This Week in Startups - Stephen Wolfram on AI’s rapid progress & the “Post-Knowledge Work Era” | E1711

Episode Date: March 31, 2023

Stephen Wolfram of Wolfram Research joins Jason for an all-encompassing conversation about AI, from the history of neural nets (7:53) to how modern ai emulates the human brain (19:33). This leads to an in-depth discussion about the pace at which AI is evolving (43:46), The “Post-Knowledge Work” era (58:45), the unintended consequences of AI (1:03:52), and so much more. (0:00) Nick kicks off the show (1:24) Under the hood of ChatGPT (7:53) What is a neural net?  (10:05) Cast.ai - Get a free cloud cost audit with a personal consultation at https://cast.ai/twist (11:33) Determining values and weights in a neural net (18:28) Vanta - Get $1000 off your SOC 2 at https://vanta.com/twist (19:33) Emulating the human brain (23:26) Defining computational irreducibility (26:14) Emergent behavior and the rules of language (31:49) Discovering logic + creating a computational language (38:10) Clumio - Start a free backup, or sign up for a demo at https://clumio.com/twist (39:38) Wolfram’s ChatGPT plugin (43:46) The rapid pace of AI  (58:45) The “Post-Knowledge Work” era (1:03:52) The unintended consequences of AI  (1:11:45) Rewarding innovation  (1:16:12) The possibility of AGI  (1:20:07) Creating a general-purpose robotic system FOLLOW Stephen: https://twitter.com/stephen_wolfram FOLLOW Jason: https://linktr.ee/calacanis Subscribe to our YouTube to watch all full episodes: https://www.youtube.com/channel/UCkkhmBWfS7pILYIk0izkc3A?sub_confirmation=1 FOUNDERS! Subscribe to the Founder University podcast: https://podcasts.apple.com/au/podcast/founder-university/id1648407190

Today, on This Week in Startups, Jason is joined by Stephen Wolfram of Wolfram Research. The two have an incredible conversation about AI, including Wolfram launching one of the first chat GPT plugins, the history of neural nets, how exactly chat GPT works, how this technology is going to shape jobs in the future, and so much more. Stick with us. This week in Startups is brought to you by Cast AI. If you run software in the cloud and it's been a significant cost driver, listen up. Cast AI automates cloud cost reduction with clients saving an average of over 60%.

Twist listeners can get a cloud cost audit with a personal consultation free of charge. Visit cast.AI slash twist to get started. Vanta. Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get a SOC2 report fast. Twist listeners can get $1,000 off for a limited time at Vanta.com. Clumio. When you're building a company, don't let backups and compliance requirements distract you. Let the data protection experts at Clumio help with immutable air gap backups that put compliance on autopilot. Visit them at Clumio.com slash twist to start a free backup or sign up for a demo.

All right. I'm really excited for our next guest today. Stephen Wolfram is here. It's a founder and CEO of Wolfram Research. You might have used. Wolfram Alpha, which I guess some people call a search engine, but it's obviously much more than that. And he's a prolific author. I really don't need to introduce him all that much. I guess a great place to start would be maybe to talk about what we've seen with chat GPT and how impressive is it to you watching 3, 3.5 and 4 come out over the past year, and then we'll get into the plugins and how Wolfram Alpha is sort of plug-ins. into it?

Well, you know, I've been paying attention to neural nets since about 1980. That was when I first programmed up a neural net, didn't do anything terribly interesting. So it's, you know, and then 2012 comes around and deep learning neural nets start doing interesting things. We started putting them into language and so on. And I've been sort of tracking large language models for a while and they didn't seem that exciting. And then chat GPT came out and suddenly it was exciting and it was able to do really useful

things. And I think we still don't completely understand what allowed that jump to occur. But I think we kind of get some idea now, now that that jump has occurred, we can go back and look at, you know, why does this work? What's really happening and so on. Yeah. So for a layperson, when you type a question into chat GPT or I guess Google's bard is out, We see Po from Cora.

So many different language models are being released. What is actually happening under the hood when we ask it, hey, I have some salmon and how should I prepare it? What are my options? What is it doing actually behind the scenes? Well, I mean, it's doing something incredibly mundane. It's very surprising that it can be as human-like in its output as it actually is. Because, you know, in the end, what it's doing is it's saying,

you've typed some text, I'm going to continue that text, the way that the statistics of text on the web and in other places that it's been trained from works. So it's kind of like if you're just doing with letters, if you typed a cue, then, you know, there's an overwhelming probability that you comes next in English at least. And it's got a much more elaborate version of that. And the thing that, you know, you might think, well, you just count, you know, if you've got some phrase, you just say, how many times does that? occur on the web and what's the typical next word when that occurs on the web, that in and of itself

doesn't work because there just isn't enough text on the web. There might be a trillion words that you could find, you know, between the web and books and things, but that's not enough to be able to give you sort of statistics on what's the next word after, you know, the best thing about AI is or something. There aren't enough occurrences of that that you can sort of statistically work it out. So you have to have a model. And the thing which is interesting, surprising, is that this particular model that's the idea of a neural net turns out to give you results that are very human-like, that, you know, when it has to work out sort of how will it extrapolate from just the pure statistics of what's on the web, it extrapolates in a way that

somehow similar to the way humans do it. And, you know, I think that in the end, that's because neural nets actually work very much the same way as the sort of wiring in our brains, works. And, you know, the history of this is, you know, back in the 1940s, people knew that, you know, brains had neurons and there were, you know, well, we know, though there are about 100 billion neurons in our brains. And they've all got, you know, these little electrical devices, basically, where each one is connected to maybe a thousand, 10,000, whatever other ones. It's a big, complicated mass of sort of wiring, neural wiring. And that was what people started doing was thinking about, well, what's the kind of formal representation of that?

What's the kind of mathematical way to represent that? That was invented in 1943. And at that time, and in the 1950s and 1960s, people were like, well, what does it do if you have five neurons, if you have 10 neurons, you know, if you have 30 connections between neurons and so on and didn't do anything, it did a few things that was somewhat interesting. We didn't do anything terribly exciting. It turns out when you have 100 billion neurons, 100 billion connections between

neurons, a few million neurons, that turns out that you can capture a lot more of what actual brains do. And it wasn't obvious what that number would be. It wasn't obvious how big the, you know, how much data you would have to train with, how big the number of neurons would have to be to get sort of human-like behavior. I mean, the thing that is the other critical point is there is enough text now available on the web that you can kind of figure out the statistics. You can train the neural nets kind of well enough from that text that it can produce things

which are a good match to what would be sort of the human-like way to continue that sentence, so to speak. That's kind of, it is sort of remarkable that these systems are basically just writing one word roughly at a time. And yet, just by the way the sort of statistics works out, the whole essay or whatever ends up being coherent. And so three things had to come together. One, the corpus that it was trained on and who knew it, but that wound up being the

World Wide Web and all of these different, you know, data sets could have been Reddit, Wikipedia, the obvious one. So the data sets had to grow to a certain size. Then we had to have enough compute and enough storage to process it fast enough. And then the language model had to be written or built by somebody. Those were the three components that have been actually to make this happen. You know, the language model, there are some clever ideas, but actually between 1940 and now, there were a lot more, you know, very clever ideas that didn't work out.

And what we actually have now, the structure of the neural nets, with a few extra pieces that are kind of important, but they are kind of, they seem minor relative to the things that were tried in the intervening years. The neural net is really close to what people imagine neural nuts would be like back in the 1940s. it turns out the simple thing kind of worked. Yeah. Explain to Williamen what is a neural net and how this comes up with these connections. And it is pretty amazing that it is exactly what we thought it was.

