this post was submitted on 09 Jul 2025
125 points (100.0% liked)
Technology
72733 readers
1534 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
There really needs to be a rhetorical distinction between regular machine learning and something like an llm.
I think people read this (or just the headline) and assume this is just asking grok "what interactions will my new drug flavocane have?" Where these are likely large models built on the mountains of data we have from existing drug trials
Life sciences are where this sort of thing will shine.
Reproduceability is always an issue.
?
Reproducibility of what we call LLM 's as opposed to what we call other forms of machine learning?
Or are you responding to my assertion that these are different enough to warrant different language with a counterexample of one way in which they are similar?
Those models will almost certainly be essentially the same transformer architecture as any of the llms use; simply because they beat most other architectures in almost any field people have tried them. An llm is, after all, just classifier with an unusually large set of classes (all possible tokens) which gets applied repeatedly
A quick search turns up that alpha fold 3, what they are using for this, is a diffusion architecture, not a transformer. It works more the image generators than the GPT text generators. It isn't really the same as "the LLMs".
I will admit didn't check because it was late and the article failed to load. I just remember reading several papers 1-2years ago on things like cancer-cell segmentation where the 'classical' UNet architecture was beaten by either pure transformers, or unets with added attention gates on all horizontal connections.
I skimmed the paper, and it seems pretty cool. I'm not sure I quite follow the "diffusion model-based architecture" it mentioned, but it sounds interesting
Diffusion models iteratively convert noise across a space into forms and that's what they are trained to do. In contrast to, say, a GPT that basically performs a recursive token prediction in sequence. They're just totally different models, both in structure and mode of operation. Diffusion models are actually pretty incredible imo and I think we're just beginning to scratch the surface of their power. A very fundamental part of most modes of cognition is converting the noise of unstructured multimodal signal data into something with form and intention, so being able to do this with a model, even if only in very very narrow domains right now, is a pretty massive leap forward.
I'm not talking about the specifics of the architecture.
To the layman, AI refers to a range of general purpose language models that are trained on "public" data and possibly enriched with domain-specific datasets.
There's a significant material difference between using that kind of probabilistic language completion and a model that directly predicts the results of complex processes (like what's likely being discussed in the article).
It's not specific to the article in question, but it is really important for people to not conflate these approaches.
Actually I agree. I guess I was just still annoyed after reading just previously about how llms are somehow not neural networks, and in fact not machine learning at all...
Btw, you can absolutely finetune llms on classical regression problems if you have the required data (and care more about prediction quality than statistical guarantees.) The resulting regressors are often quite good.