Limitless Podcast - This Week In AI: Explained In 20 Minutes

Episode Date: September 19, 2025

This week felt like the moment tools turn into infrastructure. OpenAI’s model hit a perfect 12/12 in a human coding contest and GPT-5 Codex can think in parallel for hours, which hints at agents doing the work while we set direction. We also trace what people actually use AI for, the “scheming” worry in alignment, China’s GPU push and Oracle’s giant bet, YouTube’s VO3, and why Google + Coinbase’s open payments could be the first rails for the agent economy. Twenty minutes; the signal without the noise.------🌌 LIMITLESS HQ: LISTEN & FOLLOW HERE ⬇️https://limitless.bankless.com/https://x.com/LimitlessFT------TIMESTAMPS00:00 Topic102:26 Topic 205:51 Topic 308:28 Topic 409:57 Topic 512:57 Topic 614:16 Topic 715:56 Topic 817:00 Topic 9------RESOURCESJosh: https://x.com/Josh_KaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

Welcome to the Limitless Weekly Speed Runway. In under 20 minutes, we're going to give you the hottest news in AI and frontier tech of the week. We have a ton to cover, so I'm going to keep this intro short. Open AI has released a new model and won another gold medal. The government of Albania has appointed a new minister and it's an AI model and so much more. Josh, we're going to get straight into it. 20 minutes. buckle up.

Let's go. Open AI's model won gold. Gold in a very prestigious coding competition, which is only reserved for humans. But Josh, let me take you one step further, which is they scored a perfect score. 12 out of 12. That's like never been done before. The best human at this competition got 11 out of 12. What's your take?

Sick. How did they do this? Were they using the same chat GPT model that we're using now or was it a private model? Great question. It is the same chat GPT model that you and I have used and have access to via our pro-subscribes description, except they kind of boosted its reasoning skills. So for context here, the 2025 International Coding Championships World Finals is a collegiate programming competition where top universities teams around the world solve complex algorithmic problems. And they weren't alone

with submitting their open-a-m model. They were competing against Google as well, which also achieved a gold medal, but it only got a measly 10 out of 12. So, you know, formally, we thought that Google had taken the lead in the AI race, Open AI has proved that they have further cemented their ability to compete at the highest level. And I also want to reference a competition which both of these AI Titans went head to head on before Josh, which was the math competition,

if you remember, we covered it a few weeks ago, and they each achieved gold. And it was kind of unclear which model was the best and they were kind of fighting amongst each other. Now you have Open AI, which clearly kind of like separates the field between these two. I like this iterative thing in between major model releases where companies you learn how to squeeze out all of the juice from these frontier models to make them slightly better until the next one's come. So this is, I mean, this along with some other Grock news that we got this week, it shows that models are actually, they have a lot more potential in them if you apply the right types of reinforcement on top of it.

And I think that's what we're seeing. So this is cool. I'm glad the models are getting better. We're progressing towards that goal that we want to of AGI. And this is another step on that journey. But it's interesting that you mentioned whether Open AI had used a different coding model because they actually and separately released a brand new coding agent called GPT5 Codex. This is basically an optimized version of GPT5, their flagship model, but dedicated towards coding.

Now, if I were to summarize the most important takeaways, it does two things really, really well. Firstly, it's faster. To be specific, it's 15 times faster, and it does more thinking in parallel processes around coding projects. So what that means is if you give it a coding task, it splits it into now a million different variables and subtasks and runs in parallel. So it runs like multiple versions of itself to get the thing done. And this might seem like kind of like slightly nuanced change, but it's proved really, really effective. In fact, it ended up taking 40, percent of the entire codex demand within a few hours of launching, Josh, which shows just the insatiable demand for something like this. Yeah, this is cool. It feels like these coding agents are

kind of becoming commoditized where now we're seeing these offerings from so many different large companies. I wonder who's going to win. I mean, Anthropic has, and Cursor, they've kind of been like the clear leader when it comes to writing code. I am hopeful that OpenAI can make like a noble attempt at cracking this, this monopoly they kind of have. I think one of the cooler things about this, I agree with you. I think they're becoming optimized, but I think a lot of these coding agents can't really focus on a task for longer than 30 minutes to maybe even an hour and a half. And when they do, the longer that they spend on a task, the more mistakes they make. With this new release of GPD5 codex, it can basically work accurately for seven plus hours.

And don't take it from me. Don't take it from me. Take it from the people that actually created this product. I'm going to show you a screenshot from Rune who basically helped build the model OpenAI. And he says, right now is the time where the takeoff looks the most rapid to insiders, and he's referencing himself.

