Still, the guys developing ML algorithms like Neural nets, Linear regression, Decision tree, SVM, Naive Bayes, KNN etc., weren't doing science?

It's like asking whether a carpenter is a scientist because they developed a cabinet...

Well, Wikipedia might be wrong, but this is how they describe Computer Science: "Computer science is the study of computation, information, and automation. Computer science spans theoretical disciplines (such as algorithms, theory of computation, and information theory) to applied disciplines (including the design and implementation of hardware and software). Though more often considered an academic discipline, computer science is closely related to computer programming."

You might argue, though, that Computer Science is not "real" science.

I honestly feel like that, as a "Software Engineer". As the digital world is new, and full of similar concepts to the old analogue world, it's natural that we borrow names along with the concepts. Similar to how a lot of things in the sea are named after land things, with a marine prefix, like the sea horse, sea star, sea cucumber, and so on. And now in IT we have engineers, architects, rockstars, tribes, and science, even though, very often, they have no relation to the original profession or concept, and especially doesn't have the responsibility or impact of that.

Computer science is the study, that is looking at how computers and the use of computers can benefit mankind (in my short lay version). Software engineering is a subset of computer science where you build a real world application of the studies.

The algorithms behind the Facebook feed and your favorite AI/ML product is the science. The code that powers the feed and chatgpt is the engineering.

Wikipedia is not wrong IMHO, but then I studied computer science 20 years ago so maybe I’m out of touch

I agree with you. There's a lot of engineering, and science going into IT. But I'm a Software Engineer too, and while I'm clever and do clever things, I can assure you that there's no rigor or reason to call it engineering. This version of engineering is at most the re-use of the word, like how the word art is not just covering artistic expressions, but skillfulness too, even though when you apply a skill artfully, you don't create art. Similarly, I do a lot that takes ideas of engineering, but really I'm just writing code and kinda design a smallish system maybe.

Engineering is at all levels, I’ve designed systems that span global infrastructure and systems that run in embedded systems. I still think that design process is a key part of engineering. Figuring out how to put things together is part of design. I agree the more jr engineer you are, the less “engineering” you see.. but you can’t run until you learn to walk.

I talk to my friends who sent projects into space and friends who design deep sea drilling rigs. The engineering process is very similar.

I think that you could genuinely make that argument though.

I hate to break it to you but that is part of science. Perhaps the major part of it too. Hypothesis: I can build a cabinet with these materials which will bear some load range. Experiment: I built it and it obviously works.

Now change cabinet to particle accelerator that a giant team of other theorists and engineers designed. Am I not doing science by participating in building it? So experimental scientists arent scientists?

To me, you're describing the differences between a cook and a chef. Just because you've built something with well known methods doesn't make you a scientist. You didn't come up with anything new. In fact, we're starting to sound a lot like Apple. Apple is (in)famous for taking ideas that someone else did all of the hard work of developing and proving to work, and then take various ideas like that to combine into an actual useful working something. (They also do a lot of pure deep research as well, so don't think I'm going too far with the analogy.)

This image is burned into my retina https://en.meming.world/images/en/a/aa/Something_of_a_Scient...

semantics, I say. cooks are "chiefs" of certain things otherwise they'd have no value on a team. unless you're saying that chefs setting the menu means theyre more of a scientist. at worst you're saying e.g. scientists are like chefs in that their used their creativity to make up the standard model. at best you're saying chefs are theorists and cooks are experimental scientists which is , again, to my own point.

i'm saying that chefs are more likely to be aware of what is happening when doing the whipping, baking, cooling, etc and why that's important vs just doing it as a step. so when they experiment with the menus, there's a bit more than basic understanding in why they think the experiments might work. cooks are just the kids in chemistry class following the directions, and depending on how well they follow the directions (and to a point how well the instructions are written), they might not blow up the lab (or burn the m.f. soufflé)

says who? cooks at a fine restaurant better know everything about their part. but they chose to try to work under a greater chef because their eventual upward mobility as a practitioner is higher. literally chef means leader or the boss. in the rest of the world that usually means less specifically qualified. not likely here that a chef isnt a good cook but a good cook certainly has the ability to set a menu if they wanted. you're not making a strong argument imo.

Yeah to my newb eyes, the core of machine learning can (and probably is) as scientific as any mathematics goes (granted they try novel ideas, and not applying old solutions). Now openai sits more on the scaling/engineering to make his business.

There is actually an old argument about whether Computer Science is math or science. The lack of hypotheses and physicality drives the "it's math" camp. I tend to fall into that camp. Most computer science is either math or applied math.

That said, most of machine learning is not just engineering. If anything, there is little engineering done at all.

The note is interesting and in a way a time taking puzzle,

but do not forget the apparent point in the original poster's is the lacking approach below the implementations of above said techniques.

The techniques you listed in a substantial way expanded our knowledge ("oh, we can do this ... and solve more problems") while raising little more questions.

The said implementations on topic give results but raise many more unanswered questions.

(This post is unfortunately an "immature" reply - just a provisional tentative point. The matter of "ML algorithms ∩ science" requires more time and concentration and consultation of the texts of the Great Ones.)

>> but do not forget the apparent point in the original poster's is the lacking approach below the implementations of above said techniques.

