OpenAI's new "SPUD" model will change the ENTIRE Economy

OpenAI has a new, yet unreleased model they're referring to as Spud. Could this be a GPT 5. 5, perhaps a GPT 6, or is it an entirely new naming convention? It very well could be. As uh some OpenAI employees have hinted, this is going to have a brand new feature that's going to be very different from what we've seen before. So, in just a one memo to OpenAI employees, Sam Altman announces the following. Number one, OpenAI is killing Sora. Sora is gone. I mean, still there, but not for long. It's going away. The next major model, Spud, has finished pre-training. And Altman himself will be taking a step back from safety oversight and focusing more on fundraising and infrastructure. Also, behind the scenes, there's a pretty big math breakthrough or math milestone that no one's talking about. Here's the thing, these aren't separate stories. It's one pretty major story. It's such a big story, in fact, it completely destroyed my green screen behind me, but that's a separate issue. Talk about it later. Okay, so first and foremost, let's talk about the big new model rumors, a lot of which are confirmed by OpenAI. So, OpenAI has completed pre-training its next major model, internally code-named Spud, like a potato. Altman told staff that it's a very strong model and it's coming in a few weeks. He described it as something that can really accelerate the economy. So, we're looking at a release possibly later this month or maybe in April. Meanwhile, Altman is stepping back from direct oversight of OpenAI's safety and security teams. So, the safety team will now report directly to Mark Chen, and the security team moves to the scaling operations under Greg Brockman. Altman's stated reason for doing all this is because the infrastructure is lacking. We need to build more chips, more data centers, and supply chains, and building them at any quote unprecedented scale. Senior exec Fijisimo, so he renamed his organization, his place on the org chart as AGI deployment. Is this the first time OpenAI had AGI anywhere in its org chart or as a product category? So, what does all of this mean? What is the thing that connects all the dots? Well, as I've said earlier in a previous video about Sora going away, OpenAI is uh done with the side quests, and they're exclusively focusing on the one thing where they think everything is going. That thing is, you can call it the super app. It's the one app to rule all apps, if you will. It's combining ChatGPT, Codex, Atlas, the web browsing tool that OpenAI has. All of it is being combined into one product, and the Spud model, GPT 6, or whatever it is, that Spud model is going to be released soon, that will power all of it. Now, this kind of restructuring and refocusing did have some casualties. Sora was one of them. Employees reportedly were saying how they were shocked at how much compute Sora was using. So, Sora got killed to free up compute for Spud and this super app. Also, in the past, OpenAI talked about potentially shipping video generation within ChatGPT. That's completely gone. That's off the road map. And I'm sure this wasn't a light decision because they had a three-year licensing deal with Disney. Disney had a $1 billion commitment for production with AI video generation. Sora was the thing behind that licensing deal, and that whole thing is just gone. Apparently, this was news to Disney. Now, it sounds like that deal never officially closed, and no funds were exchanged, so the deal is dead, it's off the table, but kind of it seems at the last moment and somewhat suddenly. Disney, of course, put out a vague statement saying, "Oh, we respect OpenAI's right to pursue whatever they want to pursue, etc. " Now, what happens to the Sora team? Well, so we have Bill Peebles, so he's the head of the Sora division. He had some words about what's happening with Sora, what they're pivoting to. He's saying, "We need to automate the physical economy. " And that the Sora research team will now be focusing on quotes, "Systems that [clears throat] deeply understand the world by learning to simulate arbitrary environments at high fidelity. " So, we're talking about world models for robotics, similar to what Google DeepMind has been hinting at, similar to what Nvidia has been working on. And Altman confirmed this. He was saying Sora research team will prioritize longer-term world simulation research, especially as it pertains to robotics. So, in the memo, Altman is citing specifically Google and Anthropic as the competition. Of course, Anthropic and OpenAI both are racing towards their IPO this year, and a lot of the labs have been saying that they're kind of having a hard time keeping up with Anthropic. Either they've said it uh officially or just we can kind of tell from their actions, but Anthropic is rapidly gaining pace, specifically with enterprise customers. They're putting out excellent coding tools and white-collar AI tools like Claude Co-work. They also have Claude Code. Now, we have Dispatch, which allows you to control your AI agents from your phone. Altman has said publicly that OpenAI is racing to ship similar tools. And of course, they've aqua-hired the creator of Open Claw, which will hopefully help them in that aspect. So

