How is AI impacting science?
michaelnotebook.com
michaelnotebook.com
I once had a co-researcher that asked for help with his Matlab code. We were in a smell-lab together working on mice. His code was about controlling some scent valves in response to a mouse putting it's nose in a hole and breaking a laser beam. Importantly, this code was running his thesis, essentially, and he had no coding experience prior to this experiment.
I said sure, I'd love to help, but you owe me a six-pack. So we sat down one day and started going through his Matlab code. After the 18th nested 'if' statement, I had to up the payment to a case of beer.
LLMs would have helped my co-researcher out a lot. He might have actually gotten the code working at least.
Most HNers don't appreciate, I think, how difficult coding is most people. AIs are already helping bridge that gap.
Another more mundane area is in communication. Research papers are notorious for being obtuse and jargon filled, to the point where many fields are nearly speaking another language entirely. AIs can help with that as well. Not only in the writing side, but also in the reading side. Can can put a paper into them and ask for summaries, you can put multiple papers in and ask for review papers. You must be very careful, of course, but the speed increase is just amazing.
Also, a larger point is that this was fundamental research. A very hard thing. You're supposed to be getting it wrong most of the time. Then when you think you've got it right, you try to publish. Then you hope the reviewers actually read anything you wrote and catch all the errors [0]. But a lot of the time, errors sneak through and really only get discovered after a few years/decades. That's fine in the end, as science is a 'strongest-link' system where we really only care about the big and correct discoveries. 'Weak-link' systems are the opposite, they are governed by the small little falsehoods and errors.
[0] Very unlikely, reviews almost always only review to see if the paper is of a high enough quality for the journal to include. I want to be clear on this point: Reviewers do not re-do experiments and typically do not go back through your math, let alone the code you used. Now, each field is different and has different standards, and those vary wildly. But generally, reviewers aren't there to make sure the science is right.
Otherwise you'll be at the mercy of whatever autogenerated code that you happen to be using, since you literally don't have the skill to check if it's actually working.
The other alternative is to work with a programmer.
Our 'salary' was ~$24k/year. So, just enough to stay alive and for a beer on Fridays.
If my co-researcher had the coding abilities he needed, he would have left.
And, apologies if I wasn't clear enough here, we were grad students.
My co-researcher had never taken calculus, let alone any coding. Trying to get him up to speed with all the various skills he'd missed out on would have taken another 4 years of full time 'remedial' education. Of course, there was no time for that, let alone budget.
I totally agree with you though. And that's why I really think that AI would have helped out a lot. It's not perfect in any way, but your velocity through all the pitfalls of learning is a lot faster with the AI helping out.
Which is not to say it can't be useful, but don't expect a cure all.
Writing fresh code is easy because you just express your ideas. What's hard is hitting all the contingencies appropriately; it can take years of 'burn in' to get right.
Reading code is easy - especially if you don't know the language well - because someone has already assembled all the correct syntax, etc. What's hard is that the code may be convoluted from years of monkey-patches to handle all those inconvenient contingencies, to the point of obscuring the core ideas. eg, making you search through ten nested indirections to figure out how something actually executes...
On the writing side, LLM's make it much easier to have the idiomatics of the language easily on hand. (eg, I never ever remember the exact way to use 'super' in python... but ML assistance is happy to fill in the parens as soon as I start them.)
What I’m getting from these comments is (as expected), most people are happy to take the easy assumption filled route rather than actually applying rigour.
We were a neuro lab, so coding was seen as something similar to a hammer: Bang away with it all you want, but at the end of the day, it's going back in a dark drawer where it belongs with the other tools.
I think the gap is quickly narrowing on the need to fully comprehend most
computer programs for most use cases, ...
I wish it could, but it appears not to.
As an example, ask it to code up something from a "less popular domain" (e.g., Clifford algebra).
You might notice that it appears to not have enough data to make good guesses. Some "AI" hype folks appear to move the goalpost then.
However, maybe the stock market and some companies benefit from all this sensationalist stuff.
I wish I had a HAL 9000 beside me, but I am sorry Dave, I cannot let you do that.
We are basically at the point where we could build the fembots from Ex Machina or David, the little boy from A.I., if we were just a little better at robotics. We'll see the teddy bear Supertoy from the latter movie under peoples' Christmas trees inside five years.