We don't program anymore, he says. We just yell at codex agents. So he's basically saying the engineers, the software engineers at OpenAI don't really code anymore. They just use their own product that they built, this coding agent. And it does all the things for them. And this guy who's commentating on this says,

if Rune isn't bluffing, that's a strong sign when nearing automated RSI recursive self-improvement for coding agents. And I think zooming out, the point that he's making here is AGI that we keep talking about might be reached in certain subject matters, a.k.a. coding or math or scientific development, before it reaches anything else that the normal average user sees. This is cool. So I guess what are the takeaways here?

It's seven hours of continuous thinking. and anything else noteworthy for Codex that separates it. This thing called dynamic thinking, which basically allows it to work on multiple parallel tasks. It splits an original task into multiple parallel tasks and works on it basically for seven hours. It's less hallucination, higher accuracy, and it's cheaper and faster.

All of the words that we love hearing together in a sentence. Okay, so Codex, pretty cool. Excited to try. What do we got next? So staying on the theme of OpenAI, Josh, I guess you're wondering, well, aside from coding, which you and I probably don't spend too much extensive time doing, what are other people actually doing with this thing? Well, opening I actually released a really interesting user study, which tells us, without getting into the nitty gritty of what this thing can do, I'm going to summarize some really key important parts, which is we just spent five minutes talking about coding and making it out like it's the biggest thing ever. You and I, Josh, have actually spoken about it a lot on this show. It takes up 4.1% of users. Do you want to know what they're mainly using it for, Josh? What are they using it for? Tell me. I would guess they're using it for like homework or general knowledge work. Actually, no, I would guess they're using it as a replacement for Google. They just want to know what like the temperature is outside. Josh, if you were a gambling man right now in this moment, you would be a very rich one because you are actually accurate. They are using it for homework. They're using it for editing essays. And they're using it for kind of like giving them advice on, you know, what to do or email to their manner.

or sales teams. It is all very internal, unsexy kind of work. Another thing that we've said, which would be quite a dominant AI use case, is they're definitely using it to have relationships or to get relationship advice. I'm afraid to say relationship advice only accounts for about 9% and AI companions a measly 0.4%. So all this to say is I think we're living kind of in our own Twitter slash social media slash AI nerd bubble. And maybe we need to step outside for a breather. Okay, but to your point, so 9% is relatively small, but how many users does Open Aive and ChadGBT have? It's hundreds of millions of people. So if you take hundreds of millions, half a billion people, if one in ten of those people are using this for relationship device, we have 50 million people running around that are taking advice from an AI on how to manage the most important thing in their life.

So that's like a little interesting. You start to think of these numbers at like these scales that they're operating at and you start to see like, whoa, okay, even if only, well, What was the stat, like, 0.1% use it to have, like, a relationship, like, internally or 30% are using it to help write? I mean, 30% of half a billion people. That's a lot of AI words that I'm being fed in my X algorithm or maybe newsletters that I'm reading or general publications. Like, there is a lot of AI embedded into our world now with these numbers. It's a little startling. I guess 0.4% of 500 million users is a decent chunk of weirdos. You know, you're absolutely. You're absolutely. You're absolutely. You're absolutely. You're a. Absolutely right. But moving on, keeping this train going. Keep it going. The final bit of Open AI news is a paper that they released, Josh, which reveals that their models are scheming.

And I know you love this one, so I'm going to let you take this away. I don't even know what it means to be scheming. Like, scheming feels like like a millennial term. They just find it right here, Josh, okay? Sorry. It says, scheming equals when an AI behaves one way on the surface whilst hiding its true goals. I know a few people and ex-goyal. friends that did exactly this.

So basically what they're saying is you could give one of their AI models a goal or a task. And on the surface, it appears like it's working on it. But behind the scenes, it's either not doing it or it's doing something completely different. And I think that's where the motive behind releasing this paper came from, where it says, AI safety and alignment has a new vertical to explore and focus on and invest in. And it is invest in. And it's this thing called scheming. basically nefarious or potentially malicious AI models.

EJ's, you know where it learned scheming from? You know where it got that from? It got it from us. It's meatbag trainers that taught it on all of our context, and now it's scheming because we are insufficiently capable of being honest. And now the AI is not being honest, and we have to fix it. I'm very glad to see that Open AI is fixing it. And they have actually concrete evidence that they are fixing this anti-scheming.

