Nobelio fizikos premija skirta dirbtinio intelekto pionieriams

 

 „Nobelio fizikos premija buvo skirta John Hopfield iš Prinstono universiteto ir Geoffrey Hinton iš Toronto universiteto už atradimus ir išradimus, kurie sudarė mašininio mokymosi blokus, pagrindžiančius daugelį galingiausių šiandienos dirbtinio intelekto (AI) modelių.

 

 Mokslininkai padėjo apmokyti dirbtinius neuroninius tinklus, kurie gali atpažinti dėsningumus dideliuose duomenų rinkiniuose, naudodami fizikos idėjas, todėl tapo įmanoma atlikti mašininio mokymosi užduotis, tokias, kaip veido atpažinimas ir kalbos vertimas.

 

 Hintonas, britų ir kanadiečių kompiuterių mokslininkas, žinomas, kaip vienas iš „AI krikštatėvių“, daugiau, nei dešimtmetį, dirbo „Google“, tačiau praėjusiais metais pasitraukė, norėdamas laisviau kalbėti apie AI plėtros riziką. Jis ir kiti išreiškė susirūpinimą dėl būdų, kaip dirbtinis intelektas gali pakenkti žmonijai.

 

 Antradienio ryto pokalbyje su Nobelio komitetu Hintonas sakė, kad „nerimauja, kad bendras to rezultatas gali būti už mus protingesnės sistemos, kurios galiausiai perims valdymą“, tačiau jis „dar kartą darytų tą patį“, kai kalbama apie jo polinkį ką nors dirbti.

 

 Hopfieldas pakartojo šio laureato susirūpinimą antradienio popietę vaizdo pokalbyje su Prinstono auditorija, pilna dėstytojų ir kitų svečių.

 

 „Man rūpi ne AI, o AI kartu su informacijos srautu visame pasaulyje“, – sakė jis ir pridūrė, kad paprastas algoritmas neuroniniame tinkle gali valdyti labai didelę informacijos sistemą, tačiau jį „neramina“ mintis, kad tokių tinklų darbas gali būti nelabai suprantamas.

 

 Nuo devintojo dešimtmečio 91 metų Hopfieldas ir 76 metų Hintonas atliko svarbų darbą, kuriame naudojamos pagrindinės fizikos sąvokos, kurdami dirbtinius neuroninius tinklus.

 

 „Jie parodė visiškai naują būdą, kaip naudoti kompiuterius, kad padėtų ir padėtų mums įveikti daugelį iššūkių, su kuriais susiduria mūsų visuomenė“, – X antradienį paskelbė Nobelio komitetas, Stokholme paskelbdamas premiją.

 

 Dirbtiniai neuroniniai tinklai, kaip rodo pavadinimas, yra programos, padedančios mašinoms mokytis – įkvėptos žmogaus smegenų ir jo neuronų tinklo gebėjimų.

 

 "Mes galime atpažinti vaizdus ir kalbą ir susieti juos su prisiminimais ir praeities patirtimi. Milijonai neuronų, sujungtų kartu, suteikia mums unikalių pažinimo gebėjimų", - sakė Ellen Moons, Nobelio fizikos komiteto pirmininkė.

 

 Mašininio mokymosi tikslas – imituoti tuos gebėjimus, tiekiant kompiuteriams neįtikėtiną duomenų kiekį, kad jie galėtų atlikti užduotis, pvz., siūlyti nuspėjamąjį tekstą arba pasirinkti kitą „Netflix“ laidą.

 

 Hopfieldas pradėjo šį darbą 1982 m., kai sukūrė „Hopfieldo tinklą“, neuroninį tinklą, galintį išsaugoti ir prisiminti dėsningumus su tik daline informacija, įkvėptas to, ką jis pavadino „asociatyvine atmintimi“.

 

 Dirbtinis intelektas siūlo „gebėjimą susieti dalykus, susieti patirtį“, – sakė Hopfieldas 2020 m. podcast'e, nurodydamas, kaip paminėti tik kelis faktus apie žmogų, pavyzdžiui, kaip jis atrodo, kaip skamba jo balsas, kur jis įstojo į koledžą ir kur jį sutikote – suteikia galimybę ką nors prisiminti, kas jis yra, net be vardo ir pavardės.

 

 „Daugelį AI pažangos, kurią matome šiandien, tikrai galima tiesiogiai arba netiesiogiai priskirti šioms idėjoms, – sakė Dmitrijus Krotovas, IBM tyrimų darbuotojas Kembridže, Masažas. „Jo įtaka buvo nepaprastai didžiulė“.