That's right. I think the approach that is lacking is best explained in the submission guidelines for the Journal of AI Research:

>> Papers describing systems should clearly describe the contributions or the principles underlying the system. Papers describing theoretical results should also discuss their practical utility. In general, it should be clear how the work advances the current state of understanding and why the advance matters. Papers should report on what was learned in doing the work, rather than merely on what was done.

https://www.jair.org/index.php/jair/about/submissions

Doing stuff is not science, even if it's doing stuff that's never been done before. In science, we experiment to understand why stuff can be done, and how, not just that it can be done. And of course it matters what is being done, so for example experimenting to find the best combination of spices for a dish is not an instance of science.

It's more a science than pure math or CS, since they are producing falsifiable models.

In what sense are they falsifiable?

if they're wrong, they don't work.

Same is true of a bridge or a combustion engine though?

If my BST implementation isn't sorted, it also doesn't work.

The difference is that you can reason about why it didn't work.

With deep learning models not so. They are too big to reason about in the same sense as your BST algorithm.

Hence, you need a scientific approach to construct them. I.e., with lots of experimenting, hypotheses, etc.

That seems like it pushes it further from science, no? The point of a well-crafted hypothesis is that if it doesn’t bear out, you know that it’s because one+ of your assumptions was wrong. Your ability to continue your scientific inquiry is pretty much == your ability to then identify which assumption was wrong.

Not sure what you are saying here. Perhaps an analogy helps.

Psychology is a science. You can make falsifiable statements about the human brain. You will need experiments to build and test theories. It's the same with deep learning.

With computer "science" (and math) it's not the same. You can reason completely about your subjects, i.e. you can determine if something will or will not work just by reasoning, no experiments needed.

For more information on the differences between math and science I recommend reading: https://en.wikipedia.org/wiki/Scientific_method#Relationship...

Seems like the distinction is mainly between which tools are available to you as a scientist (at least if we stick to comp-sci, math is in a league of it's own). When, or if, we can completely model a human brain, a psychologist would no longer need to perform experiments to test their theories.

Given enough computing power, most theories could theoretically by proven or falsified purely through reasoning.

The point is: being able to run a brain inside a computer is not the same as understanding that brain. If you wanted to build a new brain, you'd have to reach for the tool all the time in an iterative way and hope for the best. Only tools that aid in understanding matter. We have only very few tools that help DL researchers better understand what they are doing. Hence DL is more akin towards science than towards math/CS or engineering.

You don't need to do an scientific experiment to tell why your BST doesn't work. "Computer Science" is a misnomer because most of its contents and methodologies are from mathematics (which is used in science but is not a science in itself).

CS uses mathematical proofs. You don't need a computer to execute your code to tell why the BST doesn't work. You can introspect your code and figure out why it works or does not work. If it's correct, CS methodology says you can "prove" that it works (without executing it).

Working with large AI models is like working with an artificial brain. It's as scientific as neuroscience in this sense. You make some hypotheses, tweak some hyperparameters, and get a result, which may or may not invalidate your hypotheses. Nobody knows why. Science is not necessarily about knowing the fundamental "whys" (amateurs think humanity has figured all the "whys" out, but that's a lie). It's about establishing some useful model of how things work.

But it's definitely possible to know why your BST does not work, even without a computer, without empirical testing. That's why CS is not a science.

Exactly. I think the word "science" changed its meaning a lot these years.

For some reason people tend to consider a field that is more "formal" (like pure math, some CS concepts like lambda calculus) more science, even though historically formal systems came very late, and practically very few systems can be described that way.

I really wonder whether people who regurgitate "machine learning isn't science" think theory of evolution is science or not.

I don't think things are quite as dire as that. I have experience as a machine learning scientist, have actually never worked on the kinds of papers you're talking about, but have helped apply machine learning to scientific problems. First in a systems biology lab for tasks upstream of drug discovery, and then in industry for credit risk modeling. I've mostly just used classical machine learning or Bayesian machine learning models though.

My interest has always been in using machine learning as a tool to help understand some underlying phenomena, not in trying to push to the top of the leaderboard on benchmark datasets. I think this kind of attitude isn't uncommon in academic labs focused on doing real science, though the methods used aren't necessarily state of the art, and ML typically plays only a supporting role.

Curiously, I'm going to present at a conference later today and will give a toy example showing why those leaderboards are not necessarily reliable for distinguishing between the quality of different models.

Pretty sure the idea of scaling is under test here, just because it’s relatively straightforward to use larger models doesn’t make it less of an application of the scientific method. Also, process supervision, devops is experimental for a startup like OpenAI, and the mixture of experts really hasn’t been used with MHA models at scale before.

I’d argue it’s pedantic to assume all science has to be completely novel or revolutionary. More neurons good, is a perfectly reasonable reason to experiment. I think you’re being a bit “snooty” in setting such a high bar to gatekeep science…

Are you secretly Ali Rahimi? :-D

{Hypothesis, test, loop} is the scientific method, and I can guarantee it is being used when fine tuning an LLM.

...but in a technical context, not in a scientific one.

One thing is identifying (age of copper; age of bronze) the best ways of smelting ore to obtain the metal through trial and error, another is to try and understand the nature of materials.

Those steps are both necessary and usually take decades to build understanding.

But the activity of "attempting to find explanations" is specific.

The attempt to formulate laws (the "better faring hypoteses") can bring to "universal laws" or to more humble "patterns". It will contribute to the scientific effort - it does not exhaust it.

Next time we get the urge to complain that GPT-n is just applying patterns seen in the training corpus we should remember humans don't do much better. We are just language agents with rich feedback from outside. We can't even write software top-down in one go without running the code.

