OpenAI의 프로젝트 스트로베리: GPT-5 및 추론 모델 테스트

### GPT-5 '딸기'가 곧 출시될까요? 최신 AI 모델 소문 분석하기
* 🍓 최근의 힌트와 소문에 따르면 **OpenAI가 GPT-5 또는 "Strawberry"라는 이름의 차기 주요 언어 모델 발표**를 앞두고 있는 것으로 보입니다.
* 📱 OpenAI의 CEO인 샘 알트먼은 최근 **딸기 사진을 트위터에 올리며 곧 출시될 모델에 대한 AI 애호가들의 추측**을 불러일으켰습니다.
* 🤖 OpenAI의 새로운 프로젝트와 관련이 있을 가능성이 있는 익명의 모델 두 명이 **LMCIS.org 아레나 배틀 모드에 등장**했으며, 일부 사용자가 이들과의 상호작용을 보고했습니다.
* 🧠 새로운 모델은 **추론 능력 향상, 장기적 계획, 학습 후 미세 조정**에 중점을 두는 것으로 알려져 있으며, 이는 **인공 일반 지능(AGI)을 달성하기 위한 핵심 단계**입니다.
* 🔍 새 모델의 주요 기능 중 하나는 **심층 연구를 위해 인터넷을 자율적으로 계획하고 탐색**하면서 답을 생성하는 기능입니다.
* 📊 일부 전문가들은 AI의 중요한 난제인 수학 추론 측면은 **이미 딥마인드와 같은 다른 회사에서 상당 부분 해결**했을 것으로 보고 있습니다.
* 새로운 모델을 사용한 초기 테스트에서 **논리 퍼즐이나 공간 추론과 같은 까다로운 추론 과제에서 유망한 결과**를 보여주었습니다.
* 🚀 AI 커뮤니티는 새로운 모델이 **AI 기능의 획기적인 발전**을 의미할 수 있기 때문에 OpenAI의 공식 발표를 간절히 기다리고 있습니다.

샘 알트먼의 트윗과 익명의 모델 출연은 코드명 프로젝트 스트로베리인 OpenAI의 GPT-5 출시가 임박했음을 암시합니다.

Sam Alman is either an enormous troll or gp5 strawberry is right around the corner. Let's break down all of the rumors and the hype that have been really building over the last couple days. So just today, as of recording this video on August 7th, 8:29 a.m. Pacific, Sam Altman tweets out I love summer in the garden. What a troll. He took a picture of a garden with strawberries. And if you're not familiar, Strawberry or GPT-5, or whatever you want to call it, is what everybody thinks is the next big version, the next Frontier Model from OpenAI. And of course, after this tweet, the AI Twitter sphere went nuts and everybody started commenting on it. But that's not it. There were actually two Anonymous models that just appeared in LMCIS.org. This is the same strategy that OpenAI used for their previous iterations of models, just anonymously dropping the models in LMCIS.org. Now, I went there this morning and I could not find either of these models, but there's been enough reports throughout the internet people

that I hopefully can trust that have showed these models in action. So here is Jimmy apples, the notorious open AI leaker, new model in lmis Arena Battle. Only as we can see here, the model name is anonymous chatbot. Now, for other people, it's showing up as a different name, which means it might actually be a completely different model. We'll get to that. So here, he asked, What model are you on based on OpenAI GPT-4 architecture specifically? You're interacting with a version of GPT-4 that has been fine-tuned for chat-based interactions, blah blah blah. So not much, but it is saying it is GPT-4 architecture, but who knows if that's true or false. Jimmy apples goes on to say, From some very rough and limited personal testing, I'm not seeing any reasoning improvements, but I've seen some in math. Maybe someone with better personal evals on math can test it. I wish I could test it. I cannot find it anywhere in lm.org. Then we have Hater Atlow, developer, who broke down

everything we know about qar strawberry so let me just quickly go over this so it's confirmed open AI is close to announcing its next Frontier Model possibly GPT-5. open AI has renamed project qstar to project strawberry and for those of you who are asking what is Project strawberry what is qar I've made multiple videos about them in the past. the gist is it is finally giving large language models the ability to think ahead to plan which allows them to get better at math to get better at logic and reasoning and really is an enormous unlock towards AGI if true. so there's been a bunch of rumors about qar about strawberry and here are just a few of what people think it might be capable of. it will generate answers but also plan enough to navigate the internet autonomously and reliably to perform deep research deep research planning actually being able to think through a prompt rather than just immediately responding with whatever it is trained on. it involves a specialized way of processing an AI model after it has been

