Patrick Collison: "What If You Succeed?"

**Harj Taggar** (0:07)
Okay, Patrick, thanks so much for being here. Welcome to Startup School.

**Patrick Collison** (0:11)
Great to be here.
Harj and I first met 20 years ago, and we started a company together. Sorry, am I giving away the introduction?

**Harj Taggar** (0:22)
Yeah, I thought this was my interview, but keep going.

**Patrick Collison** (0:26)
Well, we started a company together many years ago, and I learned a huge amount from Harj, so it's really fun to do this.

**Harj Taggar** (0:33)
All right. Well, actually, speaking of that, so when I think when I first met you 20-something years ago, at the time, your most impressive achievement, I would argue, was Chroma, your dialect of Lisp.

**Patrick Collison** (0:46)
Any Lisp programmers here?
Oh, wow. Okay. That was, I think I heard one whoop, which is more than I expected. But yeah, I really liked Lisp when I was in high school. Yeah.

**Harj Taggar** (0:56)
So what I was going to ask is, a prolific 16-year-old today could presumably just like prompt call to write their Lisp dialect.
Would you advise them to not do that and still do it? Is there any value in such things?

**Patrick Collison** (1:11)
I don't know. I wonder a lot.
Yeah. Like I was saying, on the one hand, it used to be really fun to write all this assembly and machine code and to optimize your instructions and layout and memory and everything, and now we don't have to do that anymore. Compilers do it for us. We don't mourn it too much. So maybe in the same way, we shouldn't mourn source code. We should just transcend the plane of instructions to Claudius at all. But emotionally, I miss it.

**Harj Taggar** (1:41)
How about, I think, just as I've been hanging out here with these students, maybe the question behind it is many of them are just wondering, what should they be learning at college? What is in this AI world?
How much should they be trying to learn and derive from first principles and how much should they just outsource to the AI?

**Patrick Collison** (2:01)
Right.
I mean, my model of this is cache, the CH, not an SH, where Jeff Dean has this famous set of numbers that every programmer should know, bandwidths and latencies and just relevant constants. You should have a reason about it as you build systems. And obviously, you're thinking of building any system or distributed system or whatever, like all lookups and all relevant bandwidths between different components are very different, right? And you're retrieving something from L1 cache is very different from retrieving from RAM is very different from retrieving across the network or whatever. And I think it's like that with knowledge for fine. Yes, you can ask the agent or something to compute something for you or to look something up for you, whatever. That's a hell of a lot slower than knowing it in cognitive L1 cache. And you can have way more round trips in your brain than you can muttering through super whisper or typing it out or whatever. And so I think even granting the full capabilities of the of the models, I still think this is a pretty like I think for a long time to come, neuronal lookups will be will be much faster.
And then look, if you look in revealed preference at what companies themselves are doing, whether they are companies like Stripe or the Labs or what have you, there still seems to be an enormous premium on cognitive ability. And so I wouldn't, I think renouncing that before there is evidence that we've saturated those benefits would be premature.

**Harj Taggar** (3:41)
I mean, are there specific things that maybe you personally, either personally or as CEO of Stripe, you still purposefully choose to do yourself and retrieve from your own cash, even though the agents would probably do a reasonably good job?

**Patrick Collison** (4:00)
I still write myself. Like I, I don't, I both philosophically but also specifically, substantively dislike the writing of the models. I mean, it's very interesting, right? Because these can prove the Jacobian conjecture, you know, whatever. And so clearly, they're capable of these monumental feats.
But somehow, I still haven't read the LLM essay that I found super compelling. Now, it is just very hard to like RL them in that domain because the utility function or something is kind of hard to define. But, yeah, I think writing is a pretty... Interpersonal communication and writing, I think, are still very fundamental and being able to reason sensibly in the multi-dimensional space of reality. And in some kind of indescribable way, I feel like the model is still kind of deficient at that. And so I've never, I've yet to send, you know, every tool is now trying to prompt me with pre-written suggestions, whether it's, you know, Gmail or apparently WhatsApp just rolled this out. And I think I've still sent zero of those in my life.