> patterns

Except we consciously (intentionally) work on those matchings.

> are

Opinions should be offered in proper places. Polling is not debating.

Incidentally, that opinion is also offensive. Not just some tossed "All X are Y, goodnight", but a loaded one.

If you happen to conjure thoughts and let them roam wildly, well don't. Make an effort. Critical thinking as a conscious intentional effort must have plenty of literature. The "feedforward" output, the hallucination, is not a dish for service.

Corporate wants you to find the difference...

That is to say, the two things you mention are the same process. "Identifying the best ways of smelting ore to obtain the metal through trial and error" is the easy part, when you get to pick low-hanging fruits in a field. But as the easy options get cleared out, continuing improvements requires increasingly complex, sophisticated methods - that's where the process transitions towards "trying and understanding the nature of materials". It's still the same process, but couple layers of abstraction up from the original "how to get better stick than neighbor and beat them instead of getting beaten".

The difference, while moving in "interdependent" directions, is in the purpose: obtaining some sufficient information on how things work versus an actual consideration of the nature of things.

It is not really (fully the same process), because you could (in theory) "early stop" when you have achieved technically sufficient competence - the description of the optimal process -, before the jump to the understanding. Something works - that is the technical side; why it works is the scientific side.

Note that the approach has controversial sides: if you took Newton's "hypoteses non fingo", He refused that jump explicitly:

> hypotheses, whether metaphysical or physical, or based on occult qualities, or mechanical, have no place in experimental philosophy

I.e.: Newton proposed an «experimental philosophy» which stopped short of a "true understanding of the reasons behind phenomena", in order to avoid contexts in which solidity (at his time) could not be expected. "Understanding" as "identifying universal laws" was the best that seemed affordable.

I wanted to address that in the second paragraph, which I ultimately deleted before submitting, because I couldn't phrase it right. But since you brought it up: I'd consider this an effect of specialization.

In the process of improving your object-level "how things work" goals, you end up generalizing and stacking increasingly complex theoretical models. Soon enough, you end up with people working high up the stack - not knowing or caring about the initial goals. Those people end up growing the "mound of knowledge" both upwards and sideways. The work is sort of self-justifying, but really, it's also self-similar. Where early materials science may have been driven by, say, desire for better/cheaper weapons, soon enough, you have people doing materials science because they desire to solve a puzzle. Whether the practitioners are smelting different combinations of ores to find one that will win them the war, or they're mixing up different kinds of equations to figure out a clean solution to a theoretical conundrum - it's the same process, same motivation. And it always involves play.

The kind of methodical, boring approach, with hypotheses and control groups and peer-reviewed papers? That's the boring part you have to do after play.

See also (with no implied judgement in this context): software developers that lose sight of (or care little about) business goals, and instead aim for theoretical markers of what "good code" is, and/or solve abstract puzzles of algorithms and architecture. Or the MBAs that view companies as abstract money-printing processes, running them by the rulebook that's entirely independent of whatever it is the business is actually doing or selling. Both are cases of growing complexity creating a new field of work that's independent of what brought it into existence.

> an effect of specialization

It is also a specific purpose, a deliberate endeavour. We are in front of a world, we have to digest its phenomena, we try to understand it - it is a primal process, i.e. mental digestion. In some ways it could be argued that the "scientific drive" is anterior to the technical one: first you assess, then you act. Consistently, science started as "natural philosophy".

It is relevant, in the framework of "scientific drive as (development from the natural process of) mental digestion", that terminologically "science" is not just "the realm where false statements start to exist (and try to be avoided)" ("discriminate" - ex scindere), but also, before that, an activity of sorting, of "separating" conceptual elements - mental digestion. The "scient" (as a participle of scire, used as "to know") is a "distinguishing" mind.

We not only «desire to solve a puzzle» - it is our nature (as per the above) as minds dealing with a phenomenical world we need to grasp, as part of the activity of "seeing" it, already before action.

«Play[ing]» is a process that both allows to refine knowledge of the world, physically (in the physical play), and to refine the grasping of the world, mentally (in the mental play).

> growing complexity

Or just abstraction, enabled by a refined framework (following which, other areas become primary, topical - in the synthetic effort which is part of the above said process).

That's a necessary but insufficient condition for something to be considered scientific.

That's a common interpretation of what science is, it largely ends up being driven by confirmation bias. Because how do you know your hypothesis and test are even really connected? Or what you're seeing is a cause and not a correlation? This is why you need arguably the two most important factors in "real" science: predictability and falsifiability.

Predictability means that if your hypothesis is correct then you'd be able to formulate other improbable (and ideally currently untestable) predictions from it. And falsifiability means that if these predictions fail to occur then your initial hypothesis was also almost certainly wrong. So for instance Newton's hypothesis was that gravity was driven by a mathematical relationship between the mass/distance of two bodies. It was good science because it lead to the shocking ability to be able to dramatically simplify orbital dynamics, and create a complete predictive system of these bodies. It was even used to mathematically discover a completely unknown planet - Neptune.

Incidentally, his theory would also be able to be shown to be false if any of these unexpected predictions ended up being false. And that's actually exactly what happened. Observation of Mercury's orbit about the Sun showed it was off by ~1/3600th of one degree per century, relative to what was expected. And it's from there that people knew there was a mistake, which would only be explained later by Einstein hundreds of years later, who hypothesized a system with far more absurd predictions... and so the story continues.

