OpenAI의 프로젝트 스트로베리: AI 추론 및 계획의 발전

### OpenAI의 프로젝트 스트로베리: AI 추론 및 계획의 발전
* 🍓 OpenAI는 논리 추론 및 계획을 위한 qar 이니셔티브를 기반으로 **Strawberry라는 프로젝트 코드명**으로 새로운 추론 기술을 개발하고 있습니다.
* 🤖 스트로베리 모델은 **AI가 미리 계획하고, 자율적으로 인터넷을 탐색하고, 심층 연구를 수행**하여 대규모 언어 모델의 현재 한계를 해결할 수 있도록 지원하는 것을 목표로 합니다.
* 💭 이 프로젝트는 AI의 추론 능력을 향상시키는 것을 목표로 하고 있으며, OpenAI 대변인은 다음과 같이 말했습니다: **"우리는 AI 모델이 우리처럼 세상을 더 많이 보고 이해하기기를 원합니다."**
* 🧠 AI 연구자들은 과학적 발견에서 소프트웨어 애플리케이션에 이르기까지 다양한 분야에서 AI 모델의 잠재력을 발휘하려면 **추론을 강화하는 것**이 중요하다고 생각합니다.
* 📊 OpenAI는 인공 일반 지능(AGI)을 향한 AI 모델의 진행 상황을 추적하기 위해 내부 척도를 개발했으며, 현재 **채팅 GPT는 5단계 중 1단계**에 불과합니다.
* 🔍 다음 단계에는 **박사 수준의 문제를 해결**하고, 사용자를 대신하여 조치를 취하고, 새로운 혁신을 창출하고, 전체 조직의 업무를 수행할 수 있는 시스템이 포함됩니다.
* ⏳ 스트로베리는 장기적인 작업을 수행하도록 설계되었기 때문에 **모델이 미리 계획하고 장기간에 걸쳐 일련의 작업**을 수행해야 하며, 결과를 도출하는 데 **며칠이 걸릴 수도 있습니다.**
* 🌐 이 프로젝트는 **AI 역량을 크게 발전**시켜 잠재적으로 **AI 연구 및 응용 분야에서 획기적인 발전**을 가져올 수 있습니다.

OpenAI는 AI 모델의 계획 및 추론 능력을 향상시키기 위해 코드명 Strawberry로 새로운 추론 기술을 개발하고 있습니다.

for those of you who don't remember, qar is the secretive project inside of OpenAI that is essentially a new way for artificial intelligence to do logic reasoning planning. All we've had are little drips and drabs of leaks out of OpenAI, and now we have more information in the form of a project code-named Strawberry. And that's what we're going to be covering today. And not only that, OpenAI has laid out the foundation for how they're going to determine when we actually hit AGI and the different levels before that. So today we're going to be covering this article from Reuters by Anna Tong and Katie Paul. OpenAI working on a new reasoning technology under code name Strawberry. Now, Strawberry is the new code name for qar chat GPT maker. OpenAI is working on a novel approach to its artificial intelligence models in a project code-named Strawberry, according to a person familiar with the matter and internal documentation reviewed by Reuters. Reuters could not ascertain the precise date of the document, but it details how

open AI intends to use strawberry to perform research and again just another extension of this whole qar Saga and again qar is all about logic reasoning and most of all planning that is something that today's large language models just do not do well. They are not able to actually plan ahead without the help of agentic Frameworks around them. So here it says strawberry models aim to enable the company's AI not just to generate answers to queries but to plan ahead enough to navigate the internet internet autonomously and reliably perform what open AI calls deep research. Now one thing I've been thinking a lot about lately is something that Sam Altman said a couple months ago. I think about deep research in two ways. There's post-training and pre-training. Sam Altman mentioned that the total set of data available to train models is not really growing that quickly and so you really have two options at that point. You can have other models create synthetic data to then train new models on but I'm not sure how well that approach is going to

work because it's all going to be derivative data. And then the other option is to allow models to do a lot more with the existing data. And what does that actually mean? Well, when a model is trained, it trains on a set of data. But models that are pre-trained in this way basically look at the data once. And that is not how the human brain works. If you think about it, if you're trying to learn something sufficiently complex, you read about it, you take your time to think about it, you reference other information sources while you're building up your own knowledge about it. And it takes weeks and months and sometimes even years to learn about a subject in depth. And that is just not something that large language models do today. But what if they could? What if they could actually take that same set of data and think about it over and over again in different ways to come up with potentially the best set of knowledge? And then post-training, it's kind of similar. When you prompt a model, it gives you the answer to your prompt.