And we just had to wait for compute and corpus of data to training data to kind of reach critical mass, I guess, or some tipping point. So okay. So what is a nill nut? So in brains and in neural nets, there are. are neurons, and neurons have this feature that, well, I was putting in the case for brains, it's rather similar for artificial neural nets. When a neuron has all these so-called dendrites, all these incoming connections that are just pieces of the nerve cell, so to speak, and a nerve cell is basically an electrical device, and when the nerve cell kind of fires,

it produces an electrical pulse, which it sends out to its outgoing wires, so to speak. neural wires, so to speak. So what's happening is roughly when there are kind of, when, in the first approximation, in the original way this was set up, when there kind of are enough incoming wires that have signals on them, then the neuron says, okay, I'm going to fire, and then it produces a signal that gets sent out to the sort of next neurons that are connected to it. Now, the thing, this idea of weights, which is a big thing that people talk about, in neural nets, that has to do with the fact that if the incoming, the sort of the, if there are incoming signals on all these various wires, it's not just there's a signal

and every signal is treated the same, each of these incoming wires has a certain weight. It might be a positive number, might be a negative number. You know, it's like a weight of 0.72, a weight of minus 0.34. And roughly all those different weights get the, when there's a signal, you multiply by the weight, you add all those things up, then there's kind of a thresholding function, and that determines whether the neuron fires and sends data on to the next neurons

down the line. That's how it seems to work in brains, and that's pretty much how it works in artificial neural nets. Listen, if you run software on AWS, GCP, or Azure, you know how crazy the bills can get. The pricing and uncertainty can make you really anxious, right? You get that sticker shock. But there is a way, to lower your bills. And the best way to do that is

Cast AI. They audit and optimize your cloud cost and your performance. Major cloud providers don't do this. Why would they want your bill to be high? They don't want you looking at the bill. They want you paying the bill. Cast AI wants to discover what could be reduced in your cloud bill, right? So they're on your side. And they're going to eliminate the stuff that you pay for, but you don't use. That happens all the time. People spool stuff up. They forget to turn it off.

And they also search for less expensive hosting options within your cloud provider, right? And then you start saving immediately. these on average cast AI customers save over 60% on their cloud spend. So if you do this right now, you could be spending 60% less and just think of what that will do over the next 30 months and how much runway you're going to add to your startup. That's all capital going back into your business, right? That's going to your team members. It's going to your marketing.

It's going to acquiring customers. And that's going to get you closer to your next round of funding. Every dollar counts in this funding environment. You know that. So before you go and sign any multi-year cloud contracts or make any drastic personnel decisions, just stop for a moment and check out what Cast AI can do for you. you. They're going to give you a personal, free cloud cost audit. And you get a personal consultation. It's free. So why wouldn't you take it? Cast.a.i slash twist to get started. Visit cast.a.

i slash TWIST and get your free cloud cost audit today. Okay. So first question is, you've got this prompt. You wrote out the prompt. You're saying, you know, the best thing about AI is or something. That has to turn into a bunch of numbers that represent kind of the intensities of firing of a collection of neurons. And so there's a certain amount of, well, there's this whole idea of embeddings. These are ways to sort of turn words into numbers. And the idea is that if you have a sort of a good embedding, then words that are similar in meaning will correspond to collections of numbers that are nearby. So, you know, something like, I don't know, elephant and rhinoceros might have a sequence of, well, it might be, let's say, a thousand numbers

that the thousand numbers that represent elephant are fairly similar to the thousand numbers that represent rhinoceros, but they're completely different from the thousand numbers that represent, you know, Jupiter or something like that. And so the first thing is you've got to grind the words up, turn them into numbers. then those numbers are used to determine kind of the intensities of this first layer of neurons, and then you go through a sequence of layers. So for chat, GPD, I think it's a few hundred, maybe 400 layers. And so what's happening is the sort of the data from the thing that at the, you know,

the initial numbers are kind of, they go into the first layer of neurons, then they go through these weights, they will get multiplied, things fire, you go to the next layer, go to the next layer, and so on, and when you've gone through those, I think it's about 400 layers, you get to another collection of numbers. And that other collection of numbers then gives you essentially the probabilities for a set of possible words that might follow. And then you have to decide, well, which word are you going to pick? Are you going to pick the word that was most probable, according to the statistics of the web,

so to speak? You're going to pick the word that was second most probable or whatever. And there are many pieces of kind of slightly black magic that go into making one of these systems really work well. There's sort of the, if you always pick the most probable word, then at least for writing like English essays, that tends to be, it seems rather monotonous. Sometimes it just repeats itself, all kinds of bad things like that. But as soon as you pick sometimes the not top probability word and the sort of a parameter, the temperature parameter that determines kind of which, how far down. the ranking words you'll pick, that seems to lead to a more lively result. As you mentioned, one other thing that's sort of a critical piece of what's worked in something

like chat GPT is this idea of transformers. And so the question is, when you have the words that it's already written, what do you do with those words? How do you feed them into the neural mat? And the question is, the one thing you kind of know about those words is they're in a sequence. They're not just, oh, there's different words in different places. And so what happens is the neural net kind of learns, it knows, given that we're going to add the next word, it says, well, the word three back has this level of importance, the one five back has this

level of importance and so on. And then it combines in a very kind of sort of bizarre way. It combines multiple different sort of patterns of how it pays attention to previous words and does the whole thing multiple times. And out of all of this comes the results from once. Now, one question is, okay, so that's sort of the setup of how, given that you are feeding in a prompt, you're feeding in text, how it will determine what text to write next. Next question is, well, you've got this whole neural net and it's got all these weights. And in chat GBT, it right now has 175 billion weights.

How do you determine those weights? You know, any collection of weights will have the property that you can feed words in, and some words will come out. The problem is if those weights are picked at random, the words that come out will just be complete nonsense. The question is, how do we pick weights so that the thing kind of conforms to the statistics of the web? And so that's this process of neural net training. and essentially what you do is you kind of you say, well, here's some text from the web,

and we know what the next word is, but the neural net doesn't know what the next word is. So have the neural net sort of guess what the next word is, and then it might get it right, it might get it wrong, but typically it will start off getting it wrong, and then you say, okay, how would you have to change the weights in the neural net to make the word that comes out be closer to write than the one that actually came out. And so you iteratively do this. That's the training process. It's just kind of tweaking all those weights.

And there's a mechanism called back propagation that kind of helps you make it not be an absurdly mathematically difficult problem to figure out how to tweak the weights so that you'll actually get the thing that. So you're kind of training it on, you've got a piece of text. you're kind of masking out the words at the end of the text. You're trying to training it so that the weights are such that the words that are at the end of the text will be the ones that when you took the mask off will really be the ones that were there. So you find some high quality piece of text. Here is the Wikipedia page.

Let's assume it's high quality and it's been vetted on China. And it just starts reading it. It gets the word wrong. You don't punish it, but you tell it, hey, you got it wrong. And then until it gets the words right, it gets a cookie or it gets – punished in some way. Well, the trick is that you're kind of, it's like an evolution process, like biological evolution or something. You're kind of gradually adapting it to get closer and closer to the right answer. And there's a systematic way to do that, and that's what the training

process ends up being. And, you know, the fact is it's trained on a trillion words. So it's trained on, you know, this whole process, if you only trained it on a million words, well, it would be able to learn some things like, you know, you follow Q and things like that, but it wouldn't learn the things that make it seem like a meaningful, you know, essay or something of this kind. That seems to require, you know, an amount of text that is, you know, reasonably, that's about what we humans have produced and put out in kind of publicly accessible form. and it's also the number of weights that you need is sort of roughly comparable to the kind of

number of words that you read in your training set. Nobody really quite knows why that is, but that's another sort of a random fact, so to speak. Listen, it's 2023. The macro picture is a little shaky. It's uneasy out there. And tech is getting hit super hard. As such, you cannot afford to lose sales for silly stuff, like not having your sock too right now. If you are unsure about your SOC2, you need to check out Vanta.

Vanta makes it incredibly easy to get and renew your SOC2. On average Vanta customers are SOC2 compliant in just two to four weeks. Compare that to three to five months without Vanta. And they partnered with over two dozen audit firms who have been trained to file SOC2 reports directly within Vanta. This is a total no-brainer. A bunch of my portfolio founders have used Vanta and they've had amazing experiences. And if you don't have Soctu Compliance, you can't close major customers. One major customer, that can be the difference between your startup thriving or going away.

So get it done right now. That is going to give you $1,000 off because you listen to this podcast. Think about it. $1,000 off Vanta.com slash twist. You got to write that down. Put it in your notes. V-A-N-T-A-com slash twist for $1,000 off your sock tube. When we compare it to what's happening in a human's brand,

and I know we actually don't have the answer to this yet. We don't understand what consciousness is. Exactly. We have theories and ideas. But when the human is asked, hey, what are the most popular desserts in America? And when the neural net is asked and chat GPT is asked, whichever version of it is asked, hey, what are the most popular desserts in America? When you look at what happens in a human and then we look at what we know is happening in the software, how close are they? do we actually think we're emulating what happens in a human brain or that we've developed a new process that is similar but maybe not exactly.

