that seem to exist and and and that um so, you know, I'm I I guess I I feel like the uh uh this question of the importance of sort you know”
By 2012 or so, image identification systems using neural networks had reached a plateau where they were only marginally better than systems from 2012; the breakthrough in capability happened quickly, then improvement stalled, showing that technological breakthroughs solve specific problems but don't automatically generate ongoing improvement.
“the things that one had by 2012 or so, we were making image identification systems and so on using these neural net ideas, they're not that much worse than what one has today. In other words, in the last 13 years or so, things haven't in that particular domain haven't improved that much. It's it's like you reach a threshold, you start to be able to do something, then that works and then you build that capability into a bunch of systems and it becomes useful, but it's not as if that capability itself, you know, just because it made that one jump doesn't mean it's going to make lots of other jumps.”
Image identification was effectively solved by neural networks by around 2012, achieving 90%+ accuracy rates, which is typical for machine learning—not 100% because even defining 100% accuracy is ambiguous (is a cat wearing a dog costume a cat or dog?).
“you get it right you know 90 something% of the time or whatever and that's kind of typically the story of machine learning... It's not 100% but it's not even clear what you mean by getting it 100% right... is that dog that has been given a cat suit or something should that be a dog or a cat?”
Inside modern neural networks are 'lumps of irreducible computation' that have been fitted together during training, analogous to building a stone wall from rocks; the result is incomprehensible but functional.
“there are these kind of lumps of irreducible computation that you find there and what's happened in the training of the AI is it's kind of fitted together these lumps of irreducible computation to do the things we want it to do. Kind of the analogy I've been using is it's kind of like it's building a wall out of rocks. It's building a stone wall, so to speak.”
Wolfram's approach to AI has been to use computational tools to amplify human capability rather than fear or prevent AI, serving as a personal model for human-AI coexistence.
“my general approach to things is use any tool you can you can. And so you know for me you know one of the things that's funny I was just realizing this actually people are kind of sort of [1:17:00] say well if the AIs get to be really smart they'll you know what will that be like for humans so to speak?”
When Wolfram showed Marvin Minsky (an AI pioneer) a demo of Wolfram Alpha shortly before release, Minsky was initially dismissive (having seen countless failed question-answering systems), but after seeing it work, became excited and promoted it at the event.
“I was a couple of weeks before we released Wolf from Alpha. I happened to see Marvin Minsky, who I mentioned earlier... So I say to Marvin, we got this cool new thing that's coming out, you know, let me show it to you... he's like, changes the subject. Not interested... I said, "Look, Marvin, you know, you should look more carefully. This time it actually works." And so he types a few more things and then he's like, "Oh my god, it actually works." And he's running around this event that we were at telling people, "You've got to see this."”
The design of a computational language (like Wolfram Language) is of highest intellectual and market value, more valuable than implementing it, because the design is what enables new ways of thinking about problems.
“for me the thing that is really of the highest value is what I've spent lots of effort on which is kind of the functional design of the language... there's now I don't know what is a thousand hours of actually what's involved in doing that um out there in the world... you could take that spec and you could reimplement it. Good luck with that.”
Wolfram pursued interesting work over higher-paying work at key decision points in his career, suggesting that for some people, intellectual interest outweighs financial incentive.
“I mean one's attitude about uh uh you know I've lived a life where I like to do interesting things things I find you know fulfilling I've you know I'm practical enough that that activity has made me a very decent amount of money but it is not you know for me there have [1:55:40] been a vast number of forks in the road where it's kind of like do the more interesting thing. Do the thing that makes more money. I'll always pick the more interesting thing”
Stephen Wolfram's preferred mode of scientific work involves diverse projects spanning biology, machine learning, and mathematics in rapid succession, a capability enabled entirely by having built automation/computation infrastructure over decades.
“it would be completely inconceivable to go across all those different areas... without the technology ology tower I'd built and without a certain amount of scientific knowledge that that sort of I've accumulated... write something about biology one month and about machine learning another month and about foundations of mathematics another month.”
