In a sense, the decision is somewhat made for us in that we're developing next-generation stuff that LLMs don't know how to auto-code at @spritely. We are working on core infrastructure that needs to be carefully thought about and written. LLMs introduce a lot of errors and aren't good at doing this kind of work on their own.

And the goal was always that our work is there to be lifted from, to spread outward, the way people have long drawn from the well of the MIT / Stanford research labs in CS for decades, but for decentralized networking today

But doing it now, in this way, in this environment, it's just really depressing and demotivating.

@cwebber

“Code/Software has no value since it can be duplicated at no cost”

The first time I heard this was 20 years ago and this has always been true.

@cwebber For what it’s worth I think that we are eventually going to recognize “needing to throw massive computation at things” as a symptom of language and discoverability shortcomings that we’ll find better ways to address. We already package utility up in libraries and deterministic generators, but finding and learning what resources do what remains difficult.

I think there’s still a better future out there where solving new problems is still a non-captured contribution to the common good.

@cwebber LLM is the past, written in stone. You'll never get a picasso or an andy warhol out of LLM, because it's a blender for existing data.

Your post reminded me of something else. For a time I thought that ML could be good for turning video into 3D skeleton animation. But that only work if I want realist animation. If I want waluigi doing a 100m sprint, I need an animator. Which need to train. Like picasso did train in classical art.

No art under AI, unless it's more of the same.

@cwebber I've been thinking recently about how people are shown to lose skill at a task when automation is introduced to aid that, because they quickly come to rely on the tool rather than trust their own judgment. There was a study, for example, at how the introduction of AI detection models for cancer screenings actually lowered the overall accuracy rate, because no one was checking the machine's output.

Anecdotally, I've seen the same thing with programming, where people will defer to Claude when a problem is encountered rather than thinking critically about it themselves. And so I suspect we'll reach a point where actual skill at this job will be in demand, because people who rely on vibe coding won't be able to reliably fix the issues that they've introduced.

On a more personal level though... the reason I do this is because I enjoy the craft. Solving problems elegantly with code is deeply rewarding to me. And so I very much don't understand why someone would vibe code to begin with. Do they not like the work? And if not, why did they choose this career path? But yeah I agree that the whole situation is super discouraging.

@cwebber yes, it will be slower to move without genai and i also find that depressing. Even more so because one feels rushed and less engaged. It is both more stressful and more boring.

But, one will hopefully not always have to use the latest frontier models. I don't expect to see super human AGI so at some point older commodity models should be good enough.

At that point I can start crafting again by providing a scaffolding/spec for the LLM to implement.

A prompt like "Write an operating system" is not a spec nor is an unstructured back and forth dialogue. I expect to write specs in a formal language - not unlike a programming language, which is then pre-processed by some standard LLM on a low temperature setting ...

I wonder what that language will look like...

By that time, of course, LLMs will have tought me emough about running a non-software business that my lifelihood no longer depends on coding.