And what is that overlap? If there are two circles of consciousness and answering questions versus a computer answering it, how much do they actually overlap? Well, I mean, I think what's happening in LLMs is fairly close to what happens in brains. There are some things missing in current LLMs. You know, brains have this in a computer, there's typically, you know, there's CPU, GPs. It's processing a lot of data. And then there's memory.

Most of the time, the memory of a computer sits doing nothing. Just sits there storing what it's storing. In human brains, every neuron both computes and stores things. So we have a little bit of an advantage, at least, for right now in that regard. Also, the way that something like chat GPT works, it just feeds forward. You know, you feed it in the prompt, and it'll kind of ripple through the neural net. It'll say, okay, the next answer is this. in our brains, we're pretty sure that there's some kind of feedback loop.

And maybe the feedback loop is similar to the one that chat GPT effectively has, or an LLM, effectively has where it sees the prompt it's got so far. It adds a word, that becomes the new prompt, and it can kind of feedback that way. So when we ask this question, hey, what are the most popular desserts in America? And the first thing that comes to mind is ice cream. And then ice cream triggers me to think, well, apple pie, of course. and then apple pie and ice cream trigger, whatever the next thing is.

Yeah, maybe. Yeah. Look, I think that the thing to understand is, when it comes to sort of a computational process like how brains work, it's, there's a lot of detail that in the end doesn't matter. I mean, it's just like saying, if you want to fly, do you need feathers, do you need flap, you know, or do you just need wings? Turns out you need wings.

pretty much, unless you're a drone with rotors, but there's the little wings that happen to go around. But, you know, the details of feathers and so on don't turn out to matter. And so, you know, in the case of brains, our brains have a lot of detailed sort of stuff in them that's, you know, things like the, you know, the glucose that's supplying energy to the neurons and things like this, which is obviously different from the electronic case. But at a sort of computational architecture level, I think it's surprisingly close. And what's sort of remarkable is people have kind of known roughly what this is like for, you know, what is it, you know, 80 years or something.

I mean, this is, that part is sort of unsurprising. Now, I would say there's, well, there's a lot more to say about how a computer works as compared to how a neural net, the sort of brain-like neural net works. The thing about computers and computation in general is it can sort of, it kind of goes much deeper than what a neural net can do. Because what happens is you have some computer, it can, for example, it can go in a very sort of tight loop figuring out what's what's the result of a computation. None of that stuff is happening in something like, you know, an LLM like neural net. It's just rippling through saying what's the next word and, you know, it just ripples through and so on. And there's this kind of concept that's sort of a concept that I kind of invented in the 1980s called computational irreducibility.

It's kind of the feature of sort of deep computation. Because you say, okay, what is the computation? You have certain rules and you're going to just apply these rules over and over again and you see what the results are. Those rules might be the sort of the way the CPU or the computer is set up. They might just be some rules about black and white squares or whatever else. But the way it works is sort of the essence of the computation is you just keep applying these rules over and over again. You see what comes out. So the question then is you're applying these rules.

There's a certain number of times you have to apply the rules to get a certain result. The question is, can you jump ahead and see what the result will be more quickly than just following all those rules? And what turns out to be the case is there are many situations where you can't jump ahead. It means you have to do this computation. If you want to get the result, you have to actually go through the steps of the computation. And when you have computational reducibility, basically the neural net is, it's too shallow to be able to deal with that. It can go a certain distance.

Like if you ask chat GPT right now, you know, match parentheses. You've got open, open, open, open, open, close, close, close, close, whatever. You're trying to make it just make sure that the number of closed per ends, you know, that close perands match the open parameters. It can do it up to a certain point, and then it sort of says, well, it doesn't say this as such. Actually, I haven't asked it, but maybe I should. Why it fails. It might have something interesting to say.

But basically, it's just sort of run out of layers of neural methods. It just can't represent that deeper computation. And so there's this sort of in, there's this world of computation which include irreducible computations, that you just can't shortcut. You just have to do the computational work. And then there are these shallower things that are what we humans are using most of the time when we're generating language, probably a lot of the thinking that we do works that way. And so there's sort of a difference between how you can do things in principle with computers

and how things work in something like an LLM. And you might say, do you care about irreducible computations? Well, the answer is, for example, in nature, many things that go on in sort of the physical world, if you want to work out how they work, you kind of have to do irreducible computations. Those things weren't sort of made for humans, so to speak. I mean, our language and things like that is sort of made for humans in some sense. But nature just is what it is, and it can, to work out what it's going to do can involve these irreducible computations. And then it's up to us to try to simulate them in some way or to try to figure them out here

in this, yeah, let's talk about emergent behavior. like are we projecting into it that it's learning in some way or evolving in some way at this point or is it truly with so many people using it now and the reinforcement learning that's happening and then all these plug-ins putting in do we get the sense that the model is learning at some faster pace now and that this concept of hey maybe we're not in control of it do you believe that that's kind of the moment we're in right now because as you were saying earlier like This thing is kind of surprising us right now. So I'm kind of wondering what the next surprise will be.

Let's pull that apart a little bit. So first of all, sort of emergent behavior. One of the things that that typically means is you put certain rules for how a system works and what the system does is much more complicated than the rules that you put in. And that's what happens in irreducible computations all the time. That's kind of the thing that probably makes nature seem so complex to us as it's full of these irreducible computations, where the rules are quite simple, where the actual behavior seems to us very complicated.

Now, the question of what LLMs are doing, and to what extent that's sort of an emergent thing, first of all, just to clarify one thing. So in the present state of things, the actual little chat sessions that people are, are having with these LLMs, yes, they're being stored, they will be used for training, but it's not an immediate loop. That's not something that's been done technologically as an immediate thing. It's more of a sort of a long-term process. So it's not like, you know, every person who types into it is getting smarter and it's going to, you know, it's going to take over the world as a result. My chat with it is independent of your chat with

it. It has threaded chats together. So it's learning, it's applying the model in each of our individual threads, but if we both started asking it about desserts, it wouldn't suddenly be like, oh, wow, two people on two different coasts in the United States are talking about dessert, and let's pull that all into our knowledge. But that is coming, obviously. Yeah, yeah, but that just doesn't happen to be here yet. I mean, that's just a technological, that's a technological privacy, you know, policies, et cetera, et cetera, et cetera, kind of issue, but that's, I think that's an interesting one, actually. I just, I've never heard anybody actually have this conversation, but what is the ethical,

right thing to do if a hundred people right now are talking about, you know, this new pandemic that they're seeing and it's trying to put together that information to maybe warn us, a pandemic is actually happening. There's 100 people talking about it in this region and they're spaced out at this distance and this is the qualifier of a pandemic starting. Right. I mean, this is something obviously one's already seen, you know, from things trending on various, social media and, you know, search queries and things like that, that's already a thing of it like that. I mean, the question of, you know, how private is your particular chat session and your

particular, you know, I don't know, psychological counseling session with the chat part or whatever else. Yeah, who owns that? Right. Well, versus what, I mean, there's the same thing that happens in, you know, medical stuff all the time, which is, you know, to know an aggregate what the results, what happens medically to lots of people is a huge societal value. Yet, you want to keep the individual records of individual people private. So there's kind of, you know, you want the aggregate to be something

that can be mined, but you don't want the individual things to be separately minable, and that's a whole, whole technical can of worms about how you can do that and to what extent you can do that and so on. I think the same thing will probably happen here. Coming back to this question of sort of what, why does the LLM work? What is it really doing? In what sense is it emergent? What's going on? I think the thing that is probably for me, sort of the biggest kind of aha feature of what we've seen with Chat GPD is the fact that probably language is not as complicated as we thought it was. I mean, language is kind of the pinnacle of our species as sort of collective achievement in some sense.

And so we think it's a very sophisticated, complicated thing. But we already know there are certain rules of language. Like we know, you know, syntactic grammar. We know, you know, a typical sentence has a noun, a verb, you know, a noun, might be an adjective and a noun, things like that. We know this kind of, these kind of structural regularities to language. Well, I think what's happened is that in these LLMs, what's been discovered is that there are many more regularities in language than we had sort of classified before. So there many, it's kind of like language we know from sort of the structure of sentences about nouns and verbs and so on. We know there's sort of a construction kit, the sort of puzzle pieces that you can put together.

