Things Chris Says on Podcasts

No, and that's where the sort of danger lies. Um, and AI are fantastic bullshitters in that regard. We all know that large language models were system prompted into being sycophantic for a while, and they've largely undone that, but the effect was, oh, hey, users like it when they are flattered, therefore make the AI flattering. Just horrible amounts of bullshit. But when we look at it from the user's perspec- perspective, if there is somebody who does not care about what they're-- whether what they are doing is true, and I realize that's a load-bearing word, or whether it is just effective, is a huge problem when it comes to, uh, using AI and abusing AI and abusing the output of AI. Because is this adding technical debt? Sh- sure, but hey, it raised our stock price for a quarter. Um, we all know the story of, uh, Google, and I don't remember the VP who decided to put paid results up in front of real results, and now another VP, could be the same one, who put the AI results in front of real results. Mm. But that certainly raised their stock price for a couple of quarters and has utterly ruined its credibility for users, right? Yeah. You now have to go several pages deep to get to the real references. I mean, the AI is…

yeah. Almost, almost no one will remember what this is. But back in the day, right, Letraset was really, uh, was a printed sheet of letters that you could purchase off the shelf, and they came in these little plastic sheets that you could lay down on top of a sheet of paper and rub the top of the plastic, and it would push the letter onto the paper. But that would create horrible baselines, and if you're setting a ton of text, it was a nightmare. So- On about the '80s, we had phototypesetters, which were complicated machines that had these transparent disks, each of which was dedicated to a single typeface, not even a font, but a single typeface. And you would have what was essentially a markup language. You would sit and you would code that markup language into this ancient machine, and it would tell things like, "Here's the size. Here's the position of this body of text. Here's the column width." Just like CSS now. Mm. But then you would tell it to go. You would load up some photographic paper, and then a, uh, a strong light within the machine would… Uh, it was a pinpoint light. Um, would… Uh, the machine would position that transparent thing, turn the large…

Seven years required experience. Yeah, exactly. Um, and the, the first thing I realized when I used Claude Design, 'cause I was throwing a problem at it just to see what it did and how it worked, was it sat and chewed on the problem that I presented it for a while and then dumped a bunch of screens on me. And I had this dumbfounded, stupefied moment of like, I cannot… I have no hook for understanding this, for appreciating w- what are the design trade-offs that got us here? Why is this good? And because we're user-centered designers, why is this good for the sort of people who would be using it? Was completely absent. Um, and- I mean, I had a great deal of sympathy for our clients or stakeholders who sometimes get just dumped screens on them, and without, you know, the hand-holding necessary to sort of n- explain why it's good, why it's good for users. What I wound up doing with Claude Code was I taught it about pair design, and I'm like, "Okay, don't just dump screens on me. That's stupid. We're going to pair design this thing step by step, going from user needs and technological constraints through scenarios." And over the course… And it did it, and it started doing pair design with me, with me as the generator.