The ability to reason or understand seems to be not there in current LLMs.
See my other comment:
https://news.ycombinator.com/item?id=38819544
Simply put: "fake it 'till you make it" = probabilistic guesses given a large data set = seems to fool us people believing that an LLM "thinks" and "reasons".
There you go. Try it with 4. Still doesn't work? Let's revisit when 5 comes out. You will lose the argument eventually, it's only a question of when.
with deriving the geometric product (Clifford algebra) based on a set of definitions and axioms (e.g., the distributive property). Unfortunately, it failed, making numerous errors along the way.
Also maybe try asking it something that ordinary humans can be expected to accomplish, rather than complaining that nobody has invented Ramanujan-as-a-Service yet.
Also: Further, "AI" beating masters of chess also seems to be a product of "beat a human with probabilistic guesses given a large data set".
That was true for chess. Lee Se-Dol's defeat by AlphaGo was different. That wasn't supposed to happen. That was a combination of probabilistic Monte Carlo methods and... something else entirely.
That's really when the train of "AI can do X but not Y" left the tracks, IMO.
There you go. Try it with 4. Still doesn't work? Let's revisit when 5 comes out.
So, it "thinks" and "understands", it does not make (probabilistic) guesses?
Definition: guess
to give an answer to a particular question when you do not have all the facts and so
cannot be certain if you are correct
To your other point:
Also maybe try asking it something that ordinary humans can be expected to
accomplish, rather than complaining that nobody has invented
Ramanujan-as-a-Service yet.
You will lose the argument eventually, it's only a question of when.
"To understand" and "to think" are two very different things. Understand means more or less to encode and compress effectively from a perspective of such a system - there's quite a bit of evidence that they do that.
As for "thinking" that is impossible for LLMs as thinking is an action - and LLMs aren't agents that can plan and take action.
Actually AlphaGO and AlphaZero were agents capable of thinking - just in a extremly simplistic world which is the game of Go, Shogi or Chess. But they had a world model (which was fully known as it was for simple games) and a way to plan the action they will take by evaluating what impact will they have upon the world and how beneficial it will be for them.
Just that extending that system/agent to the real world is very hard.
A flippant way to answer the question might be to ask, "What makes you so sure you're not arguing with an AI right now?"
A less-flippant answer is simply that we would need to agree on what understanding and thinking mean before we could debate whether AI is doing it now or can potentially do it in the future. I doubt we'll come to agreeable terms there, but I'd start by suggesting a simplistic hierarchy of cognition:
0) Memory: the ability to store information. Trivial.
1) Knowledge: the ability to retrieve that information associatively.
By that I'd imply a requirement that knowledge must be retrievable by indexing at a conceptual level. Knowledge isn't "I have four bananas and the third is ripe"; that's memory. Knowledge is "Bananas are edible fruits." It's more or less unquestionable that this is satisfied by embedding memorized items in a high-dimensional vector space where conceptual similarity corresponds to distance measured by some norm or another.
2) Understanding: the ability to cast new knowledge in terms of existing knowledge. This is where the increasingly-popular compression-as-intelligence metaphors start to come into play. Anytime you use an embedding scheme to compress data, you are exhibiting understanding. When Stable Diffusion encodes information about entire works of art with one or two bits of actual information, which it can later use to synthesize new works, that's "understanding" IMO.
3) Reasoning: I'm fine with the definitions of inductive reasoning (applying specific understanding to the solution of general problems) and deductive reasoning (deriving specific outputs from general understanding.) AI models perform induction whenever they recognize and encode patterns, and they perform deduction anytime they solve a specific problem using those patterns. At this point the burden of proof falls on anyone who claims these processes aren't happening.
So that leaves "thinking", which is a topic in philosophy rather than neuroscience or compsci, and forms of reasoning that go beyond the simplistic inductive and deductive models. Hypothesis formation is interesting, for instance. That's one area where someone using a GPT 3.x model might walk away with their negative preconceptions regarding AI being fully validated, as you did, while someone using GPT 4 might walk away with a furrowed brow and a dazed look. Maybe, once the previous bases are covered, Occam's Razor is all you need to perform abductive reasoning. If not, what else?
These are all interesting points of debate that go much deeper than "LOL, it can't even construct proofs using Clifford algebra." If you're satisfied with that, then again, there's not much room for the conversation to progress.
