Alexey Guzey asks:
How much have you thought about AI and when will AI be able to do scientific research both cheaper and better than us, thus effectively obsoleting humans?
My first reply: I guess that AI can already do better science than Matthew “Sleeplord” Walker, Brian “Pizzagate” Wansink, Marc “Schoolmarm” Hauser, or Satoshi “Freakonomics” Kanazawa. So some humans are already obsolete, when it comes to producing science.
OK, let me think a bit more. I guess it depends on what kind of scientific research we’re talking about. Lots of research can be automated, and I could easily imagine an AI that can do routine analysis of A/B tests better than a human could. Indeed, thinking of how the AI could do this is a good way to improve how humans currently do things.
For bigger-picture research, I don’t see AI doing much. But a big problem now with human research is that human researchers want to take routine research and promote it as big-picture (see Walker, Wansink, Kanazawa, etc.). I guess that an AI could be programmed to do hype and create Ted talk scripts.
Guzey’s response:
What’s “routine research”? Would someone without a college degree be able to do it? Is routine research simply defined as such that can be done by a computer now?
My reply: I guess the computer couldn’t really do the research, as it that would require filling test tubes or whatever. I’m thinking that the computer could set up the parameters of an experiment, evaluate measurements, choose sample size, write up the analysis, etc. It would have to be some computer program that someone writes. If you just fed the scientific literature into a chatbot, I guess you’d just get millions more crap papers, basically reproducing much of what is bad about the literature now, which is the creation of articles that give the appearance of originality and relevance while actually being empty in content.
But, now that I’m writing this, I think Guzey is asking something slightly different: he wants to know when a general purpose “scientist” computer could be written, kind of like a Roomba or a self-driving car, but instead of driving around, it would read the literature, perform some sort of sophisticated meta-analyses, and come up with research ideas, like “Run an experiment on 500 people testing manipulations A and B, measure pre-treatment variables U and V, and look at outcomes X and Y.” I guess the first step would be to try to build such a system in a narrow environment such as testing certain compounds that are intended to kill bacteria or whatever.
I don’t know. On one hand, even the narrow version of this problem sounds really hard; on the other hand, our standards for publishable research are so low that it doesn’t seem like it would be so difficult to write a computer program that can fake it.
Maybe the most promising area of computer-designed research would be in designing new algorithms, because there the computer could actually perform the experiment; no laboratory or test tubes required, so the experiments can be run automatically and the computer could try millions of different things.
Not in the foreseeable future.
https://www.nature.com/articles/s41587-023-02103-0 (free version https://rdcu.be/dwSSy ).
See also “How hard is cognitive science” by Rich and colleagues and “Reclaiming AI as a theoretical tool for cognitive science” by van Rooij and colleagues
Those are interesting articles, to be sure, but note that the conclusions drawn in them apply to any prospective scientists bound by the limits of classical computation. In other words, they apply just as much to human beings as they do to possible AI scientists. That is, unless you believe that the achievements of human scientists are meaningfully dependent on performing actions which fall outside the scope of classical computation.
” I guess the computer couldn’t really do the research, as it that would require filling test tubes or whatever.”
You know that automated liquid handling is a thing, right? See also things like https://www.nature.com/articles/d41586-024-00093-w , for AI + experiments. There is, in fact, a lot of work being done on this!
What is research? Is it measured by publications, “original” findings, quality of output, novelty, accuracy of some sort, …?
Some of these can already be done by AI, some probably never will be. I don’t think the question is well formed. When can AI replace humans doing X, where X is precisely defined, could be an answerable question. AI combined with robotics could probably accomplish a lot. When I am teaching MBA students, I tell them that anything that can be precisely answered is something an AI could do – the things humans should concentrate on are the things that cannot be automated. Unfortunately, too much of our educational systems are designed around teaching things with known answers – precisely the things an AI could do (admittedly with errors, just as humans make such errors).
“My reply: I guess the computer couldn’t really do the research, as it that would require filling test tubes or whatever. ” There can be a lot of technique in lab science that AI would have trouble with. To pick an example from long ago, a neurophysiologist friend of mine spent about a month learning how to dissect a nerve from lobsters and stick wires in them to record the neural control of patterned muscle contractions. Similarly, my virologist wife dealt with lab preparations that took a long time to learn how to do, although things like that don’t get described in methods sections.
On a related question, can AI write comments to a statistics blog?
As I understand, there is already a big history and literature on “optimal experimental design” (Mark Pitt and Jay Myung have worked on this in cognitive psychology). As you and other commenters point out, a computer can already do a good job of that if (a) the parameters of the experiments are easily defined and (b) there is a set of putative mechanistic models that allow one to derive predictions about outcomes from the parameters of the experiments. In that setting, you can do an explicit cost-benefit analysis and determine the “optimal” experiment to run that will maximize expected information gain inversely weighted by cost.
Stan or JAGS or whatever can do that already, so I’m not sure what modern AI methods can add to that approach. I guess you could use a deep neural net to try and infer mechanistic models by looking for patterns in experimental parameters and outcomes. I remember people were doing that with complicated regression models 10-15 years ago, with little success. A more promising approach might be to use AI to derive predictions for complex but pre-specified models. If deriving predictions is prohibitively difficult (e.g., no analytic solutions, expensive simulations), you could train a neural net to extend the predictions of an existing model to new parts of the parameter space.
