OpenAI's Project Strawberry: GPT-5 and Reasoning Model Tests

Sam Altman's tweet and anonymous model appearances hint at the imminent release of OpenAI's GPT-5, codenamed Project Strawberry.

### Is GPT-5 "Strawberry" on the Horizon? Analyzing the Latest AI Model Rumors
* 🍓  Recent hints and rumors suggest that OpenAI may be close to announcing its next major language model, potentially named **GPT-5** or **"Strawberry"**.
* 📱  Sam Altman, OpenAI's CEO, recently tweeted a picture of strawberries, sparking speculation among AI enthusiasts about the **upcoming model**.
* 🤖  Two anonymous models, potentially related to OpenAI's new project, have appeared in the LMCIS.org Arena Battle mode, with some users reporting **interactions with them**.
* 🧠  The new model is rumored to focus on **improving reasoning abilities, long-horizon planning, and post-training fine-tuning**, which are key steps towards achieving **AGI (Artificial General Intelligence)**.
* 🔍  One of the key features of the new model is its ability to **generate answers** while also **planning and navigating the internet autonomously for deep research**.
* 📊  Some experts believe that the math reasoning aspect, a significant challenge in AI, may have already been **largely solved** by other companies like DeepMind.
* 🧪  Initial tests with the new models show **promising results** in challenging reasoning tasks, such as **logic puzzles and spatial reasoning**.
* 🚀  The AI community is eagerly awaiting an official announcement from OpenAI, as the new model could represent a **significant advancement in AI capabilities**.

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

The Hype Around the Possible Release of GPT-5 "Strawberry"

Sam Altman's cryptic tweet featuring strawberries has sparked rumors of the upcoming release of GPT-5, the next Frontier Model from OpenAI.

Two anonymous models have appeared on LMCIS.

org, following OpenAI's previous release strategy, adding to the speculation.

The rumored capabilities of GPT-5, including advanced reasoning, planning, and autonomous internet navigation, have contributed to the growing excitement and anticipation for its potential launch.

AI models are traditionally trained and then fine-tuned later, but OpenAI aims for consistent fine-tuning and learning.

The project, known as Project Strawberry QAR, focuses on long-horizon tasks and post-training fine-tuning for optimized performance.

Despite this effort, there are doubts about whether it will outperform other advanced AI models in math reasoning, as demonstrated by DeepMind's recent success.

The Importance of Strategic Thinking in AI Algorithms

AI models need to develop strategic thinking and long-term planning to improve their reasoning and output quality.

There is a growing need for prompting techniques that encourage models to explain their reasoning process.

Twitter is abuzz with speculation about "Project Strawberry" and its potential impact, with one account serving as a prominent hype man for the project.

The new GPT-4 model is only accessible through the Arena Battle prompts, making it difficult to obtain.

When a new model is used, it provides a step-by-step breakdown of the problem, offering multiple answers.

The logic and reasoning problem is presented with clear step-by-step explanations for each model's response.

Misconceptions About Microwaving a Marble in a Glass

The experiment involves placing a marble in a glass and microwaving it, with the misconception that the marble's position will change.

The video suggests that the marble will move from the bottom of the glass to the table outside the microwave, which is a misunderstanding of the principles of gravity.

The creator plans to conduct more tests, but it's important to approach such experiments with accurate scientific knowledge.