doesn't really think about it. It just gives you the first answer that it can come up with. But what happens if it has time to think, and it's actually able to iterate? We achieve that through a number of different prompting techniques today, like Chain of Thought and even mixture of agents, in a sense. But large language models by themselves can't do that, and that is what qstar, and now codenamed strawberry, is aiming to do, apparently. So here OpenAI actually commented on this story. So OpenAI company spokesperson said in a statement, We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a practice in the industry with a shared belief that these systems will improve in reasoning over time. Now I've actually heard a few things about what qar could actually be, and that includes everything from planning to deep research, but also being really good specifically at math. And so let's keep reading. A different source briefed on the matter said OpenAI has tested AI internally that scored over a

90% in a math data set, a benchmark of Championship math problems, but they couldn't determine if this was the strawberry project. So still we're kind of only getting little leaks and little pieces of information, but it is interesting nonetheless. On Tuesday at an internal alliance meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. And this article is also tied to the release of information about how OpenAI plans to basically gauge where they are in the AGI timeline. And I'm going to go over that in a moment. Now here's a really important paragraph, and because it is very true for how models work today. Large language models can already summarize dense text and compose elegant prose far more quickly than any human. The technology often falls short on Common Sense problems, whose Solutions seem intuitive to people like recognizing logical fallacies and playing Tic-Tac-Toe. Now we've seen these models get a lot better at these logic and reasoning problems, you know if you've

watched any of my videos. I run them through a series of logic and reasoning tests. And over the last 6 months, and especially just the last four months, we've seen tremendous progression in the large language model's ability to solve some of these problems. But they still do fall pretty short of getting it right consistently. AI researchers interviewed by Reuters generally agree that reasoning in the context of AI involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions, and work through challenging multi-step problems reliably. Now a few things here. AI that can plan ahead, that is not how LLMs work today. You are required to either prompt them to do this, but even that goes so far. But if you're able to give them memory and wrap it in an agentic framework, then you're starting to achieve the ability to plan, especially when you combine multiple models together. Next, reflect how the physical world functions. Now, Yan LeCun has talked a lot about large language models being World models, World simulators, and that the Transformers

architecture alone is not enough to achieve that. He believes that other people think that the scaling laws and just using Transformers models. The same thing we have today is enough to be able to have world simulations. And he gives a number of examples for how large language models are not able to really simulate the physical world today because he basically says language alone is not enough to model the world. An example that he gives is say you're starting at the North Pole and someone tells you to walk 200 m in One Direction then turn left walk 200 M and then ask you will you ever cross over your first point. Now when you're thinking about that in your head you're not actually using any language whatsoever you're using spal reasoning and World modeling in your head. And this is super easy for humans to figure out but it's actually quite difficult for large language models to figure out and that's his point. And Nexus says they want this new strawberry model to be able to work through challenging multi-step problems.

reliably something again multi-step planning multi-step problems llms generally struggle with. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. And if we do have this greatly enhanced planning and reasoning from AI, we might actually be at that point of AGI. And if you watched my video about Leopold Ashen Brenner's situational awareness paper, you know that AGI is just one step before super intelligence, and this is probably one of the ingredients to hitting AGI. And here's a quote from Sam Alman said earlier this year: The most important areas of progress will be around reasoning ability. And right here, they actually mention Yan Laon, for instance, one of the pioneers of modern AI. Yan Laon, who works at Meta, has frequently said that LLMs are not capable of human-like reasoning, which is what I just mentioned. Now, OpenAI has teased a number of new technologies. Sora

gp4 O's voice and they keep delaying it or having no firm date of when they're going to release it. In fact, GPT 40, which was really just a much faster version of gp4, is the only recently released innovation that we've had from OpenAI in quite a while. So we're kind of due for a big release from OpenAI, which is exciting, but we haven't had one yet. In the article, they say, In recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company's pitches. So a little bit more on Strawberry now. It includes a specialized way of what is known as post-training OpenAI generative AI models are adapting the base model to hone their performance in specific ways after they have already been trained on reams of generalized data, one of the sources say. Now, that's interesting because if it required training a new model from scratch, that would be an

