OpenAI's Strawberry AI: Insights on Upcoming Orion Model

OpenAI's new model, Orion, leverages Strawberry for advanced reasoning and reduced errors in AI applications.

### OpenAI's Upcoming Model: Insights on Strawberry and Orion
* 🔍  OpenAI is reportedly on the verge of releasing a new model named **Orion**, which follows the development of **Strawberry**, a technology aimed at **enhancing AI's reasoning capabilities**.
* 📈  Recent articles reveal that OpenAI demonstrated Strawberry to U.S. National Security officials, marking a **significant step towards transparency** as advanced AI becomes a **national security concern**.
* 🧠  Strawberry is designed to **improve complex problem-solving**, allowing AI models to engage in **multi-step reasoning** rather than simply providing immediate responses.
* 💡  The model aims to generate **high-quality training data** for Orion, addressing the **scarcity of publicly available data** and potentially **reducing the hallucinations or errors** in AI outputs.
* 💰  OpenAI's subscription revenue has **tripled to 283 million monthly**, indicating a **strong market demand**, although the company still faces **significant operational losses**.
* 🛠️  As competition in the AI space intensifies, with rivals like Anthropic and Google's DeepMind making strides, OpenAI's success hinges on the **effective launch of Orion** and its **unique capabilities**.
* 🔄  There is speculation that users may eventually **toggle Strawberry's advanced reasoning capabilities on and off** depending on their needs, allowing for **tailored responses in various applications**.
* 🎉  The excitement surrounding these developments mirrors the hype seen in other tech industries, as enthusiasts eagerly await the **official unveiling of OpenAI's latest innovations**.

apparently, OpenAI is on the verge of releasing their brand new model, and now we have a ton of additional information about it, including the fact that OpenAI may actually have shown this new technology to the federal government. So, we're going to go over all of that right now. Alright, so we have two main articles from The Information that were posted over the last couple of days. First, OpenAI shows Strawberry. So, we've talked a lot about Strawberry, talked a lot about QAR. I'll link those videos in the description below, and I will drop a link to the full article in the description below. OpenAI shows Strawberry AI to the feds and uses it to develop Orion. And Orion is new to me, and apparently, it is their next Frontier Model. Now, here's the thing: if you were ever doubting Jimmy Apples, the infamous OpenAI leaker, look at this tweet from all the way back on November 24th, 2023. Let's conquer the cosmos mood curious Jimmy. So, with that, did he actually know? And then as of

today yeah been waiting since last year for this mood patients Jimmy and he's referring to the information article about Orion so maybe he knew about it so let's read a little bit about the article so this article by Stephanie palazolo starts with in case you were wondering why Sam Alman cryptically posted a picture of strawberries earlier this month the answer almost certainly has to do with strawberry a mysterious technical breakthrough that could help open AI models complete complex tasks such as math problems that conversational AI has traditionally struggled with so just a quick catchup qar strawberry it's all referencing the same thing slower thinking models models that can actually plan ahead think through problems multi-step reasoning and actually do much better at math and logic and reasoning these models don't just return the first token they predict they actually go off and think about it and actually do long-term thinking in mid July Reuters reported on the existence of strawberry and this morning we published this piece with even more details and we will go over that in a

moment here is the most interesting bit. This summer, his team, Sam Alman's team, demonstrated the technology to American National Security officials, said a person with direct knowledge of those meetings, which haven't been previously reported. By demonstrating an unreleased technology to government officials, OpenAI could be setting a new standard for AI developers, especially as advanced AI increasingly becomes a national security concern. The demonstration could be part of OpenAI's push to be more transparent with policymakers who could cause the company problems if they feel threatened by its technology. And scanning down a bit, and maybe also take a shot at Meta Platforms for releasing open-weight AI that China and everyone else can access. So you all know how I feel about this. Open source is the way to go. To think that a single private company is going to be able to protect secrets, to protect the weights against an adversary like China infiltrating and getting that IP, it's just so unlikely that they will be able to protect it forever. Right here, Meta CEO Mark Zuckerberg says it's