what we still don't know about the new model is the parameter count, whether it's a reasoning or non-reasoning model, whether it's multimodal. I mean, we can guess at these things, but no details have been made available yet. We don't know what it's going to be called, whether it's again GPT 6 or 5. 5 or something completely different, nor how it compares to the other models. From how they're talking, it almost seems like a kind of a different thing entirely. But, based on a some moments quote of saying this can really accelerate the economy, that's not really how Altman talks about regular product releases. For incremental updates, that's not really the wording that he would use. And the AGI deployment naming convention, the org chart rename, I don't think that's accidental. It seems like they're getting ready to call something AGI level. And of course, kind of the Altman power move is to get himself away from security and safety to have those departments report to different people, and himself kind of freeing up more bandwidth to focus on chips and power and supply chains, etc. Now, here's what was kind of lost in this developing story, what was happening concurrently to this, and it was kind of just buried because of how big this news was. So, while OpenAI is announcing Spud, or revealing, announcing, whatever the word is, doing this whole thing, Terence Tao, a Fields Medal winner, we've talked about him before, he's considered by many to be the greatest living mathematician. Before, he's already talked about how AI, somewhere in late 2025, it seemed like it really accelerated and made a breakthrough and was able to help more and more with various mathematical proofs and theorems. It solved a lot of the Erdős problems, which are very complicated problems specifically selected by originally the mathematician Erdős, but now Terence Tao is kind of the spiritual successor to that, keeping that alive and is selecting problems and finding whether or not they've been proven or if any progress has been made towards solving them. And right around the release of the latest Opus and the GPT models, there has been some noticeable and notable improvements in how well they're able to do this autonomous theorem discovery. So, Terence Tao publishes a paper recently, and there's this quote somewhere in this paper. I would like to read it. So, again, this is considered by many the greatest pure mathematician alive. Here's a direct quote from one of the papers that he just published. He's saying, "After experimenting with Alpha Evolve," right? So, that's Google DeepMind's LLM-driven, so it's LLM plus a harness around it, you can say. So, you can kind of say it's the advanced AI agent for scientific or math discovery. Tao says, "I was led to a way to prove the toy model's integral bound by splitting it into two lower bounds. " So, he had this problem, he sort of approached it by splitting it into two sort of parts, and he's saying ChatGPT proved the first and I proved the second. Right? So, kind of like let that settle in a little bit, right? So, the greatest mathematician alive today, he's now collaborating with AI to split a proof that he's doing into two parts, and AI is carrying one of those parts. So, what is Alpha Evolve? It's Google DeepMind's evolutionary coding agent. It's built on top of Gemini, so you have whatever model of Gemini you want to use. When they first published it, I think they were using like 2. 5, I believe, but you can switch them in and out. You can even have an ensemble of them, as Google put it. And it sort of iteratively goes through a lot of different solutions, tests them, see if they're better. So, it can refine code, math, and a lot of other things. Google DeepMind, when they first announced it, they posted how to improve their data center architecture efficiency, and improved some of the hardware that Gemini was getting trained on, as well as actually improving Gemini's own training process, right? So, kind of like improve itself, you could say. And this model, this agentic harness, this AI agent, whatever you want to call it, it has already improved mathematical bounds previously thought to be best known. So, Tao used this technology, Alpha Evolve, to find the correct approach to solving this problem. So, he used its evolutionary tree search to find the correct approach. So, rapidly kind of brainstormed a lot of things, tested them out, and said, "Okay, this is looking like it's a good approach. " Then, he splits the approach into two problems, handles one, and ChatGPT, that's what he referred to it as, so we don't know which model it is, but you know, it's an OpenAI model, or assuming it's the latest model, or maybe the reason that he chose to use ChatGPT instead of specifying the model may be because he has some advanced access to a previously unreleased model. It's probably not the Spud model that we're talking about here because reportedly that got trained or finished training yesterday. So, March 24th was the day that I think it was officially finished training. So, it wasn't that, but some other OpenAI model was what helped them do the proof for one half of this problem. Tao also maintains a GitHub kind of wiki with 100 AI-solved Erdős problems. And in January of this year, so GPT 5. 2 Pro solved three Erdős problems in a single week like shortly after coming out in January of this year. This increase in its abilities in math is a very recent. So, I would guess

