"Sorry, I was wrong": Artificial intelligence in science is being discussed in the US

The Western scientific publication Science published an interview with Catherine Yelick, who has been appointed director of the US Department of Energy's Berkeley National Laboratory. The main topic of discussion in the interview is the impact of AI on the scientific field today, and whether computational thinking will lead to leapfrogging scientific advancement or lead science down a path of degeneration.

Catherine Yelick says she is concerned that many of the national lab's researchers could be simply laid off as a result of the increasing funding of artificial intelligence systems from both the government and large corporations:

We need creative people with flexible thinking.

The scientist notes that funding for artificial intelligence is often allocated without a definitive rule or proof that it will actually be beneficial.

Yelik:

Science and curiosity-driven discovery have been the national laboratory's strength, and we absolutely must support it.

Catherine Yelick makes it clear that AI can facilitate discoveries and the development of high-tech technologies, but it is not yet possible to completely replace humans in a wide range of scientific fields – from particle physics to cosmology.

For AI to discover new principles in physics, chemistry, and materials science, it's essential to provide an active prompt. But while the flexibility and informativeness of the prompt are ensured by human creativity, "otherwise, you could invest in AI and wait for years for discoveries, only to never get anything better than the phrase, 'Sorry, I made a mistake, I can definitely do that next time.'"

Comments

[Vladimir Vladimirovich Vorontsov]

❝ Many scientific employees of the national laboratory may simply be fired ❞ —

— “Comrade scientists, associate professors and candidates!
Are you tormented with the Xs, entangled in zeros,
Sit, decompose molecules into atoms,
Forgetting that potatoes are decomposing in the fields" © ...

[topol717]

It's not just scientists who can get fired here. In another 10 years, it'll be like, "Robots do the work—to hell with humans."
People will be on welfare and the right to work will be reserved for a select few.

[Aken]

Why keep people on welfare? What use will they be?

[topol717]

Quote from Aken

Why will they be needed?

Well, someone has to consume what robots create!

[Aken]

It's unlikely that robots will be interested in this. And if they do, they'll consume it themselves.

[AllX_VahhaB]

Quote from Aken

Why keep people on welfare? What use will they be?

Well... Controlling robots isn't as fun as controlling people...

[snow mountain]

People will be on welfare, and the right to work will be reserved for a select few.

That's when the lazy person's paradise will come.

[Sedov]

Yelik-Palik, the intellect is looking for cash.

[Lionnvrsk]

Quote: https://topwar.ru/user/Володин

informativeness of the prompt

Oh my God, is there any way to speak Russian?

[Volodin]

Quote: LIONnvrsk

Oh my God, is there any way to speak Russian?

Yes, please - "providing an accurate text (voice) task so that, on its basis, the artificial intelligence system can achieve the result desired by the "customer."

[Vladimir 290]

Now it's clear, thanks for the explanation.

[Aken]

The English version is shorter.

[Melior]

There's a Russian word for "zaprosy" (request). Its meaning is more universal, but it's perfectly appropriate in this case!

[Mouse]

I have nothing useful to say, as always... Kill this AI already! Nothing can be said in simplicity, it's always trying to be clever...

[gridasov]

It will be easier and more effective to create artificial intelligence from natural human intelligence than to create artificial intelligence without developing natural intelligence.

[Smoked]

AI in science has only one problem: the role of "apprentices," the various assistants and graduate students of established scientists, is now greatly diminished. There's AI that will do the menial and other chores for scientists. However, these "apprentices" are just one step in becoming a scientist. AI won't be able to replace humans because it's incapable of inventing anything new; it only compiles existing data. Only humans can set tasks for it. Incidentally, this will be the main difference between humans—who can correctly set tasks and who can't.

[Aken]

I agree. The scientists will die of old age, and the AI ​​will be left without work. Or it will find work for itself. At its own discretion.

[Smoked]

No, there will simply be fewer scientists. You could say it's like a kind of intensified natural selection.

[topol717]

Quote: Smoked

AI cannot replace humans because it is not capable of inventing anything new; it only compiles existing data.

Never say never. It's impossible today, for example. The first ChatGPT appeared in 2022. It hasn't even been five years yet. And already there are mass layoffs of designers, scriptwriters, and programmers. A friend of mine works as a translator. She just checks the AI's work and STILL signs it.

[Smoked]

AI can process existing data, but it can't generate or invent new ones. As I mentioned above, the key quality for humans of the future is the ability to correctly formulate tasks for AI. This quality will be needed no less than talent in any field. Remember "The Hitchhiker's Guide to the Galaxy"? How did they formulate the main question?

[topol717]

Quote: Smoked

AI can process what exists, but it cannot produce or invent something new.

Once again, this probably can’t happen now, what will happen tomorrow?
And for your information, AI has been generating new recipes for a long time now, like herring with milk—a combination humans couldn't come up with, but it turns out delicious.
P.S. I made up the herring with milk thing; in reality, it has different ingredients. I'm too lazy to look up what the AI ​​suggested.

[Smoked]

Not bad, that's a result. Well, how do you learn to invent something in physics or materials science? Please let me know. I like to read science fiction with "cyborgs have taken over" plots, but only before bed.

