Computational Conversations EP#9 | Electricity vs. computation: What can we learn?

Hi, I'm John McLuhan.

And I'm Conrad Wolfram.

And this week's Computational Conversation podcast is on the comparison contrasting the revolutions of computation and electricity and the changes that they brought about to society. So, Conrad, you've used this comparison of the rise of electricity in comparison to computation in talks, and I thought we could use today's conversation to elaborate on that. So perhaps you could start by explaining the key idea and why you used it.

Well, I think it was sparked because I was invited, in 2021, I hope I've got the date right, to give a talk at the Royal Institution, and particularly at a maths institute called lims, in the Royal Institution that is occupying Faraday's apartments, as they then were. And Faraday ran the first motor in 1821 at the Royal Institution. And so this gave me the idea. I think that was the reason for this talk happening at the time it did, but. Or the sequence, or I forget the whole rationale for why the talk was then. But anyway, it somehow correlated with this motor running. And so that gave me the idea of thinking about electricity and its effect on society and what Faraday in particular had done and what Maxwell subsequently did, and just thinking about that as against what we are trying to do with computation and the sort of parallels and differences between those two major societal changes.

So what do you think the big parallels are?

Well, they're both very significant technological revolutions. They both had elements that existed before they were sort of put together into a whole. Many elements of electricity were known, I mean, right back effectively, when they once recalled this, but to rubbing fur on amber and static electricity in the Greeks and so forth, right the way forwards to more recent discoveries in the 18th century. What was interesting about Faraday particularly was he put together electricity and magnetism. He put many of these elements together in a consistent form.

Form.

And the other thing that was interesting, I mean, the real talk actually at the. At that moment was about, you know, where are the trajectories on, do we have further to run with this trajectory of computation? Are we going to get more juice out of this revolution than we've so far got? Where are we at on that, on that whole sort of trajectory for computation? How does it compare electricity? And my point was, well, you know, with 200 years on from this running of the motor, and still we're doing a lot of new things with electricity all the time, and I reckon we've got a very long way to go with computation.

So to what extent do you think that the fact that it was in practical terms a new discovery in Faraday's day. I mean, as you say, knowledge of it had gone back some time, but there came a point where suddenly you could generate electricity on demand and that was understood. Whereas computation we've been able to do for a long. It's, it's just emergingly, it become emergent as it become more powerful. But there's never been a. We've now got something new to play.

With and well, we had computers at some point which have emerged over the last, you know, I mean, you know, it depends what, what, where, where you want to draw the line, the sand. I mean, obviously the, the idea of it was, was a, essentially, I mean, it's modern incarnation, Turing. But in practical terms, most people have had their hands on computers for a few decades and even less than that in some cases. So the idea that you can just switch on computational power for most people is very new. And certainly the extent to which it's been done is new and the fact that it's so grown in capability is new. I think the other thing in your.

Analogy, the kind of rubbing fur on things was kind of like hand calculating useful and very limited purposes, but not for the kinds of things, things that you could really do with it. And that's the same thing that's happened with computation. You think that it was useful for very limited things like making your hair stand on end.

Yeah. And it wasn't really. Yeah, correct. I mean, there were bits which were known about it, but it wasn't pieced together in a useful way. And in particular the motor obviously was an absolutely critical use of electricity. A fundamental change in what could be done. And things like light bulbs, you know, I mean, electric light change, a huge number of things aside, they stand out.

As two killer applications of electricity that would have crossed on immediately and been, you know, instantly understood as being important. Do you think computation has those killer applications or are they yet to emerge?

Well, I think we've had pieces like data science, I would say actually the current tranche of AI. I mean in a sense all of AI is, is an application of computation. You can argue it different ways. Obviously there's a massive branch is the current machine learning LLMs and likewise we're obviously doing other sorts of AI with sort of more deliberate computation. I think that one interesting step is to think about computation as one entity, which is something to some extent we pioneered 36 years ago with Wolfram. I mean, it's like at the time people were thinking about algebra systems. Could a computer do algebra, and then separate. People were thinking about deploying it, as was common on numerical calculations. And there were graphical things starting, et cetera, et cetera. And there were programming languages. And the idea, though, to think about all of this is we're trying to assemble everything about computation in one place, its representation. What you do, all the algorithms in a sort of consistent way, I think is still not quite there. And I do think that has a big parallel to both Faraday and subsequently to Maxwell, in the sense that Faraday pieced these things together and that piecing together has a huge power, which the disparate pieces, even very powerful disparate pieces like motors, don't quite do in the same way. And we're still learning that we're still getting bits of electricity, the applications of which are pieced together with other things in a way that we hadn't quite understood.

