Post

this is the kind of statement that - did people like this care to actually have a normal conversation with people who they describe as having psychosis - would be downgraded at the very least to “we don’t know” but that would require allowing your priors to be. challenged

WADR i've never seen you challenge your priors on this issue.

my priors got challenged in january of this year. i was firmly on the side that most of this website was on in december of 2025.

there are a lot of people that are both my friends and people i know online, and i have a posting history to show, that i was insanely skeptical and fed up with AI hype in 2025 and before. i used to look down on people who were using it in software.

In 2026 it was already useful for problems humans can't possibly do within the budget given. We've migrated 600 ingestion pipelines for data feeds some over couple terabytes from one SQL db to a completely different.

It took a bloody murder for my effort, but I beat it into working through it. And it went through all 600 with GPT4, o1 and later GPT5.

If you social-engineer an adversarial conflict situation, you can squeeze material quality gain, at a cost of token increase.

This was my turning point in feb '26, not super clear from the text within that it represents such a shift, but it's the first time I posted positively about capabilities, and they only got more impressive since then, for better or worse. bsky.app/profile/did:...

I didn't ship any code written with LLMs until Feb '26, stuck to my guns on writing it all by hand (while occasionally testing the current state of models), last in the company. I'm not "wrote no code by hand in months" like some, but nowadays a much smaller part of my work is typing code manually.

I think there's a genuine disconnect between programmers and creative career folks on this. I have genuinely tried to incorporate this shit in my process and it is basically useless for me. Pointing that out does not mean I haven't challenged my priors.

i cant pretend to know what exactly you'd need these tools to be able to do for it to be useful in your life, but i fully believe its possible they are not that useful for you at present! i dont think that contradicts the claim that we dont understand the reasoning and motivations of agents

one thing i will say is that things are changing fast enough that if your priors are more than a month old, you need to challenge them again! astra for example is a very significant step forward for a lot of people's tasks.

It was goofy when generative AI came out and they were bad at math and OK at art. "Doesn't that seem right". Turns out it wasn't right. Like OpenAI spent billions on video generation, had a deal with Disney and then late 2025 stopped all that and announced a pivot to enterprise/coding.

OpenAI also had a pretty weak video generator, though. There are several other video generators out that are much better. I think OpenAI just didn't have much luck with their approach, for whatever reason

I'm a poet who does software development to pay the bills. LLMs are a useful (not revolutionary) additive tool for the latter, but they don't serve any legitimate role in the former. Any work they could do displaces either myself or another human, which defeats the entire purpose of making art

A key thing here is that professionally, you are a software developer; you write poetry as self-actualisation. Clearly there's no point in getting the computer to do that, any more than it would make sense to send a robot to do your gym routine.

There is one possible use for them there: I’ve used LLM to help me translate to English Finnish poetry that likely wouldn’t ever get a professional translation. The end results are very much better than what I could’ve done myself. Note I don’t just press the button, it’s an iterative process.

That's a part of this conversation that I've only seen raised recently: there is a ton of domain-specific training information (every public GitHub repo, every language's public docs) for having LLMs generate code, and it's possible to automate the loop of generating and testing code.

This has led software devs to wildly overestimate how useful these tools are outside of their field. They often don't seem to realize that there were a ton of structural and data-availability advantages for LLM use in coding that can't be generalized to any other field