inevitable that China will get it one way or the other, and I agree. All it takes is one slip-up in security, which we're humans. We make mistakes. One slip-up, and then China will have the weights anyways. And it's not just about the weights; you also need the compute to power it. You need a lot of other things as well. I did a full review of Mark Zuckerberg's thoughts on open source, and it overlaps with my own quite a bit, so I'll drop that video in the description below. Let's keep reading about strawberry and why it matters to Orion. And again, strawberry qar, it is different from Orion. Oran is reportedly their next Frontier Model that is going to leave everybody else in the dust. One of the most important applications of strawberry is to generate high-quality training data for Orion, open ai's next flagship large language model that's in development now. Why is that important? Much of the training data on the internet has already been used. There's almost none left that is public, and

easily accessible now. It's all behind paywalls and authentication. That's why OpenAI has been doing deals with different publications and Reddit and all these other companies. And that's why Twitter's dataset is so important to X.A. But Sam Altman a while ago hinted that maybe more and more training data isn't actually necessary. There's actually two other approaches that are possible. One, that they can actually do a lot more with existing data. And the second path, which it looks like they're talking about here, is actually generating synthetic data using one model for another model. Now, there's been a lot of doubt that that is a sustainable model because if you're creating data with one model to be used on another model, it's basically derivative data. A lot of people think that large language models aren't actually creating new knowledge; it's simply outputting what it already knows. So if it's only able to output what it already knows, how could it really train a model to be that much better using strawberry could help Orion.

reduce the number of hallucinations or errors it produces. Research ERS tells me that's because AI models learn from their training data, so the more correct examples of complex reasoning they see, the better. So what could be going on here is they have a set of data, and Orion and Qstar are doing long-term thinking and multi-step planning and basically thinking through the data to make sure it's accurate and then producing new data that is highly accurate. But there's also a push within OpenAI to simplify and shrink strawberry through a process called distillation so it can be used in a chat-based product before Orion is released. So that tells me that the technology behind strawberry might just be so slow it's unusable in a consumer setting. And that would make sense, and I'll actually touch more on this in a bit, but the gist is strawberry and QAR takes time. That's a feature, not a bug. It actually takes a lot of time. It thinks through it just like humans would. When you ask someone something they don't

immediately spit out the first thing that they think of, or at least most people don't. Instead, they take time to think through it. They might take notes, especially hard questions, reasoning, logic, math. These are things that we take our time with. We write down thoughts, and we iterate on the result. Here, the information admits we're not sure what a strawberry-based product might look like, but we can make an educated guess. One obvious idea of what strawberry could be an actual product is incorporating strawberries' improved reasoning capabilities into Chachi BT, and that's the obvious one. However, though these answers would likely be more accurate, they also might be slower. That means that strawberry might be ill-suited for applications where users expect immediate responses, like OpenAI's search GPT search engine, but ideal for less time-sensitive use cases, like fixing non-critical coding errors in GitHub. Now, another good use case for this is Agents. If I have an agent working for me 24 hours a day, I really don't need a response immediately. I can

give it these superpowers powered by Qstar, powered by strawberry, allow it to go do its thing and then bring me back the best possible response you could imagine. A not-so-distant future where chat GPT users are able to toggle strawberry on and off depending on how sensitive their requests are. So that's the first article from the information about Orion. That's the first time I've heard of it, so hopefully, I'm bringing it to you for the first time too. Now, Chubby from X, who is an amazing follow if you're not already following him. What Chubby touches on here is that strawberry is slower because rather than just responding immediately with whatever it thinks is the next correct token, it actually takes its time and does what is called system two thinking rather than just system one thinking. Now let's read more about strawberry. So this is another article from the information. Open AI raises to launch strawberry reasoning AI to boost chatbot business. So it is reported that Open AI is looking to raise even more capital, which is insane.

They've already raised so much, but you know what? It costs a lot of money to build these models, at least for now. Its researchers are trying to launch a new artificial intelligence product they believe can reason through tough problems much better than its existing AI. Here, they just touch briefly on QAR. We've already discussed it quite a bit. When given additional time to think, the strawberry model can also answer customers' questions about more subjective topics, such as product marketing strategies. To demonstrate strawberry's prowess with language-related tasks, OpenAI employees have shown their co-workers how strawberry can, for example, solve New York Times connections, a complex word puzzle. So, if it were given the puzzle and just asked to output immediately what it thought, system one thinking, then it's not able to solve it nearly as effectively as if it had time to iterate and use maybe tree of thought or chain of thought or any of these other more advanced capabilities where the model can actually look ahead, plan, and test things out, come back, and test other things out as they see things are working or not.

