AI가 주도하는 과학적 발견: 인류의 미래에 대한 시사점

### 과학적 발견에서 AI의 미래와 인류에 미치는 영향
* 🤖 진정한 과학적 발견을 할 수 있는 인공지능의 출현은 **인공 초지능(ASI)을 달성하기 위한 전조**입니다.
* 📈 레오폴드 애쉬 브레너의 '지능 폭발' 개념은 **AI가 스스로의 발견을 통해 스스로를 개선**하기 시작하면 인간의 지능은 뒤처질 수 있다는 것을 시사합니다.
* 🔬 에이단 토너 로저스의 최근 연구 논문은 **1,800명이 넘는 과학자들에게 AI 도구가 어떻게 도입**되어 기존의 노동 집약적인 과학적 발견 과정에 혁신을 가져왔는지를 강조합니다.
* 💡 구글의 알파폴드와 같은 발전으로 AI는 이제 **단백질 접힘을 예측**할 수 있게 되어 이전에는 불가능했던 과학적 혁신의 잠재력을 크게 높일 수 있게 되었습니다.
* 📊 데이터에 따르면 **AI 보조 연구원이 44% 더 많은 자료**를 발견할 수 있어 **특허 출원이 39% 증가**하고 **제품 혁신이 177% 증가**했지만, 그 혜택이 모든 연구원에게 균등하게 분배되지는 않았습니다.
* 😔 생산성 향상에도 불구하고 과학자의 **82%는 창의력 감소와 기술 활용도 저하**를 이유로 직무 만족도가 떨어졌다고 답했습니다.
* 🔄 인공지능이 일상적인 업무를 대신하면서 인간 연구자들은 의사 결정과 창의성에 집중할 수 있지만, **인간의 만족도와 지능에 대한 이해의 미래에 대한 우려**도 커지고 있습니다.
* 🧠 코딩과 과학적 발견을 처리하는 AI의 잠재력은 인공 지능 시대가 다가옴에 따라 **인간의 지적 능력을 기계에 맡기는 것**이 어떤 의미를 갖는지에 대한 중요한 질문을 던집니다.

인공지능은 과학적 발견에 혁명을 일으키며 연구 생산성을 크게 향상시키는 동시에 인간의 성취에 대한 우려를 불러일으키고 있습니다.

What do you think is going to happen when artificial intelligence is making true scientific discoveries, actual new science that humans did not discover themselves? This is an absolute prerequisite for AI achieving ASI, artificial superintelligence. Now, if you remember, I made a video about Leopold Ashen Brenner's situational awareness paper, and in that paper, he specifically called out what is called the intelligence explosion. Would then unquestionably be an intelligence explosion, and the intelligence of man would be left far behind. Thus, the first ultra-intelligent machine is the last invention that man need ever make. That is the premise of the movie The Matrix. Very scary stuff when AI is able to make its own scientific discoveries and then reapply it to itself, which is basically self-improvement. And then, at that point, you deploy hundreds, thousands, millions of these AI agents that are recursively self-improving, and that is the intelligence explosion. And what does that actually look like when it is applied to the scientific field? What does society look like when science is

solved. That's what we're going to be looking at today. This is the research paper Artificial Intelligence, Scientific Discovery, and Product Innovation by Aidan Toner Rogers out of MIT, and it just came out just about a week ago. And what Aidan did was give artificial intelligence researchers to over a thousand scientists in one of the largest US scientific research and discovery labs. And if you remember a few months ago out of Sakana AI, Sakana Labs out of Japan, they released something called the AI scientist, which was kind of a glimpse into what this might actually look like. That is AI that is able to propose theorems, go and run tests, iterate, fail, iterate, discover new science. And right here it says, To provide evidence on these questions, I exploit the randomized introduction of an AI tool for materials discovery to 1,800 scientists in the R&D lab of a large US firm. So what does the traditional scientific discovery process look like? Well, it's a lot of manual work, of course, and that's really where AI can come in.