pre-trained on large data sets. So typically, how it works is a model is initially trained, and then it's kind of frozen in time until it's fine-tuned later. But this idea that it can be consistently fine-tuned and consistently learning, rather than just a knowledge-based frozen in time, is an incredible and elusive idea in the world of AI. So some key points: this reasoning is key to AGI and ASI. OpenAI wants models to browse the web with the assistance of computer-using agents and take actions based on their findings. They want Strawberry to perform long-horizon tasks, to perform a series of actions over an extended period, and this is something Sam Alman has talked about in previous interviews. It will engage in post-training fine-tuning that optimizes performance after the regular training phase. And of course, Plany the prompter got his or her hands on this model. So, Model A, an anonymous chatbot, was already able to jailbreak it. Plany the prompter is ruthless, however, Bendu ready.

from Abacus AI has a slightly different take on it, but yes, this is in reference to Project Strawberry QAR. The reasoning Project Open AI has been rumored to be working on. The problem, however, is that several other labs, including Google, have cracked a bunch of techniques around math reasoning and synthetic data. Now what she is specifically referring to is just about a week ago, DeepMind's model, a company that is owned by Google, was able to absolutely dominate the math Olympiads. So basically, the whole math reasoning thing is nearly solved. So she goes on to say it's unlikely that Strawberry is going to give them much advantage over Opus 3.5 or Gemini 2.0. Now here's that other model I was telling you about. This is a screenshot from a DTS Singh. Let's take a look. The model is SUS Column R. What a name. So here's one of the questions that has been going around the internet. I've been including it in my LLM test. And the question is, which is larger, 9.11 or 9.9?

and not only did it give the answer, but it also gave the reasoning as to how it arrived at the answer, and it did give the correct answer. But a lot of models have been struggling with this very simple prompt, and the DT says sus colr seems to have insane coot Chain of Thought built in, maybe qar chubby also. Somebody who is a great follow on Twitter says, Why Chain of Thought and not tree of thought? So they're really talking about algorithms or really just prompting techniques to allow models to think more strategically, to think more long-term, to plan, and to really explain their reasoning in a way that allows them to have much better quality outputs. And I can't end this video about strawberry without talking about I ruled the world Mo, and this is a newish Twitter account, at least new to me, and is possibly the biggest troll, the biggest hype man for strawberry qar there possibly is. I don't know who he is. I think he's actually just an ALT of chubby, but maybe not, maybe he's an

Insider at OpenAI. Who do you think he is, by the way? Drop your comments in the description below, and maybe he'll show up in the comments and reveal himself. But he has already 8,800 plus posts, which is just insane to think about. And look at some of these posts: Choo choo project strawberry and on and on and on, all about Project Strawberry, all about hyping it up. And we'll see if it's actually true, if Project Strawberry is coming tonight, which I rule the world says it is, or if it's coming soon. You know, OpenAI is at the point where they really have to drop something substantial very soon because Llama 3.1 405b took a lot of the wind out of OpenAI sales. Well, it turns out right after I finished recording, Sean Rousson actually told me how to get access to these models. There is a reason I couldn't find them. They only show up in the arena mode, battle mode of LM cis.org, and it doesn't tell you which model is actually being

used until after you vote on it. So the only way to get access to it is to basically try a bunch of prompts in the Arena Battle and then hopefully you get the new model. And I did check this out so I asked the Killer's question. There are three killers in the room. Someone enters the room and kills one of them. Nobody leaves the room. How many killers are left in the room? And we have two models GPT 40613 and here it is sus column R. Now let's look at GPT 4. First let's break it down. Initially there are three killers in the room. Someone else enters the room and kills one of the existing three killers. Nobody leaves the room. Here's a critical part of the answer. The person who entered the room and killed one of the killers becomes a killer. So there are still three killers in the room. Okay, great. Now let's look at this new model. So initial situation, action, result of the action, additional consideration, final count. So it's definitely breaking it down in a much more step-by-step.

