AI: The Next Great Disappointment
By Jan Zheng
Jan Zheng builds AI tools for researchers at labspace.ai. You can find Jan on X at @yawnxyz and on LinkedIn.
So, I have a confession: At one point I actually believed ChatGPT could revolutionize science.
I mean, it has superhuman intelligence. Infinite knowledge. I mean, it could just connect all ideas across all of science, simulate all the possibilities, and crank out discovery after discovery, at a pace humiliating even the most diligent of Harvard postdocs. But it turns out it can’t even humiliate Stanford postdocs.
Dario Amodei, the founder of Anthropic AI, wrote a 20,000 word essay singing the praises of how AI will one day make the world a better place. I sipped the Koolaid, got all excited, and scribbled a bunch of notes about all the ways ChatGPT could improve biology, not in ten years, but today. I just had to write my very own essay!
But what am I, a prude? A dilettante? Of course I wasn’t going to actually write my own essay myself. I fired up ChatGPT, gave it all my notes, and asked for a draft. Then another. Then many, many more. One after another, they were all terrible.
What a joke.
I wanted an essay in the likes of David Foster Wallace. Instead I just kept getting something clunky and awkward and wet, like David Sedaris. What a mess. I couldn’t trust ChatGPT anymore. If it couldn’t write my essay, how was this thing supposed to do anything to science? I had to think of something else. I had to write it myself.
I wanted a good hook. But honestly without AI, my cognitive capacity for deep creative thought is mostly vestigial at this point. So I relented, and reached for the AI: “Hey ChatGPT, write a captivating hook for my essay about how great tools like microscopes reinvent microbiology. But make it really really good!”
“Certainly!” it replied.
So, it starts off not with one, but with two hooks. It tells a story of a pair of rivals, who both lived around the 1600’s, who both separately invented the microscope, and who both are credited for inventing microbiology. And, you can’t make this up—both of their last names end with Hook. That’s it. Unsubscribe.
But it turns out, after lots of Googling, Chat was actually right. The first Hook, Robert Hooke, was a prominent scientist. He published Micrographia in 1665, a masterpiece in “the history of microscopy and scientific observation.” His book featured illustrations of many tiny things, from insects to plants to textiles. He even invented the term “cells” to describe the compartments in the structure of cork. His work in biology turned him into a scientific celebrity, and made him the centerpiece of the newly formed Royal Society of London—the most prominent scientific institution of its time. Until the other hook came along.
On September 7, 1674, just a few years after Micrographia, Antonie van Leeuwenhoek, a Dutch cloth merchant with absolutely no scientific training, actually invented modern microbiology.
In an unsolicited letter to the Royal Society, he first described a previously unknown creature, much smaller than a mite, much smaller than anything Hooke had ever described, “with green and very glittering little scales.” Over the next few years, he wrote hundreds more (still unsolicited) letters to the Society, describing thousands more previously unknown creatures. Animalcules, he called them, and they could only be seen through his own, meticulously polished, obsessively made microscopes.
Throughout his letters, Leeuwenhoek described in minute details every single animalcule he ever came across. These were creatures we now call bacteria, protozoa, blood cells, sperm cells, yeast cells, and fungi. The only thing left out of his letters though, were any details of his microscopes. Robert Hooke, at peak desperation, learned Dutch just to pry Leeuwenhoek of his secrets. That didn’t work. And actually, nothing worked: Leeuwenhoek never wrote down or shared with anyone any details about his microscopes, even after his death.
Leeuwenhoek wasn’t a mad scientist, or some evil genius mastermind. He was an entrepreneur cloth salesman who created a successful business and retired in his 30’s. With his wife’s passing in 1666, he needed something to occupy his mind. He picked up a copy of Micrographia, which had been published the year before, and became more and more obsessed with smaller and smaller things.
Micrographia taught Leeuwenhoek how to make his first microscope. These were less like advanced scientific instruments, and more like middle school science fair projects. Regardless, he made them at home, and brought them wherever he went. He’d take them to the white cliffs of Dover: “I observed that chalk consisteth of very small transparent particles; and these transparent particles lying one upon another, is, methinks now, the reason why chalk is white.” He’d take them to his backyard pond, where he’d look at plants and wildlife.
Over time, he’d gradually improve Hooke’s design, so he could see smaller and smaller things. And within just a few years, he’d become the first person to ever see a microbe.
There’s a poetic, almost pathetic irony behind the Hooke vs. Hoek rivalry. While Hooke’s book taught Leeuwenhoek how to build a microscope, Hooke never actually built one himself! For him, the design was just a toy. It was a small glass bead sandwiched between two metal plates. You’d put the thing you wanted to see on one side, smash your eye against the other, and then you’d just look through the glass. To Hooke this was “offensive to my eye.”
Instead, Hooke preferred his multi-lens microscopes. These were much more complex, took up an entire table, and could, in theory, see a lot more. These were Microscopes for Adults. He never even got to learn that Leeuwenhoek used his own offensive “toy” design to beat him at his own game.
***
So, I have another confession.
Eight years ago, I left the prosperous tech world, to make an impact in healthcare. It wasn’t really out of generosity or out of a need to make the world a better place. I just wanted to make something cool. Creating apps is easy, but creating apps that save lives? That’s brag-worthy.
