spion

Fullstack SWE. ex-Apple. Prefer insightful discussion to debate. Rust, TypeScript, Effect, SolidJS, localfirst, devops, keto, stats/science, audio/DSP

[spion @spion]

@josevalim or at the very least, explicitly marking they know this operation can fail as its unchecked, which marks that code for higher attention (unwrap / bottom / unsafe)

[spion @spion]

@josevalim People who appreciate statically typed languages generally accept the cost of (some) false positives and type annotations because they find immense value in the compile-time assurance of zero false negatives.

[spion @spion]

@perrymetzger This is not about recognizing good art from bad - its about wanting to find out more about the interesting person behind it and realizing there is no interesting person behind it, but a smear of people that served as a training source for a generative model.

[spion @spion]

@perrymetzger I did see that - someone aptly pointed out that people are equally disappointed if they've been tricked into buying a fake Monet, even if they didn't know its fake. So yes, people care about a connection with the artist.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted Its been a mix: many artists and some open source maintainers in particular have found it harmful so far.

[spion @spion]

@perrymetzger I have no such problem with brilliant software that solves my problems - don't care too much if it looks a lot like other software or whether the code / architecture is particularly inspiring or innovative if it still does its job very well.

[spion @spion]

@perrymetzger For me, art is inseparable from the artist, and because LLM generated art is often dominated by the mix of other artists work that hasn't been properly attributed, I find it incredibly difficult to connect with it.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted ... and that we want a fair approach that continues to support creative work while also enabling AI driven innovation.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted There is also an implicit claim that a LLM reading a book is equivalent to a single person reading a book. That's also not true, and the above discussion I think shows there is more work to be done to understand how we want to treat generative AI training...

[spion @spion]

@perrymetzger Yes it can be, and it can be a source of connection, or expressing your own style, or resonating with others, or getting inspired. But LLM generated software has substantially less of those aspects in many ways.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted Libraries are not the closest equivalent though - a company buying copies for employees that then do productive, metered work - thats much closer.

[spion @spion]

@perrymetzger Right. So while for a programmer, the act of pure productivity and bringing something to people can continue to be a source of joy, because its still something potentially immensely useful for many, for an artist it can indeed be (somewhat literally) soul-crushing.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted I am not saying that. I'm just saying that if we are equating the LLM to a single person reading a book (e.g. lets say a company employee) then that math doesn't check out when we run thousands / millions of LLM instances in parallel while utilizing that knowledge.

[spion @spion]

@perrymetzger I gave the example with fans that shows the above claim is not really true. People engage with art fully (artist included) far more often than they do the same with programming.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted Thats why I'm proposing the example of a company buying a single copy of e.g. a programming book for thousands of employees to use for their day job, as the closest equivalent.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted We are trying to find an equivalent to humans reading from a book and learning. To make that equivalence full, we should both consider the reading part and the utilization part.

[spion @spion]

@perrymetzger @themistern Its absolutely as creative of an activity as writing poetry. (Although perhaps not as much now with LLMs). Its just not the point when earning a living or solving a problem, because most users of most programs don't treat said programs as art, but more practically.

[spion @spion]

@perrymetzger @ARomeoSierra @MattisRedacted So why does the LLM count as a single human when reading books, when it can run millions of sessions in parallel when writing?