Future of Science and Technology Q&A (May 29, 2026)

**Stephen Wolfram** (0:01)
Hello, everyone. Welcome to another episode of Q&A about Future of Science and Technology.
Let me see, we have a whole bunch of questions here.
Let's see, Julie asks, is human oversight becoming more important, not less, as AI improves? Well, that's one of those questions that's partly about the psychology of what people think they're getting from AI as a tool. I mean, I think that when people's expectation is, lots of things will go wrong, I really got to look at what's happening, then people will look at it and not use it for things where it will be a problem if something went wrong. At the point where people say, it's just going to get it right. That's where you run into trouble. I think the thing I'm seeing, it depends on the task.
There are a number of things to say about, I would say in the area of coding, what I'm seeing is that people who really understand the big picture of what's happening, these are the people who can make good use of AI and know how to break down a problem to the point where the pieces can realistically be done with AI. People who just like throw the whole problem at an AI, that's not usually going to go well. Let me break that down another time, which is to say when the thing you're trying to get the AI to do is to make something that just has to look right, then maybe that can work. If it's just make me a form on a website that roughly looks like this, it may very well be able to do that. What's underneath may be a horrible mess and may have lots of crazy things that you wouldn't like if you understood them. But the thing you want is just the thing you're seeing and maybe it can do that well. But when you're building a bigger sort of tower of functionality, it isn't really good enough to have something where the pieces have to be modularized by you, and then it depends on what you're doing. In some cases, it's kind of essential that you understand what the piece does. That's a place where people I think get into a fair amount of trouble where it's just like, oh, the AI wrote this piece of code. It's 300 lines long. I don't understand any of it. Maybe the AI can tell me, you ask the AI, what's going on in this piece of code and tells you something and you still don't understand that. This is a big place where our technology, Wolfram Language, is really important because we built kind of the highest level language that's ever been made, a language which really is trying to deal with computation rather than just pure programming, so to speak. And it's a place where, having the right Wolfram Language code, you get something where if it does a good job, you can actually understand it and you are expected to understand it, so to speak. It's not something where it's just gobs of low level code which no human could be expected to understand. So I think that happens to be a critical point for our technology, the ability to have code that you understand. I mean, essentially, one of my statements about our technology, it's what you need if you want to understand what you are computing. If it's okay that you don't understand what you're computing, but it just looks roughly right at the end, then maybe you don't need us in the middle. But if you want to understand what you're computing, you absolutely need us.
There are many cases where you do need to understand what's being computed. Even cases where people wouldn't have used our technology in the past, but where now they're trying to make a piece of a big system, and they're going to build a lot on top of that piece, and they've got to understand what that piece is doing, or the whole tower is going to topple over.
But I think the thing that I'm observing is that people who understand what they're trying to get are at a big advantage. People who were in the trenches programmers, who just knew how to turn the crank to get out more code, those people are not in good shape. In the time of AI, because what becomes important is, did you know what you were trying to get? Do you understand what happened? And how well can you modularize things to the point where you're asking the AI to do something that can plausibly do? I mean, something I've seen a bunch of recently is the following phenomenon. Is people who are like product managers, for example, who maybe wrote code 20 years ago, but haven't been doing that for a long time, say, oh my gosh, now we can actually with our own fingers write code that implements something. Probably not production code, but something where we can kind of see what the consequences of our ideas are. Well, it's interesting that that's exactly what we achieved for technical computation 40 years ago, when we introduced Mathematica. It went from the point where if you wanted to understand the consequences of kind of a technical idea that you had in the past, the only way you could do that with a computer is you find a programmer, you get the programmer to do that thing of programming it. They come back in a few weeks and they show it to you, and you've probably forgotten what you asked them to do. And the iteration loop is very slow. What happened with Mathematica, a now Wolfram Language, actually very quickly among people like kind of the high-end physicists and mathematicians and so on, and then spreading to other people was the people realized, oh my gosh, I can actually have the process of computing things be something directly under my control, as sort of the principle who understands what you're trying to do, and that makes the whole thing a vastly more efficient loop. That's something that has sort of extended now to a bunch of areas that involve kind of user interfaces and construction of websites and kind of general, sort of more general kinds of, I would say, shallower code that's now been extended with AI. That feeling that you can sort of go from an idea to an implemented thing. I have to say that was sort of my whole point in building Wolfram Language and its distant predecessor from 1979 SMP, was exactly to have the ability for myself to have that loop of going from sort of idea to implementation as efficiently as possible. And I think that I kind of have felt that I've sort of been living what we would now think of as sort of the AI dream for more than 40 years now, because for me, at least a lot of the kinds of things I'm trying to do in science and research, in technology development, I can go very directly from an idea that I have to a thing that's implemented with the tool of Wolfram Language. And as I say, that's now sort of that experience of get, you know, go from the idea to a thing is extended to more people and people who understand sort of what, understand the idea and then can tell whether the idea got correctly implemented or got, once they see the implementation, it's very common, you have an idea, once you see it implemented, you realize, wait a minute, that wasn't quite the right idea. This part of it is great. That part of it doesn't work so well. You know, you want that kind of tight loop. And if you're the person originally coming up with the idea, or originally conceptualizing the thing, that's the place where having that tight loop really makes sense. If your job had been in the trenches, kind of grind the cogs, so to speak, well, then you probably, being asked to figure out the idea, you're gonna say, that's super difficult. We don't know how to do that. And that's not our job. And well, so there's then a big advantage if you do understand what sort of the big picture is. I think that the, as I say, in answer to the original question about human oversight, I think that it really is a question of the expectation. I mean, if you expect that your self-driving car, for example, is really gonna do the right thing, you just don't pay much attention and you let the thing do its thing. If you're always like, I know it's gonna make a mistake in this or that case, like it's gonna make, this seems to be a bug in current times of these things, that making turns that aren't, non-allowed turns on red lights and things. But so you know that's a thing which goes wrong, so you pay attention to that. If you are lulled into a sense of security by the fact, oh, it seems to be working perfectly, then that's where you run into real trouble.