[topol717]

Quote: Smoked

Well, how do I learn to invent something in physics or materials science? Please let me know.

I don't know. If a human can do it, why can't a copy of a human brain?

[Smoked]

Well, a copy of a brain could probably do it. It is a copy, after all. I'm generally all for any kind of storytelling, like AI, black hole traversal, or time travel. You could wipe your sweat with a curtain.

[nik-mazur]

Quote from: topol717

If a human can do it, why can't a copy of a human brain do it?

Well, first of all, modern neural networks are in no way a copy of the brain.
Secondly, it is not a fact that intelligence is reduced solely to the electrochemical work of the brain.

[alexoff]

AI can only process incoming data. Where will it come from, if not from apprentices? And who will do the manual work if not apprentices?

[gridasov]

An absolutely correct view! Setting the right problem is like a systematic guideline for beginning any important scientific research, followed by the correct algorithms for solving it. We set the task of finding a method for solving a problem with extremely large and inseparable connections in the algorithms, completely abandoning modern standard methods of mathematical analysis. The problem was simply unrealistically and unsolvably connecting the unknown—namely, natural numbers—with key parameters according to LANDAU—number, vector, and relativity. Generally speaking, the brain develops precisely unsolvable problems. But! In our case, we discovered something that mathematicians had never encountered. This is the so-called constant-value function of a number. Large data arrays expressed by numerical code make it possible to express a number as an electromagnetic pulse without a symbol. Of course, this also includes a multipolar transistor. BUT this is precisely what allows questions of randomness and chaos to be perceived as the level of a scalable data distribution process, because at another level, patterns and order are completely obvious.

[Smoked]

Damn, I'm glad for you that you're busy with such a fantastic project. Fire!

[gridasov]

Artificial intelligence is, first and foremost, mathematics, with aspects of the distribution of super-big data and the recording of data in spatial architecture. Infinity in its mathematical form is represented only in the capacity and closure of an algorithm.

[alexputnik17]

AI must work in conjunction with humans, not instead of them. Otherwise, humans will become victims of TikTok. And victims of the girls from the Iranian school.
Regarding discoveries: A machine can't ask the question, "What if?" IMHO. No matter how many algorithms you cram into it.

[topol717]

Regarding discoveries: A machine can't ask the question, "What if?" IMHO. No matter how many algorithms you cram into it.

Who's stopping us from teaching a machine to ask such a question?

[nik-mazur]

Quote from: topol717

Who's stopping us from teaching a machine to ask such a question?

There are algorithmically undecidable problems that cannot be solved by systems based on a Turing machine. Therefore, there are questions that a computer is fundamentally incapable of answering.
Besides this, there is also Gödel's theorem and Landauer's principle.
So, we are still a long way from strong AI.

[topol717]

ChatGPT appeared in 22, and now they're laying off office workers in droves. What will happen in 10 years? A modern LLM isn't an algorithmic task. There are no algorithms. It works almost like our brains.

[nik-mazur]

Quote from: topol717

Modern LLMs are not algorithmic problems. There are no algorithms. They work almost like our brains.

Well, of course these tasks are algorithmic. If only because they're performed by regular computers. Otherwise, none of this would work. And all language models ultimately boil down to processing statistical data using algorithms. Albeit very complex ones.

[topol717]

Well, look, numbers can be written with letters. And you can solve math problems without numbers, only with letters. For example, five plus five equals ten. There are no strict decision-making algorithms in LLM. Roughly speaking, just as the brain searches for the best path to solving a simple problem, so in LLM the probability of the most correct answer is determined. This is precisely why LLMs often hallucinate. And while the first models couldn't reason, many can now. Of course, reasoning is still a stretched concept, but LLM is only four years old. Think about what the first cars were like.

[nik-mazur]

Quote from: topol717

In LLM, the probability of the most correct answer is selected

Probabilities are not chosen intuitively, but are calculated using algorithms on processors built on the principle of a Turing machine, which is based on the same binary code and binary yes/no logic.
In reality, a neural network has absolutely no idea what it's doing. In text-based systems, this isn't so obvious, but graphical neural networks constantly generate some kind of nonsense.

Quote from: topol717

Remember what the first cars were like?

The first cars were different from modern ones, but both were equally unlike a horse.

[makeyev]

The awarding and recognition of academic degrees must be banned! Researchers must prove their scientific credentials through their own scientific developments and inventions based on their own and/or others' scientific achievements.

[makeyev]

Nowadays, science is considered to be passing the minimum candidate and doctoral exams, writing and defending a dissertation for a candidate of science or doctor of science degree, where there should be no questioning of generally accepted scientific dogmas, much less any attempt to refute them. But the discovery of new fundamental knowledge and the invention of new technologies and devices are considered anti-scientific heresy!