You can't go, you know, verb, verb, verb. That's not a possible sentence. You know, it's got to be a noun verb, noun type thing or something like that. So we know that there are these sort of puzzle pieces you put together. And I think what's happened is that there are sort of what's been discovered by LLMs, in effect, is that there are a whole collection of other puzzle pieces that don't just deal with parts of speech, but they deal with little fragments of meaning and language. And there are things that can be put together meaningfully and there are ones that can't be put together meaningfully. And, you know, we have one example historically of where this kind of thing was discovered,

and was discovered 2,000 years ago, which is the idea of logic, which was presumably discovered by Aristotle. And, you know, in a sense, Aristotle was doing a humanized version of sort of machine learning, because what he did, presumably, is he took all these arguments that people made, all these pieces of rhetoric and so on, and he said, what's the pattern of how arguments work? You know, if you say, all men are mortal, Socrates is a man, therefore Socrates is mortal. That's a certain pattern. You don't have to be talking about Socrates, you don't have to be talking about mortality,

you could substitute in any kind of thing there, but that structure is a meaningful structure that you can put into something that you say. And he lifted from that this kind of idea of logic of, you know, and oars and knots, and, you know, this implies that and so on. And that becomes one of these kind of semantic regularities of language. That's one that we know. There are a bunch of others, I think. And the LLM has basically found them. And we've been a bit negligent in the last couple of thousand years not looking for these things.

I mean, there was a little burst of interest in the 1600s, but it kind of died off. And then people had, I think people kind of thought it was too hard, and they were a little bit proud of the fact in the 1950s, it became clear that this kind of grammatical structure of nouns and verbs and things, how that worked in lots of different languages and people all kind of excited about the way that it had been figured out that that worked. And so they didn't really look for these other things in a serious way. I think so that's kind of the, I think that's what's sort of a science fact that was discovered. And in the end, once you know that that's the science fact, it's sort of puzzle pieces being

fit together, it all seems a bit less miraculous, so to speak. And so we're figuring out, or the models figured things about language that we just maybe haven't been looking for. And we, as humans, with language as the pinnacle of our existence, whether it's poetry or science or, you know, any number of arts or debates, it's kind of how we mitigate the entire world. It's how we make decisions, these debates that occur, presidential debates, Congress, Senate, at your dinner table, who are you going to vote for, how are you going to raise your kids? We maybe have valued this as something super magical, but with the corpus being actually kept somewhere, the internet,

and then the ability to process it so quickly with these new GPUs, you may have just figured out, hey, this actually isn't all that complicated. Right, but I think what we learn is that sort of the essence of meaning is something that is, which is the thing that we represent with language, as sort of a calculus in a sense, a formal structure of how meaning works. Now, the fact is some aspects of that, well, somebody like me or perhaps me in particular, you know, my lifelong project basically has been sort of figuring out how to make things computational. And one of the things that, you know, is my kind of long-term project is to make a computational language,

a language that can represent things in the world in a sort of precise formal computational way. And that's what the thing we call Wolfram Language is, it's kind of started off as Mathematica and kind of evolved into Wolfram Language over the last 35 years. But the kind of the idea there is to take things in the world like, I don't know, two cities and you're asking, you know, what's the distance between them or these kinds of things and have a precise kind of formal representation of those things, that is sort of both writable by humans, readable by humans, readable by computers, executable by computers.

And the fact that, I mean, it's been my kind of last 40 years, basically, and I've spent building up this kind of language to represent things computationally. And in a sense that the language represents a lot of kinds of things that are very useful to talk about in the world. It doesn't happen to represent kind of everyday chit-chat type conversation. But, you know, and chat chvety has sort of added that as another element of something that we can see how it fits together with language. But, but. You're going to say, no, you were going to say.

No, I mean, the thing that, you know, people ask, for example, does chat Chb-T understand what it's talking about? well, it just has these rules that say how the next word goes in. It doesn't. You could, I mean, that's how we work too, probably. And you can ask, do we understand what we're talking about, so to speak? And there isn't, but it is in a sense doing a very shallow computation. Kind of the idea of computational language is once you have something represented in computational language, you can kind of go all the way, you can compute whatever you want with it.

You can do irreducible computations, you can do all sorts of things. And so, you know, the thing we did a dozen years ago with Wolfmalfa was natural language understanding where you go from small fragments of human language to computational language. And once you can do that, that's a sense in which you have true understanding. You've got natural language, you turn it into computational language. Once it's computational language, you can compute anything you want from it. So that, in a sense, is true computational understanding. understanding, so to speak. And that's a different thing from what sort of a raw LLM feels with.

And that's, by the way, what the plug-in that we just worked on with OpenAI, the Wolfram plugin for ChatGPT, that's what it's achieving, is being able to connect this kind of LLM layer to this sort of what I think of as kind of computational bedrock of what one can compute from. and there's all sorts of implications. Did you know that today is a major holiday in the tech world? That's right. It's World Backup Day, March 31st. So we have a few reminders and tips from the folks at Clumio.

First, make sure your data is protected. Obviously, remember, because your data is in the cloud doesn't mean it's automatically backed up. And your data is your responsibility, right? Think about that. It's not your cloud providers. It's yours. Second, don't let backups and compliance requirements distract me. Your engineer should be focused on product innovation, not compliance audits.

Third, take control of your cloud costs. You'd be surprised how much of your storage comes from backup snapshots and replication that you're doing with your expensive cloud provider. Clumio will help you with all three of these points with a turnkey, data protection, service that's air-gapped, immutable, and cost-optimized. They have saved customers. Listen to this, over 30% on backup costs while putting security and compliance on autopilot. So you're going to, you know, save money and you're going to save time.

What's better than that? I love these kind of services. Whenever I'm looking at a business to invest in, I'm like, does it save people time? Does it save them money? And does it make them laugh or entertain them? I'm not sure Clumio is going to entertain you, but it's going to save you time and money. That's two of the three major business models in the world. Two of the three great value propositions for consumers.

Save them time, save them money. Clumio. All you got to do right now is go to CLUM-I-O-com slash twist. Just start a free backup or sign up for a demo. that's clumio.com slash twist, CLU-M-I-O-D-com slash twist. Write it down now. If you were asking,

hey, what's the distance between these two cities or what are the similarities of this elephant and rhinoceros? Oh, both mammals, both gray, skin color, whatever, both formidable, whatever the words are that are coming up.

Chat Chipt-T actually seems to be getting things wrong if you ask it for numbers or equations doesn't seem to get it right very often. And so is the idea here, chat GPT can start discussing and maybe summarizing what's the difference between these two things or the distance between these two cities or the difference between these two cities, but then Wolfram could actually give the correct answer computation. You know, we didn't know how well this would work, but it actually works rather well.

I mean, we have very conveniently, we have, I mean, in the technicalities of the plugin, it has two different endpoints inside it. One of them is going to Wolfmalfa. Wolfram Alpha takes natural language input, takes small, fragments of natural language, and Wolfram Language is a precise computational language. And sometimes what ChatGBT is doing is taking this big lump of text that somebody might have given as a prompt or the thing it's trying to write. And it does surprisingly well at crispinning that up to the point where it's either a small fragment of natural language

that can be sent to Wolfram Alpha, or it's a piece of Wolfram Language code that can be sent to the Wolfram Language interpreter, a tricky thing that happens is, in both those cases, particularly the Wolfram Language case, sometimes it gets it roughly right, but it isn't exactly right. But then we actually run the code, and we can see what happens. And then we tell ChatGBT, BT, well, it didn't quite work. Why don't you rewrite it?

And it does. And so it goes through several. Oh, fascinating. So just a simple thing, like, what's the distance? If you ask Wolfram Alpha, everybody probably knows this who's listening, you asked at the distance between Los Angeles and London, it's going to give you a really broken down tight answer. But if somebody were asking that in a less precise way,

Wolfram, the plug-in could then mitigate kind of the wordiness or how it's buried. If you wrote a very poetic description of what you wanted, you know, Wolframel was built for people who kind of have a question to ask. They write it in a natural language way, but they're kind of direct. They just ask the question. They don't say, you know, I'm having a whole thought about, you know, going from here to there on an elephant and I'm wondering, you know, how many, you know, steps of the elephant have to take and this and that and the other. What ChachyPT does pretty well is to boil that down into something which turns into distance between this and that divided by stride length of an elephant. Wow.

That type of thing. I haven't tried that particular thing. I'm not sure that that particular thing with elephants, I'm not sure about. But, you know, the other part, definitely. But so, you know, the other things that can be done, once you're computing, you can do things like have ChatGPT produce, you know, call the Wolf and Plugin, generate graphics, you know, do we have lots of real-time feeds of data. So do a histogram, do a chart, whatever it happens to be. Right, right. And then it could be from the weather in some particular place where it's the current weather and, you know,

or the current stock prices or whatever else. So it's able to, and it has a sort of precise computational way to kind of figure out what to say about the world, so to speak. You know, what the LLM does very well is to take this complicated mass of natural language, boil it down into something that becomes a sort of a precise thing, and then it takes back the results. and sometimes it'll just generate a picture that comes straight from us, but sometimes it'll knit back the results into the essay that it's writing.