The combination of LLMs and computational systems like Wolfram Alpha is very powerful: the LLM generates text (including deciding what to compute and generating queries), calls the computational system as a tool, receives the results, and weaves them back into the text output.
“what LLM kind of are readily able to do is is quite different from what you can readily do with sort of raw computation. the combination is really powerful and that combination you know one of the things that will eventually happen although I don't yet know how to do it is to make a more sort of fine grained uh integration of those things because right now the LLM is going along it's generating a bunch of text then it generates kind of the the things it needs to call our system as a tool gets the results back then keeps going and uh uh and takes those results and sort of weaves them into the text that it's writing.”
Inside AI systems, there are 'irreducible lumps of computation'—chunks of irreducibly complex calculation that can't be simplified or understood by humans. These are fitted together (like rocks in a stone wall) during training to achieve desired outcomes like distinguishing cats from dogs.
“there are these kind of lumps of irreducible computation that you find there and what's happened in the training of the AI is it's kind of fitted together these lumps of irreducible computation to do the things we want it to do. Kind of the analogy I've been using is it's kind of like it's building a wall out of rocks. It's building a stone wall, so to speak. It's taking these lumps of computation that sort of happen to fit in and more or less correspond to the thing you need to tell cats from dogs or whatever, and it's putting a bunch of those together so that you get something that is uh kind of uh that that achieves the objective we want”
Wolfram lives with incomprehensible results from his programs every day and is constantly humbled by things he didn't predict. This is normal and expected in scientific work.
“when I'm working on those kinds of things practically every day I'll be sort of humbled by the fact that I imagine what the thing is going to do and then it does something I didn't imagine”
Throughout history, humanity has coexisted with phenomena in nature (tornadoes, diseases, wild animals) that we don't fully understand but have learned to manage through prediction, protection, and adaptation. Future coexistence with powerful AI systems will follow the same pattern.
“in the natural world, there are tornadoes and things that don't align with what we want. we end up, you know, being able to predict them and, you know, having tornado shelters and things like that to kind of exist alongside them. And no doubt there will be things like that that come out of the computational universe, so to speak.”
Wolfram started his first company (which later became Inference Corporation) in the early 1980s focused on symbolic computation systems, but venture capital pushed the company toward expert systems development, causing a strategic pivot that Wolfram did not initiate.
“I started my first company that was sort of aimed at doing sort of mathematical computation kinds of things through a a series of probably not great business decisions of deciding that, you know, I should bring in other people to run the company and things like this. We ended up getting venture capital and the venture capital was like these expert systems things, they're amazing. you should go chase that particular you know shiny direction so to speak. So the company kind of pivoted to having well a division of the company doing expert systems kinds of things”
Wolfram's company is about 800 people, which is surprisingly small relative to the amount of software it produces, because the team has leveraged automation extensively and built tools recursively (using previous automation to enable new automation).
“my company is what about 800 people which is really kind of tiny comp relative to what we've been able to produce... Um, it's uh, well, it, you know, I think relative to the amount of software we've produced, how has that been possible? Well, it's because we've automated the heck out of things. And, you know, we're building on sort of a, you know, we built this tower where we're kind of recursively able to do more because we've automated the last thing we were able to do.”
Wolfram is planning to use AI to analyze millions of physics papers thematically to extract insights relevant to understanding the experimental consequences of his physics theory, something that would be impossible without AI's capability to process large volumes of text and extract semantic patterns.
“For example, I'm interested. We made a big progress in fundamental physics about 5 years ago and there's a big question about whether there are experimental consequences of this theory of physics that uh uh that are made um that uh whether they're experimental consequences that one can figure out what they are. Maybe there are experimental consequences where the experiment was already done and people just didn't know how to interpret it. Well, there are millions of physics papers out there in the world. I haven't read all of them. I couldn't read all of them. So, a question is, can I use AI to essentially thematically analyze all of those papers”
People often claim they can't do something because they lack money, but Wolfram observes this is frequently an excuse rather than a real constraint—many cases show that even wealthy people with plenty of money still fail to do things because of lack of initiative/confidence, not money itself.
