← Back to context

Comment by echelon

4 days ago

Anyone who hasn't tried AI for coding absolutely should.

This is the future.

I don't think we'll ever manually write code again. It's just so much faster.

> I don't think we'll ever manually write code again. It's just so much faster.

If velocity was the most important criteria, well, we could always write tech-debt faster, we just chose not to.

Unless the LLM/agent is carefully curated, it will produce tech-debt faster than it can fix it.

For some products, it seems not a problem - you just want to validate PMF on a product (of course you'll have a new problem now, which is that everyone with $20 to spare can do the same).

For others, a longer-life product is preferable. We shall have to see how things shake out. My best guess would be that we have more useless stuff that is free or close to free, and fewer useful stuff that is free or close to free.

> This is the future.

NFTs and crypto were also the future.

> I don't think we'll ever manually write code again. It's just so much faster.

More work for the people who like to fix tech debt.

  • > NFTs and crypto were also the future.

    I find this the least convincing argument ever. Its only a gotcha if you assume all/most of the people excited about one were excited about the other. Personally I never met a real person who gave a shit about crypto, much less nfts. But AI interest is everywhere, with it roughly 50/50 in my life of people who are uneasy with it vs use it regularly.

    I don't disagree about monumentous amounts of tech debt and risk being created. Its my hope for my own job and skills being relevant going into the future. I do like playing with it, understanding it as a tool. But it is just a tool, not a machine god, and regularly fallible.

    • > crypto, much less nfts

      Actually I met one guy who was somehow deep in NFTS when this Boring-Ape-NFT took off and he told me how much money he has now (on paper) - then they were vaporized and he lost everything.

  • You need to try it before you knock it.

    I was a doubter. This will literally work 100x faster than you. It can one-shot 1kLOC across dozens of files in mere minutes and understand the context.

    You'll need to pay back a lot of those performance gains in reviewing the code, but the overall delta is a 2x speedup at minimum. I'd say it's closer to 4x. You can get a week's worth of work done in a day.

    A human context switches too much and cannot physically keep up with these models. We're at the chess take off moment. We're still good at reviewing and steering.