4 ms·

I enjoy working with interns, you can see them learn and they are always making new mistakes. I get a return on the effort of training them. They might even con

I enjoy working with interns, you can see them learn and they are always making new mistakes. I get a return on the effort of training them. They might even convert into full time employees and take some of my workload. I don't get that feeling from the LLMs. They have about the same skill level of an intern, but they don't _learn_. I can't offload any work to them, and they take the same level of effort to manage. I'm not in this job to manage interns. I'm in this job to solve problems. Training the intern is a payment by current me for future me. This isn't like going from books to Internet search or StackOverflow. It doesn't provide an immediate benefit to me, and I don't benefit in the future from my contributions. I'm not seeing the share-alike vibe necessary for scale. I want tools that make me faster and more efficient. That's how I keep increasing my wage. Maybe if I could _see_ the AI learn from my training? Maybe if I saw benefit from the effort? However, right now, I'm paying for the benefit of training the LLM.

n4r9 2y ago
I have a similar take. As yet I'm unconvinced by claims that using an LLM and correcting the output is faster than just writing the code. What happened to "reading code is harder than writing it"? With an intern or very junior developer you have hope that in a few months/years time they'll be reducing rather than adding to your workload. With an LLM, either you're using it for really rote or boilerplate tasks (fair enough) or you spend so much time looking for subtle errors that it's not worth it. Plus, much less fun than actually crafting something yourself.
anguspmitchell 2y ago
I think the intern analogy is right. The problem is that there’s always a communication cost to outsourcing something. And you can’t outsource a whole project to an LLM, like a human intern. You’re just outsourcing one micro-, sub-task after another. To use the human analogy, it’s more like you’re standing over their shoulder and telling them what function to write, one after the other. And they’re SUPER fast with small, pure functions, but they get confused with anything else. Is that a faster way to program? Maybe? If you can fluidly decompose things into small functions in your head and the problem can be solved that way? I don’t know though, I find myself using chatGPT for bigger meta questions much more frequently than the Copilot autocomplete

n4r9 2y ago

I have a similar take. As yet I'm unconvinced by claims that using an LLM and correcting the output is faster than just writing the code. What happened to "reading code is harder than writing it"? With an intern or very junior developer you have hope that in a few months/years time they'll be reducing rather than adding to your workload. With an LLM, either you're using it for really rote or boilerplate tasks (fair enough) or you spend so much time looking for subtle errors that it's not worth it. Plus, much less fun than actually crafting something yourself.

anguspmitchell 2y ago

I think the intern analogy is right. The problem is that there’s always a communication cost to outsourcing something. And you can’t outsource a whole project to an LLM, like a human intern. You’re just outsourcing one micro-, sub-task after another. To use the human analogy, it’s more like you’re standing over their shoulder and telling them what function to write, one after the other. And they’re SUPER fast with small, pure functions, but they get confused with anything else. Is that a faster way to program? Maybe? If you can fluidly decompose things into small functions in your head and the problem can be solved that way? I don’t know though, I find myself using chatGPT for bigger meta questions much more frequently than the Copilot autocomplete