“I see happen a lot I remember well you know you see people saying I can't do that because I don't have enough money to do it. Sometimes that's true but a lot of the time it's just not true... it's just an excuse for for, you know, I don't really have the initiative to do that thing.”
Wolfram sees AI as a tool that helps him turn ideas into computational reality faster. He doesn't see it as replacing his thinking but amplifying it, and he extends this to humanity: AI tools will let people pursue more of what they care about, not eliminate their agency.
“for me, talking to the AI to have it help me write a piece of code or something like this is a pure, you know, it just it just helps that process. Now there's another thing that I've just started to do which I'm not quite sure how well it's going to work out but it's this...talking to the AI to enhance my capability to think about things”
The design of Wolfram Language is his highest-value intellectual contribution (compared to implementation), representing decades of thought about how to represent computation and knowledge in a way that's both powerful and practical.
“what I've done in my life probably the thing of highest value I think is a design of our computational language in other words the implementation sure that's valuable it's cost a huge amount of money to do it um and you know people use it you know every day all the time um but for me uh it's actually an interesting thing you should say that because for me the thing that is really of the highest value is what I've spent lots of effort on which is kind of the functional design of the language.”
Eliezer Yudkowsky's AI doom argument consists of three claims: (1) AIs will optimize nearly any objective given resource constraints; (2) AIs will have a wide range of possible objectives; (3) Most possible objectives leave no room for humans; the third claim is questionable because it's unclear what it means for a non-human system to 'have' an objective.
“His theory is this. His theory is uh AIS will be able to optimize the doing of almost anything. I agree with that... The up to the constraints of the physical world... second statement is AIS will sort of have objectives that are kind of a wide range of different objectives if you can define them as having objectives ives. I think that's kind of true, but I'm not sure what it means to say that they have objectives. Well, that so then the next claim would be most of those possible objectives don't leave room for humans. That's a much more bizarre claim, I think, because it's like saying that, you know, it's very it's difficult to define this notion of an objective for something that doesn't have the kind of thinking that that isn't humanlike”
People often blame lack of money for failures that are actually about lack of initiative or confidence; Wolfram encountered a CEO about to make $50 million unable to do things a young person without money wanted to do, showing barriers are internal, not external.
“I had a a charming case actually many years ago now. I I kind of have a hobby of doing uh kind of CEO counseling and um of advising companies and so on. And I also [2:17:40] have always found an interest in kind of mentoring kids. And I I kind of noticed at some point that those two categories, both of which I found interesting, were the the category of kids and the category of of CEOs, they're types of people who believe that anything is possible.”
When ChatGPT was released, even its developers didn't know it would work as well as it did, and if they had known, they probably would have constrained it further; many of the design choices were exploratory rather than production-ready.
“I remember chatting with the folks who worked on it just after it came out and sort of my obvious question was did you know it was going to work they were like no Um, and in fact, probably had they known it was going to work as well as it did, they would have tried to constrain it in a lot more ways than they did.”
Wolfram and team have built systems where LLMs can call Wolfram Alpha and Wolfram Language as tools, using English as a transport layer between natural language requests and computational systems, enabling LLMs to accomplish computational tasks through delegation.
“one of the things that we had from Wolf from Alpha was a system that could take text and compute from it. So that was a pretty nice combination because you can take the things that Chad GBT is producing as text... And the thing we realized well very immediately actually was that you could use English effectively as the transport layer between the AI and our computational system and you could have chatbt kind of call wolfam alpha as a tool.”
Wolfram had a conversation with Eliezer Yudkowsky (a leading AI alignment researcher), and after understanding his doom scenario, Wolfram believes the reasoning is incorrect because it rests on a faulty notion of AI objectives.