incredibly big monetary and time investment by OpenAI. So if this technology is able to take an existing model and do post-training on it, that's going to decrease the time to release and decrease the cost of training it tremendously. Strawberry has similarities to a method developed at Stanford in 2022 called self-taught Reasoner, or STAR, and we've covered that in a previous video. STAR enables AI models to bootstrap themselves into higher intelligence levels via iteratively creating their own training data, and in theory, could be used to get language models to transcend human-level intelligence. More on Strawberry: it is aiming to perform long-horizon tasks. The document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time. And this is something Sam Altman has talked about, and that's both pre- and post-training. Basically, rather than the model simply taking your prompt and spitting out the first thing it thinks of as soon as it can, it will actually take seconds, minutes, hours, even days potentially to

Think through it, come up with the best possible answer, and then give it to you. And that might actually be a setting that you can give to the model, saying, Hey, take as much time as you need, or I need this within 10 minutes. And the longer you give it, the better the result will be. But currently, there is not any really known technology that can do that. And that would be an enormous unlock if it could be done like that. And imagine that in a multi-agent system, you might have a few agents working together in real-time, and then they task a long-thinking agent to go off and take its time while it's doing other stuff. And then that long-thinking agent comes back, reports what it's found to the other agents, and then they put it all together and deliver you whatever that deliverable was. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of Kua, a computer-using agent that can take actions based on its findings, according to the document.

one of the sources so that's it great reporting by anat Tong and Katie Paul thank you for putting this great article together now I want to show you one last thing. This is an article by The Verge and this is about how open AI will actually determine how powerful its AI models are and you're going to be a little bit surprised to hear its current iteration of chat GPT is only level one of five. So they really think that the current version of chat GPT is really dumb and that's actually exactly what Sam Alman has said in the past. He has said the current version of chat GPT is really bad and in fact it will be the worst it ever is right now. So as impressive as it has been over the last year to see chat GPT and its iterations in performance, we are still just at the very beginning. Open AI has created an internal scale to track the progress its large language models are making towards artificial general intelligence. So chat GPT today level one and open AI claims

it is nearing level two, though, and level two is defined as a system that can solve basic problems at the level of a person of a PhD. Level three refers to AI agents capable of taking actions on a user's behalf. In my opinion, this level three is going to change humanity when you can have agents that are working on your behalf and actually accomplishing useful tasks 24 hours a day, not simply answering world knowledge questions. That is a tremendous productivity increase for humanity. And when Apple made their Apple intelligence announcements, that's kind of what they were just scratching the surface of, where you can get Siri to do things on your phone, which is super interesting. And I have a lot of thoughts about what the architecture of this needs to be, but it is certainly, in my opinion, needing to be local. And that is because for an AI agent to actually accomplish things on my behalf, I want it to have access to all of my private information, and thus I don't want it leaving my phone. I definitely don't want

to send it to OpenAI so all of this needs to be done locally on the phone or on any Edge device. Now, here's the thing, and I have a video that I'm creating all about this. I actually strongly believe in a future where we're going to have very narrowly defined, very vertical models that are small, hyper-efficient, low-cost, that can run on edge devices, but more on that in a coming video. Next level four involves AI that can create new innovations, so basically scientific discoveries, math discoveries, and so much more. And if you remember the chart of Leopold Ashenbrenner's situational awareness paper, that would be when we reach AGI because at that point, when we reach AGI and AGI can go do research about itself, self-improve, that is when we're going to have the superintelligence explosion. Then level five, the final step to achieving AGI, is AI that can perform the work of entire organizations of people, and this ranking feels very similar to how they ranked autonomous driving vehicles, so level

zero level, one level, two, and then so on. Where all the way at level five, you basically don't even need to pay attention. The car's just driving you. So that's it. What do you think of QAR? Do you think strawberry is the next iteration of QAR? Do you think it's coming soon? Is that GPT-5? Let me know in the comments. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

Q-Star LEAKED: Internal Sources Reveal OpenAI Project "Strawberry" (GPT-5?)