Obviously it could be completely wrong, but it certainly _seems_ to comprehend geometric algebra more than I do.
English as a programming language makes this all work.
Some of the comments here actually freak me out.
It's kludge after hot patch after voodoo incantation. I once worked on a computer that under no circumstances could it be put on the web. It was running Windows 98 in ~2017 and was only used for a single program to control an instrument. I know for a fact that it's still online and running. It was slow as all get out, but no one had any kind of time, let alone gumption, let alone skill, to try to bring it up a millennium.
On old hardware/software, that's normal, and will continue to be normal, simply because the value is in the instrument, and money spent trying to get the software/drivers to work on more modern systems would be better spent on buying a new and better instrument. The only way this would change is if we force all hardware vendors to fully open-source (and ideally upstream) their code (look at the whole IoT ecosystem).
One still needs to find that balance between smart, functional / practical and maintainable. That’s the art of coding.
This describes me quite well. During the last few days I was working on getting the matrix for an integral operator, using Chebyshev polynomials. Numpy has regular and Chebyshev polynomials, but the documentation is a bit of a mess. I asked ChatGPT for help. It didn't work that well. But it got me started. It took a bit, because Numpy has actually 2 APIs for all their polynomials, and the ChatGPT code was mixing them. Once I figured that out, I decided to stick to a single API, and I finished the code snipped in about half an hour.
Without ChatGPT it would have been much more difficult. How many people out there are aware of the Chebyshev polynomials in numpy? Certainly nobody in my circle. ChatGPT was aware, but it was also hallucinating quite a bit. Still, it helped me get the job done.
what I'm saying is that there's no quantum computing; but I actually mean to say I simply do not understand whatever 'quantum' means in the context of computing given my own opinions of what computing really is (I have a philosophical opinion).
in my view (as distorted and twisted as it is), what quantum computing devices really do is sensing or measuring. computing is classical in nature and no amount of hand waving will change this.
whatever happens to logic in a qubit? if both A and B are anything between 0 and 1, what's the negation of such a thing?
The problem is the pop sci explanation where qbits are everything between 0 and 1 at the same time is super wrong and misleading.
Your qbits are in states that have "amplitudes". You do things that change the amplitudes, and that happens to do computation. From the amplitudes of your qbits, you can figure out what the probabilities of observing different outcomes is.
It's not that the qbit is anything between 0 and 1. It's in a precise state, that results in percentage probabilities for different outcomes.
The negation of true 80% of the time is false 80% of the time.
https://quantum.country/ is a pretty approachable starting point if you’re curious. The math is no more intimidating than any other skilled discipline.
"aren't you just too stupid to understand?"
maybe I am... maybe I just think differently with another brand of depth
what I wonder now, given another comment about amplitudes, is what happens to combinatorics? (and alphabets, and languages as sequences of strings from finite alphabets) but it's just simpler to dismiss me as foolish idiot.
If your qubit is "all values", the negation is the lack of any value, e.g. "a measurement that results to 0".
However, in most cases, you will have a segment within the possibilities. E.g. one qubit can hold "vertically polarised light at frequencies between X and Y at phase P", then the negation could be one of "horizontally polarised light at that frequency and phase", or, more common, same polarization and frequency but the phase is 180 degrees off P. That way, if you add them together, you get the cancellation. Another negation is "all other possible combinations".
However I must note that when you call a qubit "all possible combinations" that's not a typical qubit value.
Think of it like this, "binary" is "0 or 1", but a "binary value" is only one of either 0 or 1. A qubit value is, depending on your interpretation, photons of some polarisation, frequency and phase, or electrons of a particular spin, or some other combination of those. Some of those can have multiple combinations, for example a qubit value could be the combination of a bunch of photons at different values. Or, an electron with an "unknown spin". Or an electron with an unknown spin that's the complement of another electron at another unknown spin, but when you measure either you will reveal the other, and so on. So, only some qubit values will have "all possible combinations" i.e. "when observed, it could literally be anything".
The internal of "what was the value before you observe it" is, well, controversial, and to most people, almost irrelevant, and even, "not good to consider".
I think von neumann cant handle purely parralel algorithms efficiently. Even gpus. I think thats whats where the "all qbits interacting simultaniously for n steps after being put in initial conditions" is about. Its purely parralel.