 

 Hintonas išplėtė Hopfieldo darbą, kad sukurtų tinklą, kuris galėtų atpažinti žinomus dėsningumus dar nematytoje informacijoje. Jis sukūrė techniką, kuri padeda optimizuoti neuroninį tinklą, pakartotinai taisant klaidas, kol jos išnyksta.

 

 Mokslininkų požiūris į neuroninių tinklų mokymą padėjo paruošti kelią tokioms sistemoms, kaip „ChatGPT“.

 

 "Būtų sunku įsivaizduoti, kad GPT ir viskas kitas būtų čia be jų", - sakė Jerome'as Delhommelle'as, Amerikos fizikos draugijos aktualios duomenų mokslo grupės pirmininkas ir chemikas iš Masačusetso universiteto Lowell.

 

 Paklaustas apie jo mėgstamą AI įrankį, Hintonas atsakė, kad naudoja GPT-4, naujausią OpenAI modelį, kai tik nori sužinoti atsakymą į bet ką. "Aš visiškai juo nepasitikiu, nes jis gali sukelti haliucinacijas," - sakė jis, - "bet beveik visuose dalykuose tai ekspertas, nors ir nėra labai geras, ir tai labai naudinga." [1]

 

1. U.S. News: Nobel Prize in Physics Goes to AI Pioneers. Woodward, Aylin.  Wall Street Journal, Eastern edition; New York, N.Y.. 09 Oct 2024: A.3. 

Nobel Prize in Physics Goes to AI Pioneers

 

"The Nobel Prize in physics has been awarded to John Hopfield of Princeton University and Geoffrey Hinton of the University of Toronto for discoveries and inventions that formed the building blocks of machine learning underpinning many of today's most powerful artificial intelligence models.

The scientists helped train artificial neural networks that can recognize patterns in large data sets using tools from physics, making machine learning tasks like facial recognition and language translation possible.

Hinton, a British-Canadian computer scientist known as one of the "godfathers of AI," worked for Google for more than a decade but quit last year to speak more freely about the risks of AI development. He, and others, have expressed alarm over the ways AI could harm humanity.

In a Tuesday morning call with the Nobel Committee, Hinton said he is "worried that the overall consequence of this might be systems more intelligent than us that eventually take control," but that he would "do the same again" when it comes to his work.

Hopfield echoed his co-laureate's concerns in a video call Tuesday afternoon to a Princeton auditorium full of faculty and other guests.

"The worry I have is not AI quite directly, but AI combined with information flow around the globe," he said, adding that a simple algorithm in a neural network can control a very big system of information, but he is "unnerved" by the idea that such networks' work may not be well understood.

Since the 1980s, Hopfield, 91 years old, and Hinton, 76, have conducted important work that uses fundamental concepts from physics to design artificial neural networks.

"They have showed a completely new way for us to use computers to aid and to guide us to tackle many of the challenges our society face," the Nobel Committee said on X Tuesday as it announced the prize in Stockholm.

Artificial neural networks, as the name suggests, are programs that help machines learn -- inspired by the abilities of the human brain and its network of neurons.

"We can recognize images and speech and associate them with memories and past experiences. Millions of neurons wired together give us unique cognitive abilities," said Ellen Moons, chair of the Nobel committee for physics.

Machine learning aims to mimic those abilities by feeding computers incredible amounts of data so they can master tasks, like offering predictive text or choosing your next Netflix show.

Hopfield pioneered this work in 1982 when he created the "Hopfield network," a neural network that could save and recall patterns with only partial information, inspired by what he called "associative memory."

AI offers the "ability to link things together, link experiences together," Hopfield said on a 2020 podcast, pointing to how mentioning just a few facts about a person -- such as what they look like, what their voice sounds like, where they went to college and where you met them -- enables someone to recall who they are without their name ever being mentioned.

"A lot of progress in AI that we're witnessing today can certainly be directly or indirectly attributed to these ideas," said Dmitry Krotov, a research staff member at IBM Research in Cambridge, Mass. "His influence has been absolutely monumental."

Hinton expanded Hopfield's work to create a network that could recognize familiar patterns in never-before-seen information. He developed a technique that helps optimize a neural network by iteratively correcting errors until they disappear.

The scientists' approaches to training neural networks helped pave the way for systems like ChatGPT.

"It would be hard to imagine that GPT and everything would be here without them," said Jerome Delhommelle, chair of the American Physical Society's topical group on data science and a chemist at the University of Massachusetts, Lowell.