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

OpenAI의 프로젝트 스트로베리: AI 추론에 대한 새로운 접근 방식

OpenAI는 자율적으로 심층 연구를 계획하고 수행하는 AI의 능력을 향상시키기 위해 프로젝트 스트로베리라는 새로운 추론 기술을 개발하고 있습니다.

이 프로젝트는 논리 추론과 계획에 초점을 맞춘 qar 이니셔티브의 일환으로, 현재의 대규모 언어 모델 기능의 공백을 메우는 데 중점을 두고 있습니다.

목표는 인간의 뇌가 시간이 지남에 따라 학습하는 방식과 유사하게 AI 모델이 학습 전 단계와 학습 후 단계에서 기존 데이터를 철저하게 분석할 수 있도록 하는 것입니다.

OpenAI의 프로젝트 스트로베리와 AI 추론의 미래

OpenAI의 프로젝트 'Strawberry'는 AI 모델에 인간처럼 추론하고 수학이나 계획과 같은 업무에서 탁월한 능력을 부여하는 것을 목표로 합니다.

이 회사는 인간과 같은 추론 능력을 갖추고 세상을 더 잘 이해할 수 있도록 AI 기능을 개선하기 위해 노력하고 있습니다.

AI 연구자들은 AI의 추론 능력을 발전시키기 위해서는 미리 계획하고, 물리적 세계를 이해하고, 다단계 문제를 해결하는 능력을 향상시키는 것이 중요하다고 생각합니다.

AI 모델은 언어에 의존하고 공간적 추론을 사용하지 못하기 때문에 물리적 세계를 시뮬레이션하는 데 어려움을 겪습니다.

AI의 추론 능력을 향상시키는 것은 복잡한 문제를 해결하고 AGI로 나아가는 데 매우 중요합니다.

OpenAI는 고급 추론 능력과 AI 모델의 전문화된 적응을 약속하는 새로운 기술인 "Strawberry"의 출시를 예고하고 있습니다.

스트로베리 프로젝트: 혁신적인 AI 기술

OpenAI의 Strawberry 프로젝트는 스탠포드에서 개발된 STAR 방식과 유사한 사후 학습을 구현하여 AI 모델 학습에 소요되는 시간과 비용을 줄이는 것을 목표로 합니다.

또한 스트로베리는 장시간에 걸쳐 최상의 결과를 도출하는 장시간 작업 수행에 초점을 맞추고 있으며, 이는 현재 AI 기술 환경에서는 독보적인 역량입니다.

이 프로젝트는 이러한 기능을 사용하여 AI 모델이 Kua 에이전트의 도움을 받아 웹을 탐색하여 자율적으로 연구를 수행할 수 있도록 하는 것을 구상하고 있습니다.

AI 모델 진행을 위한 OpenAI의 확장성

OpenAI는 인공 일반 지능을 향한 AI 모델의 진행 상황을 추적할 수 있는 척도를 개발했으며, 현재 채팅 GPT는 레벨 1이며 레벨 2를 목표로 하고 있습니다.

레벨 2는 박사 수준의 기본적인 문제를 해결할 수 있는 시스템으로 정의되며, 레벨 3은 사용자를 대신하여 작업을 수행할 수 있는 AI 에이전트로서 생산성을 크게 높일 수 있는 수준을 말합니다.

사용자를 대신하여 작업을 수행할 수 있는 Siri와 같은 AI 에이전트의 개발은 인류를 위한 잠재적으로 혁신적인 변화의 시작이라고 할 수 있습니다.

AI의 미래: 엣지 디바이스에서 AGI까지

연사가 곧 공개할 동영상에서 예측한 것처럼, AI 개발은 엣지 디바이스를 위한 작고 효율적인 모델로 나아가고 있습니다.

다음 단계의 AI 혁신은 새로운 과학적, 수학적 발견을 창출하고 AGI에 도달하며 궁극적으로 조직 전체를 대체할 수 있는 AI를 개발하는 것입니다.

이러한 발전은 자율 주행 차량의 수준에 비유되며, 사람의 개입 없이도 AI가 업무를 수행할 수 있는 능력으로 정점에 이릅니다.