So is the journey then to move from a world of light bulb shops and motor shops to a world of electrical engineers, and is that the path you see for computational?

Well, indeed. And I've often joked, and I don't know if I've got my timing right here, but if we were to put up a sign, the equivalent of us at the moment putting a big sign saying, we're at Wolfram. Do computation. That's what we do. People are a bit bemused by that because they say what we want is sort of data science or modeling, or we want some very. Or AI, whatever that means. I think mostly what they mean by that is machine learning. I think if we were in the 1870s and we put up a big sign saying we're electrical engineers, I imagine people would also be likewise, a bit confused. And most people were looking for the specific applications, and they didn't get this big picture. And they weren't used to people who were thinking about the big, big picture to deploy whatever they had. So essentially that. That's what we're sort of trying to. Trying to figure out.

So one thing that to me stands out as perhaps different in. In the two journeys is that electricity, by its nature, really, really wants to be standardized. You know, apart from the Tesla vs Edison AC vs DC argument, basically that electricity, in its distribution around the world has had to. Had to, in order to be efficient, become one unified way of doing it. And in that case, Tesla won with.

You should explain that a bit more. Remind me what had happened there with that argument.

Well, Edison was fixed on direct current, which is great for batteries, and Tesla had, I think, understood the idea that transformers and Therefore, the distribution over long distances were, was a lot easier if you have alternating current, because transformers only work when there is a change in current. So the oscillation makes it very easy to step up and down voltages, which was the big win of AC over dc. But Edison was a big showman. He was there first. And my belief is that, at least the urban legend is that Edison popularized the electric chair as an execution method because he wanted to associate AC with being dangerous. And he was putting out the idea that alternating current was a dangerous thing. Everyone should go with DC as it happens, as I said, Tesla's won. So now all electrical distribution around the world, everywhere that isn't a standalone device with batteries in it has alternating current flowing through the circuits. We couldn't have two power stations on different systems flowing into any kind of grid system that shares the electricity because they're just not compatible. So you're kind of forced to make a choice one way or the other. I don't think that's true for computation. You can easily make a piece of software that can stand alone, just like a battery can be on a kind of different voltage than your mains. And it's sort of missing that force to say we have to pick a side and say there's going to be a unified way of representing these things in the way that at least grid distribution has been forced to unify in electricity.

Yeah, I'm not sure that's true of data. I mean, if you're swapping data or models and you want doing it in a meaningful way, I mean, you can do things with electricity to step it up and down and have different setups, but it's extremely inconvenient to do it in all sorts of ways. And I think the same is true of computation to some extent. I mean, yes, you can have your own local thing where it works totally differently to somebody else's. You can represent it in different way. You can have different function names and operations. You can have data structured or even not structured very well, but actually making those things common has a huge power to it. And I think one of the things we've noticed right over the last few decades ourselves at Wolfram is that things that we didn't sort of predict were common. I think if we Talked about this 20, 30 years ago, I think the idea that sort of functions, data notebooks, you know, machine learning, representation, sound, everything else were sort of interactive application, everything were all the same object type, and that really worked and that that representation, symbolic representation, really unified those. And you could do all sorts of weird things between these different, apparently different domains. I think we weren't even sure of that to the extent we are now. But actually there's a huge power you've got from that. And that's one of the reasons we were so able, we're so able to coexist and make a sort of better hybrid solution, I think, of AI than in many cases than the pure sort of statistical LLM direction that's around. So I'm not sure I buy that. I think it's what one sees as the power at the given time is sort of critical, I suppose.