It could be the case that information cant be extracted before it is computed, and that for some algorithms the peak efficiency serial version is equal to the best efficiency purely parralel version. It could also be the case thats not true and that classes of quantum algorithms could be better.
At the end if the day the way i see it substrate is just an engineered means of applied ops per time.
The only way I ever found an access to understanding quantum computing is by doing the math, as other pop-sci explanations don't really reflect what is happening (at least for me).
Materials science and chemistry (e.g. predicting how drugs may behave in our bodies) will benefit, or other computationally heavy "try out many permutations and see which ones fall though" tasks may become a bit easier. Climate modelling, supply chain logistics, financial modelling, and of course AI.
Modern quantum simulators running on supercomputers is an exponential (O(k^n)) kind of operation. But a quantum computer can simulate any quantum effect in just one step, O(1) time. Because ya know... a quantum computer is just a computer that has isolated quantum effects and allows a programmer to control them easily.
Yes and OpAmps / analog electrical circuits were way faster than digital computers for decades because of this.
When you know that the current of a diode is related to the exponent of the diode's voltage, you can do silly things like calculate logarithms using OpAmps and diodes in a mere nanosecond or so.
Most 1980s synths used OpAmps as the basis of the math / calculations for signal processing, because digital computers just weren't fast enough to compete yet. Those circuits still work today.
Alas, digital computers are too cheap, accurate, and fast these days so nearly everything is digital now.
----------
I guess your point though is that maybe a quantum computer (or OpAmp / electrical computer) can only physically simulate the effects that the hardware contains.
Ex: an electrical computer cannot simulate BJTs unless it has a BJT somewhere. So it would still come down to the computers physical load out.
Similarly, a Quantum simulation on a quantum simulation computer would only have some subset of primitives implemented.
In the case of codebreaking, where QC may have an advantage, AI is unlikely to provide an alternative because modern AI is all based on finding probabilistic patterns and cryptography is explicitly designed to be resistant to that kind of attack.
To program a serial algorithm in a quantum computer would be to miss the point. You could encode the interactions between clumps of qbits to function as nodes in a wave simulation.
There would be few enough nodes in the wave that it would be not equivalent to a true wave simulation. It would be a discretization, but now it functions as an analog asic like proxy for the real thing. If the groups of qbits are smaller than the true nodes in a wave surface your computation would be faster than the real thing, but lower precision.
In a classical computer this would have to be done with buffers, or a very narrow set of parralel deterministic programs, otherwise impossible. (examples: a subset of cellular automata rules, gravity sort)
Is any of that right or am i completely off base.
Quantum is quantum. We did a few quantum gate simulations in my college back in the day.
The gist is that entangling bits (say b0, b1, b2, b3 are all entangled), has special properties. You can almost perform retroactive computing. you can send these values to quantum gates to define properties, like b0 + b1 = b2 * b3.
Later, you unentangled the values and they will 'snap' into some true state. Like b0,b1,b2,b3 == 1111, but the values 0001 or 0010 are also possible.
There is nothing 'parallel' about this. It's just a quantum operator using the properties of entanglement / disentanglement for the purposes of computation.
-------
I'd say that quantum is closer to retroactive computing than parallel. Operations seemingly go backwards in time thanks to entanglement.
That doesn't mean quantum is useful in all algorithms, but it seemingly has applications in cryptography and 3SAT / NP Complete and... Lol simulations of quantum effects.
--------
Ultimately, Quantum is only useful if you get enough bits at a time entangled. For now, it's faster for standard computers to brute-force all 2^32 bit combinations rather than using lasers+diamonds to entangle 32-bits and perform the operations.
Quantum needs more entanglement and at lower costs to become competitive. But maybe it happens. At very least, quantum computers do exist but who knows how long it will be before they are competitive.
Notice that the GP is about the other way around. Solving NP-hard problems will absolutely make quantum computers useless.
And in the case of physics simulation (the actually valuable use for a quantum computer), there's no guarantee that you can't find a probabilistic solution.
The bigger concern of companies like that is if someone bankrolls dirt-cheap FPGA’s or Adapteva-style cores with open architectures targeted by a toolchain like Synflow’s or Cray’s Chapel. From there, domain-specific applications (esp optimized kernels) targeting those tools. Then, MPP-style hardware as cheap as commodity servers to scale it up and out. I’m talking engineering the thing by combining proven strategies, not doing new research.