Here, the information talks about the OpenAI business and its sales of LLMs to corporations. The ChatGPT subscriptions have roughly tripled to 283 million in monthly revenue compared to a year ago. That's insane, though its monthly losses are likely higher than that, as reported by the information. And I've already explained that they are probably still losing money, but they're not going to go bankrupt anytime soon. The company is privately valued at 86 billion, but OpenAI prospects rest in part on the eventual launch of a new flagship LLM it is currently developing, codename Orion. So, we already talked about that. Now, why is this next model so important? Well, open source pretty much caught up with GPT-4. Llama 370B, Llama 345B is a fraction of the price. You can run it locally, you can fine-tune it, and it's nearly as good or as good as GPT-4 for the majority of use cases. And not only that, we have the CLAD models, we have Grok 2, we have more Grok models.

coming we have perplexity for search, so the competition is heating up really quickly, and OpenAI really needs to launch something incredible to jump ahead because intelligence is being driven down to a cost of nothing. Here, they talk about the same things that we touched on in the last article, but let me just read it again. OpenAI is also using the bigger version of Strawberry to generate data for training Orion, and said a person with knowledge of the situation that kind of AI-generated data is known as synthetic, something we've touched on a lot on this channel. It means that Strawberry could help OpenAI overcome limitations on obtaining enough high-quality data to train new models from Real World data, such as text or images pulled from the internet. And here, apparently, OpenAI is going to be launching agents soon, so Strawberry could aid upcoming OpenAI agents, that person said. Using Strawberry to generate higher quality training data could help OpenAI reduce the number of errors its model generates, otherwise known as hallucinations. Now, one of the biggest

blockers. One of the biggest hurdles for artificial intelligence in general to be adopted within enterprise settings and more critical settings is the fact that it still hallucinates. There's a bunch of things that you could do to reduce hallucinations, whether that's improving your prompts, having multiple agents talk to each other and verify, kind of like an agentic system, and pulling in information from the internet to verify, and of course doing your own verification. But at the end of the day, if we want large language models to be fully autonomous and to really run at the scale that we believe it can, it really needs to reduce hallucinations to nearly zero. So, the CEO of agent startup Minion AI and former Chief Architect of GitHub Copilot says, Imagine a model without hallucinations, a model where you ask it a logic puzzle and it's right on the first try. The reason why the model is able to do that is because there is less ambiguity in the training data, so it's guessing less. So, from Sam Altman at an event back in May, We feel like we

have enough data for this next model, so they are close, and I really cannot wait to see the next thing that they bring out. I have said this a lot; it's really hard for me to imagine that a private AI company is going to have some completely unique research or unique tech that allows them to make a 50-100% improvement on what's already out there. Maybe 10%, maybe 20%, but a large leap from what is already on the market today is really hard for me to imagine just because of the way that the scientific community operates. All of these papers get published, and unless OpenAI research scientists are the only ones in the world who actually thought of a new idea and it didn't leak out anywhere, it's just very unlikely that they're going to have some completely new technology here. They talk about why solving math problems could be so lucrative for OpenAI as a business. AI that solves tough math problems could be a potentially lucrative application given that existing AI isn't great at

math-heavy fields such as Aerospace and Structural Engineering. Now, that isn't necessarily true. Google's DeepMind team actually got, I think, silver at the math Olympiads, so it's definitely possible today already that AI can be extremely good at math. And right here, they reference what I just mentioned. Google DeepMind said its AI would beat most human participants in the International Mathematical Olympiad. Another major rival, Anthropic, said its latest LLM could write more complicated software code than its prior LLMs could and answer questions about charts and graphs with improved reasoning capability. So again, competition is there, and specifically about Anthropic, the cloud model is seemingly the favored large language model for coding use cases. Everybody seems to be using Cursor, the IDE that is AI-native, in conjunction with the cloud models. It continues with to improve models' reasoning. Some startups have been using a cheap hack that involves breaking down a problem into smaller steps. Though the workarounds are slow and expensive, it's kind of weird to call it a cheap hack.