I see what you're saying when it. To stretch the analogy a bit further, it feels like data standards like JSON and XML are a little bit like the way that we've standardized on a fairly small number of socket types to plug into walls that, you know, you've got about three in Europe and one in the US. I don't know how many there are worldwide. Not very many. Like there aren't that many different data formats for sharing structured and unstructured and rectangular data. But the idea that the thing that flows through it is sort of so fundamentally unified, it feels like we're the only ones saying that. But it lacks the driving force that electricity had to make that pervasive. And certainly electricity has lots of competing standards when you think about the number of different battery types you can buy and as I say, different socket typ.

Pipes to some extent. But I mean, they're actually compared to what they could be. Given the, the fantastic range of use of electricity, I would argue they're quite a small number. And you know, it's awfully easy with plug. You know, we all have plug converters when we travel right. Nowadays where it's very, very easy to. Yes, there are a few standards for electric plug sockets, but they're very few. I mean, compared to the number of countries there are in the world and everything else, they basically, you know, I think in Wolfram Alpha we actually enunciate the different sorts. And now you even get plugs which do everything right. You even can you put any plug socket from anywhere in. And they work like in airplanes often, for example. So even though it's not 100% standardized, it's standardized enough. It's very easy to go between them with very little effort. So I actually think it's amazing. I think it's also amazing that electric circuit diagrams you can read. I mean, if you know anything, you know, I mean, the, the symbols are all the Same, it's pretty much universal how those are written down. That's worked incredibly well for the most part, et cetera, et cetera, et cetera. So I think we're underplaying the amount of standardization that there has been electricity, and I think overplaying a little bit how, how obvious that needed to be was as against computation. So at the moment we're in this, and one, let me. I think another part of this picture is, and this is probably more true of computation for every individual piece you want to do. It seems that for that one piece alone you can optimize the interface, the setup, the representation better than you can by going across the whole thing in one go. But of course, that's only for that one item. If you try and work it for all the different aspects of computation you want to do, it's vastly better if you have this general representation and approach. But so for anything one's trying to do at that moment, it seems that the narrow approach might be better. Now, whether is that true originally with electricity? Possibly, I don't know.

Yes. I wonder whether all of those differences were ironed out by these different local areas bashing up against each other and having to thrash out their differences versus insightful people mapping out larger areas in one go. I don't have history on that, I'm afraid.

Yeah, well, I mean, I think Faraday put these huge number of different things together and invented new, you know, discovered new whole things. I think that my understanding, and I'm no expert on this, of Maxwell was, you know, and his collaborators was, you know, he managed to put mathematize this into a sort of structure of these four equations and other things he did as well. Possibly that's slightly the reverse of what we've been doing. We've said, right, this is the structure in which we want to try and fit things. And it turns out that structure has worked very well for representing a large number of things over multiple decades. So maybe that's a slightly different way in which this has looked.

So maybe we're more Maxwell than Faraday, that a lot of computations had already been invented, but we're the ones trying to make some kind of unified representation and language to pull it all together.

Yeah, although don't forget the various automations and algorithms things we have invented that plot, you know, sort of fit in there and the representations as well. So it's, it's interesting to compare them. And I think the. I mean, another thing which is interesting about this whole journey is when you start to Assume that this new power is everywhere. So I think people were very deliberately thinking of electricity as something that you could deploy in these very specific cases. And then at some point it's sort of like, oh, maybe we could electrify everything. I mean, I still remember the wind up windows in the car, right? It's like, well, obviously just put an electric motor and make them electric windows and manual toothbrushes versus electric toothbrushes. So it's like we have a toothbrush, let's make it an electric toothbrush. So I think we're starting to enter a little bit this. You know, we've got a subject or we've got a way of thinking about something, let's make it a computational version of that way of thinking about something. See what happens with that. Deploy computation on anything, whether it's biology or political campaigns or whatever, to optimize the decisions.

