Lenny's Podcast: Product | Career | Growth
00:44:31 · Lenny Rachitsky, Keith Rabois
[00:44:31] S1: Yeah. So that's the other part of my insight recently is just as AI makes it easier to build, it's kind of expanding from the middle out. And what remains is figuring out what to build and making iterating on the idea. And then it's at the other end of the spectrum, which is distribution, getting anyone to pay any attention to what the heck you've done, because again, there's so much happening every single day. [00:44:48] S0: Yeah. Cutting through the clutter. I mean, that's always been critical. I mean, when people are pitching me as an investor, it's one of the things I'm dialing into immediately is how the hell is this going to cut through the clutter? It's one of the reasons why I don't like to take customer feedback into account because by definition, when you put something in front of a customer, that's not a proxy for the real world. In the real world, you have to cut through the clutter while they're going to their Barry's bootcamp, while they're doing their job, while they're raising their kids, et cetera, et cetera, while they're walking, while they're on the subway. An isolated fake experiment doesn't give you actionable insights and often is directionally wrong. [00:45:26] S1: Interestingly on that line, there's a few companies that are launching, Simile is one, that is simulating humans. I don't know if you heard about this company. They basically are building AI models of the actual people so that you can simulate your marketing launch and your product experience with people before you launch.
00:45:26 · Lenny Rachitsky, Keith Rabois
[00:45:26] S1: Interestingly on that line, there's a few companies that are launching, Simile is one, that is simulating humans. I don't know if you heard about this company. They basically are building AI models of the actual people so that you can simulate your marketing launch and your product experience with people before you launch. [00:45:44] S0: I don't know if I've seen that specific company. I have seen one or two pitches. My general question as a refrain on those type of companies is what are they training on? Because again, if they're not training on the right data, it's dangerous to say you're simulating humans. [00:46:02] S1: Yeah. Yeah. That is the question. Oh, man. What a weird world. Along these lines, actually, you have this, I don't know if you call it a hot take that AI content is going to surpass human content and that's just the future. [00:46:13] S0: Oh, that's inevitable. [00:46:15] S1: I was talking to someone and then she was saying in the Chinese TikTok app, it's just all AI videos now and it's very good. She actually enjoys watching these videos and these stories.
00:46:15 · Lenny Rachitsky, Keith Rabois
[00:46:15] S1: I was talking to someone and then she was saying in the Chinese TikTok app, it's just all AI videos now and it's very good. She actually enjoys watching these videos and these stories. [00:46:26] S0: I think there'd be a binary sort maybe. I could see products and content thriving that is clearly still human-generated and that there's some desire for authenticity. Just like, for example, this piece of art is a Warhol. Anybody could create this now, high fidelity, but there's still an appreciation for this was created by Andy Warhol. I think there's going to be a curated experience with a premium of provenance that you know is human-created. And then there's just going to be a rank filtering of what's the best content and whether it's AI-generated or not. It's a little bit like the Ben Thompson strategy thing of curation versus algorithm. There's going to be two polls and the curation may be for human created stuff. And then there's going to be the algorithm, which is just what's the best? [00:47:18] S1: I 100% believe that. That makes so much sense. I was just thinking the other day, it's so interesting that AI is getting very good at video. Videos are actually fun to watch that are AI-generated. And then images are getting really good, but writing is still really bad that is AI-generated. And it's ironic that it's called a large language model. It's all context, text, text trained. And it's just like that's the thing it's least good at, which is really weird.
00:47:18 · Lenny Rachitsky, Keith Rabois
[00:47:18] S1: I 100% believe that. That makes so much sense. I was just thinking the other day, it's so interesting that AI is getting very good at video. Videos are actually fun to watch that are AI-generated. And then images are getting really good, but writing is still really bad that is AI-generated. And it's ironic that it's called a large language model. It's all context, text, text trained. And it's just like that's the thing it's least good at, which is really weird. [00:47:44] S0: Well, I think part of that's an economic decision of token rationing by the LLMs. They basically are, when you prompt, they're trying to make their economics work within a bell curve distribution. And so for example, if you prompt an LLM to write something that's very short, the quality is significantly better than if it's a page to chapter to book. I think it's tokening ratioing versus actual quality. [00:48:17] S1: You would think though, I don't know if it was Hemingway of just like, "If I had more time, I would've made it shorter." It's more work. Yeah. [00:48:23] S0: But then you have ... I think the writing, right now what works best is short, many short examples and then- [00:48:32] S1: Tweets, basically.
00:48:32 · Lenny Rachitsky, Keith Rabois
[00:48:32] S1: Tweets, basically. [00:48:33] S0: ... picking, editing, reprompting. It's a little bit like when I used to be a litigator 25 plus years ago, the hardest part of writing a brief or all the art and all the magic was the first paragraph. If I could write that first paragraph really well, the chance I would win the case and convince the judge would go through the roof. So what I would actually do is I might have three weeks to write a brief. I might spend the first week or more walking around the office, thinking through, "How am I going to nail the first three sentences?" Once I figured out how I wanted to frame those first three sentences, I could write 30 pages in two days, but getting that right could take a week or two. Sometimes it would occur to me in the shower or literally the shower or in the middle of a run. It's like, "Oh my God, here's how I distill it." Then I could just sit down and power through the rest of the brief. And I think it's a little bit like that. The powering through the rest of the brief works well, you still need the first three sentences. [00:49:36] S1: Is there a story from that time in your life that would be interesting to share? I didn't even know about that.
00:49:36 · Lenny Rachitsky, Keith Rabois
[00:49:36] S1: Is there a story from that time in your life that would be interesting to share? I didn't even know about that. [00:49:40] S0: So yeah, not everybody knows. I spent the first four and a half years of my professional career as a law clerk and then litigator, which occasionally comes in handy truthfully. Sometimes trading off business risk against legal risk is a valuable thing to do. Sometimes it's not accidental that many of my best investments are in financial services in heavily regulated areas because I think I can sort of do the legal risk assessment in my own brain. This was definitely true back in the day when I invested in YouTube. But I think most things that lawyers learn are very inconsistent with being entrepreneur. So when you're graded as a lawyer, you learn every exam in law school virtually is issue spotting. So you get credit for identify ... There's fact patterns and identify all the issues and then resolve them. So you learn to identify everything that can go wrong, moderately useful, but very not useful as an entrepreneur. I mean, sure, you can identify all the reasons a company could fail, not that difficult to do. The art is solving those. So not the best training. And then you also measure your productivity by hours worked. Literally, you bill per hour. It took me two years, maybe a year and a half after converting to technology to stop tracking all my time. I used to literally write down in my notebook, "Oh, half hour of this meeting, 20 minutes on this, blah, blah, blah." I just couldn't get that out of my brain.