Anna Gressel: Yeah, I think it is a great topic, and there’s probably no area that we’ve seen be as affected from a job perspective as coding. We’ll talk about that, but it’s really kind of amazing, you know, what people can do and how important it is. And I mean, I was just reading a post online about someone who was, you know, coding. I mean, it’s, you know, it’s stuff that as a non-coder, you would not think would be easy to do. And now with AI, there are a lot of opportunities opening up. So, lots to dive into today.
Katherine Forrest: Yeah, so let’s get into it. So we’ll talk about what’s happening with coding capabilities, some of the cyber issues that using AI models to code raises. And another topic that is really important is, while we’ve got some shift where some of the—I’m going to call them newer coders—are being replaced by certain AI models, there has to be a way to continue to train humans for the ongoing human oversight that is still really necessary. So let’s get into some of all that.
Anna Gressel: Yeah, it is a lot of micro, a lot of macro. I think we have some ground to cover today. But first, you know, I think it’s worth spending a moment on kind of why this is all coming up now and why, with generative AI, is this even a discussion we’re having. And just to start with one of the fundamentals of this, not to fully answer that question: when generative AI models were released, one of the important data sets that were used to train even early models were code. And that included GitHub and also other code that was underlying websites. And a lot of that code got tokenized and ingested into even those early generation GenAI models.
Katherine Forrest: Right, I mean, you tend to think when some of the, I’ll just use an example, but if you had a scraper or a crawler that would go out and it would grab text, it was also grabbing what was around text. And what was around text from certain websites and just in certain applications was code. And it was everywhere, so that got pulled in. So you had both data sets of code, but you also had code that was actually code underlying applications that were actively being used. And all of that actually taught the models what they know about code. So we can think about it in terms of the model learning how to read and write in different coding languages.
Anna Gressel: So, Katherine, I think an apt way to pick up all of this discussion is to talk about your own forays into AI coding, because I know you have a little bit of a story to share with folks.
Katherine Forrest: An “apt,” I like that. Apt, app, you know, you get that? Yeah? Well, you know, my coding story is so pathetic, but it does show how really anybody can do this. And so what I did at one point, just opening up ChatGPT, was I thought, well, “oh, can you really code? Could anybody really code?” So I said in natural language—which is just sort of, for me, it was just using an English prompt query in the query bar—I said, “can you draft for me a ping pong program?” And it said, “sure,” and as it does, very enthusiastically. And it proceeded to, in Python actually, create a ping pong program. And it—you know, you can sort of watch it just generate the code right there in front of you. And then what you do is you block copy it, and then you put it into your computer and you would then execute. And then it created like a little executable ping pong program. So, first of all, I’m not a coder. I have zero experience in coding, but I was able to create this tiny little program that literally had like a ball going back and forth, or like a little sort of pixel across the screen going back and forth that you could sort of use one of your backspace and then forward space sort of things to sort of hit. I don’t want to tell you how good or bad I was at it, but I do want to tell you that it worked. So, anyway, that’s my story of how really just using a natural language prompt can allow you to code.
Anna Gressel: But Katherine, I want to ask the question that all of our audience members, I’m sure, are asking themselves, which is like, are you a secret ping pong game whiz and we didn’t know this about you?
Katherine Forrest: No, oh my God, I’m so terrible. I’m so terrible. And I’m one of the people who would get so frustrated with the ping pong balls that I would squish them.
Anna Gressel: Oh no, that was me as a kid for sure.
Katherine Forrest: Right? So I think that I gave birth to two children who probably, when they’re trying to do beer pong, they’re like challenged, you know, and it’s like it’s all my fault genetically. But anyway, that was my foray into learning how to code through an AI tool, and it was really, though, pretty phenomenal because the natural language query allowed me to create something which I could then literally block copy into executable code.
Anna Gressel: Yeah, and I think one of the things that’s so impressive with AI coding is the step change we’ve already seen from the early models to where we are now. And that includes the fact that AI models can now not only write the code, right? Before they could certainly do that, but now they can check the code, trial run the code, fix errors in the code and present it as a finished product. And some of that is about the agentic capabilities that are now being built into these coding products.
Katherine Forrest: Yeah, so explain that a little bit, why the agentic capabilities are sort of useful for this.