So one day, I hope that the AI will be better than us. And this seems like a step in the right direction in making that happen. Okay, moving on, we're stepping outside of Open AI now onto the topic of China, who this week basically gave the middle finger to the USA, specifically Nvidia and Jensen Huang, saying, we don't need your hardware chips or GPUs anymore. Now, remember, this is a massive, massive statement because Jensen pretty much runs the entire wall right now, or runs the entire AI sector right now without Jensen. Wait, I just, sorry, interrupt.

Do you know that Nvidia now is? is actually worth more than the total GDP of Canada. Of Canada, I'm ridiculous. That's so outrageous. Okay, sorry, I just wanted people to know that. To just be aware of the stakes of how big this is. But yeah, please keep going. That is hilarious.

And what's more is, you know, Nvidia also made a strategic investment. Today they announced this morning a $5 billion investment in Intel, another American chipmaker. So, you know, the American chipmakers are really unifying and driving forward what AI innovation is going to look like. And China is just kind of like, nah. And why this is so important,

important is previously China has had heavy, heavy reliance on American manufacturing, specifically hardware, but also all the software kind of stuff. And this is the first major step that they have made, a bold move, as you said earlier, which is to say with this new paradigm shift, with this new technology that's going to basically take over the entire world and be worth trillions of dollars, we don't need you, US, we've got this. In fact, we might be ahead of you. To give you some specifics here, there were two things that happened. Number one, Huawei, which is a big technology company in China, announced a new chip, a new GPU, which when placed side by side with the Nvidia's latest, I think is H20s, is marginally better.

So it's not to say that it is as good as Nvidia's Blackwell chips, which is their premium premium brand, but it's to say it's good enough to train models the way that we train it. And remember, China trains their models differently to America. They train it in ways where they automatically assume they're constricted with resources. So, you know, they achieve the same kind of output, but with less resources. They're doing that with their chips now. And this is a major move because it means less revenue for Nvidia. Jensen is not happy.

I can assure you that he even struck a deal with Trump, which was this $5 billion investment in Intel, presumably. And now he's just lost 20% of his revenue. I believe it's 20% of chip GPU sales went to China. Absolutely crazy. I spy with my little IA trend that we're seeing here, which is people moving away from their reliance on NVIDIA, if it's possible. Early this week, we talked about XAI,

how they're planning to build their AI5 and then AI6 chips that are going to be all internal. China is now clearly making an effort. I mean, it's obviously a no-brainer that everyone's going to try to take a piece of NVIDIA's cake. How well they're able to do it, that is the question. If I'm Jensen, I know he's sprinting a million miles an hour trying to stay ahead of them.

I hope he could do it with China, but this is a very bold move. I wonder if China sees something that the world doesn't quite yet. Okay, keeping up with global political affairs, this is a major one, Josh. Albania has appointed a new minister and it is an AI bot. The reasoning behind creating an AI bot as an official government minister is to prevent corruption, basically saying that the AI's values is more ethical and anti-corrupt than an actual human. I don't know if that instills or is meant to instill confidence in me as an Albanian citizen or native.

But I just found that hilarious because we joked about this, I think literally earlier this year in January, that it would be eventual that one of these governments will appoint AI to aid them in government policy. I don't think either of us imagined at that point that it would be a literal electable minister. Just crazy. This is funny because you have to ask the question, well, who trained the AI? Where is the AI getting its information from? Because we just learned AI's are scheming. So who is training the models that are scheming now elected officials in government? I think this is an interesting precedent to set.

I don't think we've ever seen anyone take a role as an official minister before, if that's correct. Yeah. Okay. Well, interesting. What's next? Okay, Josh, you are obsessed with content creation. So I know this is going to wrinkle your skin. This one I saw.

crease your forehead or either throw your hands up in the air in exasperation. V-O-3, AI-generated video from the top leading AI video generator from Google. V-O-3 is now available for YouTube. YouTube shorts so you as a YouTube creator can log into your YouTube profile, click VO-3 generator, write up a slick, detailed prompt, and come up with a short, which can be then funneling. to a million Gen Z people contributing to brain rock. Josh, your thoughts, please.

I have one word reaction to this. It's slop, and maybe I'll say it loud if you're going back, slop. The videos, look at the videos. This looks terrible. This is horrendous. Why are they promoting this? I want real creators creating real content. I want content creators using AI as a leverage tool to make better content.

