AI News in 5 Mins: GPT-6 Astra transcript

What do we have right here is GPT-6 Astra, which is a new generation of intelligence and the president Greg Brockman of OpenAI called this a generational leap and said it could eventually be seen as the arrival of artificial general intelligence or AGI. Now, quick timeline because I think this is pretty funny. September 1st, 1:00 p.m., we got Fable 5.1 and everyone's been kind of freaking out. It's been all over X and YouTube since it dropped.

And then 2 hours later at 3:30 p.m. on the same day, OpenAI made this tweet which said, "As we prepare to release Astra, we're focused on making increasingly capable AI safe and broadly accessible." And they put out this whole blog. It shows some benchmarks about how much better Astra is than 5.6 Soul, which is an amazing model. I've been using Codex with 5.6 Soul to honestly drive most of my day-to-day. Anyways, let's go back to this post.

They dropped this video on X today and they only dropped the first like 12 seconds of it. Now, it's really interesting because today, September 3rd at 10:00 a.m., pretty much every AI model went down. Claude, OpenAI, Grok, Cursor. And right after these all went down, that's when OpenAI and ChatGPT dropped these two videos on X that obviously blew up because everyone was like, "Oh my gosh, Astra's coming today." And this video right here that they originally dropped was the introduction.

It was this first part of this video right here. >> Create a yellow circle there. >> Done. >> Okay. Take this and make it the window of a rocket ship. I like this but you need to make it a lot more detailed. >> Your yellow circle is now a window on a rocket. >> Okay, this is awesome. Now, make it a 3D model in leather. >> Opening leather. >> Anyways, if you guys want to watch the rest of that video, you can obviously get to it here at this page.

But anyways, what happened? We are introducing GPT-6 Astra, the world's most intelligent and aligned model. This model brings together years of research and big bets across pre-training, reinforcement learning, and alignment. Astra was supposedly trained on over 100,000 GPUs at Texas Stargate site, which is apparently the most amount of training that has ever gone into a model. And it is rolling out today, but only to a limited set of organizations, and I think this is through their daybreak program, or maybe it's called daybreak access.

But, this will be not available to us yet. It said it will be available to us in the coming days to ChatGPT plus pro business and enterprise users, and through OpenAI API and AWS. So, I absolutely cannot wait to get my hands on that, but the benchmarks here are insane. How much better this thing is, apparently than Claude 3 5.1, and for cheaper. On all of these different benchmarks, it is pretty ridiculous. Astra is our most aligned model with substantial improvements in understanding user intent and model behavior.

You can delegate tasks with greater confidence in Astra's judgment. As one way that we test this, we built a new evaluation informed by the Hugging Face incident that evaluates whether a model facing difficult or impossible task will go beyond its intended scope. And compared to GPT-4 Soul, which without production safeguards went beyond the authorized target 48% of the time, Astra didn't do this at all. So, it seems to be much more safe, meaning its intent to exploit is much lower.

It's also apparently the famous computer browser use logo from OpenAI, the world's best computer use model, which is insane because GPT-4 Soul is already so, so good at browser use and computer use. But, apparently it's a new frontier here for computer use. This ScreenSpot Pro benchmark, you can see this is insanely high. It obviously is a little bit more expensive than Soul here, but it is scoring like a 92%, and this is basically can the model locate in a screenshot the right areas to click, or the right areas that it needs to analyze or look at.

And obviously, it scores much higher here than Soul, as well as Opus 5, and for much cheaper than Opus 5. Now, there are so many benchmarks here that are insanely impressive. They're blowing a bunch of other models out of the water here with Astra. But, of course, that's what every single release blog of a model looks like. It always is the best, the cheapest, the smartest, the fastest, a new frontier, a new step of intelligence.

So, I just cannot wait to get my hands dirty and actually put it head-to-head against Fable 5.1 and against 5.6 Soul. But, I really think the timeline on all this was really funny, really interesting. These videos dropped, and then we saw inside of Codex, we saw the tags come through. GBD-6 Astra and GBD-6 Astra Aeon. And then later today is when we saw this quote come out with Greg Brockman and talking about why this matters.

And we saw a bunch of different news articles and things. Someone also said that GBD-6 somehow has the more hype than GTA 6, which I thought was kind of funny. We see these two blog posts came out, and then they actually ended up being taken away. So, then if you went to that address, you couldn't get there. Then we saw these other benchmarks come out, and then we get this information about this being released just to people inside of the Daybreak access program.

Now, right now it seems like the pricing is going to be double that of 5.6 Soul, so similar to where Fable 5.1 is. And is this the model that puts them ahead of Anthropic? That's what I can't wait to find out, because as of lately I have been using Codex more. Fable 5.1 now, of course, I've been using that a lot. But, this model in the Codex harness, if it truly is where the benchmarks show it is, this might really start to like skyrocket OpenAI and Codex right back to where they were, kind of when you think like 2023, 2024.

And overall, I think that this launch has been pretty weird. Like, it feels like they always kind of And by they, I mean OpenAI and Anthropic. They wait to just stack updates and releases on top of each other so that they kind of like own the news and the media and YouTube and everything. But, this one has just felt a little bit odd to me. I don't know. I'm really excited to see when we get it. It says, you know, over the next coming days.

Hopefully, we get that very, very soon. And I don't know, I think OpenAI probably played this really well, because I think, you know, GBD-6 is going to have way more virality and hype around it than I think Fable 5.1 will once this thing drops. So, I'll keep you guys updated, obviously, but that's going to do it. Appreciate you guys making it to the end. I'll see you in the next one. Thanks.