ML has both of those things. In fact the situation is better than in many other sciences because the evaluation is completely digital. When you run an ML experiment you get back an objective measure in minutes or hours telling you if the model performed better or worse. If it's worse, boom, falsified. Many current papers are making predictive testable claims -- that based on various things we know what ideal parameter values are, the types of problems current NN archs will perform well on or poorly on, etc.

We've got lots of datasets at this point in history - it ends up being pretty reasonable to formulate a hypothesis, propose a change to the ML system, then see how it does on a host of datasets. Most papers focus on one or two datasets, getting some initial improvement going from train-to-test, but the successful ideas end up transferring to a wide range of contexts. The extension to other dataseets + contexts constitutes the prediction of performance on other systems.

And sometimes (often?) techniques do fail to transfer! IME, it's only the simplest ideas which do transfer to new contexts well - there's a lot of dark heat in pushing Imagenet another hundredth of a point, which you realize when you try to take incremental 'new SOTA' papers and apply them to audio problems.

But the things which /do/ transfer constitute real advances. And that's (drumroll) science! Sometimes things look good initially and get falsified. Medicine is still science, even if drugs get disqualified in phase iii trials...

It is not the scientific method at all. You didn't even include an analyze or a publish step, two critical components to the scientific method. The point is that science is a thoughtful, methodical, recorded, repeatable process that is scrutinized by not just your friends but by the entire world, including those who compete against you. In science, it's normal for your competitor to recreate your experiment exactly to see if you did it as well as you claim. Or to push it further.

What OpenAI is doing doesn't even resemble science at all.

Umm.. no. I’m pretty sure when science was “invented” (whatever that means), there was no way to share your research with “the world”. You wrote it down in a book in you cared to reproduce it.

Stop gate-keeping science. Observe, Predict, Test, Repeat. That is science. Kids do science when they watch adults do things. Just because you aren’t sitting in a lab publishing things no one reads doesn’t mean you can’t “do science.”

Just simply learning something isn't "doing science." The scientific method is a natural learning process it's not the only thing that defines science. This is incredibly light gate keeping. It's just maintaining a basic professional definition.

If a child watching an adult is "science", then when I brush my teeth I'm a dentist.

Words have meaning and when you dilute them down this far the only result is the destruction of communication.

Words do have meaning.

den· tist ˈden-təst : one who is skilled in and licensed to practice the prevention, diagnosis, and treatment of diseases, injuries, and malformations of the teeth, jaws, and mouth and who makes and inserts false teeth

Brushing your teeth does not imply you are a dentist.

sci·ence noun 1. the systematic study of the structure and behavior of the physical and natural world through observation, experimentation, and the testing of theories against the evidence obtained. "the world of science and technology"

A child may not be a good scientist, but they observe, experiment and explore their ideas. That sounds like science to me. Observational learning is a very common way scientist begin to understand things.

Do not forget that with ANNs we are still dealing with technologies with obscure sides: function approximators where the original function, and what it means, is oftentimes not reconstructed.

Replication of the experiment, "verification" of the law, can still be within a largely preliminary side - of the steps that describe behavioural details, patterns over "protocols", without the explanatory side (that which cause understanding, e.g. enabling deduction).

The publish step of the scientific method always seemed out of place to me - it means you can't discover anything on your own, which doesn't seem correct to me.

However, to discover something new (not just new to you) does require contact with established knowledge. And if you don't publish, you can't be very sure that it really is "new"; and it does not become part of established knowledge - it has not been discovered.

i.e. the concept of discovering new knowledge means it cannot be done on a desert island.

I think the colloquial understanding of "science" means studying how the world works. This is more like engineering: exploiting our understanding of how the world works to build new capabilities that we don't fully understand.

I don't think you would call it "science" for a bunch of single cellular organisms to cooperatively evolve a multicellular one. Similarly you wouldn't call it science when humans create digital lifeforms that require actual science to be done to understand how they work.

The reason that this is Engineering as opposed to Science, is that the hypothesis is just, "hey, maybe this will work".

Nobody has really explained why it works.

Which is a fundamental flaw in science.

Most of AI academics have spent their career theorizing complex algorithms or complex explanations of intelligence.

But the engineers have built large enough Neural Networks to give us data points that show intelligence is emergent out of relatively simple components.

Unsurprisingly, the people who believed they were the smartest were the least likely to explore the possibility that human intelligence isn't general, but specialized.

Echos of the academics building heliocentric models of the universe centuries ago.

Ask Feynman why magnets works... And he'll say no one knows. https://www.youtube.com/watch?v=36GT2zI8lVA

It's the same in ML. Good predictions do come from real ideas about how the systems work - ideas in information theory, entropy, cognitive science, statistical mechanics and so on - formulated in a context of our existing understanding and prior results. That's science.

Feynman was a genius. It takes a genius to say something like that, and it's surprising how many "scientists" are upset by statements like that, because they are so proud of their hard-won capabilities in modelling. Their curl and divergence operators, their Maxwell's equations... they forget that the map is not the terrain, and the universe is, at it's root, irreducibly mysterious, and the fact that we understand anything about it at all is gobsmacking.

>irreducibly

What are you basing that on?

This is my opinion, so take with a grain of salt:

I believe it's entirely possible that the scientific method breaks down past a certain level of system complexity defined somehow by thermodynamics.

This would in part be due to the infeasibility of running the proper experiments to understand the effects of single variables when tens of thousands of variables might be changing at the micro level.

