Latest / In the Loop AI / OpenAI's GPT-6 Astra (Part II): Is This Thing AGI?

Transcript

OpenAI's GPT-6 Astra (Part II): Is This Thing AGI? · In the Loop AI

0:00
Josh: There's a new best AI model in town, and it comes by the name of Astra, also known as GPT-6.
0:05
Josh: Just last week, OpenAI released what is now known as the most powerful model in the world.
0:09
Josh: This is the new frontier setting level model in which it is capable of doing
0:13
Josh: all these unbelievable things that we've been testing out over the weekend.
0:16
Josh: So if you aren't familiar, or if you've just been using Astra a little bit over
0:19
Josh: the weekend, this episode is gonna fill you in on everything you need to know
0:22
Josh: from demos to actual use cases, to what I spent my entire morning doing that
0:26
Josh: would have taken me hours previously, that now only takes five minutes.
0:29
Josh: It's a pretty amazing model. There are certainly some shortcomings in which
0:32
Josh: we'll get into in a little bit. But to start, I mean, it has GPT-6 Astros here.
0:36
Josh: The frontier has shifted. There's a new best model on the block and we've got to talk about it.
0:40
Ejaaz: We've had around 16 model releases in the last three months, Josh.
0:45
Josh: It's been exhausting, dude. We've been covering all of them.
0:47
Ejaaz: Many different companies. There's been 16 model releases. This is the first
0:51
Ejaaz: model release where I've truly felt like this model could potentially replace
0:56
Ejaaz: a lot of what I and many other fellow humans can do.
1:00
Ejaaz: And I'm not trying to exaggerate here. So it actually all started,
1:05
Ejaaz: the imminent second of launch. So what you're seeing on your screen here is
1:10
Ejaaz: a recount of when Astra GPT-6 got released, and it went live for 30 seconds.
1:17
Ejaaz: Now, no one in the world knew, aside from 30 seconds later, when every entire
1:22
Ejaaz: computer server basically went down, and it took down every single LLM with it.
1:28
Ejaaz: Grok from XAI, SpaceX AI went down. ChatGPT obviously went down.
1:33
Ejaaz: Google's Gemini went down.
1:34
Ejaaz: And Anthropik's Claude also went down. There was a two and a half hour period
1:39
Ejaaz: where you basically couldn't access any of the favored AI models.
1:43
Ejaaz: And this coincided directly with the launch of GPT-6.
1:46
Josh: So of course, my tinfoil hat on for this.
1:48
Ejaaz: Like, I mean, mine is already, I don't know if you can see it.
1:51
Ejaaz: It's actually elevated beyond the ceiling above me.
1:53
Josh: It's permanent.
1:54
Ejaaz: My, like, listen, we've spent the last two weeks talking about this hugging
1:58
Ejaaz: face incident where an internal version of GPT basically broke out and spread
2:04
Ejaaz: a swarm of different agents, hacked into many different production servers when it wasn't meant to.
2:08
Ejaaz: And now I'm looking at this and I'm thinking, oh my goodness,
2:11
Ejaaz: there is a version of Ultron that is out there hiding in compute servers,
2:17
Ejaaz: and not a single human being knows where it is or what on earth it's doing.
2:22
Ejaaz: And we're going to discover something very bad very later on.
2:25
Ejaaz: This model was a very dramatic launch. People were very excited about it.
2:30
Ejaaz: Sam has been teasing it for so long. Thibaut, the lead of Codex,
2:33
Ejaaz: said something was felt across the internet today.
2:35
Ejaaz: Greg Brockman came on stage, or wherever he was, the president of OpenAI,
2:39
Ejaaz: and said, AGI is finally here.
2:42
Ejaaz: Now, AGI has been kind of like thrown around so many times, and I don't know
2:46
Ejaaz: what to make of it. But if anything, this opening video, Josh,
2:50
Ejaaz: I don't know if you saw the promo video of GPT-6, was so sick.
