Build for changing models, not today’s limitations

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 25:29

   You know, the thing about the way the technology is moving is like every model release or every couple of weeks now, because it used to be months and before that it was years. But now I feel like every couple of weeks you get to a set of new standards of what you can do with just models generally. And once you get there, there's a tendency for people to think, OK, what are the things that don't work?

2. 25:48

   What are the things that work? and you build kind of like a scaffolding around that. And I think the thing I keep seeing back and back and back from many products and many companies is they spend too much time on building the scaffolding and the processes and the orchestration for the limitations of the models and the things they can do only for you to kind of become incredibly irrelevant the moment a new model just like leapfrogs you all together.

3. 26:15

   And everything you build, you have to throw it to the trash and it's months and it's expensive and it's also culturally and like frustrating for people to like do that all the time so i would say that the biggest thing more than like thinking about the cost is how do you make sure you don't get stuck into that suboptimal point how do you make sure you can just keep on you know preparing yourself for the next thing and the next thing and the next thing that i would say is way difficult way more difficult than kind of optimizing for whatever moment we have right now i've seen you uh describe it in the past as like you know there's there's a product company or there's a company that's building the systems to kind of keep shipping products because things will change really fast. I also saw you posted recently on social

4. 26:54

   media that 2026 has been so crazy in AI that how could you possibly really plan too far out? So how do you, how do you think about this and how do you keep your, your team from overloading or freaking out? Like everything is just going to completely change every week and this is no way to live. Yeah, yeah, yeah. Um, well, I think you've, um, you need to kind of like thrive in that. first of all you need to thrive personally and culturally and professionally in like a world like that i think personally there are some people who might not feel there's too much change and like you just personally like freak out and like you it's you get paralysis you can do anything it's like So it's kind of a self-selecting thing where, like, are you ready for the thing?

5. 27:34

   Because, like, it's going to be hard to predict. But I think I personally thrive on that a lot. So I like it. The second thing I would say is, like, really reminding everyone of sunken cost. You know, like, engineers in some cases, or I think, like, many people love to stay very close to things they've worked on for so long because they've worked on for so long. and like there's someone at some point who has to make a call and be like we're just like going to throw it away and move on like the fact that it's not helping us like it doesn't matter it's great that you spend all the time and it's perfectly beautiful code but like who cares like move on you know and the ability to move on but also learn from that it becomes really important so i think like that idea of sunken costs and and just like personality wise being

6. 28:21

   adaptable and being able to change becomes crucial and the third thing is again long-term strategy like there are things you're gonna you know are gonna be there at some point so instead of project sometimes providing the team with very concrete objectives what you can do and i think what we try to do at some point is you provide them with like boundaries you know and where they can like move and experiment and do things but it's less about being prescriptive about this particular thing and more about like well we're heading there right so if you think about video generation video generation three years ago two years ago was a very early stages i mean the first model was we released was two and a half years ago and it was it was by today's standards really bad by two years and a half it was like the best ever and a lot of assumptions could have been made

7. 29:07

   around how you write the best text prompt to get the thing that you want right eventually to for us it was like well that's a suboptimal thing to work on because there are going to be other control mechanisms around the corner that we need to craft and create that'll be much more impactful than kind of knowing obscure prompts because i don't know if you remember this but people used to write incredibly complex prompts with like words and combinations and everyone was saying the the new job is just prompt engineer like know the world's most elaborate prompts yeah that for me was like look um you can spend all the time you want on that but realize that like no prompting will just become more natural and like you don't have to do all the obscure things you have to do. And that's where we're right now. You can just say this gravity one very

8. 29:49

   naturally, provide the right references artistically and get very close, almost there in a couple of minutes. And so, yeah, it's reminding ourselves of that long-term kind of perspective.