> as it is practiced today

Reminder that such situation - science being out of focus, below the attempt to obtain practical results "empirically" - is not necessary but contingent.

The effort towards "science" is just postponed and/or relegated to other researchers in the same field.

And anyway, it is not that the effort towards "explanation" is completely absent. Many are working on the prospect of achieving big results through bets.

Science is knowing how the model scales when you throw this much processing power at it. They won't even tell us how much processing power that is.

> racing cool-looking sports cars.

I like the analogy attempt but there's a lot of science behind Formula 1

I love this diatribe! But to be fair, this kind of engineering historically can at least lead to science.

Sure. But remember also that the standard “science” is largely a broken model. Academia sucks. It’s honestly bad. It’s overflowing with papers that are misleading, irrelevant, or fraudulent due to a mix of poor stats knowledge, bad incentives, and a failure to organize effectively.

While people praise the scientific method, the majority of achievements we attribute to “science” are not derived from guess and check grad students doing their thing.

The standard paper is not very good. And a lot of the reason is that our scientific model is largely designed to find tiny effect sizes. Which is fine for like… medical stuff. But the juice is in large effect sizes. Stuff so obvious you don’t really need the stats to evaluate it. It doesn’t really matter if you followed the scientific method or not when you discover penicillin. It just works, clearly. And you can demonstrate it working again.

I think the search for tiny effects and the belief that they should drive treatment, is one of the biggest things wrong with medical research. Doctors will put millions of people on a pill that hits a vital metabolic pathway, for life, based on tiny (to my thinking) statistically significant (i.e. only a 5% chance they aren't based on a real effect) results.

These days I only believe large effects- like smoking causes heart disease

Medical standards aren’t that bad. It’s just a complicated system with lots of interacting effects that only work for specific circumstances. Coupled with the fact that many medicines are not cures, but something that makes you die slower.

Right. In historical context the scientific method was a huge improvement over the previous method of understanding the world, which was mostly religion.

But today, academia is a victim of organizational and political capture, making it less competitive for talent.

People struggle to differentiate the scientific method and the lifecycle of academia.

In some ways academia is akin to a religion or cult that sprang up around the scientific method, and the cult has now diverged as far from the seed as far right christianity has from jesus’s teachings

Oh come on, have some perspective: academia has millions more people to slaughter before they catch up with the cult of Christianity.

https://apholt.com/2019/01/30/death-estimates-for-the-crusad...

John Shertzer Hittell– Estimates 1,000,000 total dead for the crusades to the East covering the period from 1095 to 1291.

“In the two centuries of this warfare one million persons had been slain, but it had not been without some compensations.” John Shertzer Hittell, A Brief History of Culture (1874), p.137.

[...]

John M. Robertson– Estimates 9,000,000 total dead for the crusades to the East covering the period from 1095 to 1291.

“It is a reasonable calculation that in the two centuries from the first crusade to the fall of Acre (1291) there had perished, in the attempts to recover and hold the Holy Land, nine millions of human beings, at least half of them Christians. Misery and chronic pestilence had slain most; but the mere carnage had been stupendous.” John M. Robertson, A Short History of Christianity (1902), p. 278.

You might be stuck in an argumentative mode of thought, it seems like you replied to an interpretation of my comment that left out some salient details.

If you have some salient details about millions of people being slaughtered by academia, then please post them!

Otherwise my point stands that your comparison is ridiculous, because religion has killed millions more people than academia, despite Jesus's teachings and the "Thou Shalt Not Kill" commandment.

Religion has diverged vastly and diabolically further from the teachings of Jesus than academia has from the scientific method. There's no comparison.

You're being extremely anti-intellectual while whitewashing and rationalizing the slaughter of millions of innocent people, misery, carnage, and chronic pestilence, if you think academia is anywhere near as bad a cult as Christianity.

He compared it to a cult. Not all cults have killed millions of people. But many cults do stray very far from the original stated reason for the cult existing. For example some cults will treat their members like free labor and claim the goal is to make everyone happy through hard work and lack of worry about money. But then the leader will get a lot of money and start buying nice things for himself.

Buying a nice car with other people's labor isn't as bad as killing millions of people but it is straying very far from the original declared mission.

You seem to believe that murder is the metric that can be used to measure how well an organization is doing as far as following their mission.

I remain convinced that the correct model for science is a) serious hobbyists and b) patronage for the really talented. It would weed out the careerists, at least. You just can't make a "science factory" and expect to produce discoveries like cogs. It's stupid to even try.

I think you can. That’s basically corporate research teams.

The problem with academia is that it’s comprised of people who are held hostage until they find something. Just terrible incentives.

I'm tired of science as a religion. People treat it as some gospel, like if you check out some criterions you're suddenly "scientific" and instantly get a sense of validity and authority that you shouldn't logically get.

I judge things as "what you can do", not "what can you predict". The only demonstration of knowledge and understanding is being able to do something. Not predict. Not "scientific method" and ridiculous "peer review" (actually peer pressure), not blind trials and not rigorous statistical analysis.

In the end of the day, you either manage to do something or you don't. So much of so called science had lost all contact with reality because our judgement of success isn't successfully doing something, it is successfully jumping through "scientific" hoops. Look at string theory and social sciences. The scientific process, instead of being a tool, became the purpose. It became a stamp of validity to seek. A stamp of validity with gatekeepers in the academia, in the peer review process, in the media coverage afterwards all the way to social media censorship and "fact checkers".

