Transcription
The details coming out about chat GBT5 are simply extraordinary. Open AAI are in this to win. And when you mention GPT5 now, people sort of roll their eyes like, when's it coming? There is an army of leaks coming out of it now. I think it's not overhyped. I think it's underhyped.
This morning, I was testing our new model and I got a question. I got emailed a question that I didn't quite understand. Uh, and I put it in the model, this GPT5, and it answered it perfectly. And I really kind of sat back in my chair and I was just like, "Oh man, here it is moment." And I got over it quickly. I got busy onto the next thing. But it was like, I mean, this what kind we're talking about. I felt like useless relative to the AI in this thing that I felt like I should have been able to do and I couldn't, and it was really hard, but the AI just did it like that.
Yeah. It was a weird feeling. It's now in its final testing phase and a big part of that test is to make sure it's actually safe, given all of its abilities, to release to the public. It was due out in July. Sam Alman announced a delay. It's now due out in August, so a few weeks away. So, a model at the very frontier with world-changing implications is coming out within a few weeks. The first thing to know is that GBD5 will be smarter than everyone on the planet.
GBD5 is the smartest thing. GBD5 is smarter than us in almost every way. You know, you know, there are these moments in the history of science where you have a group of scientists look at their creation and just say, you know what? What if, what have we done? What if maybe it's great, maybe it's bad, but what have we done? Like, maybe the most iconic example is thinking about the scientists working on the Manhattan project in 1945, sitting there watching the Trinity test and just, you know, this thing that had it was a completely new, not not human-scale kind of power. Everyone knew it was going to reshape the world. And I do think people working on AI have that feeling in a very deep way. You know, we just don't know. Like, we think it's going to be great. There's clearly real risks. It kind of feels like you should be able to say something more than that. But in truth, I think all we know right now is that we have discovered, invented, whatever you want to call it, something extraordinary that is going to reshape the course of human history.
Dear God. Leaked OpenAI document suggests that chat GBT5 has been, is being tested in sensitive domains. Testing this model to the limit to understand how it can perform in tasks like mitigating biological threats, dangerous biotechnologies. Right.
I'm not convinced I want it playing with biotechnology, analyzing vast data sets of scientific literature in order to try and find patterns, looking for interdisciplinary connections in different fields to find connections that humans might have missed. It's even been suggested that it could be testing for new materials by simulating the atomic structure. One of the most cited examples is that it's looking at drug compounds to try to simulate new drugs, simulating millions of different drug compounds and their behaviors to try to find new medicines. Bearing in mind, it knows everything. It has all of our data sets and it now has a context window or memory to process what we think will be millions of tokens, which is an absolutely enormous amount of data within one context window. So all of that information in its head and it's looking for connections in science to discover new science. Now, testing here is obviously important and they're also developing safeguards for misuse, for if someone prompts chat GBT5 to produce malicious malware information or biological threat information. So this is great if you're a nefarious villain. Also anticipated that GBT will incorporate Sora 2. Open AAI knows that it needs to compete with the best of the best, which is currently V3, which can create these amazing audio included polished videos.
Of Troy. They've just claimed the winner's trophy.
Let's park it inside. The idea is that Sora 2 will be able to create videos from text which are high quality and have audio and integrated into the new system GPT5. This is because GBT5 is to be multimodal. It has to have, as far as possible, everything in one place with the model deciding when to do deep research, for example, without you having to scroll through this chaotic complicated system to choose which individual model is best. They will want to put as much of that together, including the new Sora 2 system. The next thing is its autonomous agentic behavior. The idea that this would happen in GPD5 was met with some skepticism. But we have just had the launch of chat GPT agent. Today I asked agent to book me an itinerary to go to a tennis tournament in like Palm Springs next year. Find me an itinerary that works around the semi-finals because that's.
It can take whatever you want.