There's another workflow that's really quite interesting right now, which is, you know, we only learned this workflow in last two weeks, so it's very fresh kind of thing. The pace is crazy right now, right? I mean, it's amazing how when everybody in the world becomes enamored by something and says, oh, let me try to break it, let me try to fix it, let me try to, you know, stress test it. It's really incredible what the hive mind,

of just consumers and scientists, developers, and everybody in between trying to break this thing or jailbreak. Yeah, well, I think the big thing right now is to, I think one big thing is just understand workflows, understand use cases. Yeah. And understand kind of how to think about what to do. So like, for example, you might, you know, here's a thing that people have discovered. You can say it gives some results and you can say, you know, do you think that answer is right? and it turns out then it will that question turns out it's better at answering that question probably than it generating the answer in the first place so nobody knew that was going to be the case

the other thing that's just totally bizarre is the the whole business of prompt engineering of being able to say you know things like if you look at the prompt for the wolf and plug-in it's just, you know, we've been steadily adapting it, but it's full of, you know, we put pleas into the sentence and that makes a difference. We put, you know, don't do this, do that, don't do this. Here are examples of what you should do. The fact that any of this stuff works is really remarkable. And I think we're, you know, the sort of theoretical description of how the neural network works,

we're pretty far away from being able to say, given that theoretical description, this is how you should put commas into your prompt to whatever. That's a big distance at this point. At this point, prompt engineering is kind of a bit like animal wrangling, I think. It's kind of like you don't really know is this animal and it's flapping around. And it turns out if you pull on its ear, it will do this. And we don't really know. If you're trying to get this Mustang and you're trying to tame it, yeah, be careful.

But yeah, walking up to it quietly and like we're taking steps and just trying to get the Mustang and corral it and get it to put a saddle on. Maybe it's going to work, maybe it's not. It's pretty amazing. I always think, and then I'd love to get your thoughts, since you've, you know, basically helped create this category here on the impact on society, humans, and how quickly that happens. Because this feels qualitatively different, the pace that this is happening, then, I don't know, automation of software.

There was this concept, oh, my God, you know, TV comes out. Everybody's going to get a PhD because you can just turn your TV on. You have all this free time. You just turn it on. You're going to learn. or oh, Wikipedia came out, the internet's out, and oh, MIT's putting every course online, Coursera, you know, this one, Star-X, everything. Okay, everybody's going to be able to go to MIT or Harvard.

It turns out, well, human motivation is such that maybe everybody doesn't want to take the time to take all these courses that are freely available on YouTube today, which is just mind-blowing for a Gen Xer who thought, wow, whatever they're teaching at MIT and Harvard, that's locked up at the Cyber Tower. Now it's literally available for free, and it has 300 views on you. YouTube right now instead of $3 billion. So what happens now in society, realistically, when a whole swath of things that people are getting paid for, copyrighting is one example, journalism is another example, certain aspects

of journalism, research, and then design. I was reading a Reddit thread recently. Somebody said they went from spending three or four weeks to make a character in a video game, and now it takes them two or three days, but they kind of feel bad about it because it's not as artistic, but they're going to be able to make characters in games, you know, already 10% of the work effort, which means like you're not going to need as many designers, logos, whatever it happens to be. Is this, does this concern you or are you in the camp that humans always find more work to do? Because this seems to be moving at a faster pace than anything

we've ever seen. You know, I actually just sort of was curious, so I kind of studied what happened to jobs in the U.S. over the last 150 years. And, you know, things happen that are fairly dramatic and, you know, in technology, but actually it takes a generation before it fully works its way through the system and you fully see the effects. But I think here, the things that are happening is there's a lot of kind of cases where there's sort of semi-boiler-plate text that people generate. or there's text, so semi-boilerplate that people have to understand.

There's a bunch of people who do that. And this is a, you know, this is a really good way to do it. So how will it actually work? I mean, so let's say you are filing some, you know, you're writing some proposal, you're filing some, you know, compliance type thing or whatever else. You have certain points that you know you have to make. But dressing that in a whole giant essay is something that you used to have to do. used to take a lot of human effort to do that.

You just say, here are the main points I want to make. Go make an essay out of this, using sort of background foundational facts. And it'll make an essay. And then goes over to the other, you know, whoever's going to read it. And, well, they might actually read it as a human or they might feed it to their LLM, which will sort of grind it down. And they will have given their LLM a prompt that says, look for these kinds of things. And so they will extract the information that they want from that, which again might turn into, you know,

three bullet points. And so it's kind of like, so what this is, it's like an interface. It's like we had, you know, graphical user interfaces. I just started calling these louis, linguistic user interfaces. Yeah, I like it.

Louis. It works. Because, I mean, it's kind of like the, you know, it is a convenient transport medium. You know, an essay is a convenient transport medium for information, particularly when the two sides aren't really quite aligned. I mean, it's like fill out this form, check this box, check that box. then you can easily sort of transfer it from one side to the other.

But when, you know, each side doesn't really quite know what the other side is looking for, this is a convenient way to transfer information. Now, that means there are people and professions that have been, that are quite kind of knowledge worker type professions, but people have assumed are like, oh, nobody's going to automate these knowledge worker type professions. Yeah, it's not possible, right? Right. But that turns out human judgment.

Yeah. Right. Turns out that that's not true. And it turns out, and so, you know, and one of those areas is, well, for example, one area is programming, where, you know, I have to say, if people had paid attention to stuff we've been doing for the last 40 years, they wouldn't be in this particular pickle. Because, you know, the whole idea of the computational language that we've been building is that all of that boilerplate stuff that exists in lower level programming languages, we already automated that. You know, when you say, you know, geo distance or something between two cities, we've already automated all that stuff about, you know, pulling lat long from databases and figuring out, you know, the, you know, spherical geometry of blah, blah, blah, blah, blah, blah, all that stuff,

which, you know, you write it in, I don't know, Java, Python, whatever else, it's a big slab of code. Or maybe you pull it out of some library here that doesn't, you know, interface with some library there. This has been kind of the low-level kind of manual labor-type programming that, now, it's not to say that there aren't millions of people who use our computational language. So this is, and none of this applies to them because they already know how to kind of do things at this higher level. But there's an awful lot of programming that has been done using programming languages. And one thing to make clear is that, you know, what is a programming language?

It's a way of kind of letting a human tell a computer in the computer's terms what the computer should do. You know, the computer has a certain memory, you can make an array, you can have variables, you can do this. But those are things that are sort of in the computer's terms. Kind of the whole idea that, you know, I've pursued for the last 40 years or so, is to have a language which is kind of a bridge between how we humans think about things and what can be done computationally. so that we're kind of representing things at a human level rather than at the level that happens to be convenient for the computer. There's a lot more work for the people who build the language to do that, but that's why I just pent the last doing.

When we look at it, is this going to be a slow change? I remember when I got my first loft in New York, and it was the 90s, and they had manual elevator operators. They would take it to your floor in this old building. And I remember the, you know, 10 years later, they got rid of them and they put in automated elevators. Elevator operators as a concept took 50 years to kind of deprecate over time. I think there's like a couple left in America.

The Hotel Del Coronado in San Diego famously kept their old elevator and their elevator operator because it's charming or whatever. Ice cutters. I think I've been in that hotel. Yeah, maybe I even know. There's like an old guy who's in there. It's the one from Some Like It Hot, the famous film. And it's quite charming.

But ice cutters, refrigerators, switchboard operators, you know, operators generally lamp blighters, all this stuff has gone away. But it took time. So when we look at this, does this feel like programmers are going to become 10 times better? And yeah, we'll just get more accomplished in the world. Or does it feel like this is going to wipe out swats of jobs really fast? And then what do you think that does societally? I think some things will go fairly quickly in this particular case.