Even $100 million put into such systems would deliver more value in more areas than $1 billion put into quantum computing. If not, we’d at least get huge speed-ups with the parallel architectures for a while and then the QC folks eventually deliver something better in some areas. I’d be happy both ways so long as one is in my building or it’s in the cloud for $2.30/hr. :)
Honestly, I'm not sure of the long term solution here. Conferences and journals may have to introduce a system where you need an existing member of the community to "vouch" for your work to be allowed to submit (they don't have to say it should be published, just worth reviewing).
While I think it’s probably fair for people with English as a second language to use LLMs as a writing assistant (for example), there definitely needs to be some kind of author disclosure statement at the very least. This could be similar to how more papers are now including contribution statements when there are many coauthors.
Then you say rewrite the final paragraph to be slightly shorter and incorporate a certain example. Yada yada.
Likely you can even do this "write your response in a style that is not typical for an LLM"
Fold prediction is an incredibly useful tool for scientists and genetic engineers to help design new proteins, but it doesn't magically solve molecular or cell biology. Designing new functions and mechanisms is still going to involve a huge amount of labor and brute-force experimentation.
So at least according to him we've lived in that world for the last 5 years.
As a person working with / tangentially to people in the same field I would say that it's made things faster and more scalable, but protein structure isn't the be all end all of things. Researchers use AlphaFold a lot for filtering potential candidates, but that is only one step in a lot of steps. A SNP mutation -> protein structure change -> functional change is already difficult without then considering that the vast majority of mutations that create function change in humans are not in exons, so something like AlphaFold (in the form I'm familiar with) would be useless for those as well.
Eventually though an AI system that can go mutation->function change is entirely possible, although it is much much further in the future. In that case though I think you'll be quite close to a future where combined with things like CRISPR therapeutic treatment for all heritable disease would be possible.
To me, "solving" structure prediction (explicitly acknowledging there are areas where AF doesn't make accurate predictions), is a clear and satisfying win, although it still doesn't answer some of the fundamental physics questions around folding.
I am glad to see the existence proof but want to see more outcomes; in particular, I'd like to see a lab-in-the-loop that actually produces something of high value (higher than the cost of building the lab-in-the-loop).
AlphaFold is a huge advance and people are extending it to try to tackle this (eg AlphaMissense) but to say it solved protein folding is hyperbolic.
Quoting a PhD immunologist from Texas, who uses AlphaFold daily.
Just my ¢¢ half-sanity.
[1] Isn't a single hydrogen diatom just 1nm wide? How much smaller can transistor gates be safely assembled? Didn't we physically stop getting physically smaller around 14nm?
I think its like saying "we know how to make bricks and iron and bolts now, shouldnt it be easy to make a full scale functioning replica of the greater tokyo metropolitan area?"
Even if you came up with a spec for a complex tissue or just a fluid that functions as a standalone chemical factory, you would need to fabricate it. However many specific protiens from your library, in specific ratios, mixed and maybe even... positioned.
In short protein folding is just the first fundamental step towards this print any biological design out of proteins world that I guessed you are alluding to.
So my takeaway is that we just need sufficiently advanced quantum computers that can do molecule-level simulation on a large cell, THEN we'll be there.
Microsoft has an initiative called AI4Sciencie (https://www.microsoft.com/en-us/research/lab/microsoft-resea...) which published a fair amount of SDE/diffusion-based method to solve scientific problems
But it goes to show just how fast we are currently progressing.
We might see some interesting combos.
EDIT: Maybe use a sample from the Humies as benchmarks for the techniques, including new LLM’s. Let people try every approach in parallel mixing the best of each.
ChatGPT can lend a helping hand
Keep in mind that proteins and deducing the structure of them (which in turn would help deduce their function) may not be a good thing. Proteins can also be poisons, prions, etc.
But beyond that, we should not have such easy to access knowledge. It's just too much power for our current greedy, short-term thinking.
But beyond that, we should not have such easy to access cutting technology. It’s just too much power for our current greedy, short-term thinking.
Moreover, I actually agree with you in some ways, even though you meant to be sarcastic. Modern cutting technology, which certainly should include deep-well oil drilling, IS too much.
