.. That's like saying no finite binary number can exactly represent pi, so why not just use 3.

I know that use case is trotted out in the advertisement, but noone creating synthetic ground truth for training robots is going to use a prompt-to-video world generator thing. They will continue to use physics simulation and rendering which are actually correct, not just correct-ish-looking.

I think this story will be repeating in very many communities where it turns out the joy really came from the journey rather than the destination.
It's fairly similar to the recent reaction from the math community.

as much bad publicity as possible for Flock's business model. altho we'll likely still end up with just half measures.

don't kid yourself that llm's can't produce original results. rearranging and mashing up existing stuff is a huge part of human mathematical thinking. "Formal" math is grounded in manipulating lexical symbols, which llm's are quite good at.

zing !

i saw this story up on the main /. feed (rather than here in the firehose) but now i don't.
i feel like this happens every now and then.
what mechanism is at play ?

hugo aside, i just read the earthsea graphic novel and have been planning to write the author a thank-you letter. it's super well-done. the attention to detail and the pacing and the choices of what to include and what not are done with real skill and care. her son was involved in the process. and the artistry is superb. it's been a while since i spent minutes looking at a page that's just a seascape.

^----- THIS
is the main take-away here.

it seems to me that the little blurb on most title pages like "cannot be reproduced by any means electronic or otherwise" would cover ai training, but it sounds like this ruling found otherwise.

well right, your closing sentence is the clincher : we're using too much plastic.

labeling that suggests it gets recycled when in fact it doesn't is sustainability theater.

wups, /. filtered out my angle-brackets.
should read: ... from "99% sure it's (the right thing)" to "99% sure it's (something not even close to the right thing)" ...

the summary and article here present a delightfully uncluttered surface, but for folks wanting more detail there's a possibly related 2019 paper which shows a variety of image classifiers switching their output from "99% sure it's " to "99% sure it's ", due to literally a single altered pixel in the input. i doubt it's exactly the same thing as whatever this paper turns out to be about but it gives a feel for the problem space.

https://arxiv.org/pdf/1710.088...

> .. But it can't assess authorial intent.

idk if you're using the same popular code-writing tool that i am, but it's way better than human programmers at documenting the whys in addition to the whats and hows. additionally it's a snap to have it curate the architectural and design decisions currently being introduced. the oral record going forward is going to be --much-- more robust than in the past. as far as assessing the authorial intents of existing code, it's pretty good at that too.

i don't love this situation, and i mourn the craft i've spent several decades honing, but it credit where it's due.

i think you're reading a regurgitation viewpoint into what i'm saying. i put 'think' in quotes because i'm still wary of describing them as thinking, but yeah i totally agree they're way beyond regurgitation at this point. i mean, huge portions of human reasoning are language-based. the entire field of formal mathematics is founded on symbol manipulation, for example. so it makes sense that being good at working with language can result in reasoning.