“I was having a a conversation with a person who's kind of a one of the leading AI doom folk uh chap called Elazi Yukovski um that the um he had a long conversation and I I think I finally understood his his view of kind of the the the scenario of doom and and honestly as I as I said I just don't think It's right.”
Recent developments allow even faster translation from thought to code through AI assistance, shortening the loop between idea and implementation further.
“It's actually got even shorter recently because we built this notebook assistant system that is based on LLMs and other technology that we made um that allows one to kind of even more efficiently go from kind of a thought [1:18:04] you have to computational language code that can actually run.”
Wolfram has given CEO coaching/mentoring and noticed that both kids and CEOs share the belief that 'anything is possible,' whereas many other people feel 'stuck in this track,' suggesting that possibility-belief is a critical difference between people who pursue innovation and those who don't.
“I kind of noticed at some point that those two categories, both of which I found interesting, were the the category of kids and the category of of CEOs, they're types of people who believe that anything is possible... there are lots of other people who don't believe that anything is possible.”
In 2009, Wolfram released Wolfram Alpha, which takes natural language input, translates it into precise computational language, and returns computed answers; this was viewed as an AI system, but most people in AI at the time believed question-answering systems would never work, having failed for decades despite multiple approaches.
“in the mid warts I decided okay it's time to actually try and do this and so that led to this system that uh came out in 2009 called wolf from alpha... AI was absolutely dead at the time. Everybody thought nothing like this is going to work. People had tried to make question answering systems with various kinds of AI techniques, statistical, symbolic, whatever. They tried to do that for decades. It had never worked.”
Japan launched the Fifth Generation Computer Project in the early 1980s with the goal of leapfrogging ahead of American and European AI research, using Prolog and particular problem-solving methodologies that ultimately did not work out well.
“Japan was viewed as a country where kind of oh it just copies American technology... Japanese government had this this great idea. They said let's do a research project that's going to jump ahead of everybody else. It was called the Japanese fifth generation computer project and it was a project in I forget when it started early 80s sometime... their methods well they were using particularly a language called prologue which is sort of a lispish kind of thing but it had a particular idea about problem solving which doesn't work out so well in the end.”
The history of text generation goes back to the 1940s-1950s and cryptanalysis, where Claude Shannon developed information theory to understand the statistical regularities of language, enabling the idea that language could be generated statistically.
“started I guess in the 1940s when people were doing crypton analysis in World War II... English isn't just a random sequence of letters. English has certain statistical regularities... a chap called Claude Shannon worked out this thing he called information theory... wrote this paper in 1948 I think introducing information theory and this idea of sort of the statistics of things like language”
John McCarthy coined the term 'artificial intelligence' at the 1956 Dartmouth Conference and subsequently invented Lisp as the language he believed was necessary for implementing AI, particularly symbolic mathematical computation.
“the term AI was coined by John McCarthy who I knew we didn't get on that well but that's a different story... in 1956 there was this conference at Dartmouth where John McCarthy coined the term AI and uh one of the things John McCarthy did in the in the years right after that was invented this language called lisp”
In 1943, Warren McCulloch (a neurophysiologist and psychiatrist) and Walter Pitts (a young mathematician) published a foundational paper on the logical theory of neural nets that laid out an idealized mathematical framework for understanding artificial neural networks, serving as the foundation for all subsequent neural network research.
“the next probably big step was 1943 uh chap called Warren McCulla and chap called Walter Pittz... They worked together wrote a paper about kind of the logical theory of neural nets. That paper is kind of the foundation of everything that's been done since.”
With the notebook assistant system (based on LLMs and other technologies), the path from having an idea to writing executable computational code has become even shorter, allowing faster conversion from thought to running code.
“it's actually got even shorter recently because we built this notebook assistant system that is based on LLMs and other technology that we made um that allows one to kind of even more efficiently go from kind of a thought you have to computational language code that can actually run. So it's a funny thing because I I guess I've lived this perhaps ahead of where other people would have lived it.”