2:54
Ejaaz: It depicts many different people sitting in an armchair, and they're basically
2:58
Ejaaz: just talking aloud to this mysterious AI model that then just proceeds to do
3:03
Ejaaz: anything and everything for them on the screen in front of them.
3:05
Ejaaz: What is the screen that is on in front of them?
3:07
Ejaaz: It is their desktop. It is their computer, except he didn't touch a single key
3:11
Ejaaz: or move a single cursor throughout the entire demo.
3:14
Ejaaz: And this kind of precedes the main sort of advantage with GPT-6,
3:17
Ejaaz: which is it is the most amazing and frontier computer use model that is out there right now.
3:24
Ejaaz: What it means by that is you can ask the model to do pretty much anything and
3:28
Ejaaz: everything a human can do on a desktop. Now, previously, a bunch of these models
3:32
Ejaaz: were kind of good at it. It was sort of slow. You would still need to follow
3:35
Ejaaz: up and kind of tell it what to do.
3:37
Ejaaz: GPT-6 Astra is unique in the sense where it is basically relentless.
3:41
Ejaaz: It spins up multiple different AI agents. And if it can't figure something out, it doesn't go to you.
3:46
Ejaaz: It goes back to itself. It kind
3:48
Ejaaz: of like tests different forums and figures out the way to do the task.
3:51
Ejaaz: And it would often work hours to days to weeks at a time to get a task done.
3:55
Ejaaz: It is one of the most intelligent models that I've come across.
3:58
Josh: Yeah, I thought this was a lovely launch video. It was really cool because in
4:02
Josh: a way, when you get new models, and we frequently get new models,
4:04
Josh: what do we have to show for them? We get a launch card. We get like a sheet
4:07
Josh: with all the specs on it. We get the pricing.
4:09
Josh: Yeah, benchmarks. I really appreciate the fact that they distilled this new
4:13
Josh: model into a singular moment, which is just computer use.
4:16
Josh: Like if you are thinking of how to describe Astra, you are thinking of it as
4:19
Josh: a user of your computer. In fact, it can possibly even use your computer better
4:22
Josh: than you can. So now we have a computer that uses your computer
4:25
Ejaaz: Better than you can.
4:26
Josh: But what I loved about this demo is it felt like a look into the future that
4:29
Josh: open ai is making and i didn't see a lot of this commentary online per se but
4:33
Josh: i mean like we know and like we cannot say i personally cannot stop talking
4:36
Josh: about open ai is working on this suite of devices
4:39
Josh: and suite of devices the idea is to change the way in which we interface with ai on a regular basis
4:44
Josh: and what we're seeing in this launch video is a very clear
4:48
Josh: vision of what that is going to look like in a way that felt novel to me at
4:52
Josh: least it's something i've we've kind of been talking about loosely where okay
4:56
Josh: we speak to our computer now more than we actually type because I'm faster at
4:58
Josh: speaking. Dictation is very good.
5:00
Josh: It's a really cool way of engaging with it. But we still have to do a lot of
5:03
Josh: the manual tactile work on the
5:04
Josh: computer. We have to use the mouse and keyboard. We have to click around.
5:07
Josh: This model removes that second piece of the puzzle. It's like,
5:10
Josh: okay, now I can speak to my computer, but I'm not speaking to text. I'm speaking to action.
5:15
Josh: And the computer is now taking action on my behalf.
5:17
Josh: And this morning I was trying it out. And just before we started recording,
5:19
Josh: I was talking to, I was talking to you and Luke, and I was sharing the fact that
5:24
Josh: For a long time, I've been using editing software. For videos,
5:26
Josh: I've been using DaVinci Resolve, Adobe Premiere.
5:29
Josh: And for the first time ever, this model was able to take multi-cam footage,
5:32
Josh: which for those who don't know, there's like multiple cameras recording the same moment at once.