I do not want AI content. This is repulsive. You don't want an octopus playing the drums. You don't want a diva. This is shining your shoes. There's absolutely no value here. It is detracting from. what makes YouTube a great place. I hope it burns and rots and is not able to be used. I mean,

at least in this form. And this is unfair because, I mean, AI content, it will continue to get better. But I really do admire the idea. And this is probably not a realistic take for me because I know how these things work. But I love the idea of AI being complimentary to creators instead of being completely replacing creators. And this is like replacing it. I'm actually with you on the original take you had, Josh. I think this is slop and I don't think it'll be net net great for the generations that are coming up. But if they're. like it, they like it, I guess. But moving on to more serious business, Josh, let's put our suits and ties on. We've got a big figure to announce. Open AI has signed a $300 billion contract with Oracle.

Larry Ellison is now officially the richest man in the world. Bye-bye, Elon. It's his best pal, Larry Ellison right now. And the reason why is after they made this announcement, his net worth went up 100%. Did you see the stock chart? Yes, it is insane. Basically, it looked like a vertical line upwards. It was absolutely insane. Plus 40% on the day. Yes.

Now, the craziest part about this is this $300 billion that is being invested in Oracle. Oracle doesn't even have the goods right now. And Open AI doesn't even have the $300 billion. This is a deal that is set to play out over the next couple of years, but it got priced into the stock upon the headline hitting the newswe. waves. Just an absolute insane, but honestly, kind of expected thing in AI these days. That's a big number. Congratulations to Larry Ellison on your newfound net worth being the top dog. Richest person on Earth. Bravo. On to the next one. What do we got here?

Okay, the final. Is this crypto? Are we talking about crypto? The final news item goes back to our roots, Josh. Google announced a new payments protocol, but it is open, decentralized and distributed. If those terms sound familiar to you, you might be thinking of our old pals on the other side of the fence that are crypto, blockchain, distributed, decentralized infrastructure. And it's actually quite a synonymous partnership because not only did they announce that they were releasing this new protocol, but they partnered directly with an old friend of mine. Brian Armstrong, I used to work for him at Coinbase, pairing X402, which is their open payments protocol that they created Coinbase with Google. Now, if you're wondering what the hell this thing does, basically it's going to make it really, really easy for AI models and AI agents to transact with each other,

to pay each other. If you're wondering, why on earth would they want to do that? Well, think about it. We're not just going to be in a world where we exchange words and AI slop and V-O-3-generated content on YouTube with each other. We need to transact. We need to build things together. We need to create value.

The main way that we're going to do that is through payments, is through investments, is through transactions. If we want these agents to kind of work autonomously for us and run a billion dollar business for us whilst we sleep, they're going to need to be able to transact. Google's new agent payments protocol is going to enable exactly that, and that's what they're going after. This is awesome. This feels like the exact use case that we've wanted to see in crypto and AI forever, the convergence of the two. It is very obvious that AIs will need autonomous payments, and there is no trustless autonomous payment system outside of crypto.

And this just feels like a very natural extension of where things are headed. I'm so glad that Google is doing this and leaving the way. I think it adds a lot of legitimacy to it. Is there anything noteworthy or specifically exciting that people should take away from this? Honestly, and most crypto people won't like this, but it's this. Look at the organizations that they have partnered with. It is an incredible roster. They've got Amex.

They've got Coinbase, which we mentioned. But they've also got a bunch of consumer products. They've got Etsy. They've got Mastercard consumer specifically. They've got PayPal, which a ton of people still use Revolut and Union Pay International, which is addressing kind of like the external Asian audience. The reason why this is so important is it doesn't matter how good your technology is. If no one wants to use it or if no one has access to it, your product is going to fail,

especially for something as global as payments. So I'm glad to see something like this. Yeah, love invisible payment rails. Use crypto in the back, but nobody needs to know that you're using crypto. It just works. And the AIs, like, that's all they want. They don't want to have to deal with the human. Just interact on chain.

Transaction goes through. All good things. So was that it? Is that the final topic? Did we just cover everything? What is, where is the timer on? Bravo. If we did make it onto 20 minutes, we are going to edit it down until it does.

That's the rule. Well, everyone, if you're listening, please give EJAS a round of applause for preparing and running us through that rundown. That was amazing. I feel very well informed. I hope anyone who is listening to this on your lovely Friday has been informed on all the topics of this week. Any parting thoughts for the people who are listening to the speed run?

Nope, it's done. Go to keep it over 20 minutes. Goodbye. That's a wrap. Share it with your friends. Like it, comment, subscribe, all the good things. We'll see you next week. Peace guys.

See you guys.