handy. Traditionally, scientists discover materials through an expensive and time-consuming system of trial and error, conceptualizing many potential structures and testing their properties. Now, let's talk about that for a second. Google changed the world, and recently, the researchers from Google won Nobel prizes for AlphaFold. That is basically being able to predict protein folding so that we can predict proteins. Prior to that, it was basically trial and error, or you have to throw a lot of compute at it, and we were really only discovering a fraction of the amount of proteins out there. But with AlphaFold, and then AlphaFold 2, and then more recently, AlphaFold 3 just came out, we can essentially predict the entire universe of potential proteins, and that unlocks a lot of potential scientific discovery. Back to the paper, the AI technology leverages developments in deep learning to partially automate this process. Trained on the composition and characteristics of existing materials, the model generates recipes for novel compounds predicted to possess specific properties. Imagine we now have access to materials that are so exotic, and maybe we would

have never discovered, but they have properties that make them lighter or cheaper to produce, or stronger, or any property that is required for the problem at hand. We now have access to any of that, and that means that problems which were previously unsolvable in the world because of our just limits on what materials we could produce are now completely solvable, thank you to the sponsor of this video, weights and biases, weave. I'm a huge fan of weights and biases and everything that they've done, so I'm very excited to talk to you about weave. Weave is a lightweight toolkit for tracking and evaluating LLM applications, built by the incredible weights and biases team. You can use weave to log and debug language model inputs, outputs, and metadata, build rigorous Apples to Apples evaluations for language model use cases, and organize all the information generated across the LLM workflow, from experimentation to evaluations to production. Weave is easy to implement with only two lines of code, gives visibility, and helps with debugging, makes evaluations easier and faster. You

get visibility into costs and results, and you can version your prompts and models. Plus, it seamlessly integrates with all of your favorite APIs and libraries, including OpenAI, Anthropic, MSTR, Coohere, LangChain, LlamaIndex, and more. Visit Weights and Biases, check out their Weave product. I'll drop a link to it in the description below wb.me/MB. I'm very proud to have them as a sponsor. Thank you again, and now back to the video. And here is now what the new process looks like: idea generation, so an idea for some kind of novel material or some type of scientific research, and that is now being done by AI. The idea generation is no longer done by scientists. Then they have candidate materials, prioritization of those candidate materials, testing. When you get false positives, you go in this iterative loop trying to find the actual winning materials, and then they have a viable material. They file a patent, they come up with a new product prototype, they have an improved product prototype, and then so on. And over here in this gray, this is

commercialization. But this entire process, which traditionally is just done by humans in a really manual and inefficient way, can now be done by AI. And what happens when that happens? The number of research breakthroughs skyrockets, the number of patents skyrockets, the productivity of the top researchers skyrockets. I'm going to touch on that little nuance in a moment. So many things are possible when AI is able to actually help in the scientific research process. Now, some of the actual numbers, listen to this: AI assistant researchers discover 44% more materials, resulting in a 39% increase in patent filings and a 177% rise in downstream product innovation. These compounds possess more novel chemical structures and lead to more radical inventions. But it's not the same for all researchers. The top scientists in the field benefited the most. They had drastic improvements in their discovery process, while the bottom third of researchers really saw no change at all. And so, as the paper describes, the technology has strikingly disparate effects across the

productivity distribution. The bottom third of scientists see little benefit. The output of top researchers nearly doubles. AI is able to automate 57% of idea generation tasks, reallocating researchers to the new task of evaluating model-produced candidate materials. So, rather than the top researchers spending their time coming up with potential candidates to test, they just take the top candidates that the AI model suggests, and then they're spending their time doing the actual testing. And I think that's a really good analogy for the companionship between AI and humans. For the foreseeable future, AI is going to take over more and more of the mundane, repetitive kind of grunt work. Then, humans are going to take on the more hands-on decision-making and morals and judgment and those types of tasks. But it wasn't all good, and here's the problem. Scientists actually reported a significant drop in fulfillment in their job. Listen to this: 82% of scientists report reduced satisfaction with their work due to decreased creativity and skill underutilization. So, even though they

were able to have such greater outputs. They didn't feel as satisfied with the work they were doing. Now, this is quite telling for the future of humanity. If all of our creative tasks are being outsourced to AI, are we going to lose satisfaction? Maybe, but I also think it's going to free us up for doing the creative tasks that we actually want to do. Now, let's take a second to talk about the different stages of artificial intelligence because based on this paper and based on some other papers that I've been reading lately, I think we're at the precipice of the next stage of AI. So first, we had level one, which is basic chatbots, AI with conversational language. That is the GPT series of models. Then we have level two, reasoners, human-level problem solvers, and that is when you get to the later stage GPT models, and they are really good at reasoning, and they essentially pass the Turing test easily. Then we have level three agents and systems that can take actions, and we're just starting to

