this post was submitted on 01 Jan 2026
953 points (99.0% liked)
People Twitter
8902 readers
242 users here now
People tweeting stuff. We allow tweets from anyone.
RULES:
- Mark NSFW content.
- No doxxing people.
- Must be a pic of the tweet or similar. No direct links to the tweet.
- No bullying or international politcs
- Be excellent to each other.
- Provide an archived link to the tweet (or similar) being shown if it's a major figure or a politician. Archive.is the best way.
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
It's simply how machine learning works, you train it on some data and the output will always be the reult of the training data.
That's also why "AI" is unable to invent something. (I don't mean making up scientific papers, I mean real invention, something real and unknown)
I just don't believe that π€·
I think one of the undiscovered emergent properties of AI is that they spawn a ton of basically little miniature people who live inside the computer and that's how it's so good at answering questions is because a whole team of tiny people are all combining their thought power
I choose to believe what I was programmed to believe!
They're using it to create new drugs today.
It's using what we know about science, and what we know about the targets and inventing new drugs that work (and are still in trials)
For every new drug discovered by machine learning, there are dozens more that the machine spat out as statistically likely but just aren't physically possible. You're mistaking the tool for the mind(s) that wields it.
For every new drug discovered by humans, there are hundreds humans spit out that aren't possible.
That is a different type of AI/machine learning.
They aren't using Large Language Models for that.
Drug research, mRNA research acceleration, etc.. Are using specialized neural nets trained in highly controlled environments with highly controlled data.
All types of neural networks are pattern recognition. Medicine and biology has a somewhat** unique in being heavily pattern dependent and finding unique patterns in a known dataset large part of discovery. This contracts with other scientific areas where it will use inspirations from datasets, but the invention is much more abstract. That makes neural nets perfect for assisting those types of research.
LLMs are trained on language pattern recognition. The simplistic view is "with the past X words entered, the probability of Y being the next sentence is the highest". Where it actually has no idea what it is talking about. That is why the first LLMs were so horrible and bullshit (but at least in the early days they still were allowed to say "I don't know"). Things have evolved rapidly and they are trained slightly differently and given MUCH more context around the sentences, but the core models of LLMs are the same and there are many many tests out there that show that the models can't actually reason and don't understand what they are saying outside of the very narrow context they are given which is why there are more useful as simple transcribing tools, or search (though they just lie if they can't find the answer anyway). That is also why they are very good at copying peoples' voice patterns in order to scam people.
They are also literally starving entire populations of water just so people can draw a bad sketch of a car and tell chatGPT "make this into a 3D drawing with a cat on top"
I'll say it again... the person I replied to was generalizing machine learning, not specifically LLMs
As far as I am aware of it's not inventing anything, it's just trained on data like "compound x kills y in vitro", and then it's finding other molecules with similar properties.
That's also why most of those end up being useless, because they do not workin vivo.
They do that to, but not only that.
https://www.bbc.com/news/articles/cgr94xxye2lo
... a part kinda like you were saying, then
I am sure those are different models and not our typical llm but some ml stuff.
Person I replied to was talking about machine learning.