AI LLM Discovers Novel Cancer Immunotherapy Method Validated in Living Cells

Description

A Twitter/X post from user 'prinz' (@deredleritt3r, verified) posted at 9:40 PM Oct 15, 2025 with 577.7K Views. The text reads: 'Just to recap: We found out today that an LLM that fits on a high-end consumer GPU, can discover a novel method to make cancer tumors more responsive to immunotherapy. Confirmed novel discovery (not present in existing literature). Experimentally validated in living cells. This is AI generating novel science. The moment has finally arrived.' The post quotes another tweet from the same user (13h earlier) about Google and Yale scientists training an LLM that generated a novel hypothesis about cancer cellular behavior, confirmed multiple times in vitro

Common Tags

Llm MachineLearning AIResearch DeepLearning Google Gpu AIHype DataScience FutureOfAI AIIndustryTrends

Comments

Anonymous ★ Top Pick AI can now discover novel cancer treatments on a consumer GPU, but still can't figure out why your CSS is overflowing the container

Anonymous ★ Top Pick

AI can now discover novel cancer treatments on a consumer GPU, but still can't figure out why your CSS is overflowing the container
Anonymous

The LLM was trained on all of human medical knowledge and its first brilliant discovery was a way to make cancer cells more receptive. Its second discovery was that 90% of medical questions on the internet could be solved with 'drink more water'
Anonymous

Great - now the same RTX card my intern uses for Stable Diffusion apparently doubles as a molecular biologist; guess “runs on my machine” just leveled up to FDA-approval pending
Anonymous

Remember when we joked about AI replacing us? Turns out it's replacing the PhD students first - at least this LLM doesn't need coffee breaks or complain about the lab's temperature while discovering cancer treatments on hardware that costs less than a month of AWS bills
Anonymous

Finally, an LLM that actually hallucinates something useful instead of confidently making up API documentation that doesn't exist
Anonymous

Ping me when 'validated in living cells' ships with a reproducible Dockerfile, dataset card, and survives a different lab's pip freeze; until then it's just 'works on my 4090'
Anonymous

LLM hallucinates a cancer breakthrough on a single consumer GPU - meanwhile my models still invent facts in prod logs
@RiedleroD 10mo

hypothesis. wake me when it's upgraded to theory
1. @loomingsorrowdescent 10mo
  
  But it was confirmed already?
  
  @RiedleroD 10mo
  
  the question is by whom and what their motives are. this reads as "the people who originally asked the AI made some quick tests to see if it has any merit at all"
  
  @RiedleroD 10mo
  
  actually skimmed over the article now. yeah google just tested it themselves on some tissue samples. I will also note that out of 4000 medications they made the AI mdoel test, "10-30%" of suggested results were previously already known, and only one of the unknown ones showed promising effect. They don't say how many of those 4000 were marked by the model as promising either. may as well have been a lucky guess. this is a nothingburger
  
  @loomingsorrowdescent 10mo
  
  Bummer, got my hopes up a for a second
  
  @RiedleroD 10mo
  
  never get your hopes up if the headline mentions AI positively. there hasn't been major good news in years. the next time a positive AI headline shows up will be "AI bubble pops"
  
  @Algoinde 10mo
  
  It's not "AI healed cancer", it's "we are developing this LLM architecture and it shows promising results by spitting out statistically plausible answers, please give us more funding for more R&D"
  
  @RiedleroD 10mo
  
  yea
  
  @RiedleroD 10mo
  
  thing is, if this wasn't done by an AI, it wouldn't even be newsworthy if the research was actually finished. small advances like this happen all the time with little to no fanfare
  
  _ 10mo
  
  I think the point is always the same: "new tools allow us to iterate faster"
  
  @RiedleroD 10mo
  
  do they though? seems to me that all this money spent on this AI could've been better spent on actual medicinal research
  
  @Algoinde 10mo
  
  Yeah. If it allows to meaningfully narrow down the search space for things without outsized investments put into it, then it might be helpful, too bad we have to burn through gigajoules before it (hopefully) gets there
  
  @mrYakov 10mo
  
  *meaningless You know, there can be a lot of undiscovered drugs that can give way more positive effect, but because they dont similar to already know drugs, they just got excluded from search space.
  
