Earlier this month, OpenAI released a repository of hundreds of mathematical results resolving a variety of problems in pure mathematics. Overnight, entire research agendas fell; dissertations were rendered outdated. It’s the latest sign of a trend that’s been building all year: agents can post-train language models, solve Kaggle competitions, and more.
As a researcher, it’s impossible not to feel excited about agents that can automate experiments. I think of all the dropped projects and side quests that I had to cut from my thesis amidst deadlines and funding constraints, brought back to life with a few million tokens. It’s a researcher’s dream, but I’m also apprehensive: autoresearch democratizes the ability to get results – not the ability to do better research. I’m concerned that we have conflated the two.
To those who are first encountering research, the struggle of grinding for results becomes almost optional. It feels like the end of human research as we know it, when better results are one prompt away. I’m sympathetic to hill-climbing: there’s a dopamine hit that comes from chasing numbers on a fancy dashboard, because it feels like progress – seeing scores go up, loss curves go down. But I see this as intellectual hedonism.
My Ph.D. advisor, Jenna Wiens, works in machine learning for healthcare. Her talks always open with a slide where a cartoon version of her throws a laptop with a well-trained machine learning model into a hospital. It’s a metaphor for the underappreciated gap between performance and usefulness. In general, the bottleneck to real-world deployment is not how well did you climb that hill, but rather: if you climbed this hill, what does this mean for us? Who does it help, and who does it hurt?
Better for you if you take me off
I am not the first to invoke the parable of the whispering earring in the context of AI. Imagine an earring that whispers helpful tips in your ear: “Better for you to set the learning rate to 0.00251342.” It is always right: every experiment and hero run magically converges. Good results are no longer in question; with a Whatever Agent Ultra Max subscription, you never come to a meeting short on quantitative results or new experimental ideas. With statistical language and theorem names baked in, you can even perform scientific rigor: you can beat your advisor to suggesting a “Bonferroni correction for multiple hypothesis testing,” “permutation testing,” and more. What a precocious-sounding Ph.D. student you would be, had you enrolled ten years ago!
In the original story, the earring’s advice is always right, and the people who wear it live remarkably successful lives. But when they die, the part of their brain responsible for deliberation has withered from disuse.
Keep asking why
I’m reminded of Terence Tao’s recent talk on “Math 2.0,” mathematical research in the age of AI-driven proofs. In one particular vignette, he imagines an AI-developed cancer treatment, a previously-unknown cocktail of drugs with no known mechanism of action, that passes Phase III clinical trials. Would you take it now? Or would you prefer that some human expert understands the mechanism of this drug?
A common response: “Aha, we have plenty of drugs where we don’t understand the mechanism of action!” And thus autoresearch emerges victorious!
But this is a strawman. I’m sure many readers would take the drug! The question is whether we should stop asking questions. The goal of research cannot end at expanding what we know. It is also about growing what we understand. People took aspirin for decades before anyone fully understood how it worked, and once we did, we learned it could also prevent heart attacks. Being a good researcher means you keep asking why, even after something works. Understanding gives you hints about what questions to ask next, which hill-climbing with an agent won’t give you.
Towards the end of my Ph.D., while I was grumbling about industry versus academia, my advisor told me: “Trenton, you should understand that the Ph.D. is unique. It’s the only time where you have the bandwidth to think really deeply about problems like this, and go on these intellectual side quests.” So, why do we do research? These slow moments and struggles are precisely that which builds human understanding, and gives research its value.
When we forget this, we lose the beauty of research. If your goal is to produce the most impressive results or the most papers – then I will not be very convincing to you. Stop reading this, type /goal into your favorite agent, and laugh your way to papers galore. But perhaps you recognize in this the tragedy of excising human curiosity from research.
What autoresearch should be
I do not write to criticize autoresearch; I work at an autoresearch company! The strength of autoresearch is in its power to explore and build on ontologies and knowledge bases formed by humans over millennia. Surely there are missed insights that our agents can help us see, or lemmas in the literature that a human eye has overlooked; piles of disorganized data begging for a researcher’s eyes, waiting to unlock discoveries. Autoresearch can already create an incredible amount of economic and societal value.
I am here to criticize the uncritical application of autoresearch that I see all too often, and the excesses it reinforces: this self-induced academic Kessler syndrome, where low-quality vibe research inundates all fields faster than we can understand, and renders good work impossible to conduct amidst an increasingly inscrutable literature.
I’m sure I sound like those who warned of the dangers of novel-reading, or of Google’s deleterious effects on attention spans. But I argue the stakes are higher with research. Research is the heart of human knowledge production: not merely how we process or identify information, but how we produce knowledge. Despite its flaws, academia, the historical training institution for researchers, shapes the process by which humans produce knowledge. It is this knowledge that drives human innovation, survival, and welfare.
In the hands of a capable researcher, autoresearch is not just a rote result generator, but a powerful tool for self-pedagogy. In its best instantiation, it’s an always-on collaborator with an infinite wealth of knowledge to bounce ideas off of, with infinite hunger for more experimentation. It’s a research team ready to tackle a wave of challenges and analyses, to dig into every anomalous pattern at hand.
When humanity writes the story of autoresearch, it must be a story of human empowerment. I want this to be a world where no research ideas die from lack of funding, time, or personnel, where the bottleneck on scientific discovery is on the creativity and daring of human-agent teams, not resources. It is critical that we steer the next century of human research towards an abundance of discovery, not an abundance of papers. This is a value that my colleagues and I at Quadrillion Labs hold dear.