Things Kris Says on Podcasts

Yeah, which is like, I think, you know, anybody that works at one of these labs has a kind of their mind warped a bit, has a distortion, a distorted view of reality because they are working in the middle of this thing. It's like, it's also coming up from like kind of one of the conversations I was having with ChatGPT, where it's like, if you're someone that investigates airline crashes, like airplane incidents crashes, like, and that's your job, you're going to have a distorted view of reality where you think that airline, like airplane crashes are much more frequent than they are, say, compared to, say, car crashes, which are by like, I think, two orders of magnitude like more likely to happen. Yeah. And so it's like, even if you know it's still like less likely, you still think it's more likely than it actually is because your localized experience is, you see this all the time. You see this every day because that's the environment you live in. And I think this is also what's happening to people that are working at these frontier labs. They are just sitting in a stew of all of this stuff and hearing all of these things and hearing the propaganda and doctrinization and all of that. And so their view of the world becomes severely warped around what is happening within those companies, even if it, even if like what's happening within those companies is not going to go the way the companies believe. And this happens at most jobs. Like most places people go into, they get sucked into like the kind of cult-like system that is most companies, where it's like you become like really on board with it, or you're like, you know, you learn to hate your competitors and you're like, our product's better, or even when it's objectively not. Like, we've seen this play out over and over and over again. And just because this is like a novel thing, people think that those same rules don't apply. And it's like, but they do apply. It's the same thing. Like they're going to think that their product is the best thing out there, even when it's not. They're going to oversell their product, even when it underperforms. And they're going to find a way to spin it to make it seem like it didn't actually underperform when it did because they're businesses, they're companies, and they're collectives of people that have egos.

I think, I think, too, we've lost the plot on the, because I ran into this when I was building something, this thing that has taken for like 5 billion tokens or more. Part of it, like, I had sat down and did a design with GPT, my chat GPT. And I was like, okay, I'll have Claude build this thing. So I took the design and gave it to Claude. And then Claude went through and I was like, well, there are several problems with this design, blah, blah, blah. And I said, here's one thing that's probably not going to work correctly. And I was like, that doesn't seem right because this one did deep research. So it presumably confirmed that this API that exists could actually work. And then I was like, maybe it doesn't. Maybe the other agent got it wrong. But then I went back and I was like, oh, wait, you didn't actually look at how the tool works. You just made an assumption from some docs pages, but you didn't look at the right docs pages. So you got confused about the capabilities of what this tool could do. And then you tried to come up with a solution for that. And then you came up with this whole big plan that I did not want to read because it's so freaking long. You spit out so much. It's too wordy. Anyway, but I was like, oh, it would have been better if you just looked at the tool, but you instead looked at some docs page and then got yourself confused and then made some statements based on that. It's like, I think there's too much now. There's too much writing, right? When you have paragraphs and paragraphs of comments, like this is how I actually think that this project has taken like 5 billion tokens of cash read is because there's so many comments mixed with so much code that every time it wants to read something, it's got to like hoover up a tremendous amount of stuff in order to understand what's happening. So then it can write some code. And then it, it's just like, we, we've, we've definitely lost the plot there. I think there was a point in time when, like, it does make sense a little bit that like comments would be useful, but like the LLMs can just read the code, right? They, they actually can read the code and understand it. Like people can't do it very well, right? If I sit down and I try to read some esoteric code, good luck. Like if it's really like gnarly, I'm probably going to stumble over. It's going to be much nicer if there's just a comment there. But like LLMs are very good at understanding gnarly code. If you hand it, it's like, no, this is how this thing works. They can like reason through, especially more modern LLMs can like reason through it. So like, I don't even think we need it for writing code in that way. Like, I think ironically, we're at the point where the thing that was built to generate text, the thing we needed to do the least is generate text. Because like, I think we all forgot how expensive words are. In the same way that like lines of code are expensive, I think words are expensive. And when you only have like a 250,000 token context window or maybe like a high and a million, but it really doesn't really go as far as I think people think it does over the course of like a long conversation. You can't just be like wasting words on stuff. Like that's extra context you're eating up. That's putting pressure on the system. So it's not going to perform as well as it could if things were more concise. So I think the idea that we can just generate the docs with LLMs, it's like only if you understand how to do concise writing, which is a very hard thing, right? We know this. There's the famous quip that everybody talks about of like, sorry I wrote you such a long letter. I didn't have time to write a shorter one. Where it's like, actually getting down to concise, compact writing takes a lot of time, takes a lot of iterations and rounds. You have to choose what things you include, what things you don't include. It's a difficult thing to do. But I do think that probably the problem here, like a big part of the problem here is, as I kind of alluded to at the beginning of the show, the literacy crisis in that like people can't read anymore. And I think if you can't read, you probably can't write, which is probably a lot of this. And like, to be clear here, when I say like people can't read, I don't mean that they literally like cannot read words. Like that's not what the literacy crisis that I think anybody that's seriously talking about it is really talking about. It's not like people can't pick up a book and read the words. It's people can't pick up the book, read the words, and then comprehend what it actually says over more than maybe a paragraph or two, right? Like people struggle a lot to comprehend not even long-form writing, but short, but medium, short to medium term, like length writing, right? Where it's like a few hundred words, people are struggling to really understand what is actually being said in those words. And I think that's potentially why a lot of people who are not like the like critical or close readers can probably consume a lot of this LLM writing and think that it's fine. And probably also why people that generate LLM posts think that they're fine. Because I think there's likely a degree of they themselves can't understand why this writing is not good. Right. They don't understand where the problems are because they can't comprehend what they've written themselves outside of the concept of it that they have in their mind. Right. I'm very sure that when anybody is writing, that's just a concept of a thing that they want to communicate. I don't know if people can then write it and then read it and then reconcile whether what they wrote is what is actually on that page or not.

