They do actually admit that about constrained decoding somewhere in the docs. They argue it’s useless in practice because when the constraints actually kick in it harms the output too much and that it’s better to just error and retry in those cases.
That does align with my experience, though we’re not using anything close to frontier for these sort of tasks.
I am interested if it can actually improve on that. As an engineer i like the elegance of guaranteed output, but the retry works pretty well in practice.
Start with a single reviewer who is invested in the topic has been helpful for me. Occasionally it is hard to even find one person with how scatter brained everyone seems to have become lately, but depends on who you work with I guess...
> next to none percentage wise gets much royalties anymore
While artists get paid less then many would like, royalties make up roughly 2/3rds of revenue of all the streaming giants, which is a huge amount of money. This recurring monthly, mostly passive income is still helpful for smaller professional bands in between touring.
When the indie band Xiu Xiu pulled out of Spotify many were quick to say it didn't matter that much except for exposure, but in interviews they pointed out that they lost a significant portion of they're yearly income from doing it just from the streaming royalties.
Many will say touring is where the real money is, but costs are rising and what I've heard its not as lucrative for smaller bands as it was in the past.
Some links:
Spotify's own report on payouts, obviously a biased editorial, but good data nonetheless (for example 13K artists made 100k/year from spotify royalties alone) https://loudandclear.byspotify.com/payouts/
A transparent report on tour costs from a fairly popular indie band (notably they could make more if they were less principled or were a smaller band, but I think this report is still eye opening in terms of general costs) https://stereogum.com/2497822/los-campesinos-break-down-what...
BERT came out in 2018 and that’s a pretty important inflection point. It didn’t cause a pop culture frenzy, but in NLP circles it was a ‘magical’ improvement.
From technical POV that's true, but that was still a niche area at the time, mostly ignored in the broader tech community. So wrt. the broader tech world discussing harnesses, I'd still use November 2022 as the reference point.
In linguistics this sort of thinking comes from 'basic color term' theory, which lays out heuristics for deciding if a word for a color in a given language is 'basic'. 2 things going against these blue-green terms are:
* They refer to specific objects (a duck and a stone), eventually these referents can be transcended though, like with the case of orange.
* Their frequency is roughly similar to each other (along with cyan, aqua, etc.), so there's no one term for this range (e.g. there's no doubt in a corpus of English that red is the basic color term for its spectrum).
— As Vice President, Kamala Harris was a key proponent and promoter of the 2022 CHIPS and Science Act, which aims to boost U.S. semiconductor manufacturing.
This is exactly the process happening in the music space with Suno. Go to their subreddit, they all talk about how they only listen to ‘their’ songs, for the exact reasons you list.
It is very different with music. Music and images fall into "just shit something and I don't care what is is" category.
Most people prompting for things in this category will be satisfied with anything, they might not admit, but the degrees of freedom the model has is infinite. Now when you pin the output, let's say a character you generated, and ask for modifications WHILE KEEPING lots of characteristics, you reduce the degrees of freedom from infinite to a small, very constrained, set of states. There are workarounds but natively llms can't really do this. You ask the model to rotate an image, the hair becomes blue and the sword becames an axe.
With music this is much more pronounced because most people are musically illiterate, so even the basic mistakes while dragging characteristics over diffs becomes invisible. It's an interesting phenomenon I agree, but it says more about lack of taste and illiteracy of the common individual.
But on the point of "thinking hard", with music and artistic production in general, individuals (human with soul, not npc) crave for ideas and perspective. It is the play, the relationship between ideas that are hard to vocalize and describe but can be provocative. Because we cannot describe or understand, we have no choice other than provoke into another a similar contemplation.
But make no mistake, nobody is enjoying llm slop. They have fantasies that now they can produce something of value, or delegate this production. If this becomes true, instantly they lose and everyone goes directly to the source.
Art is specifically about communicating the inconceivable, cannot be delegated. If the tool is sufficient to produce art, then the expression is of the tool itself, now they ARE.
That does align with my experience, though we’re not using anything close to frontier for these sort of tasks.
I am interested if it can actually improve on that. As an engineer i like the elegance of guaranteed output, but the retry works pretty well in practice.
reply