So I think one thing where that ubiquity has been kind of interesting is if you look at automotive, where in the early days of cars, in the kind of 1890s, it actually wasn't clear whether it was going to be combustion, steam or electricity. There were electric cars right back at the beginning, and that seemed like a settled argument for a long time. But now we've got this ubiquitous energy supply in electricity, we've got other motivating reasons going on and what was previously dismissed as resurfaced as a viable way of doing things. The only analogy I can think of in computation is the, the recent wave of AI where machine learning techniques and even the things that are used in generative AI, those ideas have been known since the 60s and 70s and they were academic curiosities for computational philosophers, I guess, to think about. And now they've reappeared as the ubiquity of the computation becomes available to make them viable again. Do you think that's a pattern that we're going to see play out where computational ideas that weren't viable and were later dismissed reappear.

People also reimagine the combinations. So I mean, one we've talked about quite often is the funny business with numerics and algebra and, well, more generally symbolics. You know, before computers, people spent a huge amount of effort trying to get around doing lots of numerical calculations because they weigh, you know, by hand, they're incredibly time consuming and therefore it limits greatly what you can do. And so there were hundreds of years of tricks, basically. I mean, sophisticated ways to logic, to reason, using symbols rather than numbers, and to come out with results which you couldn't Otherwise have come out with, with the computing power available. And I think a standard example of this was, I mean, I hope this is correct. But wings on airplanes, for example, I believe conformal mapping was used to work out, you know, aerodynamic wing shapes because it could be done symbolically and you could work out a lot more than you could before computers about how to model, how to make the optimized wing enter. Computers, which can do numerics relatively well even from the get go, but can't do anything like symbolics. There's no way to do that, at least not effectively for a long time. So everybody then starts doing everything numerically, completely forgetting about the symbolic approach. And there are a lot of pluses, but there are also a whole bunch of negatives. I mean, they don't quite understand where things come from. They can't actually look at the thing. Certain things go badly off track, et cetera, et cetera, et cetera. But then at some point, and certainly even before mathematica emerged in 88, although I think we made it a lot easier to use, you know, you could do algebra on computers, but all techniques have been really, I mean, you either did it by hand, completely separately, or they'd been forgotten in terms of making it on computers. So therefore doing it numerically on computer was much faster than doing it by hand with algebra. It's been a long, it's been quite a lot of effort to get people to rethink that, to realize that that combination is good and you can switch between them and it's pretty efficient to algebr very well. And so, for example, as you know, inside some of the functions in Mathematica, for example, we do some symbolic analysis before we do the numerics to make the thing more efficient, which other systems are separate don't do. I mean, for example, for plotting a function to work out where discontinuities are and so forth.

There's nice examples in optimization as well, where you can work out the derivatives of functions that allow you to accelerate the numerical optimization routines.

Absolutely.

Actually converses was part of our story as well. I don't know if you were aware of this, but back in our early days, as we were adding numerical evaluations of all kinds of special functions, the fact that we could generate the symbolic series expansions symbolically in Mathematica allowed us to generate the code to generate the numerical versions. So there was kind of Mathematica already in the early days was being built using Mathematica, mixing the symbolic and numeric approaches.

So I think something that's happening now, which we're seeing right play out in front of us is we've got this fantastic, you know, machine learning has been incredibly successful in all sorts of ways and its latest sort of, you know, amazing application, that of LLMs, they're not the only game in town, a bit like the numerics that came before them. They do a fantastic job on certain things, but we shouldn't forget the deliberate computations that have been around and I think we're seeing in what we're doing and others. But combining sort of Wolfram deliberate computations with our inbuilt, you know, with the machine learning and LLMs produces answers that each individual thing isn't able to produce in some cases. So I think that's a sort of example forward. And I mean, it's funny to take the electricity, obviously we have hybrid cars as well as pure electric cars at the moment and pure, you know, combustion engines and that's a sort of funny, funny mixture. Now, I must say, I'm not particularly a fan of hybrid cars in that case. I am a fan in many cases of hybrid computation, but I certainly think one wants to look at the hybrid approach in different cases to see whether it's relevant because an individual approach.

Well, it's certainly a case, I think, where it does make sense, where you've got low emissions for inner cities, but you have range when you go out on the country roads.

