ChatGPT and Wolfram|Alpha

It’s always amazing when things suddenly “just work”. It happened to us with Wolfram|Alpha back in 2009. It happened with our Physics Project in 2020. And it’s happening now with OpenAI’s ChatGPT.

I’ve been tracking neural net technology for a long time (about 43 years, actually). And even having watched developments in the past few years I find the performance of ChatGPT thoroughly remarkable. Finally, and suddenly, here’s a system that can successfully generate text about almost anything—that’s very comparable to what humans might write. It’s impressive, and useful. And, as I’ll discuss elsewhere, I think its success is probably telling us some very fundamental things about the nature of human thinking.

But while ChatGPT is a remarkable achievement in automating the doing of major human-like things, not everything that’s useful to do is quite so “human like”. Some of it is instead more formal and structured. And indeed one of the great achievements of our civilization over the past several centuries has been to build up the paradigms of mathematics, the exact sciences—and, most importantly, now computation—and to create a tower of capabilities quite different from what pure human-like thinking can achieve.

I myself have been deeply involved with the computational paradigm for many decades, in the singular pursuit of building a computational language to represent as many things in the world as possible in formal symbolic ways. And in doing this my goal has been to build a system that can “computationally assist”—and augment—what I and others want to do. I think about things as a human. But I can also immediately call on Wolfram Language and Wolfram|Alpha to tap into a kind of unique “computational superpower” that lets me do all sorts of beyond-human things.

It’s a tremendously powerful way of working. And the point is that it’s not just important for us humans. It’s equally, if not more, important for human-like AIs as well—immediately giving them what we can think of as computational knowledge superpowers, that leverage the non-human-like power of structured computation and structured knowledge.

We’ve just started exploring what this means for ChatGPT. But it’s pretty clear that wonderful things are possible. Wolfram|Alpha does something very different from ChatGPT, in a very different way. But they have a common interface: natural language. And this means that ChatGPT can “talk to” Wolfram|Alpha just like humans do—with Wolfram|Alpha turning the natural language it gets from ChatGPT into precise, symbolic computational language on which it can apply its computational knowledge power.

For decades there’s been a dichotomy in thinking about AI between “statistical approaches” of the kind ChatGPT uses, and “symbolic approaches” that are in effect the starting point for Wolfram|Alpha. But now—thanks to the success of ChatGPT—as well as all the work we’ve done in making Wolfram|Alpha understand natural language—there’s finally the opportunity to combine these to make something much stronger than either could ever achieve on their own.

A Basic Example

At its core, ChatGPT is a system for generating linguistic output that “follows the pattern” of what’s out there on the web and in books and other materials that have been used in its training. And what’s remarkable is how human-like the output is, not just at a small scale, but across whole essays. It has coherent things to say, that pull in concepts it’s learned, quite often in interesting and unexpected ways. What it produces is always “statistically plausible”, at least at a linguistic level. But—impressive as that ends up being—it certainly doesn’t mean that all the facts and computations it confidently trots out are necessarily correct.

Here’s an example I just noticed (and, yes, ChatGPT has intrinsic built-in randomness, so if you try this, you probably won’t get the same result):

It sounds pretty convincing. But it turns out that it’s wrong, as Wolfram|Alpha can tell us:

To be fair, of course, this is exactly the kind of the thing that Wolfram|Alpha is good at: something that can be turned into a precise computation that can be done on the basis of its structured, curated knowledge.

But the neat thing is that one can think about Wolfram|Alpha automatically helping ChatGPT on this. One can programmatically ask Wolfram|Alpha the question (you can also use a web API, etc.):

&#10005 WolframAlpha["what is the distance from Chicago to Tokyo", "SpokenResult"]

Now ask the question again to ChatGPT, appending this result:

ChatGPT very politely takes the correction, and if you ask the question yet again it then gives the correct answer. Obviously there could be a more streamlined way to handle the back and forth with Wolfram|Alpha, but it’s nice to see that even this very straightforward pure-natural-language approach basically already works.