Sequence prediction—the ability to predict what text will come next given a sequence of input text—was extremely poor in 2021 and attempts to use it for code autocomplete or similar applications did not work well.
“could you do what was called sequence prediction could you do a better job of knowing given that you have a piece of text that starts this way what will come next and what existed even in like 2021 times like that of sequence prediction was really cruddy. It really didn't work well at all. You know, we tried to use it a bunch for doing oh things like predicting uh pieces of code for autocompletes, things like this really did not work well.”
Wolfram believes an AI tutor targeting algebra skills is likely to work (though he hasn't yet verified) because unlike raw LLMs, a dedicated tutoring system can maintain long-term context about a student's learning progression and provide appropriately scaffolded instruction.
“We have a big project right now to build an AI tutor. You know people have been trying to do sort of computerized education for 70 years and it's basically always failed to do you know one can get computers to help but to be the prime teacher has never worked. It looks promising this time around. I'm not saying for sure it will work. We'll know. We'll probably release it in a few months and then we'll know whether it works or not.”
Speech recognition in the 2010s solved a major challenge by having neural networks go directly from audio input to text output, rather than using the previous pipeline of breaking speech into phonemes, recognizing patterns, then statistically reconstructing language—this is similar to how image recognition got solved by deep learning.
“people started just trying to use neural nets to just go straight from the audio you hear to the text that's being generated. And turns out it worked... just as image recognition got solved, so kind of speech to text kind of got solved too.”
Electricity's discovery came through Volta and others observing that frog legs kicked when given electric shocks, establishing that nerves operate electrically, which combined with knowledge of neural networks suggested the brain might implement logic electrically.
“people knew that nerves were electrical because actually the way electricity was discovered was by Voltaar and people uh kind of noticing that frog legs kicked when you gave them electric shocks. So people kind of knew it's electrical and then they kind of knew there's a network of nerves in the brain. And what did that mean? Well, by the 1870s, people were talking about how there might be it might be implementing logic in the network of nerves in the brain in some kind of electrical way”
By the mid-1950s, people were implementing neural networks on early computers and special-purpose machines, and the perceptron emerged as a simplified neural network that could perform statistical image recognition.
“by 1946, there were starting to be electronic computers. By the 19 mid 1950s, people were implementing kind of this idea of neural nets on early computers and on special purpose computers that they'd made particularly for doing neural nets. And particularly there were this idea of the so-called perceptron”
In the 1980s, expert systems became the dominant AI approach, based on the idea that computers could replicate expert knowledge by encoding rules written by human experts in specific domains like geology or medicine.
“the thing that was sort of the dominant theme of AI in the 1980s was these things called expert systems. And the idea was you would have sort of a rules-based way of describing the world that you would learn from an expert. Somebody would essentially write the rules based on some expert in geology or some expert in medicine or something like this would write the rules and then the computer would be able to do whatever the expert could do”
Moats in software come from many sources: network effects (like social media), unique intellectual property (like a computational language), or other defensible advantages; no single business model dominates.
“there are so many different kinds of moes. I mean there are moes to do with I've got this you know social media platform and all your friends are on it. You should be on it too. Yeah. There's um there's a you know we've got this sort of unique intellectual property that that we've um that we've been able to build”
Wolfram predicts that robotics will be the next major AI frontier to be solved, as robot manipulation and picking up objects have proven surprisingly difficult despite being easy for humans; there is limited training data for robotic tasks compared to billions of words on the web.
“my guess is within, you know, the next big thing that will get solved in AI is robotics. Um, and you know, it's been super difficult to get, you know, robot hands to to pick things up and, you know, be able to pack boxes with whatever stuff you want from a warehouse, things like this. Um, it's it's something we humans manage to do fairly easily. It's something that it's been a little difficult to get training data, you know, for the stuff that humans write while there's, you know, a trillion words on the web”
There were two primary approaches to AI that emerged in the 1960s: the symbolic approach, which sought to represent the world computationally and reason about it, and the statistical approach, which attempted to extrapolate from observed data without explicit world models.
