Post-training favors tasks with short feedback loops

Transcript

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

1. 41:30

   And number four. Yeah. Can I ask, we've seen this real shift, I think, in emphasis, but also in resource allocation from pre-training to post-training. How do you feel about that? And do you think that is the new normal? I don't know. I think post-training works because there is a lot of tasks that people do in real life that have fairly finite sort of feedback mechanisms. So coding is like a great example, right?

2. 41:56

   Like you can do coding with reinforcement learning pretty well because, you know, you can run code in sandboxes. And a lot of labs are running modal sandboxes for that exact purpose. Start to look at cyber security is another area. Math, obviously, with Lean, I mean, we've seen some fantastic product. I think it's kind of incredible, all these like OpenAI proofs the other day. So law, maybe to some extent, you can, you know, do like eval.

3. 42:16

   So I think post-training right now, like there's like such an incredible focus on it because you have all these tasks that traditionally only humans can do but are actually well within reach for agents and for LMs. But I don't know if it extends beyond that. I don't know. There's a lot of tasks out there that have way longer feedback loops that I don't know if post-training quite yet is within reach.

4. 42:37

   A lot of people ask me, what's going to happen with software engineers in the long term? And I think we're probably going to automate everything that an agent can do within 24 hours or a week or something like that. But I kind of feel like engineers, they work on a longer timescale and they have feedback loops that are longer. So I have a feeling like reinforcement learning, we're going to start to run into some limitations around that.

5. 42:56

   But I've been wrong for so long, so maybe I'll be wrong. It just keeps going and eventually we can do reinforcement learning on things that take six months for a human. I don't know. What does the day look like in three years for an engineer? I mean, I'm extremely bullish on software engineering. I've been coding for 30 years and I've seen so many different changes with like programming languages and the cloud and so many different frameworks.

6. 43:17

   And every single time we make programmers more efficient, it just increases demand for software. There's so many parts of the world that desperately need software. And so I'm very, very bullish on software and the need for software engineering. And I think one of the biggest misconceptions from the outside is that developer is about translating from English to code. That has never been the job of software engineers.

7. 43:37

   It's to understand the requirements in the first place and be kind of an interface between the humans. You're building for humans at the end of the day and trying to figure out what does that even mean in terms of processes and workflows. And that continuously moves up in the stack. you know back in the 60s we used to write assembly and punch cards and then you move to kobold and then to c plus plus and then to python and and now we're using agents and that's just like one like it's definitely The biggest thing that I've seen in the last 30 years, but I do think we have many, many analogs and many parallels throughout the history of software and computers. To me, that makes me feel like, you know, software engineers are going to be fine for the next 10, 20 years.