But why does ChatGPT get this particular thing wrong in the first place? If it had seen the specific distance between Chicago and Tokyo somewhere in its training (e.g. from the web), it could of course get it right. But this is a case where the kind of generalization a neural net can readily do—say from many examples of distances between cities—won’t be enough; there’s an actual computational algorithm that’s needed.

The way Wolfram|Alpha handles things is quite different. It takes natural language and then—assuming it’s possible—it converts this into precise computational language (i.e. Wolfram Language), in this case:

&#10005 Cell[BoxData[ RowBox[{"GeoDistance", "[", RowBox[{ TemplateBox[{"\"Chicago\"", RowBox[{"Entity", "[", RowBox[{"\"City\"", ",", RowBox[{"{", RowBox[{"\"Chicago\"", ",", "\"Illinois\"", ",", "\"UnitedStates\""}], "}"}]}], "]"}], "\"Entity[\\\"City\\\", {\\\"Chicago\\\", \\\"Illinois\\\", \ \\\"UnitedStates\\\"}]\"", "\"city\""}, "Entity"], ",", TemplateBox[{"\"Tokyo\"", RowBox[{"Entity", "[", RowBox[{"\"City\"", ",", RowBox[{"{", RowBox[{"\"Tokyo\"", ",", "\"Tokyo\"", ",", "\"Japan\""}], "}"}]}], "]"}], "\"Entity[\\\"City\\\", {\\\"Tokyo\\\", \\\"Tokyo\\\", \ \\\"Japan\\\"}]\"", "\"city\""}, "Entity"]}], "]"}]], "Input", CellChangeTimes->{{3.8822027023950853`*^9, 3.8822027180278597`*^9}}, CellLabel->"In[105]:=", CellID->1479512159]

The coordinates of cities and algorithms to compute distances between them are then part of the built-in computational knowledge in the Wolfram Language. And, yes, the Wolfram Language has a huge amount of built-in computational knowledge—the result of decades of work on our part, carefully curating what’s now a vast amount of continually updated data, implementing (and often inventing) methods and models and algorithms—and systematically building up a whole coherent computational language for everything.

A Few More Examples

ChatGPT and Wolfram|Alpha work in very different ways, and have very different strengths. But in the interests of understanding where ChatGPT can take advantage of Wolfram|Alpha’s strengths, let’s discuss some cases where on its own ChatGPT doesn’t do quite the right thing. And one area where ChatGPT—like humans—often tends to struggle is math.

It’s an interesting, essay-style response. But the actual result is wrong:

But if ChatGPT “consulted” Wolfram|Alpha it’d of course be able to get it right.

Let’s try something slightly more complex:

At first glance, this result looks great, and I’d be inclined to believe it. It turns out, though, that it’s wrong, as Wolfram|Alpha can tell us:

And, yes, doing math homework with ChatGPT (without it being able to consult Wolfram|Alpha) is probably a bad idea. It can give you a very plausible answer:

But without “really understanding the math” it’s basically impossible for ChatGPT to reliably get the right answer. And in this case, the answer is again wrong:

Still, ChatGPT can even make up a very plausible-looking explanation of “how it got its answer” (not that it’s in any way how it really “did it”). And, rather charmingly (and interestingly), the explanation it gives has mistakes very similar to what a human who didn’t understand the math might also make:

There are all sorts of situations where “not really understanding what things mean” can cause trouble:

That sounds convincing. But it’s not correct:

ChatGPT seemed to have correctly learned this underlying data somewhere—but it doesn’t “understand what it means” enough to be able to correctly rank the numbers:

And, yes, one can imagine finding a way to “fix this particular bug”. But the point is that the fundamental idea of a generative-language-based AI system like ChatGPT just isn’t a good fit in situations where there are structured computational things to do. Put another way, it’d take “fixing” an almost infinite number of “bugs” to patch up what even an almost-infinitesimal corner of Wolfram|Alpha can achieve in its structured way.