“there were really two approaches that well first thing in the 1960s there were two different approaches that were taken to AI and those there continue to be two different approaches today. there was kind of the the symbolic approach and the statistical approach. The um the the idea the symbolic approach was kind of the thought that you can have sort of a computational representation of the world and you can have your AI kind of figure out things in the world in a kind of computational fashion... The other approach was the statistical approach just saying forget about having rules for how the world works.”
Neural networks as a concept trace back to the mid-1800s, emerging from investigations into how brains actually work, particularly following Golgi's development of nerve cell staining techniques in the 1870s that allowed visualization of neural architecture.
“neural nets were all entangled with the question of how our brains actually work... There was a chap called Golgi who figured out in I think the 1870s how to stain nerve cells because when you look at you know a slice of brain tissue under a microscope it just looks really complicated. You can't see anything.”
Golgi and Ramon y Cajal disputed whether nerve cells formed one continuous network (Golgi's view) or were separate cells with synaptic gaps between them (Cajal's view), with both receiving the Nobel Prize together despite their violent disagreement, representing a mid-to-late 1800s debate resolved in Cajal's favor.
“there was a big sort of dispute between Golgi and a chap called Romani Kajal uh because Golgi thought that sort of every nerve that you would see these sort of nerve cells and that they were really all connected in one big net. Romani Kajal thought they were all separate cells that had sort of synapses gaps between them... even disagreeing violently about that won one of the early Nobel prizes uh together”
People discovered that nerves were electrical through experiments where Voltaire and others noticed that frog legs kicked when given electric shocks, establishing early evidence that neural computation was electrical rather than mechanical.
“people knew that nerves were electrical because actually the way electricity was discovered was by Voltaar and people uh kind of noticing that frog legs kicked when you gave them electric shocks.”
Expert systems dominated AI research in the 1980s, based on the idea that a computer could be trained with rules from an expert (in geology, medicine, etc.) and then replicate that expert's capabilities through rule-based computation.
“the dominant theme of AI in the 1980s was these things called expert systems. And the idea was you would have sort of a rules-based way of describing the world that you would learn from an expert. Somebody would essentially write the rules based on some expert in geology or some expert in medicine or something like this would write the rules and then the computer would be able to do whatever the expert could do.”
OpenAI collected a huge amount of training data from the web and elsewhere, and used transformer nets—a technical architecture where neurons connect sequentially with long-range connections capable of referring backward across a sentence—to train large language models.
“the guys at OpenAI uh basically collected this huge and and well they collected this huge amount of training data... There was one kind of technical idea which I'm not sure how significant it will be in the long view of history. This idea of transformer nets... When it comes to things like language, language has the feature that it's sequential... And there were the these things called transformer nets which are kind of are networks that are sort of that only have connections sequentially but their connections can be quite long range”
One capability that distinguishes humans from other animals is compositional language—the ability to create new sentences from words in arbitrary combinations—which animals like cats and dogs lack (they have limited vocalizations like 'sit' and 'fetch').
“one of the ones we're proud of is human language... cats and dogs don't quite have that. They maybe have, you know, sit, fetch, and so on, but they don't have the sort of compositional language that we have, being able to put words together in arbitrary combinations to make sentences”
Basic necessities like food have become cheaper and easier to produce through automation and agricultural innovations (fertilizer, crop breeding), solving what was once predicted to be an intractable population-food problem.
“you know when [1:49:23] fertilizer was invented you know people thought the world was going to run out of food because there weren't going to be enough crops produced to deal with the population increase. But then things like fertilizer, crop breeding and so on were were were invented and uh you know and that problem went away through something and because food effectively became cheaper”
In the 1970s, people thought automation would be exhaustive and humanity had almost made every invention, but the 20th-century development of entirely new categories of inventions proved this wrong repeatedly.