That isn’t science, its NHST pseudoscience. In fact these AB tests are a key enabler of the enshittification of the internet, just like NHST does everywhere it is adopted.
Eg, imagine AB testing for “user engagement”, which you measure via number of clicks, time on site, etc. You will eventually AB test your way to hiding all the info people actually want behind confusing menus and so on.
So people stop coming or switch to a competitor, then you start losing revenue, so you have to raise prices, insert ads. You keep AB testing even harder. Then, if big enough, you start lobbying for laws to help you survive, ie genetate redtape (in the name of safety, of course) nearly impossible to navigate without being an established player.
Aren’t you mixing issues of whether A/B testing is done correctly (in the sense of discovering some sort of “truth”) and whether such testing is done for good or evil? I can easily imagine A/B testing done on user engagement, discovering that hiding relevant information fools people into spending more time and money, and being used in that way. This would be an example of something I might consider an “evil” use, but one that actually discovered something about behavior. On the other hand, poorly done A/B testing, such as using p<.05 to decide that website design X is better than Y, might result in changing a website design based on noise. Such a use may not be for evil purposes – it may even be used erroneously in belief that you are helping people find useful information. But the statistical results may simply be due to mining noise or the result of ill-designed forking paths. I couldn't tell from your comment whether you are disparaging the practice of A/B testing as a statistical method or objecting to how it is being used (though from your past comments, I suspect both).
Its like this simpsons joke:
https://youtube.com/watch?v=PWFF7ecArBk&pp=ygUSZ29nZ2xlcyBkbyBub3RoaW5n
It offers no protection from making “errors” at all.* But gives the veneer of doing so, so is worse than nothing.
* But no one has done an RCT for goggles vs rivers of nuclear waste. So there is “no evidence” they wont save you.
“My reply: I guess the computer couldn’t really do the research, as it that would require filling test tubes or whatever. I’m thinking that the computer could set up the parameters of an experiment, evaluate measurements, choose sample size, write up the analysis, etc.”
On one hand, the AI craze has perhaps blinded us in the sense that we have had such computer programs for a long time. Think about time series analysis and econometrics, say, within which automatic model selection has been available from the 2000s onward.
“I could easily imagine an AI that can do routine analysis of A/B tests better than a human could.”
On the other hand, AI is already doing A/B testing all over the place.
“Computers are useless, they can only give you answers.” –Pablo Picasso
Chomsky et al NYT (March 2023) states: “ … Here’s an example. Suppose you are holding an apple in your hand. Now you let the apple go. You observe the result and say, “The apple falls.” That is a description. A prediction might have been the statement “The apple will fall if I open my hand.” Both are valuable, and both can be correct. But an explanation is something more: It includes not only descriptions and predictions but also counterfactual conjectures like “Any such object would fall,” plus the additional clause “because of the force of gravity” or “because of the curvature of space-time” or whatever. That is a causal explanation: “The apple would not have fallen but for the force of gravity.” That is thinking. ..”
and later trolls generative AI’s hollowness, in morality, creativity and logic (never to be overcome with better gen. AI):
“… Note, for all the seemingly sophisticated thought and language, the moral indifference born of unintelligence. Here, ChatGPT exhibits something like the banality of evil: plagiarism and apathy and obviation. It summarizes the standard arguments in the literature by a kind of super-autocomplete, refuses to take a stand on anything, pleads not merely ignorance but lack of intelligence and ultimately offers a “just following orders” defense, shifting responsibility to its creators.
In short, ChatGPT and its brethren are constitutionally unable to balance creativity with constraint. They either overgenerate (producing both truths and falsehoods, endorsing ethical and unethical decisions alike) or undergenerate (exhibiting noncommitment to any decisions and indifference to consequences). Given the amorality, faux science and linguistic incompetence of these systems, we can only laugh or cry at their popularity. ..”
Peter Norvic disagreed with Chomsky in 2011(?) on AI. Chomsky elicited AI as an engineering solution (e.g., Roman engineered bridges without knowledge of Newton’s science — seems to me more dialectic than indifference).
Finally, Chomsky’s rejection of probability as a non-starter wrt intelligence but strictly deterministic graphs is uncomfortable to many at best.
On routine research vs big picture research… let’s not let ourselves be fooled by what we find intuitively easy and hard. Recall Moravec’s Paradox (https://en.wikipedia.org/wiki/Moravec%27s_paradox) or Steven Pinker’s statement: “the main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard.” I’m not claiming to know that routine research will turn out to be harder than big picture research (whatever exactly that is). However, I see no reason to be strongly confident in the belief that the grunt work required in scientific labs (largely consisting of manipulating objects using fine-motor skills) will be solved by AI earlier than the effective integration of numerous sources of information and intelligent experimental design. If you think there are good reasons to believe that, what are they? Note: pointing to current limitations of AI as a reason is only persuasive in tandem with reasons to think those limitations will persist into the medium-term future.
I look forward to an LLM trained entirely on TED talks