Anna Gressel: Yeah, I think we’ve talked a lot about agents and the challenges around agents. But one of the things that is so powerful about them is that they can chain together different workflows. And some of the workflows that we’re seeing chained together are really around coding. So an example that I heard recently—I moderated a panel on agents—and one of the panelists said, you know, it’s no longer the case that we need humans to do manual code checks. We can actually have our agents run all the code checks, do the check reviews and surface any issues. So we have a pretty high confidence at that company that their AI-generated code should work well in practice because the agents have done such rigorous checks. Of course, that depends on having rigorous protocols in your company to begin with. So this always goes back to this question of how strong are your systems that you have, to start, but the agents can then execute on that in practice.
Katherine Forrest: And right now, I’m both agreeing with you and then bringing it to the present tense. You know, we have a situation where basically every major model out there can do this coding, and some are stronger in certain areas than others like with every AI capability, but you’ve got the Llama herd of models that does a lot of coding and does it really well. You’ve got Claude, you’ve got GPT-5, which in the system card—and you know I’m like a fan of system cards—in the GPT-5 system card they actually talk about it as a step change in terms of coding. So we’re really going through leaps and bounds in terms of the capabilities of these models to code all kinds of things. It’s not simply auto-completion, which was able to be done by computers long before AI was doing it, where you had sort of a string, a code string, and then you would just have the computer sort of do auto-completion if it had to be a repetitive code string for a period of time. That’s not what this is. This is sort of really from the ground up, creating the code, checking the code, testing the code and doing it iteratively, finding the vulnerabilities in the code. But there can also be some vulnerabilities, and that’s why there need to be humans in the loop, and we should talk about a couple of those.
Anna Gressel: Katherine, you’re definitely right. I mean, I think when we think about the risk surface around AI coding, vulnerabilities are the number one issue that we see come up all the time. And there are actually some really interesting reports that have come out on this recently that AI-generated code can include vulnerabilities in them. And that’s because they are drawing on information that was on the web about code. And that may include out-of-date information with vulnerabilities that have since been patched or maybe even vulnerabilities that people have kind of maliciously put out there online in the hopes that they would be picked up and used to train these models. So it’s a huge area of focus for folks in the security space right now is kind of controlling for AI-generated code vulnerabilities. And that’s a place where, you know, it’s even a bigger risk if you don’t have a sophisticated security function or a sophisticated set of data scientists that can kind of figure this out within your organization. So it’s a “watch this space” kind of issue, but certainly one that we know a lot of folks are really, really, really focused on these days.
Katherine Forrest: Yeah, you know, it’s particularly important in the area of agentic AI where you’ve got agentic tools that are utilizing APIs to sort of talk to, as we’ve already discussed in prior episodes, the outside world. And if you’ve got a tool that’s then going to create sort of a portal, if you will, to the outside, you’ve really got to be sure that things are locked down in the way that you want them to be locked down. So, you know, this is something for any company that’s dealing with AI-generated code to keep a close eye on. You know, what’s interesting is with all of this, even with some of the potential vulnerabilities, you can go online and you can see—and you read about it in the paper, you read about it in the Wall Street Journal, you read about it in Wired and Verge—you can see how much code, though, is now being created. New code is now being created with AI tools. So I’ve seen statistics, anywhere from 30% of all Microsoft code, if you go online right now, you’ll see that in a variety of publications, is created by—sorry, 30% of Microsoft code is created by AI models. And others will say that 60% of all code is created by AI models. So if you move away from just large corporations that are developers, and you’re sort of in code, the code world writ large, 60%. That’s a lot of the new code.
Anna Gressel: So transformative, and there’s no way that that doesn’t come with implications for jobs and training. Katherine, do you want to talk a little bit about this? I know you’ve been talking a lot about workforce implications of AI for a long time.
Katherine Forrest: Yeah, you know, this goes back to one of the themes of AI, which is we don’t know how it’s going to transform the workforce. And you don’t want to get all despairing about what it’s going to do to the workforce because we don’t know what new jobs it’s going to create. I always go back to my 2007 example of Steve Jobs saying, “there’s an app for that,” and nobody knew that there would be a whole world of app developers that was going to come into existence and make gazillionaires. And we don’t know what the new jobs are going to be. But what we do know is that the coding domain for low-level coders is changing dramatically. And now you still have to have these coders and engineers who are at the sort of mid-level and the sort of human-in-the-loop level who will understand what’s happening with the code or can follow it and will know how to test it and know how to see errors and know how to recognize things. So you’ve got to be able to sort of jump into that kind of training without having them start necessarily in the same way. So it’s going to be, I think, one of these really interesting areas where we’re going to see some fundamental job market shifts where new coders aren’t going to be needed in the same numbers that they’ve been needed in, but we’re still going to need coders in the loop. And so that is not a job that is going away, but other industries are going to face the same kind of sort of change transformation, and we don’t yet even know where it’s going to come out. So I’m not going to make any predictions about it. I just know that we’re going through a big transformation.
