Dear AI researchers, friends,
LLMs are not just here to help you write better papers (ahum), or review them, noooo, they are coming for your jobs now as well!
First, they came for the software engineers. The likes of Cursor, and AmazonQ are saving big tech companies (and Klarna) hundreds of billions by replacing software developers with AI tech stacks and now, they have their eyes set on you.
So, what is it about?
Are LLMs really about to take over the world of academic research?
Let’s get down to business, with a bit of impending doom.
Before we start!
If you like this topic and you want to support me:
- Comment on the article; Google appreciates that and it will really help spread the word 📢
- Connect with me on Linkedin 🔗
- Subscribe to TechTonic Shifts to get your daily dose of tech 💉
LLMs are the idea generators you never knew you needed
Stanford University recently stirred the pot with their study, “Can LLMs Generate Novel Research Ideas?”
Spoiler: they can.
And they do it pretty well.
The researchers found that LLM-generated ideas were rated as more novel compared to those from human experts.
Over 100 NLP (Natural Language Processing) researchers were put to the test. They were asked to come up with fresh ideas and then review both LLM-generated and human-generated ones, without knowing which was which. And the LLMs beat humans at the game (p < 0.05, if you’re into stats).
But before you throw your research hat in the ring, wait for it, because there is a catch.
These LLM-generated ideas were slightly less feasible. So, while an LLM might dream up a plant that solves world hunger, it does not quite nail down how to actually make it work.
But still, it is a start!
LLMs and the rise of the $15 research paper
In one of my previous articles, I wrote about the AI Scientist…
Because Japanese startup Sakana AI thought, “Why stop at just generating ideas?”
So, they developed the AI Scientist. That is a model that pretends to do it all. This thing doesn’t want to just brainstorm. It wants to write code or runs experiments, visualize data, and eventually wraps it all up in a tidy scientific paper.
And the dystopian fun of it all, is tht it does all this for around $15 per paper.
That’s right.
For the price of a couple of fancy coffees, you get a whole research paper.
Suddenly, all those sleepless nights in the lab don’t seem so smart, do they?
But rest assured, my research friends, I still have not seen the application. They have only launched a paper thusfar.
But Sakana AI isn’t alone.
I have seen several platforms that claim to automate parts of scientific research.
Elicit, Research Rabbit, Scite, Consensus, and PubTator are just a few examples. They are really powerful tools, but they still need human oversight. They are good at processing data and generating hypotheses, but humans need to make sure that the quality cannot be undermined, and that the relevance of their work is maintained.
Meanwhile, them folks at OpenAI are gearing up to drop their latest model, “Strawberry” (read the article: Exclusive insights into OpenAI’s “Strawberry” project, focused on AI agents) which is to be expected before october ! The promise of this berry is that it will be even better at reasoning.
Reasoning?
GPT-4o1 and its predecessors sometimes struggle to connect the dots. But this juicy red berry promises to be a reasoning champ. Think fewer “uh, what?” moments and more “aha!” breakthroughs.
If it delivers, we will be talking about an AI that can actually think things through; whether it is a genius research plan, a true brainstrorm or deciding what is for dinner.
Basically, GPT 4o1, a.k.a. Strawberry could be the upgrade that finally stops your AI assistant from suggesting “pineapple on pizza” as a serious idea.
LLMs in biology: the drug discovery revolution
This one is not to be left behind: Chai Discovery.
That is a startup from a former OpenAI employee, and it launched Chai-1.
This model predicts molecular structures for drug discovery, and it is aiming to speed up the process of finding new treatments. And while I am on the subject, another study found that AI models like AlphaFold have already revolutionized protein folding predictions (that is a crucial step in drug development).
LLMs aren’t just brainstorming anymore with these innovations. They are setting the stage for some serious scientific breakthroughs.
INTERMEZZO
Playing a Mental Game with ChatGPT
Let's do a mental game - a gedankenenexperiment like Einstein used to call it... you've got a bunch of people sitting in a room. They are all using ChatGPT to come up with the next big thing. You would think they would be bouncing off the walls with wild ideas, but no! They were all coming up with variations of the same darn thing, apart from the people that did not use ChatGPT. It turned out they were far more original than the others.