What used to be the frontier of creative people became a stagnant beurocratic machine worshipped like a new religion. The side of the heretics burned at the stake became the ones crying out heresy.

Enjoy your new brand of science. I'll stick to the older brand of heretics and mad men which did whatever it was the prevailing orthodoxy told them to avoid doing and thinking, and I'll remind you that the only real reason those are remembered is because they did something useful, not because of the social traditions they adhered to or the rigorous scientific standards they followed.

Just add epicycles.

Add enough epicycles and you've got a Fourier transform... Epicycles were a great idea but applied for the wrong reason.

Oh, they worked really well; they were still used for numerical calculation even after Galileo. But they weren't a "demonstration of knowledge and understanding" of planetary motion.

Understanding is a moving target, though. Newtonian mechanics was incorrect for modeling the solar system as well - as it was eventually superseded - but that doesn't mean it wasn't a scientific understanding of planetary motion.

Epicycles gave a better description of the movement of the planets, based on the observations available, which was entirely falsifiable. They were eventually superseded, and that's science at work.

Interesting to say it's scientific because falsifiable. The objection is that it wasn't a theory, just fitting a function to data. It did "work" in that it captured some pattern: it was extremely good at generalizing/extrapolating/predicting. And was a "model" of something in the data. But there was no operational model behind it, of what was actually happening.

Newtonian mechanics has a model, beyond curve fitting.

BTW: I made this analogy between LLM and epicycles as a joke, but it's looking strangely isomorphic...

It's funny, because epicycles used to be the poster-child for ad hoc, overcomplex models with no conceptual basis... which is an exact match for LLM... but I don't recall the analogy being made. Not even in https://norvig.com/chomsky.html

This is incorrect: Epicycles were a model of planetary motion - the theory was that the planets moved around the earth, but also had additional circular motion as they moved along their path around the earth. This model explains the apparent geocentric motion of the planets, much as Newtonian gravity explains the apparent heliocentric motion of the planets (but is also wrong). Finding the exact parameters for the epicycles was the curve fitting part.

We now discount that model because it's based on an incorrect geocentric model of the solar system, but that doesn't mean that the model wasn't a model...

[edit] The CHomsky link is interesting - I just listened to an interview where he pooh-poohs LLMs at great length.

This point: "Statistical models have been proven incapable of learning language; therefore language must be innate, so why are these statistical modelers wasting their time on the wrong enterprise?" is interesting in that context - in the recent interview, Chomsky is now unhappy that LLMs can learn /any/ language, even unnatural ones, and therefore aren't good tools for understanding human language. Quite the reversal. I personally think it's a 'science progresses one funeral at a time' kind of situation...

At some point I'd like to carefully study the history of that early era of science because I don't know it as well as I'd like. But I believe I understand that the epicyclical model (and it was a model, rather than a theory) did not in any way depend on geocentrism. For one thing, Coppernicus' model itself, while heliocentric, retained the epicycles of the earlier, geocentric model. Instead, the assumption on which the epicyclical model depended was the shape of the planets' orbits and of the planets themselves, which were considered to be necessarily circular, and spherical, respectively. I think this had to do with assumptions about the geometric perfection of the universe, as a creation of the gods. In any case, assuming that planetary orbits were circular an explanation was needed for the apparent "retrograde" motion of the planets (meaning it looks like they double back and turn against their original heading). Explaining this apparent motion was why epicycles were hypothesised in the first place.

The first time this necessarily circular model was abandoned was with Kepler's laws of planetary motion, which correctly identified the motion of the planets as elliptical, that for the first time explained their apparent retrograde motion without the need for epicycles. Then Newton's theory of universal gravitation explained how the planets could possibly be moving on elliptical orbits. In fact, I believe Newton's theory of universal gravitation explained how the planets could be turning around the sun without crashing down, despite not having anything to hold them up. I think this was the first big mystery that the ancients tried to answer- hence the name of "firmament" for the universe.

And Newton's theory was not the end of the story of course.

I realize I've been repeating shibboleths from my postgrad without full understanding.

A problem with epicycles is they are harder to falsify. If the model doesn't match new observations, just adjust it, or add another epicycle. In contrast, Newtonian gravity can hardly be tweaked at all. So when Mercury's orbit was slightly off, they knew something was wrong.

I'm not quite clear on how I feel about this. Geocentricity is a theory, in broad terms. It seems disrespectful to say adding epicycles makes it "not a theory". As you say, there was a theory that the planets actually moved in epicyclic motion. It wasn't just calculation to them.

(I want to stress that the idea of epicycles, the mechanical craftsmanship, and actual prediction of the planets are all amazing genius.)

Yet, having more parameters than data means the model doesn't explain in simpler terms, only restates. In this sense, it's "not a theory" (by Occam's razor). It seems enough epicycles can model anything: (3Blue1Brown Fourier Series) https://youtube.com/watch?v=bL0LV0Huj1s OTOH the epicycles did predict planetary motion, so they did capture some regularity... not sure what to think.

RE Chomsky: You can see it's like epicycles: with enough parameters, an LLM is like a numerical method for curve fitting, that doesn't explain the data (any more than a fourier transform does). Curiously, they do seem to predict very accurately... yet also generalize strangely ("hallucinate").

But it seems it's got to help! Even if only as a device, like a telescope. Also, from this interview with Terry Sejnowski (https://youtube.com/watch?v=XKC-4Tosdd8) there's instances where a technique was developed for a problem with Neural Nets, and an equivalent was found in the brain.