Anna Gressel: I mean, for what it’s worth, I feel like we are going to hit some amount of what I’ll call the “TikTokification” of AI coding. Other people might call it…yeah, like people call it vibe coding. But I remember with TikTok, you know, suddenly my little sister—who’s like, you know, much younger than me—suddenly she was editing videos, like at the level of a professional editor. She just went out and was able to figure out how to do these things. And I think we’re going to start seeing people much, much, much younger really dive into coding because there’s no longer such a barrier to entry. So it’s going to be interesting. There’ll be different generations that are hit differently with this, and it may depend a lot on what they’re trying to accomplish. Like maybe creating apps or creating fun code is something we’ll have all children really able to do. But then, to your point, some of these kinds of pieces of mastery will stay in other specific work functions or kind of generational groups. It’s just really interesting to think about how this might play out over time.
Katherine Forrest: You’re absolutely right. I mean, this vibe coding that you mentioned, V-I-B-E, vibe coding. I mean, it’s such a—what are we on? Gen X, Y, Z, you know, Gen, the great beyond, whatever it is. But, you know, with vibe coding, it’s sort of like, “well, hey, let’s do this and see if we can do that.” And there’s sort of a collaborative aspect to it where you’re working with the AI tool and you’re trying to get the vibe of a product and trying to work with the tool together. But it does allow people who literally are non-coders to create. And that’s going to open up enormous opportunities for folks. So perhaps that’s where some of the people who were going to otherwise go into actually sitting in a dark room drinking—you know, I used to say Mountain Dew, but I’m not even sure if Mountain Dew is around. I used to say, you know, people who were sitting in a basement drinking Mountain Dew were coding all night long. But now, they’re going to be able to be vibing with their code and perhaps creating incredible new products with that code.
Anna Gressel: Or people who would never become coders will be able to code. Like think about the archaeologists who might suddenly be able to create an archaeology program.
Katherine Forrest: Absolutely. And you know, this is all part of the policy shift, I think—or not part of; it happens to be coming at the same time, we’ll say, as the policy shift that we’ve seen in the United States with the AI action plan, which, you know, we talked about in a recent episode where it’s focused on, in part, sort of having worker transformation, worker education, finding out the ways in which we’re going to be able to have workers find their place in this new AI-centric technical domain that we’re creating.
Anna Gressel: Yeah, I actually had someone say to me recently that they would prefer to hire someone who is less good at the job but better at AI because they thought that that person would be able to learn faster. And I thought that was such an interesting point, that we may start seeing a rebalancing of the skill mix that we’re looking for that could happen in coding, it could happen in other domains. But this question is really going to be, you know, how important is AI on anyone’s, you know, in anyone’s job search or on anyone’s resume. And I think we’re just beginning that part of the conversation.
Katherine Forrest: Yeah, no, so true, so true. So the summary version of all of this is that, you know, you’ve got these incredibly capable models now that can do all of this basic coding, and it’s being utilized more and more. Mark Zuckerberg has made a number of statements that are publicly available. I’ll let him speak for himself in terms of his statements, but about the way in which coding is going to become an AI domain. So we’ve got this acceleration happening and a new world is going to be opening up with all of this in terms of the folks who were previously going to perhaps go into coding and may still go into coding, but maybe in a different way. So it’s a story really of acceleration, disruption, adaptation. It’s the AI story, it’s the first one that’s going through such a dramatic and obvious change.
Anna Gressel: Yeah, exciting, interesting, tumultuous and disruptive all at once. So watch this space for sure, and thanks for joining us on “Paul, Weiss Waking Up With AI.”
Katherine Forrest: Yeah, and look for the red leaves that are starting to come out in places like Maine and probably New Hampshire and Vermont and wherever our audience is, far and wide. And we’ll talk to you all next week.