“a century ago people were like yeah we've almost made every [1:43:37] invention that could be made. well turned out was not true.”
Humans have different values, beliefs, and conceptions of what should be done; some choose ascetic lives rejecting modern inventions, while others embrace technological possibilities—the future will involve different segments pursuing different paths rather than convergence.
“it could be the case that at some moment sort of our species or some segment of our species just decides we've got enough. We don't have to invent anything new... I don't think that is the way the human condition is going to play out but it's not something that I mean you know it is the case”
Reinforcement learning with human feedback (RLHF) was an additional layer of training beyond simple sequence prediction that allowed ChatGPT to actually do what users requested rather than just predict the most statistically likely continuation.
“there was an additional trick which was this idea of reinforcement learning and particularly reinforcement learning with human feedback which was went beyond just how would this sentence continue... the idea of get the thing to actually do what it's told and do things like answer questions. Um, that was a thing that was sort of a a another layer of of training”
There is now a weekly ranking website where different LLMs are rated on performance, and the rankings change practically every week, indicating a rapidly advancing and competitive field with numerous models being developed.
“We have a nice little website where we every week do a kind of a rating of of how how these various LLMs are doing. And uh it's kind of remarkable. It's an active enough field that practically every week there are changes at the top”
Wolfram livestreams software design meetings, making public ~1000 hours of material showing the intellectual work of designing a computational language, providing a record of how such design is actually done.
“we we live stream many of our sort of software design meetings. So the there's now I don't know what is a thousand hours of actually what's involved in doing that um out there in the world.”
Reinforcement learning with human feedback (RLHF) was an additional training layer on top of sequence prediction that taught ChatGPT not just to continue text statistically, but to actually do what it's told—answer questions, follow instructions. This was a separate training phase beyond the base language model.
“there was an additional trick which was this idea of reinforcement learning and particularly reinforcement learning with human feedback which was went beyond just how would this sentence continue... the idea of get the thing to actually do what it's told... that was a thing that was sort of a a another layer of of training”
ChatGPT can write code in Wolfram Language in addition to Python, and can use Wolfram Language as a way to represent computational thinking, not just as a target language for code generation.
“our wolf language, which you know I'd been building for so many years and and lots and lots of people use as a language for for doing computation as a way to represent kind of one's thinking computationally. Well, chatbt could do the same thing. it can write wolf from language code”
A fruit fly has approximately 130,000 neurons, cats and dogs have roughly a billion neurons, and humans have about 100 billion neurons; scaling from fruit fly to human capabilities shows that more neurons enable more sophisticated behaviors like compositional language.
“a fruitly for example It has 130,000 neurons in its brain. We have about 100 billion neurons in our brains. You know, cats and dogs have I don't know how many, maybe a billion or so neurons in their brains. Well, you know, we get to do some stuff. We get to do a whole bunch of stuff that fruit flies don't get to do.”
In speech recognition, the challenge was not just recognizing individual phonemes but assembling imperfect phoneme predictions into coherent text, which required statistical understanding of language structure.
“people doing speech recognition where you're trying to figure out you know we hear these speech sounds and we can sort of statistically work out oh that speech sound is roughly a vowel but then the question was well how would you assemble those those sort of somewhat imperfect things about well that might have been an L that might have been an R whatever else how would you assemble those rather imperfect kind of guesses about what those phonemes what those fragments of speech were like into something which could actually be a meaningful piece of text”
By the early 2010s, neural networks began to be applied directly to speech recognition, bypassing the traditional phoneme-recognition pipeline and going straight from audio to text, which worked much better than the previous statistical assembly approach.
“people started just trying to use neural nets to just go straight from the audio you hear to the text that's being generated. And turns out it worked. And so just as image recognition got solved, so kind of speech to text kind of got solved too.”
Wolfram founded a company in the mid-1980s that pivoted to expert systems at the urging of venture capitalists, eventually becoming Inference Corporation, which built credit assessment systems and testing systems for NASA before going public in an undistinguished IPO in the 1990s.