Yeah, but you've also carrying two systems, all of which need servicing and they're heavy and adding huge amounts of complexity. I mean, the nice thing about electric cars is they are basically much simpler in many respects. I mean, the technology to make them go far and everything else. And particularly you mentioned Tesla earlier, just to be clear, of course, that's Tesla the inventor, not Tesla the car make, although of course one is causative of the other. But it's also interesting, and this may happen again some level. I mean, electricity in a way. I mean, the idea of it, you know, there are many complexities to it, but there's a certain beautiful cleanliness to it in the. In the sort of way in which it operates. You know, way a motor works is much simpler than the way an internal combustion engine works, for example. And I don't know, we'll see parallels of that, perhaps.

So before we wrap up, I wanted to ask you about a slightly different path of electricity that. And I guess this has been repeated with all kinds of inventions when there are new things in the world. In the early days of electricity that came with A lot of quackery where you had people who took this new poorly understood phenomenon and there were electric beds that were supposed to make you healthier and electric belts and there was an electric cocktail I looked up mostly with health benefits. There was some completely nonsensical machine called out now, I can't remember now it's a spondylo therapy or something that managed to be sold to three and a half thousand clinics around America doing basically fake medicine that revitalized you with electricity. Does computation have its quackery as well? Oh yeah, just like in electricity, a passing pad or something that is longer lasting.

Well, it certainly has its quackery, right. We have huge amounts of it at the moment. We have people who are over claiming for when they compute an answer and often they're doing it. Sort of a typical mode computation is that they're using much too simplistic data computation. They're throwing out a lot of the data to make it easy for them to work out a model and so forth. And they're producing an answer that appears very convincing because it's got a number associated with it, but it's complete garbage in terms of decisions you might take from it. In fact, worse in some cases than just qualitatively thinking about it, you know, with doing primary effects and forgetting secondary and tertiary effects. So unfortunately there's a certain, and perhaps this happened with electricity. There's certain panache from the number from the computation has a definiteness, it seems so scientific that somehow it has a certain ability to persuade which in some cases way beyond what is. And then there are a lot of quackery using that in cases where it really isn't appropriate to use it because the complexity of the situation isn't matched by the sophistication of the use of the computation. So that's one form of it. Maybe there are others, I don't know. Can you think of some that it.

Sounds like you're saying that at least the answer with computational quackery is more computation and better computation rather than just don't use computation there. But I suppose the same is true of medicine. There are actually useful applications of electricity in medicine, but just not the ones that were being extolled in the late 19th century.

Well, I mean, extremely useful.

Right.

I mean there's, there's, I mean, yeah, they're. The more basic things are just buzzing electricity through different things. I think I've had that done to muscles, for example, to make, make them relax. Right. But, but obviously there's electricity applied to scanning, you know, to, To. To, you know, to make. Obviously powering an MRI scanner, for example, has many other pieces to it and you know, thermionic emission in. In X rays and all sorts of things. So absolutely there are sophisticated uses. And in the end, yeah, probably, I suppose the more sophisticated uses ended up being, you know, that they're powering something behind the scenes to do something else that has produced a fantastic result. I don't know if there's an exact parallel there. I mean, I think one, maybe the.

Best parallel at the moment is just the way everything is badged AI and just by putting the computational label on something, regardless of what's actually happening on the inside, there becomes a belief that thing is somehow better and more sophisticated in the same way that simply badging something as electrical or in the 1950s, 40s, radioactive somehow made it better.

Yeah, well, that's not. That particular piece of marketing has rather reversed itself since then. I suppose there is a. I mean, maybe there's the following difference. I don't know. I mean, the initial electrical things are sort of physical. I mean, obviously not computers are a result of electricity in a sense. Right. Where, you know, things happen in software. But there was a lot of hardware where you saw things physically move and do things. Computation, of course, is different to that. It's sort of intelligence thing. In a sense. It's beating it or either going. Offering alternative routes to human intelligence or providing human like intelligence or whatever it's doing, but it's mainly initially doing that kind of thing. So it's sort of harder to see that as an object possibly. And so that changes the sort of Persona it has, I guess, against electricity. But I haven't really thought that through properly.

Okay, well, that's been a fun conversation. So I think we'll wrap it there and we'll see you all next week.

Yes, look forward to the next computational conversation.