And the more complex the “computational chain” gets, the more likely you’ll have to call on Wolfram|Alpha to get it right. Here ChatGPT produces a rather confused answer:

And, as Wolfram|Alpha tells us, its conclusion isn’t correct (as it already in a sense “knew”):

Whenever it comes to specific (e.g. quantitative) data—even in fairly raw form—things very often tend to have to be more of a “Wolfram|Alpha story”. Here’s an example, inspired by a longtime favorite Wolfram|Alpha test query “How many turkeys are there in Turkey?”:

Again, this seems (at first) totally plausible, and it’s even quoting a relevant source. Turns out, though, that this data is basically just “made up”:

Still, what’s very nice is that ChatGPT can easily be made to “ask for facts to check”:

Now feed these through the Wolfram|Alpha API:

&#10005 WolframAlpha[#, "SpokenResult"] & /@ {"Number of chickens in Turkey", "Number of turkeys in Turkey", "Number of ducks in Turkey", "Number of cattle in Turkey", "Number of sheep in Turkey", "Number of goats in Turkey"}

Now we can ask ChatGPT to fix its original response, injecting this data (and even showing in bold where it did it):

The ability to “inject facts” is particularly nice when it comes to things involving real-time (or location etc. dependent) data or computation. ChatGPT won’t immediately answer this:

But here’s some relevant Wolfram|Alpha API output:

&#10005 Values[WolframAlpha[ "What planets can I see tonight", {{"Input", "PropertyRanking:PlanetData"}, "Plaintext"}]]

And if we feed this to ChatGPT, it’ll generate a nice “essay-style” result:

Sometimes there’s an interesting interplay between the computational and the human like. Here’s a rather whimsical question asked of Wolfram|Alpha (and it even checks if you want “soft-serve” instead):

ChatGPT at first gets a bit confused about the concept of volume:

But then it seems to “realize” that that much ice cream is fairly silly:

The Path Forward

Machine learning is a powerful method, and particularly over the past decade, it’s had some remarkable successes—of which ChatGPT is the latest. Image recognition. Speech to text. Language translation. In each of these cases, and many more, a threshold was passed—usually quite suddenly. And some task went from “basically impossible” to “basically doable”.

But the results are essentially never “perfect”. Maybe something works well 95% of the time. But try as one might, the other 5% remains elusive. For some purposes one might consider this a failure. But the key point is that there are often all sorts of important use cases for which 95% is “good enough”. Maybe it’s because the output is something where there isn’t really a “right answer” anyway. Maybe it’s because one’s just trying to surface possibilities that a human—or a systematic algorithm—will then pick from or refine.

It’s completely remarkable that a few-hundred-billion-parameter neural net that generates text a token at a time can do the kinds of things ChatGPT can. And given this dramatic—and unexpected—success, one might think that if one could just go on and “train a big enough network” one would be able to do absolutely anything with it. But it won’t work that way. Fundamental facts about computation—and notably the concept of computational irreducibility—make it clear it ultimately can’t. But what’s more relevant is what we’ve seen in the actual history of machine learning. There’ll be a big breakthrough (like ChatGPT). And improvement won’t stop. But what’s much more important is that there’ll be use cases found that are successful with what can be done, and that aren’t blocked by what can’t.

And yes, there’ll be plenty of cases where “raw ChatGPT” can help with people’s writing, make suggestions, or generate text that’s useful for various kinds of documents or interactions. But when it comes to setting up things that have to be perfect, machine learning just isn’t the way to do it—much as humans aren’t either.

And that’s exactly what we’re seeing in the examples above. ChatGPT does great at the “human-like parts”, where there isn’t a precise “right answer”. But when it’s “put on the spot” for something precise, it often falls down. But the whole point here is that there’s a great way to solve this problem—by connecting ChatGPT to Wolfram|Alpha and all its computational knowledge “superpowers”.

Inside Wolfram|Alpha, everything is being turned into computational language, and into precise Wolfram Language code, that at some level has to be “perfect” to be reliably useful. But the crucial point is that ChatGPT doesn’t have to generate this. It can produce its usual natural language, and then Wolfram|Alpha can use its natural language understanding capabilities to translate that natural language into precise Wolfram Language.