“my first company that was sort of aimed at doing sort of mathematical computation kinds of things through a a series of probably not great business decisions...the company kind of pivoted to having well a division of the company doing expert systems kinds of things and and it's actually kind of strange to realize that the company changed its name eventually to inference corporation...It was an AI company. um it was doing expert systems AI and the company built all kinds of things for for I don't know it built early credit reporting systems or credit assessment systems. It built a bunch of uh testing systems for for NASA”
In 1979, Wolfram began building a symbolic manipulation program (SMP) using the language C rather than Lisp, which McCarthy disagreed with philosophically, as McCarthy believed Lisp was the only practical language for implementing AI-like capabilities.
“in 1979, I kind of got into building my own kind of symbolic computation system, and it wasn't practical to use lisp at that time... So, I used this then new fangled language called C... John McCarthy kind of never forgave me for that.”
Wolfram has spent a significant portion of his life building technology tools and choosing 'the interesting thing' over 'the thing that makes more money' at numerous critical forks in the road, suggesting that personal fulfillment and intellectual interest are more motivating than financial gain for some personality types.
“I've lived a life where I like to do interesting things things I find you know fulfilling I've you know I'm practical enough that that activity has made me a very decent amount of money but it is not you know for me there have been a vast number of forks in the road where it's kind of like do the more interesting thing. Do the thing that makes more money. I'll always pick the more interesting thing”
Wolfram wrote a book-length explanation of how ChatGPT works and why it works, which took only a week to write but has become very popular and been translated into many languages, in contrast to other works he's spent years developing.
“I eventually wrote what turned into a little book about um what is it called? what what what is chatbt doing and why does it work which became very popular but it was kind of in a sense disappointing for me because it took me a week to write and I there are lots of other things I've written that took me you know years to write and uh the one that took me a week to write people people seem to like very much now translated into lots of languages”
Wolfram and team provided the computational backend for Siri, including voice recognition, and they observed that the voice-to-text system would often mistake 'pi' (the mathematical constant) for 'pie' (food) when processing math queries, a problem that was eventually solved.
“when we were when Siri came into the world... we were the computational backend for Siri... and we were constantly frustrated because it would send uh you know whenever people were asking about math and you know pi in math the voiceto text system was sending pie as the word that they were saying but that got solved.”
Wolfram was planning to use image processing hacks to implement functions like counting people in images, but his friend Sarah convinced him that neural nets would be able to do it, and the capability arrived within a year, making his planned project obsolete.
“I myself had been uh just about to we to use image processing to just sort of make little functions... my my friend Sarah had said no no no we're going to be able to do it with neural nets one day uh it actually arrived within the year”
The system Wolfram built is used by a few million people primarily for research and development, and an increasing number of practical systems in the world use the technology underneath, though consumers using those systems are often unaware of the underlying technology.
“a few million other people... mostly for research and development... there are an increasing number of kind of practical systems in the world that that use our tech underneath... it's always so strange when you've built a bunch of tech and then you are a consumer user of some big system in the world and you know that your tech is underneath it”
Wolfram wrote a small book in a week explaining what ChatGPT is doing and why it works, which became surprisingly popular and has been translated into multiple languages, despite Wolfram having written other pieces that took years but received less attention.
“I eventually wrote what turned into a little book about um what is it called? what what what is chatbt doing and why does it work which became very popular but it was kind of in a sense disappointing for me because it took me a week to write and I there are lots of other things I've written that took me you know years to write and uh the one that took me a week to write people people seem to like very much now translated into lots of languages”
Wolfram's team maintains a weekly leaderboard ranking LLMs, and the field is active enough that the top-ranking model changes nearly every week, indicating rapid progress and competition.
“We have a nice little website where we every week do a kind of a rating of of how how these various LLMs are doing. And uh it's kind of remarkable. It's an active enough field that practically every week there are changes at the top so to speak in terms of what the uh uh what the winning LLM is uh this particular week.”