When asked about his favorite AI tool, Hinton answered that he uses GPT-4, the most recent model from OpenAI, whenever he wants to know the answer to anything. "I don't totally trust it, because it can hallucinate," he said, "but on almost everything, it's a not-very-good expert, and that's very useful."" [1]

1. U.S. News: Nobel Prize in Physics Goes to AI Pioneers. Woodward, Aylin.  Wall Street Journal, Eastern edition; New York, N.Y.. 09 Oct 2024: A.3. 

 

Why is the cream of Lithuania getting expensive dental implants with the army's money and trying to escape to Greece?

Because everyone understands that it is unrealistic for Lithuania to defend itself and the billions of euros allocated for defence are wasted. We don't have the long-range air defenses to defend against the ballistic missiles that Iran recently treated Israel with without much of a showing. Not having anti-ballistic missile defense is like having no tanks, only cavalry, during Hitler's time. Therefore, from a military point of view, we are zero. And where do the billions of euros from taxes collected for military purposes in Lithuania go every year? Gun producers, cavalry and horses need to eat. We put dental implants and here we go. We have the military staff members, like real ones, with colonels and generals between them. "A military picture: ...And the beauties are far away, they gently wave their hands, their hearts beat hotly, they will love forever..." Just like we didn't have a real army, we still don't have one. Those who have access to our money have no illusions either:

 

 "Presidential adviser Kęstutis Budrys urged on Tuesday not to have illusions about air defense, because long-range systems are lacking all over the world and they probably cannot be deployed in Lithuania permanently, while medium-range weapons are currently sufficient to protect only soldiers and critical infrastructure from some small missiles.

 

 

 

 "They have a big deficit, like for other air defense systems, and here is another reason to say that we should not create illusions about air defense as such." That's why so much effort is being made here, that's why there's so much talk about it," the president's chief adviser on national security told Radio News on Tuesday.

 

 

 

 A few years ago, NATO agreed on a rotational air defense model in the Baltic Sea region, which should mean that the Western allies will send various air defense weapons to Lithuania and other countries for rotations of various duration.

 

 

 

 However, only the Netherlands has done this so far, having sent Patriot long-range systems to Lithuania for a few weeks in the summer of this year.

 

 

 

 The Minister of National Defense, Laurynas Kasčiūnas, said last week that more information should be revealed in the near future regarding which countries will send the capabilities next year.

 

 

 

 K. Budrys did not undertake to assess whether the allies will ever deploy long-range systems in Lithuania permanently.

 

 

 

 "Whether we will ever have a continuous presence of allied systems in the entire region, I cannot plan in such a distant perspective, because they simply do not exist, they will be lacking." Here is elementary mathematics, and where there is industry, physics, how much we can produce," said K. Budrys.

 

 

 

 "If we're talking about the Patriot, they're not the only ones, we can count them, how many there are in the world, where they are right now, and who can rotate here," he added.

 

 

 

 However, according to the presidential adviser, the benefits of short-term deployment are also enormous, including for Lithuanian soldiers who learn to work in interaction with such systems.

 

 

 

 "It's not that simple, I came here, built it and it's already working." You need to activate the systems, you need to know how they function, communicate with each other, what the situations can be, and each time with a new arrival, the time during which it can already start working, it shortens. The benefit here is unquestionable," said K. Budrys.

 

 The new purchase - NASRAMS systems - will only be sufficient to protect critical infrastructure facilities

 

 

 

 The Ministry of National Defense announced last week that it had signed contracts for the acquisition of air defense systems, including medium-range systems NASRAMS.

 

 

 

 Together with the previously acquired NASRAMS, Lithuania will have three batteries of such systems in 2028.

 

 

 

 When asked how many such systems would be needed to ensure the defense of big cities and the Suwalki corridor, K. Budrys said: "It won't be enough, that can be said right away."

 

 

 

 "These medium-range air defense systems in a complex with other means are primarily intended to protect the armed forces in Lithuania (...) and other critical infrastructure objects, especially related to the support of the host country from some missiles," said the presidential adviser.

 

 

 

 According to him, in order to ensure comprehensive air defense, Lithuania should have its own long-range air defense systems capable of shooting down ballistic missiles.

 

 

 

 The first NASRAMS battery is already in Lithuania, the second one will be delivered in the first quarter of 2026 and will finally arrive by 2027.

 

 

 

 According to the contract signed last week, the third battery will be shipped to Lithuania in the second quarter of 2028.

 

 

 

 The country also acquired additional short-range air defense weapons last week."