[nevi-me]

I maximised usage and have reached the limit. I feel like I did 2 week's worth of hobby work over the last few days.

I got Fable to write multiple plans, spent a great part of the weekend reviewing them. Then with superpowers I left must of those plans executing with little intervention over the past few days.

I struggled to get Opus to just keep going without trying to convince me that it's late.

[randsorex]

"I don't use fable, I just read about it on social media. I read it is all hype and marketing so I don't believe you"

-- 50% of this forum

[donw]

I'm really curious, what sort of work you were having it do such that you could execute multiple days of work with minimal supervision.

[nevi-me]

I maintain a transit website as a hobby, and I'm building a Flutter app, from scratch. The old pre-COVID one carried mental baggage.

I spent a few weekends building comprehensive plans, designs, user maps, etc with Claude. So it has enough context to make decisions and keep going.

One session lasted over a day, I imagine partly because Fable + superpowers feels slow. I have an app on my phone that I have been test running since Monday on the bus.

What really helps (not sure if Opus used to do this) is that Claude will run through the emulator on its own, verifying that the design aligns with the Figma design system we created.

This is all building on top of 15 years of existing backend and rich features, so it's not a "build me a transit platform from scratch" where AI can end up making bad decisions.

[jakswa]

The best project I found to throw it at was cloning llama-server's web UI essentially in one shot. I'm not sure what I'll do with 5 extra days, maybe try to imagine some complex features. I'm no longer surprised at how much seems to work in these "new brain what can it do?" test, and instead think the risk is feeling like I have to take it the rest of the way once I've sunk the token cost :| https://inkcap.click

[somenameforme]

Yeah, I've been quite skeptical of LLMs, but I was wrong. The new models are capable of generating good quality code reasonably reliably. They still do some dumb things, but it's becoming more capable of one-shotting non-trivial stuff. I've no idea what software development will look like in a decade beyond absolutely nothing whatsoever what it looks like today. Even if we get sublinear progress from now on out, the current SOTA is already enough to redefine coding.

In general I expect the value of software, as a thing in and of itself, to sharply decline. With no barriers to entry, having software that just does something competently will no longer be worth much of anything. It's unclear what that will mean in the bigger picture. It'll also be interesting if this proves correct, given that software companies are largely the ones dumping so much money into this. Another probable outcome is major damage to the support-as-a-service model which again is going to directly affect many of the companies directly enabling this.

I guess the logic is that if you control the systems creating this, you control everything they're used for. But it seems equally obvious that free/local models will catch up to the SOTA today - and eventually tomorrow, so that's not a particularly realistic vision for the future.