Post #4391927

So this is apparently needed. Dear Mastodon, While we all say "AI" regardless of whether we're talking about text, audio, images or video these are not all using the same underlying technology*. Large Language Models, LLMs, use a technology known as Transformers. They "got smart" beginning in 2017 when a paper named "Attention is all you need" was published by Google. One way of describing LLMs is that by caring for everything that comes before, you can perform the next step. Audio, Images and Video mostly use a technology called Diffusion which began in 2015 but became noticable better in quality with CLIP from 2021. That technology can be said to work by starting with noise and then seeing whether any changes you make cause an evaluation to be more or less what you want in the end. LLMs do not want to store training data in the models because that means they're less smart. Knowing the answer to a question is useless - knowing how to get to the answer is valuable. Diffusion models however don't have a problem with this. Knowing what a rose looks like is knowing what a rose looks like. Knowing what a Dali painting looks like ... yeah you get it. What has been a bit shocking to everyone is that LLMs turned out to be generally useful and not just language engines. While the intention was for them to understand text, produce text and reliably translate between languages, it seems that by using text to encode knowledge about the world we also got engines that could "think" and "reason" about the world - and that has over the last year or so caused an explosion in LLM usage in, for example, software development. This was not something first dreamed up by commercial entities wanting to sell products - but by curious programmers wanting the Next Great Tool. If there's one thing us programmers are it's _lazy_. If we can automate a task we will. *) Yes I'm taking a shortcut. There are overlapping areas, but the point still holds.