I mean, I think that I get what you're saying. And part of me is like, yeah, the thing I suggested where you have like this gate for at least the Go team where they say, here's the gate of when it's that. But I also think that that's a line that you cannot define, right? It's not a definable line to say, oh, this thing has moved from discussion to something we're going to implement. It's an emergent property of the discussion. Oh, okay. It seems like we're all in consensus here. Let's move on to just, let's move on to the implementation. Then someone pops up, hey, what about this? And then it shows back to conversation, right? It's a fluid back and forth thing. It's not what is traditionally thought of as waterfall, right? It's not like a once you get through the gate, you're good. It's like sometimes things are going to go back and something's going to go forward. Sometimes this one proposal that we're like, we're definitely going to accept it gets affected by a proposal that already got accepted and implemented. Now we got to go rethink this one, right? This is a very dynamic and fluid process. I think any attempt to like shut it down in a way or to draw a line somewhere, it's just not going to work. It's like when people say, oh, we have a dictionary that defines words, thinking that the word, the definition in the dictionary is the correct one. It's like, no, the definition in the dictionary is the emergent one. We all define words simply by using them. That's how language works. And similarly with these discussions and things we want to implement, we won't know the point at which something is ready to be implemented until we've had the conversation and figured out that we have talked about enough, this enough to implement it. And there's someone that's going to go implement it. Like, that's the gate. That's the thing at the end of the day. If the Go team wants to have a special gate there or a special quality level before they'll even look at something, that's fine. Maybe they use LLMs for that, right? To have a little bot be like, does this have enough stuff in it for us for us to use our very, very little or really resource-constrained team on? Like, is this worth it yet? That's fine. But there also might be someone else that just goes and says, I want to implement the same gets a patch and then also provides it to the Go team, right? This is the thing with the decentralized community that we have is that we just had to accept the fact that I don't think we can, we can't, we can't centralize this. We can't make this an authoritarian or oligarchic system. Yeah,