that GPT-5. 4 Pro is probably the most likely candidate here, but again, we're not sure. All right, so I know we're hearing somewhat confusing narratives online in the news. Some people are like, "Oh, AI is all fake. It can't produce anything. Blah, blah. " At the same time, the greatest mathematician alive today, well, he's got a different opinion. He's saying, "Hey, AI is getting to the point where it's a credible sort of co-worker, co-scientist capable of helping mathematicians and scientists to advance our understanding of the world, our understanding of math forward. So, Tao predicted back in 2025 that by 2026, by this year, AI would become a trustworthy co-author. " Right, so that prediction, I would say, is ahead of schedule seeing as how he's publishing papers using it two models from two of the frontier labs as, you know, trustworthy co-authors or co-scientists. So, remember not that long ago, you know, 2025, early 2025, when Anthropic and OpenAI, you know, Sam Altman, Demis Hassabis, they were kind of hinting at the fact that, "Hey, soon we're going to have models capable of progressing science forward. " Right? And there was certain portion of the population that are like, "Ah, that's just nonsense. They're just lying. They're just trying to juice their profits and, you know, improve their shares value. Blah, blah. " And also around the same time they were saying that soon we're going to see superhuman coders. a lot of the coding done by these AI agents. And certainly we can argue if that's technically been achieved, right? There's still, of course, a ton of human software engineers and stuff like that, but we're also seeing these models with GPT Codex or Claude Code or the stuff that's built on top of Open Claw, we are seeing it creating some pretty impressive things. So, we might argue whether or not those statements technically have come to pass, but it's kind of hard at this point to deny the progress that's been made, right? You got to be a little bit crazy to say, "No, they're just completely off the mark. " Right? You can say that all the people working for the AI labs, they're all lying. And I've heard people say that, very smart people whose names that you know, who've written books that who have an audience. They say, "Oh, all those people are lying to juice their profits. " Fine. What about somebody like Terence Tao, right? Are they lying? Towards what end? Usually when people lie about attributing credits for a scientific discovery, they do it by not giving credit to other people that contributed. They take the credit for themselves. So, here it seems like he's crediting these AI models for doing a lot of the work leading to this paper. So, I don't think he would be doing that if that wasn't the case. I I trust him when he says that these AI models were instrumental in his ability to publish this paper. So, again, if you've been following this for a few years, I hope you're beginning to kind of see whose predictive abilities are being proven to be more correct and whose predictive abilities are not. So, the people that constantly downplay where AI is going to be in 6 months, in 12 months, would you feel like over the last 2 years they're just like nailing all their predictions and AI is not going anywhere? Or do you feel like the people that are like, "Hey, you know, look at this like exponentially growing curve of AI abilities. I'm going to just assume it continues. " Right? They just look at the chart and they look at the line in go, "Let's assume this line continues going in the same direction it's been going for the last however many 5 years or so. " As far as I can tell, those people that just kind of follow the line on a chart, their predictive abilities of what the future are being proven right. They're being proven as the correct sort of predictions about the future. So, now that the leaders of these AI labs are saying that, "Hey, we're going to soon have these models that will accelerate the economy. " What do you predict will happen, you know, this year or early next year? Do you think it's going to be complete nonsense what they said? Do you think that the economy will be completely unaffected by the models that will start coming out? If you've been betting against the AI progress this whole time, hey, maybe this is going to be the time that you're actually right. But me, personally, I would not bet against the line on the chart continuing in the same direction, the same slope that it has been all this time. So, let me know what you think. Are you excited about this S model? Do you think it's going to be a big deal? And do you think that we're going to start seeing these AI models coming out this year starting to accelerate the economy? Whatever that means to you. Do you think it's going to have a very noticeable, measurable, strong impact on the economy? Do you feel like when the leads of these AI labs, when they say things like AI is going to be doing a lot of the coding in the future or AI is going to be helping with scientific discovery in the future. Like they've said that quite some time ago, you know, in AI time. I'm talking about 6, 12, 18 months. That's like a long time ago in measuring in AI years. Do you feel like they nailed those predictions or looking back at it does it feel like they just lied to improve the value of their company? Where are you on that timeline? Which side of the argument are you on? If you don't think AI models are contributing to scientific progress, how do you think about what Terence Tao's saying? Do you think he's maybe overstating how important these models were towards him figuring out these proofs. Or maybe you don't consider that, you know, helping with scientific progress. So, let me know.

I'm curious to know. And as always, thank you so much for watching. My name is Wes Roth. I'll see you in the next one. And hopefully I'll have a better setup than this. I'm in the middle of building out to be a little bit better. I really hope this lighting isn't too weird. Anyways, see you in the next one.