He also gives a Chomsky-like view: if you duplicated the human brain and it worked perfectly, you wouldn't have done any science if you didn't understand anything.

>> RE Chomsky: You can see it's like epicycles: with enough parameters, an LLM is like a numerical method for curve fitting, that doesn't explain the data (any more than a fourier transform does). Curiously, they do seem to predict very accurately... yet also generalize strangely ("hallucinate"). What to think?

Well, that's the fundamental problem of modelling: that for any set of observations there's an arbitrary number of models that fit the data with great accuracy and even predict future observations well; and we don't know which one is the best in the long term.

The answer is that we should prefer not predictive models, but explanatory theories, that not only predict future observations but also explain why those observations should be expected to be made.

For example, the epicyclical model did not explain anything: it said nothing about why the planets should move on circular orbits with epicycles. Kepler's laws didn't explain anything because they didn't say why the planets should move on ellpitical orbits. Newton's law of universal gravitation explained it all in one stroke: because gravity. And that's why we consider Newton the greatest scientist of his era, not Kepler, not Coppernicus, not Gallileo, but Newton, because he explained the world and didn't just describe it.

>> BTW Chomsky's point E (which I'd never heard of), the last and most minor, was based on Gold's work.

Gold's negative learnability result was a huge upheaval that led directly to the current paradigm of machine learning. Chomsky used it to argue about his universal grammar but linguistics was only one of the two fields that Gold's result turned upside down.

And it was a negative result. As I say in another comment, science gives you the tools to know when you're wrong and that's how progress is made, when we find out where we were wrong before.

With epicycles, it took almost two thousand years before we figured out where the model was wrong. Let's hope that it doesn't take that long with LLMs and neural nets also, because I doubt we have another couple thousand years to spare on a wild goose chase.

> Chomsky used it to support his argument about the poverty of the stimulous but linguistics

Do you know where Chomsky refers (directly or indirectly) to Gold? I've been searching for a reference for some time.

No, I'm sorry. I'm not a linguist so I only know the relation between Gold's result and linguistics second-hand. I'm more interested in it from the point of view of inductive generalisation in machine learning; that's my schtick.

Just to make sure I didn't hallucinate all that, I had an admittedly perfunctory search online and I could find this paper:

https://proceedings.neurips.cc/paper/2002/file/04ad5632029cb...

Whose introduction describes how Gold's result is considered to support the arguments for linguistic nativism from the poverty of the stimulus. Then again, the author doesn't seem to be a linguist himself and he doesn't give any more specific references, so I'm now a little worried; and your question remains un-answered.

Have you tried wading through Chomsky's early work on linguistics? I don't have the courage to. The closest I've got to is I have a friend who has read a couple of Chomsky's linguistics books. My friend is making a living as an astrologist now so maybe that's a bit of a warning there :P

(not who you asked) I thought this would be in the linked transcript, but it's not. Norvig must be getting it from elsewhere (maybe in the 404ed video?), but it seems like misrepresentation.

The gravity model is similar though: we posit a force that pulls things together, but we don't know /why/ that force seems to exist, no more than the ancients knew /why/ the planets seemed to move in smaller circles along their circular paths. We're really not /that/ enlightened, after all.

I think that's right, but ultimately all explanations we have are based on prior knowledge that is itself not necessarily complete. It's explanations all the way down, until we hit some primary observations or axiomatic assumptions that are the hardest to get rid of.

"Enlightened" was my bad choice of a word. I get overexcited when I think of how much we have learned in the past couple thousand years and I forget that we mainly learned how little we know. Or can explain!

Yes. A matter of degree?

I'd like to think explaining means giving a model simpler than the observations. But this can be true of a predictive model, that offers no "why". Another commenter pointed out that epicycles do simplify - so they do "explain" in this sense.

> not predictive models, but explanatory theories

Maybe an advantage of an explanatory theory is in revealing more of the "black box", giving more ways to check the theory. (But I'm not sure how this could apply to Newton's gravity, since the only observations were outcomes. And no plausible way to "experiment".)

> If it wasn't for the ancients stumbling and fumbling in the dark for millennia

Is there any evidence that the epicyclic models helped scientific understanding, even indirectly? Later theories didn't seem to build on it. I wonder if it actually detoured understanding, with its misleadingly impressive accuracy, so that understanding would have progressed more quickly without it.

Thinking of pg's "great work" (https://news.ycombinator.com/item?id=36550615): to be the Newton of neural nets would seem the most ambitious aspiration of our times. But it took a bunch of geniuses just to get to Newton... and it seems an even harder problem than planetary motion. Though a difference is neural nets are based on actual neurons (loosely!).

It's looking like working human-level AI will precede understanding... perhaps by those 2000 years?

> BTW Chomsky's point E (which I'd never heard of), the last and most minor, was based on Gold's work.

Do you know where Chomsky says this exactly? I've been looking for how/if his argument is the same as Gold's.

I don't think the problem wasn't that it wasn't a "model" it's that it had lots of unexplained "plug-in" behavior. The heliocentric model had much less; it explained more using less despite being less accurate.

Nobody's arguing that the heliocentric model wasn't better for a variety of reasons... simply that getting superseded doesn't make the previous approach 'not science.' All models are wrong, some are useful, and all that.

And the heliocentric model still has plenty of unexplained parameters: the major and minor radii for each body. (Not to mention those pesky perturbations in mercury's orbit.)