What your plan is, your prompt, and break it down into a multi-step task execution, which could be things like having it book you flights to New York at the cheapest cost. There's a lot that's already out now. So, you can say, I want you to book me a flight to Scotland, not Glasgow, because they stab people there, and I want it to be quite cheap, but it has come out with some limitations. The model is designed to refuse high-risk task execution. So there are additional safeguards at the moment. How GBT5 might alter those safeguards now that chat GBT agent has been able to run for a while remains unseen. But the intention, says Sam Alman, is that these safeguards will be gradually tapered away as we figure out the risks and how this model is actually functioning when it implements things in the real world.
This is an extremely exciting new technology. There are new risks. Uh, people learned how to use the internet generally pretty safely, although of course there are still scammers and other attacks. People are going to need to learn to use AI agents, uh, and society's going to need to learn to build up defenses against attacks on AI agents as well. So we're starting with a very robust system, lots of warnings. We will relax that over time as people get more comfortable with it. But we do want people to treat this as a new technology and a new risk surface and use all of the caution that Casey talked about.
As part of that agentic ability. Simpler versions of chat GBD5 have been created to operate within the apps within your smartphone. You will have the full version of chat GBD5 with all its abilities and cognitive abilities, but there will also be an API version that's for your apps. And then there will be, if you want it to be more efficient, there will be a nanoized version, super efficiency, but it doesn't have a lot of the model's power, or a mini version with a bit more and a bit less efficient. Has it reached artificial general intelligence? But the original definition of AGI comes from the 1950s, and that is an autonomous computer that can perform any task that a human can, including physical tasks like bringing you tea. OpenAI define it by economic value. A highly autonomous AI that is more economically valuable than any human. The problem is the definition is so poorly defined. So here lies the problem. It's smarter than any human. But every time you, every time you speak to it, it's like a new AI comes out of a box, brand new, because it's a new context window. It looks like chat GBT5 will solve that problem. The models at the moment already have a built-in memory. But what you really want if something is to become AGI, it needs to sort of build as its own person and remember everything. It will still remember key features about you and therefore be able to advise and guide you better because it will know who you are, even though you've opened a new window, a new context window. This, I believe, is actually an essential step towards artificial general intelligence. The model has to be one unified entity. It remembers its experience of the world. But the second problem remains that it's locked in a phone or a computer. We would need therefore to be truly the original definition of AGI, it would need to have hardware. It would need to have robotics. There are thousand Optimus Tesla robots coming this year, but there are expected to be millions by 2030, and they'll have grock in them presumably, but that will give it the ability to do physical tasks. So, is it then AGI? Well, other people might point out that well, does it need to be sentient? Does it need to be conscious? When you ask it a question and it gives you an answer, part of that answer has come from a data center in Texas and another one, say, in Canada. It's made up from the different data centers collectively. It's possible that they are sentient with all of these different data centers making up effectively one brain. Open AI have already proved that they can be somewhat general by winning gold at the IMO maths contest, which requires general maths. That will not, that maths ability though, will not be deployed into the model for a few months. Looks like this model will be groundbreaking. Sam Alman thinks that OpenAI will own well over a million GPUs by the end of the year. The size of the data centers being constructed at the moment is utterly bewildering. I went recently to one of our new data centers under construction in Abilene, Texas.
There's about like a approximately a 1 gigawatt facility. Huge. You know, it'll be the biggest data center ever built by the time it's done. And you stand in the middle of that and the scale of this project just hits you so big. That's like one little, that's like one little part of it.
Dude. That's like eight Costco codes.
Uh, you know, there's like 5,000 people there doing construction on it and this thing is just standing up, making progress every day and you stand in the middle of this thing.
And what were you in a chariot or whatever? Like, how do you even?
You're like in a little ATV.
Oh, okay.
Uh, it's like a dirty kind of construction site. Um, but it, the scale of this thing and then you kind of go in every room and you look at all the cables, the power, the cooling systems, rack after rack after server of servers. It's humongous. There's like they're standing up these like power plants right in the middle of it. There's.
Oh, yeah.
It's crazy.
You promise, dude.
I don't know. I mean, I have all my guesses. Like, I do guess that a lot of the world gets covered in data centers over time. Do you really? But I don't know because maybe put them in space. The compute that will come out with them coupled with the increased efficiency of these models will mean that ChachiBT5 is just the start of all of this. As ever, please do subscribe and join the club. Thank you.