A recent study actually looked at this phenomenon and proved that creativity and originality is impaired when we all start using LLM’s for creativity and innovation (read the article: Lobotomized AI)
ChatGPT is a friggin creativity assembly line that is churning out the same ideas over and over again.
In other articles, I have had conversations with people who did not agree on this, because they obsiously were supprised by the seamingly creative answers that ChatGPT gave them.
But here is the thing. A Large Language Model can only infer, meaning that it can answer to questions by drawing on the information it has learned from the training data. The response it generates seems most appropritate (read: probable), based on the given question.
Now let’s do a second gedankenexperiment.
It is 1905, and it is the Annus Mirabilis of Albert Einstein. That means it's his miracle year, where he had published four papers that would forever change the course of physics (photoelectric effect, brownian motion, special relativity, and the mass-energy equivalence E=MC²). That was the basis on which he finally built his groundbreaking theory of general relativity. This specific form of physics - general relativity - is all about gravity. Concepts like Spacetime, and curvature of space, the fact that gravity equals acceleration, and that there is some freaky concept called gravitational time dilatation, whereby people that travel with close to the spead of light, get to experience a slower passing of time (compared to people on earth). He also predicted black holes, gravitational waves.
All freakishly strange concepts at the time.
Still….
And before general relativity – we thought we knew all of physics, and that especially with Isaac Newton’s law of grativy (and law of inertia, acceleration and action = reaction), all of physics would be over and done with. Students in the 19th century were discouraged to even study physics.
Then came Einstein with his relativity and turned their world upside down.
The question - the real gedankenexperiment here - that I would like to do, is to ask you, if any other physicist had the possibility of using ChatGPT at that time - around the 1900's - fully trained on all the knowledge that had been available to it at that time - including all of the Newtonian laws - that it would have come up with this devishly different concept of special and general relativity?
You simply can’t get to relativity by simply infering based on your dataset.
For that to happen, you need a lot more.
So the answer is simply – NO
END OF INTERMEZZO
Are LLMs really ready for research prime time?
Now, here is where things get a little spicy.
Some people say that LLMs are good for reviews and suggestions but not so great at deep, and original research.
Take Meta AI’s Galactica.
No, not the battlestar…..
Galactica WAS a LLM which was designed to help researchers, but it got yanked off the web just three days after launch in November ’22 because it kept generating, well, let’s call it “questionable” content.
The fact that Meta released the demo with a “NEVER FOLLOW ADVICE FROM A LANGUAGE MODEL WITHOUT VERIFICATION” message almost suggests that the company was half expecting the reaction it has predicted.
Not quite the assistant you want in your corner.
But, despite some hiccups, the potential is there. LLMs are like overenthusiastic interns. They churn out fresh ideas all day long. But just like interns, they also sometimes need a bit of guidance to distinguish between a brilliant idea and a complete train wreck.
Reminds me of my own time as an intern.
OMG. I was such a BAD scientist.
Use LLMs to brainstorm, not replace
The trend of using LLMs to write papers is on the rise. A look at research abstracts over the past few years shows a spike in phrases like “delve into” and “explore”. And that coincided with the launch of ChatGPT-3.
Hahahahaha….
Duhhhh..
Read the article: Don’t want to be caught using ChatGPT? Than stop delving !
But the thing is, it is not about replacing humans with machines. It is about collaboration.
Well, at least, for now…
LLMs have a vast knowledge base and no biases. Except when they do, but that’s another story. They can mix ideas from fields that seem totally unrelated. But while they can suggest that your flying toaster could solve world hunger, they still need a human touch to say, “Hang on a second, is this really going to work out”?
Signing-off Marco
Well, that’s a wrap for today. Tomorrow, I’ll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ♨️
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Google appreciates your likes by making my articles available to more readers.
Your point of view caught my eye and was very interesting. Thanks. I have a question for you.