Upon insertion do you incrementally build the tree, or do you rebuild it from scratch?

How did you ensure the distribution of nodes was about the same with the chunker?

The resulting decisions are fed into the coding loop with guardrails derived from those decisions. The agent one-shots features once it goes into the coding loop.

When that happens, do you read just the spec, or do you also need to read the code? Is there a difference between "I can remember what I intended" and "I can predict what the system will do in a situation the spec didn't cover"?

Interestingly, you said the spec author must be you. What happens when you join a codebase where someone else wrote the SPEC, or where an agent wrote the code and nobody spec'd it? Is the spec still sufficient, or does the "you must write it" part mean the understanding doesn't transfer?

Not being familiar with the piece, so I looked it up. Seems like it was a protest correspondence aimed at a specific institution, not a public artwork. It doesn't appear to have been exhibited, reviewed, or discussed in print in 1840. Mostly circulated privately.

So I don't think it's a good test. Because you're asking me to point out an equivalent piece in the billions of generative art produced in these six years that would be judged by future historians to have great artistic value. If we were both back in 1840, we both wouldn't be able to point out "Self‐Portrait as a Drowned Man" either. It never posed itself as art at the time, and we'd have to know what future art historians thought.

Yes, I stand by the "lacking perspective" comment, because your judgements on photography and what's considered art are from the current modern point of view, rather than putting yourself in the shoes of those judging photography when it first came out--which is the shoes we find ourselves wearing when it comes to judging generative images, because the alternative (retrospectively judging it from the future) is unavailable to us.

Hence, my core argument still stands. In what way is photography different than generated images, where in photography is considered art because humans can exercise their creative judgement, but in generative images isn't because humans cannot exercise their creative judgement?

To argue against my own stance, the tact you should have taken could have been:

If one major component of art is the human judgement, how could humans could be credited for exercising any judgement at all when prompting, when LLMs are at the end of the pipe? And if they're at all intelligence, what's to say they're not superseding our intentions with their own?

Does the lack of an existence proof for fine art photography in the first six years mean that photography will never be fine art? Obviously not for us living in the future.

You'll need to propose a different, more honest test.

> As the photographic industry was the refuge of every would-be painter, every painter too ill-endowed or too lazy to complete his studies, this universal infatuation bore not only the mark of a blindness, an imbecility, but had also the air of a vengeance. I do not believe, or at least I do not wish to believe, in the absolute success of such a brutish conspiracy, in which, as in all others, one finds both fools and knaves; but I am convinced that the ill-applied developments of photography, like all other purely material developments of progress, have contrib­uted much to the impoverishment of the French artistic genius, which is already so scarce.

Harsh. It wasn't a fringe view either. I encourage you to read how much creatives despised photography.

This argument was transferred to film when it first came out as well. It took a long while and a couple court cases before it was accepted that it took editorial choice for a photograph and film to be good. When it was accepted that to be the case, despite the aspects of mechanical reproduction, it was also accepted into the pantheon of fine art.

So it will be for AI generated images and video. So far, I've seen nothing inherent in the properties of AI generated media as a medium that obviates a human editorial eye. Even if AI could exercise comparable judgement, we will like an image because we want the curation of a specific person.

In Grant Sanderson's interview with Dwarkesh Patel about the future role of mathematicians (36:25), he says:

> One interesting take that I've heard about what mathematicians will end up being is that it's actually more analogous to art museum curators than anything else. The AI solved the thing, so the art exists. They even know how to explain it really well. But you still want someone to help you navigate this nearly infinite space of what ideas are worth engaging with. Even if AIs were in some sense better at that, I think we would always still prefer a human that we had a relationship with, because the way we get motivated to be interested in things is a social phenomenon. If you have some specific technology you're trying to build, that might be different. But the people listening to this podcast trust your curation on what's an interesting topic in the first place. It's not that they're landing here because whatever your next topic is, that's what they wanted to understand in a prior sense. They're trusting you as a curator. So my role, and arguably that of other mathematicians, might actually just shift subtly into that curation direction of what ideas are worth pursuing. That's a lot of my job right now.

Hence, I think there will be a role for AI generated images and video to be elevated into the fine art in the next couple of decades, and we'll see them in museums and exhibitions--as well more mundane uses like what's sitting on your phone right now--just like how photographs are used today.

And if he's to be believed in 1859, photography isn't art.

But today, it's widely accepted that it is. And in all likelihood, you think it is too.

Why is that? And if you have an answer, you'd have to argue why AI generated images wouldn't follow the same trajectory.

And somehow, it's a leap of logic that just because the AI image is generated that the text is also. It's as silly as asserting because someone used a camera, they hired a ghostwriter.

Comments URL: https://news.ycombinator.com/item?id=49147141

Points: 3

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49134119

Points: 3

# Comments: 0

I know that the crew uses the typical seating time to refuel and load luggage in the underbelly, but one can dream of not having to wait to get off the airplane.

I do spend a lot of time up front specifying what I want. My prompts aren't one-liners, but rather an interview process where we work through all the open questions and ambiguity.

I haven't hit the wall that the OP talked about (when agents just can't seem to make the right changes, and it's impossible for me to go in and change things manually). I used a lot of guardrails such as plan reviews, browser-based QA, adversarial reviews, unit tests, linters, typecheckers, post-commit hooks, and formal method traces. I also specified engineering principles that steers the code base to minimize state and side-effects: functional core; imperative shell, make impossible states impossible, use pure functional style, etc.

There are times, when I can feel a part of the code base is messy without looking at it, because the agent will make recurring mistakes in the same part of the code base over time. What I found the agent was doing over time is that it's been layering state variables as requirements were discovered. So what helps is to ask it to refactor all these state variables into a single sum type. And if the state machine for it is complicated, I'll ask it to write a formal model of the state in Quint. Then I'll generate traces that get run as unit tests, and ask it to write the code against that.

So while the code base isn't exactly Brownfield, it's over a year old now. As for the code base, there's a backend and a frontend. I think it helps that I established a clear pattern I wanted. You code are like memes: agents will just copy patterns they see in the code base. When it does have to create a new part of the system, I found Sonnet-level models tend to draw system boundaries in all the wrong places. Opus is better. I don't yet know about Fable.

Happy to answer any questions about my workflow.