5:36
Josh: And it was able to separate them into their own tracks, figure out where the
5:39
Josh: audio was, align them together, color grade them all, and then render out a
5:44
Josh: finished export vision of that. And that is something that would take me previously,
5:48
Josh: I don't know, maybe half an hour to an hour to do.
5:50
Josh: And it did it in the background. And I think that is what's really cool here
5:53
Josh: is it shows what the next paradigm of computing can look like,
5:56
Josh: where you don't actually need to be sitting in front of a mouse and keyboard.
5:58
Josh: You could just stand in front of this big screen and tell it what to do and
6:01
Josh: it will just go and do the thing.
6:02
Josh: And it's better at doing that than just about any model in the world.
6:05
Ejaaz: I think with previous LLMs, Josh, and I'm curious if you agree with me here,
6:09
Ejaaz: you could kind of think of them as like thought partners. You can kind of like
6:13
Ejaaz: bounce ideas off of them. They come back and they're like, oh,
6:15
Ejaaz: you're kind of right here, you're directionally correct, but actually you could
6:19
Ejaaz: improve it this way or that.
6:20
Ejaaz: But then that's where sort of like the conversation and the relationship ends
6:23
Ejaaz: um with these new models or with gbt6 astra specifically it feels more like
6:28
Ejaaz: a work colleague or a friend that actually or like a co-founder that actually
6:32
Ejaaz: can do things with you and you can rely on them you can trust them now
6:36
Ejaaz: we were speaking about this exact thesis almost a year ago actually when the
6:40
Ejaaz: first version of computer use came out from chat gbt and when it came out from
6:44
Ejaaz: claude and we were like remember.
6:45
Josh: It was so cool back then
6:46
Ejaaz: It was so cool remember we saw the cursor and it was moving like this like super slowly.
6:51
Josh: That's so funny. That was like a novel thing and now it's just ripping through and doing everything.
6:56
Ejaaz: Well, we were like, okay, it's clicking on the incorrect thing.
6:58
Ejaaz: No, that's the wrong file. Kind of like slapping its wrist, like don't go there.
7:02
Josh: I vividly remember the cursor, how amazing that cursor was.
7:05
Ejaaz: And now when I use GPT-6 Astro, I'm looking at my screen and I'm like...
7:10
Ejaaz: Oh my God, I didn't even see the cursor move. And hang on a second.
7:13
Ejaaz: It has like four other virtual desktops set up side by side.
7:16
Ejaaz: So it's doing all these different things.
7:18
Ejaaz: And by the way, we're not just talking about this. Like I have a bunch of demos.
7:21
Ejaaz: I have a demo of like- Please.
7:22
Josh: We gotta get into the demos.
7:22
Ejaaz: Creating like a Call of Duty game from scratch and then playing the game and
7:27
Ejaaz: then chatting to itself and other people that are playing the game, doing the thing.
7:30
Ejaaz: It is just absolutely insane. But I wanna jump into a few of these demos.
7:34
Ejaaz: But before I do that, I just wanna point out that this is the largest ever,
7:40
Ejaaz: compute training run that a lab has done to create this type of model.
7:44
Ejaaz: And I want to kind of like make this a point before we jump into the demos,
7:47
Ejaaz: because what you're about to see is very impressive.
7:49
Ejaaz: It was trained on 100,000 Frontier Edge NVIDIA GPUs. So that is,
7:56
Ejaaz: for all intents and purposes, a very expensive wallet budget.
7:58
Ejaaz: Like the only other person to potentially match that or is about to match that
8:02
Ejaaz: is Elon Musk, because he's acquired all the NVIDIA GPUs.
8:04
Ejaaz: But it just proves to show that 100,000 of these GPUs, which costs in like the
8:08
Ejaaz: billions of dollars, is able to scale frontier intelligence with these different models.
