Artificial intelligence can revolutionise science
economist.com
economist.com
> Two areas in particular look promising. The first is “literature-based discovery” (LBD), which involves analysing existing scientific literature, using ChatGPT-style language analysis, to look for new hypotheses, connections or ideas that humans may have missed. LBD is showing promise in identifying new experiments to try—and even suggesting potential research collaborators. This could stimulate interdisciplinary work and foster innovation at the boundaries between fields. LBD systems can also identify “blind spots” in a given field, and even predict future discoveries and who will make them.
I have to question to the value of looking over existing papers given the current replication crises, is using an LLM to review existing papers just going to enable us to do more bad science faster? Does that do anything for us?
> The second area is “robot scientists”, also known as “self-driving labs”. These are robotic systems that use AI to form new hypotheses, based on analysis of existing data and literature, and then test those hypotheses by performing hundreds or thousands of experiments, in fields including systems biology and materials science. Unlike human scientists, robots are less attached to previous results, less driven by bias—and, crucially, easy to replicate. They could scale up experimental research, develop unexpected theories and explore avenues that human investigators might not have considered.
This sounds basically of just asking a GPT to suggest ideas for experiments?
Honestly based on what I understand of science right now the biggest way generative AI could help advanced things is by assisting in the writing of grant applications faster.
EDIT: Relevant XKCD https://xkcd.com/2341/
I think actually the second idea extends quite a bit further than that. The idea is that you enable the agent to interact with an entire collection of laboratory equipment through tool-use. The agent not only generates hypotheses and designs experiments to test them, but then also actually executes the experiments through tool-use and iterates.
This matches what I was told earlier this year when looking into what LLMs could do. I work in public health, and asked staff from a large registry what they'd like to me try. I expected something like generating reports from aggregate data (with some of that automatic exploration mentioned in the article). What they really wanted was:
1. A nice chat bot to answer questions from data submitters about reporting policies.
2. A tool to ingest federal grant announcements and filter down to those the registry could apply for.
#1 is useful because that's a lot of their job. Getting people to send accurate data in the correct format on time is exhausting. #2 helps with everything, because it could mean hiring an additional staff member. That extra person can write reports and apply for grants, clean data, answer calls, and cover when somebody's out sick. Humans are still really useful.
This is half a real insight and half total bullshit. Honestly, what we need robotics and AI for in laboratory science is just old-fashioned standardization and labor-saving.
I see nothing in today's 'deep AI' that addresses these desiderata. And I don't see the current AI strategy of accumulating only existing knowledge as the means to those ends either. Optimization of learning can only learn more facts or do it faster, not think with more creativity or innovation. Memory is but a small part of genius.
I think the problems with replication are a separate axis to the problem that there is too much being published for anyone to actually read.
AI in general — never mind LLMs, even the much ones running search engines — help with the content overload.
> This sounds basically of just asking a GPT to suggest ideas for experiments?
I think the car analogy here is that if what you've suggesting was google maps, what the article is suggesting is Level 5 autonomy with no steering wheel and an opaque wall instead of a windscreen.
I have absolutely no idea how hard such a level of automation might be to actually implement, not least because of Moravec's paradox: https://en.wikipedia.org/wiki/Moravec's_paradox
Lots of time can be saved by automating those steps (and many researchers don't enjoy it so their job satisfaction could be increased). Also the resulting output could be improved if the researcher had a well structured summary to use as the foundation of their outline.
Improve search with semantic search (search by concept not keyword) Improve refinement by preprocessing and summarizing Don't print, display clean and concise data. Summarize, cite and display.
This stops short of literature based discovery, you have to bring your own research question.
We've also had some luck finding a gap in existing research. We did a POC where we scraped pubmed and graphed study results by concept. We then used the graphed concepts to explore the conceptual space.
It seems that vitamin D protects against cancer and heart disease. It seems that vitamin D supplementation protects against cancer but not heart disease. Is this because of some previously unknown effect of sun exposure (the primary natural source of vitamin D) or is it just that people with adequate vitamin D go outside a lot more and therefore also get more exercise? Don't know, would love to read the paper if someone studies it ;- )
Yeah, and then the funding agencies summarize the applications by running them through an LLM.
It might be worse than that, not only might ChatGPT uncritically accept authors' claims or reduce the effort to produce low-quality research - given the current state of the art, there seems to be no guarantee that LLMs won't claim things about a paper that was never even said.
As a tenure-track scientist who works in ML applications for astrophysics, I disagree with this sentiment. The main issue isn't that enough scientists are using tools to search through literature or form new hypotheses, the main issue is that scientists now have to validate and sift through AI-generated outputs in order to find useful signals, rather than validate and sift through experimentally derived or observed signals.