these are discovered prompt techniques and frameworks to wrap your large language model to make them more effective, and I don't see that as a cheap hack at all. And yeah, they tend to be slower, but what if you just power it with Gro grq, and then all of a sudden you have all the benefits of these additional layers of technique to get the best response, but you're also getting hundreds of tokens per second. So, it ends with what Ilas saw. Strawberry has its roots in research. It was started years ago by Ilas Suser, then OpenAI's Chief Scientist. He recently left to start a competing AI lab. Before he left OpenAI, researchers Jacob and Simon Sedor built on SS's work by developing a new math-solving model, QAR, alarming some researchers focused on AI safety. And basically, the entire super-alignment team has left since then. So, what did Ilia really see? And last, just to add a pinch of fun to this from New York Magazine, the AI guys are driving themselves mad, and this is in reference

to Strawberry man and qar and the strawberry model Jimmy apples and all of this conjecture and speculation and you know what it's all in good fun. It's the same thing that basically every other industry does when Apple is about to release a new phone. Everybody who follows Apple starts talking about what it could possibly be and talking about the leaks and you know what we're so excited about what's to come. Why not add a little speculation and try to figure out what might come. So this article is pretty funny. It just talks about how I rule the world. Mo was responded to by Sam Alman and all of his hypey tweets which is great. And here's the awesome part. They actually included one of my videos about qar which is flattering. It's so cool. They also go on to talk about Lily Ashwood who everybody's trying to figure out if she's AI or not. We talked about that on the live stream last week. So there's definitely a lot going on a lot of fun and as soon as the

new open AI model comes out. You know I'm going to cover it in depth, and I'll give you all the information. So, be sure to subscribe to my channel. And as a reminder, I have an awesome newsletter so you can stay up to date on the latest AI trends and news, Matthewburman.com. Check it out. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

Project Orion (GPT-5 Strawberry) Imminent, Already Shown To FEDS!

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

OpenAI is set to release a new model, Orion, which is developed using Strawberry AI, a technology that aims to improve complex task completion and multi-step reasoning.

The company has demonstrated this technology to U.

federal officials, potentially setting a new standard for AI developers amid growing national security concerns.

The move towards transparency with policymakers may be a strategic effort to mitigate any potential backlash from the government's perception of OpenAI's powerful technology.

The Role of Strawberry in OpenAI's Orion Model

Strawberry is crucial for generating high-quality training data for Orion, OpenAI's next flagship language model.

With public training data becoming scarce, OpenAI is exploring synthetic data generation and leveraging existing data for model improvement.

The technology behind Strawberry is slow, leading to a push for its distillation for use in a chat-based product before Orion's release.

OpenAI's Strawberry: A New Approach to AI Reasoning

OpenAI's Strawberry model takes a slower, more deliberate approach to problem-solving, using system two thinking instead of immediate responses.

This method is ideal for less time-sensitive tasks, such as fixing coding errors or providing strategic marketing advice.

The model's ability to iterate and plan allows it to tackle complex language-related tasks, like solving word puzzles, more effectively.

OpenAI's Business Growth and LLM Development

OpenAI's ChatGPT subscriptions have tripled to 283 million in monthly revenue, but the company is still likely losing money.

The privately valued company is focusing on launching its new flagship LLM, Orion, to stay competitive against affordable open-source models.

OpenAI is using its bigger version of Strawberry to generate synthetic data for training Orion and is also planning to launch agents soon to aid in reducing model errors and hallucinations.

The Competition and Challenges in AI Development

Private AI companies struggle to achieve unique breakthroughs, making significant improvements on existing technology challenging.

Solving math problems is a lucrative opportunity for AI, with competitors like Google's DeepMind and Anthropic making strides in this area.

Startups are using innovative techniques to improve model reasoning, while Ilas Suser's departure from OpenAI to start a competing AI lab has raised concerns in the research community.

Speculation and Excitement in the AI Industry

The article humorously discusses the speculation around upcoming AI models, comparing it to the excitement generated by Apple's phone releases.

It highlights the author's video on qar and the ongoing debate about Lily Ashwood's AI identity.

The author promises to cover the new OpenAI model in depth and encourages viewers to subscribe for updates on AI trends and news.