Not only because the technology exists to do it, but also because the sort of societal attitudes and sort of, oh, this is going away, so we'll make it go away even quicker because we can kind of already see the future. My guess is that some things will happen reasonably quickly. But, you know, it's always the case that things, I don't know, in my life I've had the good or bad fortune or something to invent a bunch of things that end up being many, many decades ahead of the current time, so to speak. And so then it's maddeningly slow, how quickly, you know, maddling this slowly, things actually get absorbed. I think this one, because of the kind of momentum that exists right now, I think some of it

will go quite quickly. Now, you know, what does that mean? You look at the pattern of what's happened in all previous cases. Let's say telephone switchboard operators. You know, the fact that telephone switchboard operators existed was a consequence of the fact that telephones existed, which was a technological advance. But then automated switching came in and you didn't need a manual telephone switchboard operator. But what did automated switching do? Well, it enabled basically the telecommunications industry, and that has generated just an immense range of jobs. I think one of the things you see seems to be the case is that, you know, look at America, you know, in around even 1900 was still 1850.

It was more than half agricultural work. Yeah. And, you know, the pie chart of what people did was very, you know, there was a big wedge of agriculture. and then a few other wedges, and they were all quite big. If you look at, you know, today, it's much more sliced up. You know, the pie is in much smaller pieces. And I think that's a thing that one can expect to see as sort of more automation happens, more things become possible.

There are more niches that people can fill, so to speak. And I kind of think that what tends to happen is when one of these sort of steps of automation happens, it enables things that, and it doesn't. then enables more diversity in what people can do. It isn't because people aren't all just pushing, you know, pushing the plow or whatever for agriculture. It's like, okay, now we've got that done. So now let's look at what's possible.

And I think the thing to realize about sort of the, you know, kind of the interplay between, you know, AI automation humans, you know, you've got a raw AI. It does. It's normal. That thing or whatever else. But if you say to the AI, you know, what is your goal in an engagement? existence, so to speak. It has no intrinsic answer to that question. We humans think we have an intrinsic answer. Where does that answer come from? It comes from the whole sort of web of history.

It comes from our biology, et cetera, et cetera, et cetera. But we are pretty convinced that we have, you know, we have definite goals. We want to do this. We want to do that. Those goals tend to be things that are intrinsically coming from humans. How the goals get achieved, that's where the AIs and automation and so on come in. So, you know, you're, and what I think you see happening is that when there's a big sort of enablement of things, what becomes important is what can you do with that enablement? I mean, we were talking before about kind of use cases for LLMs. It's like, okay, now we have LLMs, now we've got to figure out which use cases do we care

about. And that's sort of an intrinsically human activity because there might be lots of, you know, an LLM could just go spinning, you know, random, words out and so on, and it might, in some weird sort of anthropomorphizing of the thing, it might have a very happy time just spinning random words out. Humans look at it and say, what the heck is that? We don't care about that. Yeah, because you need a jockey.

You need a pilot, right? Right. I mean, you kind of, yes, you need to kind of define what the direction, what the objective is. So I kind of think that that's, I'm sort of, you know, what you see over and over again is something gets automated, that enables a lot of other opportunities, sometimes, and, you know, that's been the pattern. Now, you know, it's kind of like the question of, well, will that come to an end? Kind of like, will everything that could be invented eventually have been invented?

Well, we actually know from sort of theoretical science considerations actually related to computational irreducibility, we know that in sort of a formal sense, it will never be the case that there's no more to be invented. There'll always be unexpected things that you can figure out that you can invent. So in principle, there's no limit to what could be invented. The question is, it could be the case that we humans will say, hey, we're done now. You know, everything that we care about, right. Everything we care about, it's been invented.

Right. You know, we're good from here on out. Actually, that wouldn't work because it turns out the world, the natural world and so on, will continually throw up unexpected things that we'll have to respond to. So we won't be able to get into that kind of, oh, we're done now. But in the situation where we could say we're done now, then yes, it could be the case that everything that we care about has been automated by AI's other forms of automation and so on.

And sort of then we could be in, oh, there's nothing for the humans to do anymore. But, you know, I think that for both theoretical science reasons and practical reasons, I don't think that's what's going to happen. Yeah. When we look at this paradigm shift that occurred, we had agriculture, factories, knowledge work. Now, knowledge work seems like it's going to be automated. So, you know, we put robots into factories.

We put robots and automation into the fields, so agriculture and factories. You know, we don't need as many humans involved in those things. And knowledge work, we probably won't need as many humans involved in it. So then what, if this is a true paradigm shift, what is the post-knowledge work era going to be? Is it going to be prompt engineering? What is it?

What do we call this new era where anybody can talk to a chat interface and create a product or service in the world that maybe accomplishes or solve some really important or pressing a problem? Right. Look, I think that it really reflects on, you know, what we humans do and are on a special about doing. And that might be, it might be thinking. You know, one of the things about knowledge work is it turns out and, you know, education sort of directs. people this way. There's procedures for doing lots of kinds of knowledge work. Yes, it requires sort of analytical steps. But if you say big picture, think about stuff, that's not what the typical knowledge worker is trained to do. And, you know, I think that's a great sort of intrinsic

human thing is just globally think about stuff. And that's something where I think the value of that is going to go way up. I think the value of the kind of super specialized siloed knowledge is going to go down because that stuff, you know, you can drill pretty deep into a silo using automation. If you know the kind of the overall way to think getting deep into that silo is something that is now much easier than it once was. So I, you know, I kind of tend to think that the, you know, other things that are kind of, oh, I don't know, whether it's other, in a sense, more creative, more kind of things that are in a sense more arbitrary, more human chosen, like thinking we could go in this direction

rather than that direction, we could come up with this, you know, cool, you know, sort of routine or whatever that's, you know, that entertains people or whatever. These are things which are sort of much more arbitrary. They're not things where we say, you know, here's the end point, now just go fill in that endpoint in the best way. And I think, you know, quite a bit of knowledge work has ended up being something that is kind of, we know what the end result is, more or less, we know where we're going, now just fill in the details, so to speak.

And I think that- Like a journalism job or a legal job, it's kind of wrote. It's like, okay, who what, when, where, why? Okay, talk to a couple of people. You got one side of the story, see if you can get the other side, hit publish. Okay, lawyer, what do you want this agreement to say?

What do you want to happen if people break the agreement? Okay, we're done. And what you're proposing is maybe this next era, would be the creative era of humanity. The artisanal, I don't know, maybe it's the judgment phase. I'm trying to come up with the right word, but it seems like an era where human judgment and creativity is the driving force,

not the rote knowledge work. I think that's a good possibility. I mean, I think that the, you know, I tend to be, I suppose, generically an optimist, and I kind of look at the pie chart, getting more and more fragmented, and I think about all sorts of different. people who have all sorts of different skills.

And I think about the fact that, you know, for example, in my own case, right, I've spent my time doing science and computation and some technology and so on and companies and things like that. And, you know, if I'd lived at a different time in history, the things that I've really had a good time doing just wouldn't have been available to do. And, you know, that wasn't, that wasn't part of the pie chart. Back in 1850, you know, computation and science around computation wasn't part of the pie chart. things you could do. And so I think, you know, in my kind of optimistic view of things, it's kind of that, you know, there's more pieces of the pie, there's more different

things that can be done, and there's more, you know, for different people who have different interests and skills and so on, there's more that can be, can be sort of explored. Now, you know, I think that there are, when you ask, kind of, by the way, I mean, there are just, there are sort of, there will be lots and lots of new job cash. I mean, we just got prompt engineer. We're going to have... I mean, podcast or having conversations professionally in a vertical of something you're passionate about and then the fact that I get to do that for a living, just, I mean, there was Charlie

Rose that, you know, there was Oprah, but the idea that now that there are probably 100,000 people making a living and just doing podcasts and then hundreds of millions of people listening to them is mind-blowing, it's like a little slice of the pie that nobody ever considered. Yeah, that's right. And we just got, you know, we just got prompt engineers. Love that. We're going to have AI Wranglers. we're going to have AI psychologists.

You're going to have a whole bunch of new categories. And I think that is just incredibly typical of what you see happening with innovations, particularly automations. How close is we're watching this all happen in a chat interface. Not scary at all. But I guess people just wrote a signed a petition, hey, maybe we should pause this. I think that was largely, I don't know if you saw it,

but this future of life petition. I think it was largely ceremonial, like, just probably worth us considering. I don't see anybody stopping at their work for six months. I don't see anybody stopping. I think it's a, the cynic would say it's a list of people in places that feel like they're getting left behind and want everybody else to stop for a while while they catch up. Yeah, so there is, that would be a cynical approacher or just, hey, I know that this isn't going to happen, but I just want to have it on record that I said.