AI can be useful for hypothesis generation in my field [0], and I think that there are lots of great use cases where it can be used to summarize information. However, it always comes with the possibility that it might output complete nonsense [1], so scientists who adopt these tools will have to spend some of their time verifying their outputs.
[0] https://arxiv.org/abs/2306.11648
[1] https://web.archive.org/web/20230913230733/https://www.msn.c...
Things are re-discovered across adjacent fields of study all the time. There are also a bunch of times I've come across a paper when I was a year into a project, and wished that I'd had it at the beginning of the project.
I think it might be exactly what you are looking for.
One idea that I find interesting is to combine LLMs with formal verification and theorem prover tools like Coq, Lean, etc. Any mistakes by the LLM should be detectable by the verification engine. Maybe this could be useful in automating the currently ongoing efforts of 're-proving' the existing body of mathematical knowledge with theorem provers ('ChatGPT, please take this paper and verify the proofs with Lean'). And who knows, maybe one day the machines will produce some interesting mathematics on their own. Would be curious if anyone has links to works in this direction, or blogs/content by mathematicians discussing this.
But will we maintain control of the above-human-ability, autonomous AI systems these companies are racing to build? This is the AI control problem.
If not, then "AI can automate science" isn't much of a counterpoint or reason to be optimistic -- science may be automated, but not under any human's control and not for any human's benefit. In fact, if we're in this situation, the ability of AI systems to automate science is worse news than otherwise, in the same way that the invention of science by humans was bad (or at best, very mixed) news for the animals of Earth.
Wonder how much AI is actually doing here, and how much it's just paper hype, in the same way that AI companies have been shown to actually just use human resources (until they figured out the AI part ;]), I wonder how much these were just to juice up the paper a little bit.
Would researchers have less issue publishing papers that contradict sensitive findings in earlier papers if they can pass responsibility back to the robot? (e.g. robot hypothesizes and disproves widely accepted result that years of other research is based and that anyone who challenged in the past has been dismissed as a quack.)
In fact I'd be surprised if a bunch of scientists weren't already doing that. But I'm not sure of the utility of that.
In fact, using something like our new chat-ai, we wouldn't even need to understand the connection.
For the literature review piece the key problem is that LLMs are exquisitely bad at working with even the simplest kind of scientific evidence: citations [1, 2]. They will get better, but it is not clear that LLMs can deal effectively with the very sparse kind of evidence that appears in the literature. Also, generating hypotheses isn't exactly the rate limiting step, the bigger issue tends to be when you get people with pet projects/hypotheses in positions of power that dictate funding priorities (e.g. the decades long Alzheimer's Aβ disaster).
For automation and instrumentation of labs the vision is on point and there is interest, and active work, if not large amounts of funding, to bring that vision to reality [3, 4, 5]. However, we simply don't have the tooling needed to be able to express the full complexity of experimental protocols in a way that can be verified. Sure you can write a python script to control a robot, but it is exceptionally difficult to extract the scientific meaning from that.
My PhD work was to develop a formal language for scientific protocols, and I'll be continuing to develop it, but there is still a long way to go.
1. https://doi.org/10.7759/cureus.39238 2. https://doi.org/10.1016/j.mcpdig.2023.05.004 3. https://www.youtube.com/watch?v=_gXiVOmaVSo&t=865s 4. https://doi.org/10.1109/JIOT.2020.2995323 5. https://ccc.ucsf.edu/sites/ccc.ucsf.edu/files/Marshall_W_CCC...
Our key insight is that the process of citation needs to be handled outside the LLM. They're good for text processing and summarization but as you said, the LLM itself is poor at citation.
I can imagine taking the citation tree and using the LLM to compact the hypotheses, results, etc. for each node in the tree and sticking that in the vector database could get you pretty far.
All he does is write fantastical reviews to grind an axe and harass speakers at conferences.
[1] https://www.merriam-webster.com/dictionary/revolutionise
https://www.studyenglishtoday.net/british-american-spelling....
But it remains totally unclear whether those things are achievable, and to what extent they can actually do real, useful science, rather than just exist as a novelty. Of course AI "can" revolutionize science. But the proof is in the pudding. Write an article when something has happened, rather than predicting that it will (and being wrong, like every such article written for the past 70 years).
AI existential risk isn’t “inevitable” in the same way that eg heat death of the universe is.
The costs (drug research and healthcare) should be born by by the public without negotiation, while the profits (drug pricing and patents) should be monopolized.
When my college frets that ChatGPT writes better than our B students, there are various reactions. So I had to see. I signed up for ChatGTP-3.5 and asked it "How can artificial intelligence revolutionise science?"
Huh. It didn't simply plagiarize The Economist. It gave a better answer I found easier to read.