scratch the surface there, but here is where we're also scratching the surface in what seems like in parallel level four innovators. AI that can aid invention. Now again, I'm going to reference the situation awareness paper by Leopold Ashenbrener. We have it right here, the intelligence explosion. And what does it say? Automated AI research. When AI is able to research and discover new science and then apply it to itself, recursive self-improvement. That is when we're going to have superintelligence. This explosion in intelligence because recursion is all you need. And again, the Sakana AI paper, AI scientist towards fully automated open-ended scientific discovery. This is a model that is able to come up with ideas, test the ideas, come up with conclusions, and actually publish its research. So we have the LLM idea, plan, innovation, novelty check, idea scoring, experimentation, coding. Everything that a traditional AI scientist would do now can be done, or at least there's the temp of what can be done in this paper AI Scientist by Sakana AI. But then that

brings us to our last thought: if AI is doing our coding for us, if AI is discovering our science for us, does that mean we're going to get further and further away from that raw intelligence? And if we're getting further and further away, are we able to understand it as well? And at a certain point, are we just offloading all of our intellect to AI? I don't know. These are all questions that we need to think through deeply as we approach artificial superintelligence. If you enjoyed this video, please consider giving a like and subscribing, and I'll see you in the next one.

MIT's AI Discovers New Science - "Intelligence Explosion"

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

AI가 과학적 발견에 미치는 영향

AI가 진정한 과학적 발견을 할 수 있는 능력은 인공 초지능(ASI)으로 진화하는 데 매우 중요합니다.

AI 에이전트가 재귀적으로 자기 개선하는 지능 폭발이라는 개념은 인간의 지능이 뒤처지는 것에 대한 우려를 불러일으킵니다.

최근 MIT의 한 연구 논문에서는 1,800명의 과학자가 참여한 연구와 물질 발견을 위한 AI 도구의 무작위 도입을 통해 과학적 발견에서 AI의 잠재력을 강조하고 있습니다.

전통적으로 과학자들은 시행착오에 의존해 물질을 발견했지만, 구글의 알파폴드는 단백질 폴딩을 예측하고 과학적 발견의 가능성을 열어줌으로써 판도를 바꾸어 놓았습니다.

이 논문에서 논의된 AI 기술은 딥러닝을 활용하여 물질 발견 과정을 자동화하고 특정 특성을 가진 새로운 화합물에 대한 레시피를 생성합니다.

이러한 발전은 이전에는 해결할 수 없었던 문제를 해결할 수 있는 이색적인 물질에 대한 접근성을 제공하며, 가중치 및 편향성 툴킷인 Weave는 언어 모델 애플리케이션을 추적하고 평가하는 간소화된 접근 방식을 제공합니다.

이제 AI가 과학 연구의 아이디어 창출을 대신하여 후보 물질의 우선순위 선정, 테스트, 상용화 과정을 보다 효율적으로 이끌고 있습니다.

이러한 변화로 인해 획기적인 연구 성과, 특허 출원, 제품 혁신이 크게 증가했습니다.

그러나 AI의 혜택이 고르게 분배되지는 않았으며, 최고의 과학자들이 발견 과정에서 가장 큰 개선을 경험하고 있습니다.

AI는 과학자들의 아이디어 생성 작업 중 57%를 자동화하여 최고의 연구자들이 모델 제작 후보 물질을 테스트하는 데 집중할 수 있게 해줍니다.

최고 연구원의 생산성은 거의 두 배로 증가했지만, 과학자의 82%는 창의력 감소와 기술 활용도 저하로 인해 업무 만족도가 떨어졌다고 답했습니다.

AI의 진화는 기본적인 챗봇에서 인간 수준의 문제 해결사로 발전하고 있으며, 다음 단계는 행동을 취할 수 있는 에이전트와 시스템으로 발전하고 있습니다.

사카나 AI와 같이 완전 자동화된 과학적 발견을 목표로 하는 모델을 통해 AI가 발명가의 역할을 점점 더 많이 맡고 있습니다.

AI 연구에서의 재귀적 자기 개선 개념은 지능과 초지능의 폭발적인 증가로 이어질 수 있습니다.

이는 원시 지능에 대한 우리의 이해가 줄어들고 우리의 지성을 AI에 넘겨줄 수 있는 가능성에 대한 의문을 제기합니다.