This might have been a good time to be more thoughtful. Let's talk about being more thoughtful. Do you think we are getting to a point where unintended consequences are a possibility? Again, the pace, you and I haven't seen that. There are always unintended consequences. I mean, of almost anything, you know, who thought that, you know, doing research on virology that did this or that or the other would lead to this or that or the other thing, you know, good or bad?

But, Pandemics. Yeah, for example. Don't say this podcast is going to get flagged if we actually speculate that a human created COVID seems probable. Right. But, well, maybe it was an AI. No, I don't think so.

It wasn't quite at that level. The AI wasn't quite ready to do that. Now it would probably be an AI. Let's talk about that for a second, open-mindedly here. If you were to put some prompts in and you put in the sequence of COVID, which isn't really a difficult thing to sequence, and said, come up with things more deadly, or come up with things that, you know,

instead of affecting old people, affect young people that have a longer incubation period so they're harder to recognize or stop. AI could do that today. And... You know, it's a little complicated because it turned out, one of the things that's totally bizarre is that large language models

are actually useful for understanding the structure of proteins. Yes. It has nothing to do with human language. It is, however, you know, what happens with proteins, proteins are these long strings of amino acids, which is these kind of collections of atoms, and every protein is specified by some piece of our DNA, our genome, and a protein is a string of thousands to millions of amino acids.

And actually, they don't usually get as far as a million amino acids. But it's a long string of these things, and then they fold up in certain complicated ways. and it's been a long time problem to figure out, given the sequence, how does the protein fold up? It matters a lot how the protein folds up because the way that proteins actually have significance for biology has to do with their shape. And so, you know, is there a particular hole in the protein that where some particular, you know, other molecule can fit in that hole or not? Does the protein kind of, you know, knit itself together to make a muscle, you know, all these kinds of things? So it matters what the shape is. And so the question is, given the sequence, can you predict the shape?

So then that's been a long time problem. That was, a lot of progress was made on that using, well, initially not quite large language models, but now large language models. But really what's happening there is you take the protein, you take the sequence, you want to figure out what kind of shape its protein is. You then say, well, this piece matches this protein that we've already studied. This piece matches another protein. This piece matches another protein.

Now let's figure out how to knit those pieces together, and the knitting those pieces together is something that's a little bit like this kind of puzzle piece thing that I mentioned for human language, it seems. That knitting together is something that LLM seemed to be quite good at. And so then you can do the more extreme thing that people have started to do, which is to kind of use generative AI to say, you know, given a bunch of words, make a protein that does such and such. So yes, you know, the thing you're describing, I don't know, there are a reason.

lots of issues and there are lots of computational or disability questions, actually. But in broad outline, yes, it will be possible for sure to say, you know, take this and, you know, with just a linguistic type prompt, you know, find something that does,

you know, that works a little bit differently and so on. And, you know, it pulls in perhaps something from some other, you know, genome database or whatever else. So yeah, I'm sure that will be a thing that, unfortunately, perhaps, that will be about maybe fortunately, in some cases and maybe unfortunately in others. And that's typical.

Depends on a prompt engineer and what their goal is, right? Right, right. But I mean, that's so typical of progress of all kinds. You know, you can use it to, you know, cure a terrible disease. You can use it to make a terrible disease. Right. You can make a nuclear reactor, can make a nuclear bomb. Yeah, right.

Just seemed like that, you know, in those examples, people didn't have as much access to a tool. And this tool feels like it's going to have everybody's going to have, of access to it, put a couple billion people on this thing that is qualitatively different than the number of people who know how to operate and do nuclear science. Yeah, well, right. It's also the materials you need to make nuclear stuff are not in such easy supply. There's a long supply chain to produce them.

So this is certainly much more accessible. I think we just talked to ourselves into signing the six-month ban. So many scary possibilities here. It's almost like talking about them is, I don't want to accept them in the world, but, you know, it's... Well, I think the thing to understand is when one thinks about, okay, so what are the AIs going to do? First question is, what do we want the AIs to do? You know, if we were going to define a system of ethics, let's say, for the AIs, what would we want that to be? So, you know, one thing people would say is, well, you know, let's have the AIs just mimic what humans do.

Most people would say, that's a bad idea. You know, humans do all kinds of things that. that humans, we don't think humans should be doing. Yeah, get drunk and beat each other up. Yeah, which, you know, in most cases people think is a bad thing, but sometimes people don't think that's a bad thing. And it's complicated. And, you know, I think then what it ends up being is let's make it,

let's make the AIs sort of be the way that we, that humans aspire to be. But that's a much more fuzzy, complicated thing, because it's like, who's aspirations? You know, you pick some, you know, sacred book. You pick some self-help book. Be careful there, yeah. Right. And, you know, you end up with.

So, but then in the end, it's kind of like, well, maybe some group of people would agree this is how we want the AIs to generally behave. You know, we could invent a sort of AI constitution that defines how we generally want the AIs to behave. And that's probably not, that's probably a sensible thing to do. It also happens to be comparatively hard. I think to come up with what you want to say there. And, you know, one of the things that we get to define in perhaps computational language,

perhaps better than in prompts, is kind of a sort of definition of what we want, you know, what we want the AIs to do, so to speak. And then we have to figure out, how do we do that? You know, are we going to have a worldwide, this is what we want the AIs to do? Probably not a very good idea. You know, if you have a sort of monocultured, so, of AI world, it becomes rather brittle. I mean, if something's, if it's like something wrong with the, you know, with the,

with the code, so to speak, there's something wrong with the legal code effectively for the AI's. Oops, you know, we just made the whole world follow this legal code. You know, it's going to blow everything up. So, you know, the thing that has. I'm curious how you feel about the fact that this started, at least, you know, open AI as an open source nonprofit that somehow flipped into a for profit and then flip from everybody should have access to this code.

Suddenly, Sam and the team saying, you know what, this code is a little too dangerous for everybody to see, so now nobody can see it except us. Do you think it should be open-sourced and people should see this stuff and it should be more out in the open? Or do you think it's fine for it to be, you know, programmed in a small number of people have access to the source code? I don't think it matters because I think that this whole idea of LLMs

is now kind of the genius out of the bottom. And, you know, one can make LLMs. You know, Open AI did a great engineering job. And they have a, you know, they seem to have, you know, they've achieved a bunch of things other people haven't yet achieved. And that's, you know, from a business point of view, that has lots of significance, there's lots of timing and lots of, you know, what will ramp up how quickly and so on. And, you know, I don't think that it's a question of, you know, I don't think in the big

picture, I don't think that's an important thing. Honestly, the ability for, you know, innovation is hard and you have to kind of have, you know, you have to have a certain, I mean, I know in our own case, you know, I have a fairly small company that we've been running for 36 years now. I mean, 800 people or something, fairly small by many standards. But, you know, and the fact that we are able to innovate and go on innovating is a consequence of the fact that. that we have a viable business model for the things that we do. If we didn't have that, if we just said, oh, we're going to give everything away, and then, okay, how do we feed the 800 people, so to speak? Yeah, exactly. We have to have some business model.

And I personally, you know, I have to say for myself, I prefer business models which have a directness where the people getting value are the people, you know, paying for the thing, so to speak, rather than more indirect models. Advertising models. It gives a better alignment of, of kind of what one's building with what customers actually want. Seeing what's happened here,

are you just going to build your own language model and compete against chat GPT? Well, I mean, it's something where, where, you know, obviously a company like ours is capable of doing stuff like that. Easily, yeah. Sounds like it would be a no-brainer for you. So, okay. I mean, it's one of these things where I don't think, you know,

there will be many of these things. And, you know, I think that, I mean, the thing I was saying is that I don't think, you know, it's kind of this thing where if you say, well, let's pull everybody down so that sort of nobody has the, the, either the sort of the war chest or the, or kind of the motivation to be a leader, you know, that's not really very good for the world. If you want innovation to happen, you know, you have to have a situation where, where that, you know, where, for example, some, you know, organization can decide for it. itself, what it's going to do up to a point. Because if the whole world is going to vote on, you know, what should we do next? Well, you can kind of, you know, it's very implausible that creative innovation is going to happen in that situation. So I think, you know, I'm not, again, I think in the big picture, it doesn't really

matter what these particular details are, but I think that the, you know, having, having the ability in the runway to actually, and the most of the most of the way. motivation to kind of independently innovate is kind of important. And I think that's that's borne out by the fact that, you know, a year ago, we didn't have chat GPT. Yeah. And, you know, it was, you know, it's particular people who had to, you know, who were in a situation where they could do and were motivated to do the kind of innovation that was

needed to create chat chabbt. Yeah. And it's a small number of people. It's a couple of hundred, I guess, got them to hear. So it's not a, it's not like it took Google to do it. Obviously, Google has their own, but it didn't take a Facebook Google size effort. It took a relatively modest size group of people to achieve this, and they should get all the credit in the world.

