AI is unbelievably capable at doing things now, but I still find it pretty bad at deciding what would actually be useful. That gap is becoming much more obvious to me as the models get better. You can give it a massive rewrite and it'll get through it. Features, bugs, heaps of work that would have taken me ages. Then ask it to do something valuable for the company and it picks some logging. I don't think that's a small gap. We cut thousands a month off AWS with self-run NAT instances and Docker bridge mode so we could use smaller hosts. AI could help do those changes pretty easily. I just don't think it would have decided they were worth doing. The extra work was tiny compared to the savings, and we're a small team that's comfortable managing it. That's why it made sense for us. You need that context to make the decision. The changes themselves weren't particularly hard. I wonder how much of this comes from models learning off work that's already been done. They're so good when you've given them something to work from. Choosing a useful direction without that still seems much harder. I don't really write code anymore. I spend my time working out what to build, trying it, and deciding whether it's any good. I'm having heaps more fun doing that. Selfishly, I'd quite like to keep the part where I get to decide.

The surprise usage resets from OpenAI have been pretty clever, I think. If you're in the middle of building something and they give you more usage, you're going to feel pretty good about it. I remember that sort of thing.

We cut a couple of grand a month off AWS with self-run NAT instances and Docker bridge mode so we could use smaller hosts (ENI limits) AI could do those changes pretty easily. But I don't think it would have come up with them.

I love voice typing, but I don't really get the voice operating system thing. Saying "open Codex and start a new thread" just feels more annoying than clicking it. Talking makes sense for getting my thoughts out. I don't want to narrate every click.

I don't think a model would have told me to build Vibe Typer. AI helps me ship, but getting dictation to feel good across Mac, Windows and Linux takes a lot more than generating an app. I've spent a year dealing with performance, reliability and all the weird edge cases.

I used to spend heaps of time trying to make software flexible enough for things we might need later. It was hard to know where to stop. Make it too rigid and changing it is painful. Make it too flexible and you've built a bunch of stuff nobody uses. I'm much less worried about that now because an agent can refactor it so easily. It's still work, but it's gotten so much cheaper that I don't feel the same need to solve every possible future requirement in the first version. I'd rather build what we actually need and see if it's useful. If it isn't, I haven't spent ages building all the flexibility around it. If we need it to do more, we can change it then. I'm still thinking about the design. I just don't think the hypothetical requirements need nearly as much of my time anymore. It's pretty freeing to build the simple version and know we can change it when there's a reason to.

I got my agent to compare car insurance for me. I normally can't be bothered filling out that many quote forms, so this felt like a pretty good use of it. I'd happily pay someone to handle this sort of thing. I had it go through the insurers directly and fill out the forms. Same car, same information, over and over. Very tedious for me. A pretty reasonable job to give an agent. I ended up with about 20 quotes in my inbox and a comparison table. Comprehensive ranged from about $719 to $1,000 for the same car. I probably wouldn't have found that out if I'd only bothered with the first couple. I picked the cheapest and clicked buy in the email. There was still a decision for me to make, but I didn't have to spend the afternoon collecting all the options myself. I'd rather do this than assume a comparison site is checking everything for me. They're paid for referrals, so I'm not convinced all the options I'd want to look at are going to be there. I'd happily pay $50 for that if it saved me more than $50. I think the form filling is pretty doable now. Keeping all those browser sessions working without getting blocked seems like the annoying part of making it a business.

I’ve been working 16 hours straight today. It’s worrying. Anyone else? Time for bed

based

Something is training :) And doing so on 100% solar power as well. And I'm likely going to open source it. Look forward to sharing more details soon; but will take a few weeks.