In fact, all technology has a limit beyond which we should not go. We just have to apply deep thinking to figure out what it is, or at least do our best to try.
You
Technology is fine. It’s this artificial cookie-cutter society we’re forcing on everyone. We have a mental health and education problem.
The tech isn’t the issue, it’s the people who feel the need to use it for negative purposes.
Limiting tech is just a bandaid.
Technology is not fine. Not all of it is bad, but to say all of it is fine makes no sense. Yes, we have a mental and health education problem, but it's reasonable to have rules for all societies. Limiting technology that makes getting ahead too easy makes sense because even in the most utopian and good societies, if gaining something is TOO EASY, then it WILL be taken.
It's much too simplistic to frame technology as fine. If all technology is fine, perhaps every corner store should sell automatic weapons? Can you imagine a society, ANY society, in which this makes sense?
You are talking about mental problem. Do you know what have contributed to mental problem ever since the existence of human being? Isolation and loneliness, which technologies such as telephone, email and yes, social network can help with.
Name any technology that should be banned, and I can give you a reason why it should be not.
Your point seems in agreement with this to a degree.
More importantly, you claim it to be a mental health/education issue; yet you can't show me a period of human history without wars and power struggles.
I suppose the deeper point I was trying to make was that “this time is different/special” is shallow thinking, as is believing that the real issue is with our “current greedy short term thinking”.
The real issue is the human condition, which remains largely as it has throughout our history. We need to choose a way to live knowing we will die, and then live that way.
What is different about cars and new tech is that new tech comes with the ability to track and monitor all users. So historically we didnt mind if one in ten million people wanted to ram people with their truck on purpose. But now the power to monitor everyone all the time to prevent these "one in ten million" crimes exists.
Its up to you to decide if the trade off is worth it.
Until tech can do that it hasnt gone far enough.
Like everything else, it's only good up to a point!
- I believe in basic medical care, but medical technology that permits immortality is bad.
- Likewise, is reducing mortality always good? Now we have extreme overpopulation. In short, reducing OUR mortality has also INCREASED the mortality of non-human animals, which I believe have just as much right to live as us.
- Fertilizer and agriculture? Most industrial fertilizer poisons the land and agriculture has eliminated millions of acres of wild forest and destroyed a lot of ecosystem. Agriculture is also responsible for the huge population growth that is causing global warming.
- Cure for cancer? Maybe good, maybe bad IF that cure results in significantly extended lifetimes due to how it works.
I do have MANY more nuanced answers. That is why I believe in looking at each of your points AND others individually. My answer was specifically directed at AI, which I believe to be universally bad.
Most of the technologies you listed have significant negative consequences, and it's not clear to me that those are outweighed by the positives, especially if our modern society results in the destruction of the phytoplankton of the ocean producing half our oxygen.
I suspect in the long term, it would be better to have more mortality, smaller populations, and less technology than our current world. It's not a popular view because people are emotional about such things, but it's hard to imagine how a world with 8+billion and mass eco-destruction to be better IN THE LONG TERM.
Yes, I am making an ethical judgment about the means of AI, because after some analysis I do believe that the negative effects will be far worse than the positive. True, there are good applications of it. I never denied that.
Of course, my line of reasoning is far more vast than what I wrote here. I'm just arguing a specific point here, not presenting a thesis. I do think there is more to saving animals than arguing against human growth. Of course, people need incentives.
Take your stand on AI, which is more diverse than the LLMs and diffusion models that are currently making the news. Would it be bad to replace peoples by computer vision and robotics in a waste sorting center, enabling more recycling and less microplastics in the environment due to a better sorting of the waste stream ?
I see that as a good thing enabled by AI.
Antibiotics have saved billions of peoples lives from an early death. But use of antibiotics in farm animals is putting us at risk again via resistant superbugs that are sending us scrambling for new drugs to combat them.
Or, to take the analogy to 11, should each individual on the planet have access to a button that if pushed would cause the planet to explode? I seems reasonable the answer would be no, but now how do you set up a system of rules limiting who has access to the technology to make that button?
2. The laws of physics have so far prevented us from making technology like a tiny portable button that blows up a planet. Id assume a universe with weak energy constraints like that would have yielded much larger scale organisms, or perhaps none at all.
1 is you get a scratch then die because your body rots and you can do nothing about it. May want to spend some time learning about antibiotics resistance.