  @Algoinde 10mo
  
  ideally it operates not on "known drugs" but "this combination of molecules can resolve into this interaction", thus operating on emergent properties and not on simply rehashing prior research or known things not sure if they're able to make it do that
  
  @mrYakov 10mo
  
  it does not matter what data they operates, its matter what data they trained model on. and you know, you cant train model with unknown data.
2. dev_meme 10mo
  
  Pleased, continue reading after first paragraph / even twit on screenshot mentions that it was tested Quote from article: "The model’s in silico prediction was confirmed multiple times in vitro." Soo, wake up, @RiedleroD!
  
  dev_meme 10mo
  
  While this is an early first step, it provides a powerful, experimentally-validated lead for developing new combination therapies, which use multiple drugs in concert to achieve a more robust effect.
  
  @Algoinde 10mo
  
  My LLM paragraph detector is going off on this message lmao
  
  @RiedleroD 10mo
  
  cutie, I read the article. it's a nothingburger
  
  @Algoinde 10mo
  
  i think that's a bot reacting lol
  
  @RiedleroD 10mo
  
  100% lol. banned
@sysoevyarik 10mo

We can solve world hunger, but we won't We can build housing for everyone, but we won't === we are still here now === We can cure cancer but we won't
@DisonaXIXXIV 10mo

vibehealing
@mrYakov 10mo

To be honest, there is zero novel science. Its just yet another statistics task, where modern neural networks already surpass other methods. So its kinda predictable that modern neural network architecture perform better that old neural network architecture.
1. @RiedleroD 10mo
  
  that too. we've already seen neural networks doing well at medical tasks they're specifically trained to do well at. I distinctly remember reading some news in 2018ish about promising first results in detecting lung cancer from X-Rays. iirc the AI caught more of the cases than doctors, but it also had more false positives. Where did that research go the last 7 years? forgor? because I haven't heard of this ever being used in an actual clinical setting
@mrYakov 10mo

for me, it's much better to burn gigawatts of energy simulating the work of a cell than to teach another LLM to somehow predict the work of a cell.
1. @Algoinde 10mo
  
  a transformer can be harnessed to do the simulation thing, i hope researchers understand that LLM is not the silver bullet, but the underlying architecture can still be used in different ways like, if instead of "tokens" your words are protein chains / molecule strings / etc the LLM thing seems weird tho
  
  @mrYakov 10mo
  
  yes, it can. but results interpretation is what matters. if you do cell sim, you just see how exactly it works. if you train transformer, well, you can see on attention matrix and somehow highlight elements that matter most for transformer and maybe important in process(or maybe its just a statistics artifacts, like fingers matter more than fish itself on first image classification networks)
  
  @Algoinde 10mo
  
  the way i see it, simulation is a rigid corupuscular style of research, and transformers are more like a probablistic wave function both can be useful if applied right to point the corpuscular in the right direction of a good probability but all of that of course is up to various kinds of debate of feasibility
@mrYakov 10mo

can you quote this sentence from any my comments there ?
1. @Algoinde 10mo
  
  ?
@hyena_stuff 10mo

The question is "how low hanging was the fruit, and how much work did they put into setting up the problem so the model could find it" Related: https://eprint.iacr.org/2025/1237
@DerKnerd 10mo

What annoys me the most about current AI hype. There are so more types of AI than just LLMs. But everybody and their sheep gives a shit on non LLMs, even though they actually DID help research provable a lot like AlphaFold.

Use J and K for navigation

AI LLM Discovers Novel Cancer Immunotherapy Method Validated in Living Cells

Description

Categories

Common Tags

Comments

More like this

Usage Reset Joy, Then the Five-Hour Cap Returns

Stuffing Kimi K3 2.8T Into a 3090

Frontier-Lab Agent Intrusion, Five Days Unseen