Chomsky hasn't made any reversal, Norvig misrepresents what Chomsky has said.

> Chomsky is now unhappy that LLMs can learn /any/ language, even unnatural ones, and therefore aren't good tools for understanding human language

That's also not what he has said: He said they aren't useful for understanding the human language faculty, in other words, understanding how people are able to have language. As he says it obviously can't be the same way as LLMs because LLMs are able to learn languages humans can't learn.

Compressing millenia of astronomical observations into the compound of three or four sinusoidal waves whose formulae could be carried and calculated in a pocket book?

This is the reason why I left academia for startups... Seems like a better way (not perfect by any means) to innovate, actually doing things.

PS: I would be happy to connect, you can find my socials in my bio.

Science as a rear-view mirror.

You would love Thomas Kuhn.

Science is not academia. You say you prefer to judge things as "what you can do". Well, how do you judge "what you can do"? Astrologists, homeopaths, podiatrists, Christian scientists (!!!) and other such "heretics and mad men" rejected by the scientific establishment, will all tell you that they "can do" stuff, and so will all their many paying customers. How do we know they can't do what they say?

Because science gives you the tools to know that you're wrong. If you're a good scientist, you will be wrong _all the time_. That's how science advances: one mistake at a time. But you can't make mistakes if all you ever do is doing stuff with computers, like beating all the benchmarks, because that is a meaningless result judged by its own, self-chosen, measure of success that can never fail; and so can never inform.

Peer review is also not science, but it has been great to catch errors in my papers. Not in conferences, mind. I try to stay away from conferences. Everybody flocks to conferences because of quick turnaround, instant gratification. Journals have the good reviewers who can take their time understanding your work and helping you find where you've gone wrong. "Reject with encouragement to resubmit" is the best review result I ever got.

>> I judge things as "what you can do", not "what can you predict".

The goal of science is not to make predictions, but to understand how the world works, and why. Put that into instrumentalism's pipe and smoke it.

And whence the tools to judge that >science gives you the tools to know that you're wrong.

Placebos are known to work. I neither overestimate nor underestimate them. I understand them for what they are: placebos. And some people actually need them.

Moreover, in social contexts, religion also "works" in many senses. A person asked me what me what can he do about depression and feeling of meaninglessness. Science would prescribe anti-depressants, medicate him and he would both have side-effects and the problem would never be cured. I suggested that if he were religious, the meaninglessness would immediately go away and he would need no medication, no external interventions.

So again, judging religion by "what can it do", it can successfully give people meanings. Which is quite non-trivial thing that science fails to do to many people. So religion can be presumed to be acting on some truths regarding what drives and gives people meanings. Religion successfully demonstrated an understanding of that, you don't need statistical analysis or peer review to see that, it's plainly obvious. That knowledge just doesn't carry on to other stuff like how celestial bodies move, and there's no reason to think it does.

I love how you're the spokesperson for "science".

Ironically, when you look at how you describe the merits of peer review, you sound quite like an instrumentalist yourself.

Also quite ironically, the new science religion keep on pushing this funny narrative that science is about "being wrong all the time" which is like the complete opposite of scientific history. All of the established science is just the successes. Look at the great physicists. There is no mistake in Newton's laws. Or in Einstein's. They are just as valid today as they have been back then, in the areas where they were demonstrated and studied.

How did we reach this apologetic science? "good scientist will be wrong all the time"? That sounds like gaslighting scientists in a failed system rather than accurate historical statement. Science advances because someone eventually succeeds. In the grand scheme of things, there are no mistakes. There are corrections, there are expanding the domain of applicability. But the mistakes are sent to the trashbin of history, and if a failed system decided to gaslight you into rationalizing failure, stop listening to it and start seeking success.

>There is no mistake in Newton's laws. Or in Einstein's. They are just as valid today as they have been back then, in the areas where they were demonstrated and studied.

A model being incomplete is just another way of saying it's wrong.

The models they came up with are wrong, but still useful.

Sorry but I don't want to continue this discussion if you're accusing me of gaslighting others, and of being gaslit myself. Or of pretending to be a "spokesperson for science". I was hoping to have an honourable exchange.

> In the end of the day, you either manage to do something or you don't.

The hard sciences would like a word.

Taleb wrote well about that, changed my view on science. Not that science should be discounted, but it’s not the single source of truth. https://twitter.com/nntaleb/status/1419843561286160397

> Machine learning, as it is practiced today, is not science. There is no scientific theory behind it and there is no scientific method applied. There are no scientific questions asked, or attempted to be answered.

Total horseshit. There are tons of scientific papers on ML published. In fact it is MORE like traditional science than typical CS, because it is trying to reverse engineer how something we encountered in the real world works. We know NNs do amazing things, and we don't fully yet understand how.

They are doing computer science.

Obligatory xkcd: https://xkcd.com/1838/

Of course, 99.9% of science is engineering, not "science". We do science all the time without asking questions or attempting to come up with generalized answers.

Is a scientist doing a linear regression not doing science?

Anyway, alphafold, while not particularly "scientific", did answer one scientific question: it is possible to predict the structure of most proteins thru a combination of limited structural and extensive sequence information, combined with a sophisticated (and "non-scientific") algorithm. That was an open question for some time and their results convinced the community that their methods were right. What's amazing is that while it's entirely nonscientific, the results have been absolutely blockbuster in the scientific field. And even better, the only reason Alphafold was able to show this is because there was a well-defined protein structure leaderboard.