Comment by wildrhythms
2 years ago
This was a good read, and some great quotes.
>Perversely, some machine learning enthusiasts seem to be proud that their creations can generate correct “scientific” predictions (say, about the motion of physical bodies) without making use of explanations (involving, say, Newton’s laws of motion and universal gravitation). But this kind of prediction, even when successful, is pseudoscience. While scientists certainly seek theories that have a high degree of empirical corroboration, as the philosopher Karl Popper noted, “we do not seek highly probable theories but explanations; that is to say, powerful and highly improbable theories.”
Couldn't agree more.
The Popper quote is a bit outdated: scientists nowadays explicitly pursue both things: explanatory (mechanistic) theoretical frameworks and the selection of models based just on a maximum likelihood criterion. One informs the other. What chatGPT is doing doesn't seem to me to amount to either epistemology.
Not sure it matters what you call it if you can use those predictions practically in ways that traditional scientific methods were used but are slow/expensive, e.g. drug discovery.
Plenty of famous discoveries have happened accidentally even, and we study things all the time that we know happen but we are trying to figure out why.
You can complain that the system only told you how to make a room temperature superconductor, refusing to expand on why it has those properties, but you'll be drowned out by the excited cheers of people eager to both use it and study it.
I think it's a naive quote. Sounds wise. Is actually dumb. At least broadly applied in this context.
Lots of science is done without explanations. It's useful still. A lot of genetic research is just turning one gene off at a time and seeing if things work different without it. And then you say gene X causes Y. Why? Dunno. Genetics is not unique on this. Answering questions is useful. Answering questions about the answers to those questions is useful. But it spirals down infinitely and we stop at every layer because every layer is useful.
But moreso, machine learning models do embed explanations. LLMs can often explain the principles of their claims. Look at code generating models. Or code explaining models. Simple decision trees can illustrate the logic of newton's laws as mathematical rules.
Putting up claims of things that are proof of human specialness is just a reductive drawdown similar to how we used to explain everything as God's will.
> And then you say gene X causes Y. Why? Dunno.
Now this is definitely naive. Geneticicts definitely look for an explanation why this happened. Does looking for an answer involve randomly turning on and off some stuff? Yes. It doesn't mean scientists don't look for an answer.
As I said. Some do look further. Some do not. For the particular niche of genetics research, most of the time we actually don't, and that's fine, because it's not particularly actionable whereas the base layer understanding of a genetic interaction is helpful for things like personalized medicine.
We don't shit on the scientists that decide to stop searching "wait but why" and instead answer higher level questions. Because... obviously that is not always the appropriate thing to do.
1 reply →
The point is that we know many things as facts that we cannot explain. We may be looking for the explanation but, as of yet, we don't know why many things are as they are (as in the example above).
Actually, LLMs are also a good example. We don't know why chatGPT generates apparently cogent text and answers. What we know is that, if we train it this way and do a bunch of optimizations we get a machine that appears to be thinking or, at least, we can have a decent conversation with it. There are many efforts to explain it (I remember reading recently a paper analysing the GPT3 neuron that determines 'an' vs 'a' in English)
Finally, all science is falsifiable by definition, so, what we think we know now may be be disproven tomorrow.
Emergent properties is one of the places where pure understanding tends to break down under incomprehensibly huge problem space.
For example people have been doing accidental science the start of human agriculture by selective breeding without understanding the mechanics of DNA transfer. And your right Geneticists look for answers and attempt to minimize the size of the problem space in order to attempt to find answer faster, but the staggering number of interactions that can be caused by a single gene expression pretty much require to pick one place to look at with a microscope and ignore everything else going on around it in order to get an answer in a human lifetime.
Though calling it pseudoscience would be insane.
Nobody knows how general anaesthetics work. It's a stone cold mystery. Solving that mystery might lead to a new generation of anaesthetic agents or some other useful medical technology, but nobody is particularly perturbed by our ignorance; a practical knowledge of how to safely and reliably induce anaesthesia is immeasurably more valuable than a theoretical understanding.
Science might aspire to rationality, but reality is Bayesian.
> LLMs can often explain the principles of their claims.
Tbf those explanations are often just straight up bullshit nonsense.
I don't really care if this generation of LLMs is good or not. But fwiw, that's really not the case in my experience. On its face it seems unlikely that you can argue a machine that infers what a reasonable answer would be does not have an internal representation of the mechanics and actors present in the question. Otherwise it would not work. They clearly work well beyond regurgitating specific examples they learned from.
2 replies →
For the purposes of the article, it's fine if they're bulshit, it only matter that they are there
Randomly turning genes on and off to see what they do is experimentation. It leads to a better understanding of genes. Biology is messy and complex, so it's difficult to trace all the causes and effects. But there is some understanding of the mechanisms by which genes turn into phenotypes.
Certainly. And that's great. And often its also great to simply care about, say, gene-drug interactions and not find the root cause.
> Sounds wise. Is actually dumb.
Another excellent early entrant in the 2023 Accidental HN Slogan Contest.
Please don't respond to a bad comment by breaking the site guidelines yourself. That only makes things worse.
https://news.ycombinator.com/newsguidelines.html