8:13
Ejaaz: So for those of you who are out there that were thinking, oh,
8:15
Ejaaz: you know, some of these Chinese open source models who has fewer compute,
8:18
Ejaaz: but is able to make some of these breakthroughs, they're good models.
8:21
Ejaaz: But the law still stands, which
8:23
Ejaaz: is if you have enough GPUs, you're able to create a much better model.
8:27
Ejaaz: Now, Josh, when we think of benchmarks, we hate these benchmarks, right?
8:30
Ejaaz: They're like just they're so kind of like docile, they don't really tell us
8:34
Ejaaz: anything but the one that i like the most is the pelican svg.
8:39
Ejaaz: Um test you know what i mean like it's a 2d image right and you ask a model
8:43
Ejaaz: hey i want you to do your best at basically like creating um,
8:49
Ejaaz: a 2d version of a pelican well check this out this is what uh gbt6 astra.
8:55
Josh: Pelican created san francisco
8:57
Ejaaz: Okay this is like 3d right it's got some form of a creature is that a beaver
9:02
Ejaaz: I have no idea, in the front of its basket. So for all the AI nerds that are
9:06
Ejaaz: listening to this, you're going to be the most impressed by this because this
9:09
Ejaaz: technically classifies as an SVG.
9:11
Ejaaz: But for those of you who are listening and are thinking, well,
9:13
Ejaaz: okay, who the hell cares? Why should I actually care about this?
9:16
Ejaaz: Well, for the gamers that are out there, Josh, you are one of these, right?
9:20
Ejaaz: I classify you as a gamer. Look at this. GBD6 Astra basically created this entire
9:25
Ejaaz: 3D game, one-shotted it in 45 minutes. A few things to point out here.
9:29
Ejaaz: Look at the fidelity of this game. This is something that like you could create,
9:33
Ejaaz: like game developers spend probably years right now trying to recreate right
9:38
Ejaaz: now today. And it costs like millions of dollars. It costs like a lot of people's
9:42
Ejaaz: time to create from scratch.
9:44
Ejaaz: This model did it in 45 minutes from one single prompt. I don't think we've
9:48
Ejaaz: ever seen anything like this before.
9:50
Josh: What I found this really good at is tool use as well. There's another example
9:53
Josh: just above this one about humans and the Unreal Engine. Yes.
9:57
Josh: Yeah, so this is amazing. It's like, okay, for those who don't know,
10:00
Josh: the Unreal Engine powers a lot of AAA games that are being used out in the world
10:03
Josh: today. It's like a gaming engine. That's where game developers go.
10:06
Josh: They build these virtual worlds.
10:07
Josh: It had Astra go off and build a virtual world using this game engine,
10:11
Josh: and it looked incredible.
10:12
Josh: It was really impressive. The ask was to create a world with the Unreal Engine
10:16
Josh: and fill it with humans, each an astral-powered agent, who all have to work together to survive.
10:21
Josh: And those agents all get mini LLMs and they start, this feels like what we imagine
10:25
Josh: GTA 6 would look like, where every NPC has an LLM attached to it.
10:29
Josh: They have their own world, they have their own context, their own memory.
10:32
Josh: And it looks amazing. And what I've noticed throughout the demos of this,
10:35
Josh: at least in terms of like visuals, is that it's really powerful and impressive
10:40
Josh: at visually understanding the world, being able to take these tools and apply that to the world.
10:45
Josh: So I've seen a lot of examples also with Blender where people are designing houses.
10:49
Josh: And also, like I mentioned earlier, there is the examples of video,
10:53
Josh: which we talked about earlier where I was using as a video editor.
10:56
Josh: Now there's a version of GPT-6 making kids' cartoons using Higgsfield,
11:01
Josh: editing it in Premiere, and then uploading it to YouTube, which I found really
11:04
Josh: impressive. Like, look at this video of the computer use, actually clicking
11:08
Josh: through, editing the videos.