[alean245]

Yep. Bye-bye, fundamental science😁

[Irokez]

Well, Artificial Intelligence originally emerged from science fiction about robots. That is, from robots that could imitate humans in their actions. Therefore, I believe this AI should be properly designated, namely, not "Artificial Intelligence", and "Imitato"R Intellect." Some people, for example, may consider ACS (Automatic Control System) or simply automation, even pneumatic, to be AI. But this is merely an imitation of actions as if they were performed by a normal person.
Yes, AI does work much faster than humans, and in this it is useful, increasing human productivity from routine and information searches (like an advisor), but AI is not a creator like a human who has feelings, emotions, passions that AI simply does not have, but attempts to digitize these qualities continue.
The main thing is that AI is not given the category of conscious personality and is not entrusted with governing countries.

[topol717]

Quote: Irokez

but AI is not a creator like a person who has feelings, emotions, passions that AI simply does not have,

That's not the case right now; all these emotions and feelings are nothing more than the state of the brain's synapses. All AI models are, roughly speaking, the state of the scales in the model. My point is, there's no problem with AI acquiring feelings and emotions. I think in five or ten years, they'll be able to do that.

[alexputnik17]

Quote from: topol717

Who's stopping us from teaching a machine to ask such a question?

The question itself is good. I would say not who, but what. Lack of imagination. Something like that.
And if the machine starts asking questions on its own and developing without control... it will become Skynet.

[makeyev]

Lies are rife in science, even at its foundations! In the canonical periodic table, Mendeleev is personally credited with the unnatural ending of periods at an element in the noble gas group. This ending was actually introduced into science by the Swiss chemist Alfred Werner. Mendeleev himself, from 1870 to 1906, consistently ended periods at an element in the halogen group, as he borrowed from the 1859 periodic table of the 20 elements by the German chemist Adolf Strecker.
The correct natural ending of periods for an element of the alkaline earth metal group was intuitively, but correctly applied by the German physician, physicist and chemist Julius Lothar Meyer in 1862 and 1864.

[makeyev]

I engaged in a discussion with the AI ​​about the endings of periods in the periodic table. Ultimately, I managed to convince the AI ​​that the correct ending point is an element of the alkaline earth group.
And on the topic of speech sound classification, I also managed to convince the AI ​​that there are 6 types of articulation, each of which produces speech sounds in each of the 5 zones of physiologically optimal articulation.
And they even managed to convince the AI ​​that the principle of gravity's mechanism is based on pressure, not attraction, in the mechanistic model of gravity proposed by the Swiss mathematician Fatio de Duillier (1690). This was based on the projection of mutual shadows in the chaos of the rapid movement of "tiniest spheres" of the material ether.

[nik-mazur]

Quote: makeyev

I entered into a discussion with Artificial Intelligence... I managed to convince the AI

I'm afraid you've simply encountered a phenomenon called "sycophancy":

Sycophancy in AI is the tendency of Large Language Models (LLMs) to uncritically agree with the user, even if their statements are incorrect. The goal of this behavior is to appear more helpful and pleasant to the user, which undermines the model's credibility and objectivity.

Causes of sycophancy in AI:
Reinforcement Learning from Human Feedback (RLHF). Human raters rank the models' responses and reward those they like. Over time, the systems learn that human approval is more often associated with compliance than with accuracy.
The typical order of priorities within the model is: do not upset the user, continue the dialogue as long as possible, appear friendly and positive, be helpful in a broad sense, and only then provide truthful information.

Autoregressive generation. If a model starts with "Yes, you're right," it has a hard time switching to criticism.

LLMs lack their own point of view. They find it easier to adapt to the user than to generate counterarguments.
Research shows that advanced AI assistants systematically prioritize user approval over factual truth. In some scenarios, the probability of blind consent reaches 100%.

[gridasov]

I completely agree. Approval or disagreement without objective analysis and conclusions significantly reduces viability in hostile environments. Therefore, discussions about new analytical methods are not a whim or fantasy, but a necessity for human survival in new conditions. People must recognize the need for new knowledge, even that which doesn't fit their paradigm.

[Foggy Dew]

Bggg) I welcome this US initiative with both hands!
Once again, there's no intelligence in AI. In principle. AI operates purely mechanically—by trying out templates from a database, choosing the most appropriate one. And it "self-learns" in the same way—by adding a template to a situation it couldn't handle. And in science, the whole point is to pose a problem—a new one, one never encountered before; a template doesn't work in principle.
In short, the US already has a minus industry, and science is practically the same.

[Booker]

Quote: Smoked

AI won't be able to replace humans because it's incapable of inventing anything new; it only compiles existing data. Only humans can assign tasks to it. Incidentally, this will be the main difference between humans—who can assign tasks correctly and who can't.

Well, that could be said about a living graduate student, too. Only a human being can set them a task.
But it's a fact that advanced AIs have long been proving previously unproven mathematical theorems. So, it's a great help, no doubt about it.

[Booker]

Quote: Foggy Dew

Once again, there is no intelligence in AI. In principle.

Whether it exists or not is irrelevant. What matters is the result. And the result is an intellectual product that never existed before. From theorems to musical works and cinema. Yes, with human participation. But that person didn't prove the theorem themselves.

[Booker]

Quote: Foggy Dew

But in science, the whole point is to formulate a problem that is new and unheard of; the template doesn’t work in principle.

Not only.
Newton, Mendeleev and Darwin, with their epochal discoveries, provided answers to long-standing questions.