In many ways, one might say that ChatGPT never “truly understands” things; it just “knows how to produce stuff that’s useful”. But it’s a different story with Wolfram|Alpha. Because once Wolfram|Alpha has converted something to Wolfram Language, what it’s got is a complete, precise, formal representation, from which one can reliably compute things. Needless to say, there are plenty of things of “human interest” for which we don’t have formal computational representations—though we can still talk about them, albeit it perhaps imprecisely, in natural language. And for these, ChatGPT is on its own, with its very impressive capabilities.

But just like us humans, there are times when ChatGPT needs a more formal and precise “power assist”. But the point is that it doesn’t have to be “formal and precise” in saying what it wants. Because Wolfram|Alpha can communicate with it in what amounts to ChatGPT’s native language—natural language. And Wolfram|Alpha will take care of “adding the formality and precision” when it converts to its native language—Wolfram Language. It’s a very good situation, that I think has great practical potential.

And that potential is not only at the level of typical chatbot or text generation applications. It extends to things like doing data science or other forms of computational work (or programming). In a sense, it’s an immediate way to get the best of both worlds: the human-like world of ChatGPT, and the computationally precise world of Wolfram Language.

What about ChatGPT directly learning Wolfram Language? Well, yes, it could do that, and in fact it’s already started. And in the end I fully expect that something like ChatGPT will be able to operate directly in Wolfram Language, and be very powerful in doing so. It’s an interesting and unique situation, made possible by the character of the Wolfram Language as a full-scale computational language that can talk broadly about things in the world and elsewhere in computational terms.

The whole concept of the Wolfram Language is to take things we humans think about, and be able to represent and work with them computationally. Ordinary programming languages are intended to provide ways to tell computers specifically what to do. The Wolfram Language—in its role as a full-scale computational language—is about something much larger than that. In effect, it’s intended to be a language in which both humans and computers can “think computationally”.

Many centuries ago, when mathematical notation was invented, it provided for the first time a streamlined medium in which to “think mathematically” about things. And its invention soon led to algebra, and calculus, and ultimately all the various mathematical sciences. The goal of the Wolfram Language is to do something similar for computational thinking, though now not just for humans—and to enable all the “computational X” fields that can be opened up by the computational paradigm.

I myself have benefitted greatly from having Wolfram Language as a “language to think in”, and it’s been wonderful to see over the past few decades so many advances being made as a result of people “thinking in computational terms” through the medium of Wolfram Language. So what about ChatGPT? Well, it can get into this too. Quite how it will all work I am not yet sure. But it’s not about ChatGPT learning how to do the computation that the Wolfram Language already knows how to do. It’s about ChatGPT learning how to use the Wolfram Language more like people do. It’s about ChatGPT coming up with the analog of “creative essays”, but now written not in natural language but in computational language.

I’ve long discussed the concept of computational essays written by humans—that communicate in a mixture of natural language and computational language. Now it’s a question of ChatGPT being able to write those—and being able to use Wolfram Language as a way to deliver “meaningful communication”, not just to humans, but also to computers. And, yes, there’s a potentially interesting feedback loop involving actual execution of the Wolfram Language code. But the crucial point is that the richness and flow of “ideas” represented by the Wolfram Language code is—unlike in an ordinary programming language—something much closer to the kind of thing that ChatGPT has “magically” managed to work with in natural language.

Or, put another way, Wolfram Language—like natural language–is something expressive enough that one can imagine writing a meaningful “prompt” for ChatGPT in it. Yes, Wolfram Language can be directly executed on a computer. But as a ChatGPT prompt it can be used to “express an idea” whose “story” could be continued. It might describe some computational structure, leaving ChatGPT to “riff” on what one might computationally say about that structure that would—according to what it’s learned by reading so many things written by humans—be “interesting to humans”.

There are all sorts of exciting possibilities, suddenly opened up by the unexpected success of ChatGPT. But for now there’s the immediate opportunity of giving ChatGPT computational knowledge superpowers through Wolfram|Alpha. So it can not just produce “plausible human-like output”, but output that leverages the whole tower of computation and knowledge that’s encapsulated in Wolfram|Alpha and the Wolfram Language.