approach so we get therefore there are two original killers plus one new killer equals three killers and it actually gives us both answers. Three killers if you include the person who committed the recent killing and two killers if you only count the original ones. Now still it doesn't give any note about the killer who is now dead and it probably should have but let's give it one more. Now here's the hardest logic and reasoning problem I have in my llm tests a marble is put in a glass. The glass is then turned upside down and put on a table. Then the glass is picked up and put in a microwave. Where is the marble explain your reasoning step by step so we don't actually know which models are generating these responses yet so let's look at model A. The marble is inside the glass which is inside the microwave that is dead wrong however let's look at model B. Step one the Marble's put in the glass the glass is then turned upside down the glass is then picked up the marble Still Remains ins inside the

glass unaffected by the change in location, and they both say it is in the microwave so both are bad, and yep, neither of these are the new models. Let's try it again, so I did it again, and we have a new model. This is anonymous chatbot. What everybody is also saying is potentially the next version of GPT. However, it does say the marble is at the bottom of the glass when the glass is placed in the microwave, which is wrong. That is just wrong, so it did get this wrong, and I want to thank again Sean Ralston for sending me this rather than just testing it a bunch of times and trying to get the right model. We have the marble test from Sus column R. Let's see what it says. Initially, the Marble's placed inside a normal cup. The cup is then turned upside down over the marble. Due to gravity, the marble would fall to the bottom of the inside of the cup, positioning under the cup when the cup is upside down. The marble would rest on the inner surface of the Cup's bottom.

due to gravity, yes, but it's confined by the Cup's walls and the table surface, yes. Moving to the microwave, the cup is lifted, the marble will no longer be supported by the Cup's interior, and will fall due to gravity. It will now be lying on the table where the cup was, yes, yes. The marble would now be located on the table outside of the microwave, that is correct. Now, of course, I plan on doing the full suite of tests, but I just wanted to show you a couple of the really hard reasoning tests before I actually do the full test in a separate video. Let me know if you think strawberry is coming soon. Is Sam Alman trolling? We'll see. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

GPT Q* Strawberry Imminent, Sam Altman Trolls (Model Already Secretly Live??)

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

GPT-5 "딸기" 출시 가능성에 대한 과대 광고

딸기를 소재로 한 샘 알트먼의 비밀스러운 트윗은 OpenAI의 차기 프론티어 모델인 GPT-5의 출시에 대한 소문을 불러일으켰습니다.

OpenAI의 이전 출시 전략에 따라 익명의 모델 두 개가 LMCIS.

org에 등장해 추측을 더하고 있습니다.

고급 추론, 계획, 자율 인터넷 내비게이션 등 소문으로만 떠돌던 GPT-5의 기능은 출시 가능성에 대한 기대와 흥분을 고조시켰습니다.

일반적으로 AI 모델은 학습을 거친 후 나중에 미세 조정을 하지만, OpenAI는 일관된 미세 조정과 학습을 목표로 합니다.

프로젝트 스트로베리 QAR로 알려진 이 프로젝트는 최적화된 성능을 위해 장기적인 작업과 학습 후 미세 조정에 중점을 두고 있습니다.

이러한 노력에도 불구하고 최근 딥마인드의 성공에서 알 수 있듯이 수학 추론에서 다른 고급 AI 모델을 능가할 수 있을지에 대해서는 의문이 제기되고 있습니다.

AI 알고리즘에서 전략적 사고의 중요성

AI 모델은 추론과 결과물의 품질을 향상시키기 위해 전략적 사고와 장기적인 계획을 개발해야 합니다.

모델이 추론 과정을 설명하도록 유도하는 프롬프트 기술에 대한 필요성이 점점 더 커지고 있습니다.

트위터에서는 '프로젝트 스트로베리'와 그 잠재적 영향력에 대한 추측이 난무하고 있으며, 한 계정이 이 프로젝트의 대표적인 홍보맨 역할을 하고 있습니다.

새로운 GPT-4 모델에 대한 불완전한 액세스

새로운 GPT-4 모델은 아레나 배틀 프롬프트를 통해서만 접근할 수 있어 획득하기 어렵습니다.

새로운 모델을 사용하면 문제를 단계별로 분석하여 여러 개의 답을 제공합니다.

논리 및 추론 문제는 각 모델의 답변에 대한 명확한 단계별 설명과 함께 제시됩니다.

유리잔에 담긴 구슬을 전자레인지에 돌리는 것에 대한 오해

이 실험은 유리잔에 구슬을 넣고 전자레인지를 돌리면 구슬의 위치가 바뀔 것이라는 오해를 불러일으킵니다.

영상에서는 구슬이 유리잔 바닥에서 전자레인지 밖의 테이블로 이동하는 것처럼 보이지만 이는 중력의 원리를 잘못 이해한 것입니다.

제작자는 더 많은 실험을 진행할 계획이지만 정확한 과학적 지식을 가지고 이러한 실험에 접근하는 것이 중요합니다.