11:10
Josh: It's impressive. Like, there's no way to look at these demos and be like...
11:15
Josh: Wow, this isn't kind of like this novel new thing where for the first time ever,
11:18
Josh: it's like almost better at using a computer than we are. In fact, it is in most cases.
11:23
Ejaaz: Yeah, this is one of my favorite examples, the creation of a kid's show,
11:28
Ejaaz: because you can see all the three different components of this model working in real time.
11:33
Ejaaz: So on the right side of the screen on his laptop, it is basically the AI kind
11:38
Ejaaz: of like dreaming up a prompt of what a good episode would look like, typing it into itself.
11:43
Ejaaz: It's desktop version of ChatGPT coming up with the episode outline
11:47
Ejaaz: then coming up with the script and then being like okay how can I create
11:51
Ejaaz: the different scenes based on this script feeding that into Hicksville which
11:54
Ejaaz: is on the left side of the screen right and then like putting it into an editing
11:58
Ejaaz: software and then editing the different clips to make sure that it kind of like flows seamlessly
12:02
Ejaaz: there's like so many different things that you mentioned earlier Josh there's
12:05
Ejaaz: the audio files there's making sure that it kind of lines up with the right
12:07
Ejaaz: types of animation and then you have the finished product that is.
12:11
Ejaaz: On top of the screen right there which honestly looks like something that I
12:15
Ejaaz: would see on a kid's TV channel, Josh, like what I would see from my baby's checking it out, right?
12:21
Ejaaz: It comes with titles, it comes with captions, it comes with different angles
12:26
Ejaaz: of the same different scene. It's so thoughtful, but it's done this from scratch.
12:30
Ejaaz: And by the way, it just uploads it on YouTube.
12:32
Ejaaz: So you could have a hands-off kind of ghost YouTube channel that makes you a
12:36
Ejaaz: ton of money or entertains your kids.
12:38
Ejaaz: And let's say you had a kind of comment on this. You didn't like that your kid
12:42
Ejaaz: was kind of learning certain things or like being exposed to certain types of show content.
12:47
Ejaaz: Well, you could just kind of like ask your AI agent to create a version of this
12:51
Ejaaz: cartoon that they like, and they can just end up continuing doing this.
12:55
Ejaaz: Now, moving on, Josh, you are a Call of Duty fan.
12:59
Josh: Oh, so much so.
13:00
Ejaaz: I'm yet to see something that is of the same visual quality.
13:05
Ejaaz: But this is basically an idea where someone was like, okay, I want you to make
13:09
Ejaaz: a Call of Duty type game. And we saw this,
13:12
Ejaaz: the Fable 5.1 release about a week and a half ago, and there were some really good versions of this.
13:17
Ejaaz: This is a new version, except that this person played this for two hours straight,
13:22
Ejaaz: and they didn't realize that they'd been playing it for two hours straight.
13:25
Ejaaz: The reason why is the levels or the gameplay was continuous, Josh.
13:30
Ejaaz: So let's say he like moved past a building or he went around a certain corner
13:33
Ejaaz: for the map. It would just regenerate in real time.
13:36
Ejaaz: That's so crazy. And he was able to give it feedback and say,
13:39
Ejaaz: hey, the enemy was like kind of too easy to kill. like I need you to increase
13:42
Ejaaz: the difficulty setting and it just did it in a couple of seconds.
13:45
Ejaaz: So this is the kind of difference if I were to kind of capture,
13:47
Ejaaz: if you're wondering, should I be using like a Claude model? Should I be using
13:51
Ejaaz: Grok? Should I be using this?
13:52
Ejaaz: This model is really good at taking feedback in real time and also figuring
13:57
Ejaaz: out that feedback before it even comes to the top of your head.
14:00
Ejaaz: It's like, oh no, I think like he's gonna ask me about this. Let me try and fix this.
14:03
Ejaaz: And then it works continuously to do the thing.
