StefanMay 27
People defending AI-generated “art” because AI is also used to find medical cures is like saying that smashing people’s toes with hammers should be acceptable because hammers are also used to build hospitals.
Show thread⁂ Jnk ∞May 27

@StefanThinks half the replies mixing the general ambiguous term "AI " with LLM and diffusion models are missing the point in a way that proves said point.

Ain't this a whacky time to live in

As a side-note: No, generative AI is not used to do find cures to anything. They use models focused exclusively in data analysis, and that kind of AI is proving to be an absolute game changer. No hallucinations or stealing involved. Remember how we got covid vaccines in less than a year? That was a lot of hard work and proper use of machine learning.

Show threadDavyJonesMay 27

@jnk @StefanThinks
> generative AI is not used to do find cures to anything. ... Remember how we got covid vaccines in less than a year? That was a lot of hard work and proper use of machine learning.

AlphaFold was instumental in finding the structure of SARS-CoV-2, and that is generative AI with very similar maths to the early LLMs at least.

Show thread⁂ Jnk ∞

@DavyJones @StefanThinks correct, AlphaFold was relevant; but that is not a generative AI model, and has little to do with LLMs.

Yes, both are based on neural networks, which has similar math at first. Both look for patterns and LLMs do recognise (not to confuse with understand) language incredibly well, just like AlphaFold recognises protein patterns incredibly well; but that's literally everything they have in common.

After that, both LLMs and diffusion models try to reconstruct and make up data based on the mess they just made. The data that AlphaFold returns still has to be studied by professionals to be useful, making it just a valuable tool; which gets us back to @StefanThinks 's post

May 27 at 2:21pmWeb
Show threadDavyJonesMay 28

@jnk @StefanThinks
> AlphaFold was relevant; but that is not a generative AI model

Accurate structure prediction of biomolecular interactions with AlphaFold 3 https://doi.org/10.1038/s41586-024-07487-w

> this is a generative training procedure that produces a distribution of answers
> The use of a generative diffusion approach comes with some technical challenges that we needed to address
> We note that the switch from the non-generative AF2 model to the diffusion-based AF3 model introduces the challenge of spurious structural order (hallucinations) in disordered regions