With Phage Directory, that’s what we did*.* We connected patients dying from superbugs with phage labs, and we scored our first phage therapy success within just a couple of years. That felt great. A couple of years later, we were paid by the Australian government to move to Sydney and start a full clinical trial.
In Sydney, we treated a couple of dozen patients. That felt really great. But, the work was hard: As the only “computer person,” I’d spend many long days writing code to connect data from the clinics and labs to the regulators. I wanted to expand beyond Australia. I wanted to build the app that made phage therapy possible for every single country. Alas, as the sole “computer person,” that’s quite a large task. Until AI came along.
Leaps in science have always come from better technology. Sure, there was the microscope. But there’s also gram stains, which helps us tell bacteria apart, and was necessary for discovering antibiotics. There’s transmission electron microscopes (TEM) which let us see inside viruses and bacteria and cells. There’s DNA sequencing which lets us understand the building blocks of life. There’s CRISPR which lets us edit genes and create synthetic life.
And here was this shiny new technology, AI, which promised to change all of science. Biotech though is no stranger to “AI.” We’ve been using machine learning for decades to analyze and predict and classify. But this new kind of AI, “ChatGPT”, promised so much more.
This new kind of AI was different. It was smart. It could pass medical exams. It could write grants, negotiate with vendors, even sneak into the lab and do the actual experiments. Brick by brick, first it would automate your lab. Then it would reinvent science. Then it would take your job.
But of course, ChatGPT was never actually any of those things. It’s not TEM, nor CRISPR, nor sequencing. Unlike machine learning, ChatGPT can’t even land you any Nobel prizes. And the worst part? You know, when you look at an AI picture, and it always has that glossy, oily sheen? I can always taste that sheen, even in text, and once you’ve tasted it, you’ll taste it everywhere.
***
It’s offensive to my eyes.
ChatGPT and Claude? They’re not serious tools for science. They’re toys. But I suppose toys can sometimes be useful? PCs were toys before Apple and Microsoft. The internet was a toy before email and e-commerce. X-rays were toys until a researcher convinced his wife to put her hand inside one. She was the first person to ever get an x-ray taken, and when she saw the picture, she screamed “I have seen my death!” But that’s kind of what you’re supposed to do with AI! If I could stick my (partner’s) hand in it, I’d probably try it.
If you think of it like a toy, a thing you don’t take seriously, then you can just pick it up and play with it. You can try all kinds of goofy stuff. Make it write a poem about phage therapy. Make it explain how phage therapy works. Ask it stuff you’d never ask your friends and family. It sounds ridiculous, but sometimes you’ll be surprised. And if you don’t like the answers? Just ask it again!
Ok, so if you’re looking for actually trustworthy answers, based on published, cited research, please don’t use ChatGPT. Chat is more like your Uncle Jules over Christmas. He’s very drunk, he tells great stories, and somehow nails every question on Jeopardy! He somehow knows all the answers, but is still very drunk, and if you ask him too many questions he’ll probably just start making stuff up.
But sometimes, you do want him to make stuff up. Try this out: give Chat your latest research paper, then ask it for future research directions. Add your Zotero collection, and ask it for ideas. Add a grant proposal, and ask it how you can win. Then ask it for the best collaborators for your research. Then ask it to write the invitation emails. Then ask it to start the grant application.
Are your lab protocols confusing? Ask it to clarify. Is it too long? Ask it for two versions: a concise one and an expanded one. Did you scribble notes in the margins of a protocol because the experiment went wrong? Or the protocol was wrong? Well, snap a picture, upload it to Chat, and you’ll have an updated protocol.
And while AI can’t cure Alzheimers, it sure can spot errors in your spreadsheet. Did you know “one in five genetics papers contains errors thanks to Microsoft Excel” because of the SEPT2 gene? Well, blaming errors on Excel is not an excuse anymore.
Maybe you can’t keep up with the literature? That’s ok, thousands of papers are published every day. No one can possibly read them all, and now you don’t have to! Ask Chat for a short summary so you’ll know if a paper’s worth your time. You can even ask Chat to identify flaws and errors. Ask it to rip a paper apart (try it on your nemeses’ papers!).
And this is only with ChatGPT. There’s lots more tools to play with, like Elicit, SciSpace, NotebookLM, Consensus, Perplexity... And if you can write code, Claude can tailor a mini-app to your very specific needs, in just a few minutes.
But still, there’s a single word that encapsulates how I feel about AI: disappointment. Sure, it can search and summarize millions of papers. Sure, it can transcribe hours of audio, or turn any text to any spoken language. It can even turn any spoken language into a programming language. Sure, it can probably even help me fulfill my dreams of single-handedly writing all the code that powers the global phage therapy hospital and data infrastructure and will one day be used to treat tens of thousands of patients.
But can it beat cancer? Can it figure out Alzheimers? Or terraform Mars? Can it undo our deserts and reverse climate change? Absolutely not. It couldn’t even write this essay!
So, here is my admission. I humbly admit that I was wrong. I can’t believe that at one point I actually believed that this new kind of AI could revolutionize science. At worst, it’s a pile of disappointments, false hopes, and false promises and represents the worst parts of capitalism and humanity. And at best, it’s merely just a fun little toy.