As we wrap here, I guess I have two questions at the end. How close are we? That's a possible question to answer, I know, but I'm very interested in hearing your thoughts on it, to AGI. And, you know, if you had to put a year or set a betting line over under, and then how do you know? Like in your mind, when will you know that we have something that is an AGI? You know, about that, you know, last 50 years I've been paying attention to kind of what happens with computers and all this kind of thing.

And people saying, when we can do X, then we'll know we have true, you know, artificial intelligence. You know, I've personally built a few of those X's that people have said, you know, when we have this. And then when you actually have it, people say, oh, it's just. piece of engineering. That's not true, you know, intelligence and so on. I think that the thing, I mean, it's almost the thing that you'll, when you'll know you have true human intelligence is when you basically have a copy of a human. And you can always say, oh, well, it doesn't have this attribute, you know, because it isn't mortal, it can't think this way, or because it doesn't have, you know, five fingers, it can't do this. You know, the only way you'll have something

which is just like a human is to have something really just like a human. Now, I think that the question of kind of when the, you know, it's sort of an incremental thing. It's kind of a, you know, this was a big shock, chat GPT. People didn't expect sort of this level of of humanity, so to speak, in an automated system. I think that I would say that, well, in terms of, you know, what else do you want kind of thing, you know, you'll be, you know, there was the Turing test that Alan Turing made up in 1950, which I think pretty firmly, you know, as of 2023, we can do that. That one is done.

Nobody knew when that was going to happen. and that was one of the last of the kind of standard, this is a test for whether you have true artificial intelligence. So, you know, when we can ask questions like, you know, for the things, for this set of people, when we'll be able to automate the main thing that they do? It's worth understanding that when we say, when we talk about automating things,

it's kind of like, you know, back in the day, people would handwrite this or that thing, and then printing came along, and there was just a standardized font for A and D and C. And some people would say, well, it's much more efficient. We can, you know, it's much more automated. And some people will say, well, you know, you kind of lost the human touch of the calligraphic stuff. And the same will happen here.

There'll be plenty. You know, it's like, you know, when I read a chat GPT written essay, it's very perfect. It's very kind of anodyne in some ways. And, you know, it doesn't have it. It's kind of like the rug that was made by a machine rather than by a person with those other errors in it type thing. And, you know, I think people will continue to say, oh, well, if the rug doesn't have a little errors, then it isn't really, you know, an AGR, an automated general rug or something. Yeah. Who should when these machines are,

you know, the other thing is I was thinking when you talked about how this surprised everybody. It reminds me of when Boston Dynamics made that first robot that could kind of run and do flips. And you're like, huh, I wasn't expecting that. When do these two things combine? The Boston Dynamics,

you know, parkour, a robot, and chat chip BT. and what is that going to look like? You know, one of the things that, so I mean, I think some of the things about sort of being able to create geometrical kinds of things in a kind of large language model-ish way,

you know, those things are very much coming. Some of those things already here. You know, that's important for, you know, if you're making 3D objects, you're doing animation, you know, those kinds of things. That's very, you know, insipient. very, very, very close, I would say. You know, I think that the question of, you know, using machine learning to figure out how

do you grasp, you know, how do you pick up, you know, how do you pick up a cell phone or something, that's proved comparatively difficult. My guess is that will be cracked, but it has proved comparatively difficult. I think that the whole question of sort of how robotics advances my own, you know, one of the things are surprising about robotics is it's fairly non-general purpose. With computers, the big thing that was the big sort of advance that really made computers possible was the idea of universal computation, the idea you could have a fixed piece of hardware, put different programs into it, and it would do different computations. That was an idea

originally from 1920s and 1930s. It sort of became real in the 1950s and so on. And that's what made software possible. That's what basically made computers useful. Were a processor, a video game, or an Excel spreadsheet could all be done on the same computer. Right, exactly. So for robots, that hasn't really been the case. It's not the case that you can have a general purpose robotic system. You know, people, and I think that's something I've even thought a bit about how to do that. You know, I think that's something that is conceivable.

It is tricky because the physical world is nasty to deal with relative to the informational world, so to speak. Yeah. But, you know, if that happened, and I think it will eventually happen, by the way, I should say at a molecular scale, biology has solved that problem. Biology has basically made with these proteins we were talking about earlier. Biologies, you know, you just have a sequence of amino acids and it curls itself up and, you know, sometimes it can be muscles, sometimes it can be a brain cell, sometimes it can be, you know, the things, the critical things in those different kinds of biological devices, so to speak. So biology at a molecular scale has sort of solved the universal robotics problem. But on a large scale, we haven't solved it yet. Probably we will.

And when that happens, for example, in terms of the, oh my gosh, what jobs are going to be automated, you know, another set of chunks of the pie will, you know, the things that are being done now will sort of zero out and there'll be a sort of a new collection of things that become possible. And then that's, you know, something about manipulating the physical world, which, which hasn't yet been, you know, and then what will happen is, you know, the main thing that will happen is sort of manipulating the physical world will become a problem of software, so to speak, rather than a problem of how you put different, you know, sort of pieces on the, on the hand

of the robot and so on. Now, you know, that's going to be wild when you can say to this audit, this sort of general robot, take, you know, you know, Jason's, you know, bad, to his room and it's like, okay, bags, I know what those are. There's Jason, I know who he is, he's a guest, and now I need to know what room is in, let me go query, what room is in, and I'm going to carry them upstairs. The chat GPT interface or the language model would actually be able to figure out what you meant, and then you just need a physical specimen that can actually pick up bags and do this. Right. You know, the thing to understand about that in a little bit more generality is this

whole question about, you know, the chat interface can take sort of your whole, you know, speech about what you want the robot to do, then the question is, how are you actually sure that robots going to do what you thought it was going to do? Because you just had this language thing. And this is where, you know, one of the things that I've been excited about very recently is this is where, you know, our computational language initiative is really important because once you have that, you know, you've got the thing you say in natural language, if you can generate from that a piece of computational language, that's something that is intended for humans to read and, you know, a few million people know how to read it now and probably a lot more

will learn how to read it. And they'll say, oh, yeah, that's what I wanted. You know, there's two lines of computational language I can read them. And yes, you know, yep, that's what I wanted. You know, go do it now. Without that, it can be a bit challenging. You know, you can watch the robot and you say, no, no, no, don't do that. Don't pick up the, you know, don't turn.

Yeah, baggage doesn't mean his spouse or the kids. Like, that's not the baggage we're talking about. Right, right. So, I mean, you know, and you can obviously reprompt it and so on, but there are plenty of situations in which, particularly when you're building up a bigger system and when you want to do, you know, a whole collection of things where having this intermediate layer of the sort of precise computational notation is really important. But, yeah, I think that, you know, we can expect, well, one of the things that's also

funky about something like robotics is that the world, the sort of the built environment that we have was built for humans. So, you know, we have doors we can open with, you know, with hands that are at a certain height, et cetera, et cetera, et cetera. So, you know, there's sort of a certain pressure to have humanoid like robots just because we built an environment that is suitable for humanoid robots. Now, there are plenty of environments in the world, you know, which are, you know, thrown up by the natural world that are quite unsuitable to humans.

Yeah. And where we don't tend to hang out. Yeah, right. Yeah. And where, you know, something quite different would be appropriate. But, you know, that tends to make it, you know, that's kind of the analog of you've got an LLM and it's learning actual human language. And it could learn all kinds of other things, but it actually learned human language to sort of fit into the human linguistic world in that case, as opposed to sort of the human built world, so to speak. It is crazy how fast this is moving.

And it's just great to have people like you working on it. everybody can just check out Wolfram Alpha, check out the plugin, start playing with it, and share whatever you're building on Twitter. And I really appreciate you taking the time. Are you Dr. Wolfram? Should I be calling you, doctor? I feel like I should.

You know, I've noticed, here's a basic rule. If one's doing business, if somebody calls me professor, that's really deadly. Doctor is sort of okay, but, you know, for business it's Mr. Mr. No. All right.

I really appreciate you taking the time. I know you're very busy, especially at this moment in time when everybody's really excited about the work you've done and continue to do. So thank you so much. And we'll see you all next time.

Bye bye.