Latent Space: The AI Engineer Podcast
00:27:11 · Shawn Wang, Mihail Eric
[00:27:11] S0: Right. Yeah. Yeah. [00:27:12] S1: Right. So, yeah. It's a... It it... Not really. I think they're... I haven't got... Received any, like, pushback on on, like, this idea at all. I think they're actually more just... Most of them are consumers. Right? Most of most of the students do just, end up using the tools. I I don't think that they get a lot of exposure, if any, certainly in the coursework around peeling that onion back. Right? And I think... So this is the week of the harness. So so so Tuesday, first lecture, you know, logistics, blah blah blah, all that stuff. But then first half is... The first true bit of curriculum is the 200 lines of a harness. And then, yesterday, when I had sort of the follow-up to that, is we actually took Claude Code, and then you have these proxy things, these proxy tools where you can intercept [00:27:53] S0: the request. The MITM proxy?
00:27:53 · Shawn Wang, Mihail Eric, S2
[00:27:53] S0: the request. The MITM proxy? [00:27:54] S1: Yes. Exactly. Exactly right. So use that. Use that tool, intercept the requests and responses. Nice. Look under the hood, see what's happening there. And then then just like, okay, I did say you have Claude code in 200 lines, but really, okay, here's like actual Claude code. And so what I literally did is I took Opus five fives, this model of two days old, and was like, this is... I intercepted those requests. Here's exactly what's going on. Here's the prompt. Here's the system prompt. Here's the plan prompt. Here's the compaction prompt. And then we just went through all of that and sort of showed it to the students. So now they really, really have a state of the art harness that they can understand. [00:28:24] S2: Yeah. It's nice to get those three levels. Students are very much consumers. Here's the baseline agent in a loop and then the real one. I like how you have a video walkthrough of your blog post as well. Do you see students are on like extremes of some people are power users. They consume a lot for some people. You could assume that this is kind of very straightforward if you know how agents work. How do you kind of balance that in the classroom setting? Do you tailor towards the average?
00:28:56 · Mihail Eric, Shawn Wang, S2
[00:28:56] S1: Yeah, I think it's a bit self selecting because the curriculum is all here. So the curriculum has been available since, mean, left at least a few weeks. And I guess I announced it a few weeks ago, but it had been in the works for probably least one to two months. So if you looked for it, you could probably find it. And I feel like the reality is if students are not interested in this, then they won't... They just won't show up. They choose not to show up. What I'm finding that is interesting is there are definitely students that show up, but then there's also just a lot of from, you know, broader Stanford postdocs at other departments. There's a lot of, like, interdisciplinary sort of interests where folks that are in, you know, somewhere in pathology or, like, you know, medicinal sort of fields or or even, like, random other engineering fields that aren't pure CS that are coming in and also interested in kind of getting exposed to this. Also a lot of, GSB students, MBAs, that also, wanna, I guess, learn how to not have a technical co founder or whatever the motivation there is. [00:29:52] S0: Yeah. So I I think... By the way, if anyone hasn't done the MITM proxy exercise Yeah. You should do it. [00:29:58] S1: Yes. [00:29:58] S0: You'll be shocked. Claude Code has the longest system prompt of all the coding agents, something like 20,000 tokens. [00:30:05] S1: Yeah. Yeah. Although, yeah. Yeah. That is true. That is true. Yeah. Yeah. Although they've also reduced [00:30:09] S2: the size. [00:30:09] S0: They cut it [00:30:09] S2: by 80%.
00:30:09 · S2, Shawn Wang, Mihail Eric
[00:30:09] S2: by 80%. [00:30:10] S0: Yeah. Yeah. I know. But then they grow again, then they cut and then grow again. [00:30:14] S1: It's it's a seesaw. [00:30:15] S0: Yeah. Yeah. No. I I think it's interesting. You'll be surprised, like, what goes in there that, actually, sometimes you don't use. Mhmm. Sometimes it actually hints at the internal workings of some of the tools. Anything that you found there that could be cool to talk about?
00:30:28 · Mihail Eric
[00:30:28] S1: Yeah. No. I mean, in fact, yesterday when I was talking about this, it was a little bit of this, like, one of the things that I thought was most interesting about Claude Code back in the day, and I think part of why it really blew up in terms of popularity, was their focus on simplicity. Right? So in the early days, they're among the first to really lean into this, let's fewer tools do more. Right? Lean into SQL, lean into Bash, lean into, like, the things that are that are powerful, like Swiss Army knife tools, and have that be a lot of the driver of capability. I think they did a fantastic job leaning to that. I think that's why they they were as popular as they did. I think that when I presented something like the system prompt, it was like, well, there was a different system prompt about two months ago. It is now... They have reduced in complexity, and it turns out that now what was once part of the system prompt, things like how to use even Git, right, or how to, like, set up tasks things of that nature are no longer part of the system prompt. It's actually expected to be more kind of captured in the models, like the model weights itself. So they've kind of baked in a lot of those learnings more at the model level. And that just raises interesting questions around, you know, the model versus the harness. Right? And how do you kind of think through that trade off? What do companies invest in? What should you invest in to really push the frontier of your own capabilities? And so I I think we still, even as a as a community, we seesaw between the two a lot.
00:31:39 · S2, Mihail Eric, Shawn Wang
[00:31:39] S2: I think you don't want to fault them for growing a system prompt and then cutting it. At some level, it's also how the model was, right? Weaker older models, you need to give them examples of this is good, this is bad. And then obviously you cut that away at some point, right? And that's just part of learning what has been done, right? You can apply the same towards, if you're using a local coding agent with not the strongest model, you might need to add these examples, right? [00:32:05] S1: That's right. Is that correct? Yeah. Yeah. And then, of course, there's interesting AI learnings to fall out of that, which is why, as the models get better, when are you over constraining it by adding those examples, right? [00:32:16] S0: And then how you [00:32:16] S1: can pull away the over constraining. That's, I think, part of the learning as well. [00:32:20] S0: Okay. Week two, you're going to talk about context engineering. Repet, I haven't seen, but there's many versions of spectrum development. What would you highlight to a broader audience about what you're putting in here that people should know? [00:32:35] S2: Yeah. This is like the one hour version of the course. Hey, exactly. [00:32:40] S0: Please condense everything. The answer [00:32:42] S1: is going just... That's fair, yeah. Like... [00:32:46] S0: Well, yeah. It's it's also a trailer. Right? It's like, well, if you want more... [00:32:48] S2: If you're signing up next next year or next fall for courses, you know, this is your one hour.