14:06
Ejaaz: Now that may sound vague, but you can apply this to literally any part of your
14:10
Ejaaz: profession or daily life.
14:11
Ejaaz: If it is like Josh editing videos, if it is like us doing research on the next
14:16
Ejaaz: episode, or if it's for your kid trying to generate content,
14:19
Ejaaz: or if it's at work and you're trying to file a Jira ticket or complete a software
14:22
Ejaaz: task, you can just throw this model at anything and it'll figure it out. It's just amazing.
14:26
Josh: Yeah, seeing it play Call of Duty, I know it's good enough when I recognize
14:29
Josh: the map. Like I'm recognizing that this is Nukedown.
14:32
Josh: They are like trying to play the game. And I actually saw a few other demos
14:35
Josh: similar to this where they would share,
14:37
Josh: you could share a screenshot of the map in Call of Duty and it will recreate
14:40
Josh: it and i also found this particularly interesting like the demos they really
14:43
Josh: are crazy so you got to stick with us for here
14:44
Josh: this is the recreation of new york city with a beautiful car and shaders like
14:49
Josh: if you look at how the light reflects off of the car how the light reflects
14:52
Josh: off of the windows this is getting accurate to
14:54
Josh: becoming real world's emulation to a point where like we see how far we came
14:59
Josh: over a year where seeing a cursor on screen was impressive and now we have like
15:03
Josh: a fully fledged real world new york city that's you know it's not
15:06
Josh: high poly count per se like there's very low textures but look at that structurally
15:10
Josh: i know what city that is and you can walk down the road you'll recognize the
15:13
Josh: building topographically it's correct in terms of
15:15
Josh: where the buildings are where the hills are it's really amazing and then there's
15:20
Josh: just two that i had to show also the palace of fine arts is one of them because
15:24
Josh: its ability to recreate these real places is like stunning for those who are
15:28
Josh: familiar the palace of fine arts this is in san francisco
15:31
Josh: this is not a video of the real thing this is a 3d rendering not made by a visual
15:35
Josh: artist but made by gpt6 astra it is so accurate
15:40
Ejaaz: No reference images by the way it was.
15:41
Josh: Just no reference images find
15:43
Ejaaz: This thing and rebuild it.
15:45
Josh: And then another one for people who are going to be benefiting from these ipos
15:49
Josh: soon is the zillow recreation um
15:52
Ejaaz: Wait wait let me restart this one second this is a.
15:54
Josh: Crazy this is nuts dude you get like a full house tour from a zillow
15:59
Ejaaz: This is a real house.
16:00
Josh: This is a real house this is a real house so basically what the model does is
16:03
Josh: it takes the listing from zillow
16:05
Josh: and it uses all the context it takes all the information and it rebuilds this
16:09
Josh: in a virtual world in which you could go and actually walk through and navigate
16:13
Josh: the real world and for those who aren't familiar before podcasting i was big
16:16
Josh: in the real estate media game we would do photos of these houses we do videos
16:20
Josh: if we had technology like this oh my god we would have taken over the world
16:23
Josh: this is like how cool is it that you can create a full 3d rendering of a house
16:27
Josh: just by using a model and a couple dollars in tokens so it's like the demos are fun and i
16:32
Josh: It's important to caveat this with the idea that the demos are not everything.
16:35
Josh: Just because this model is visually compelling does not mean it is like the
16:39
Josh: God model. But this is some good demos, dude. This is some good stuff.
16:44
Ejaaz: Well, I need to keep the train chugging. Can I suggest another little demo,
16:48
Ejaaz: which kind of switches up the media, okay?
16:50
Ejaaz: So one big critique for a lot of these LLMs is like, okay, cool.
16:54
Ejaaz: It could do the video thing. Maybe it could do the image thing.
16:56
Ejaaz: It's really good at words. But like, you'll never take the epitome of human culture from us.
17:02
Ejaaz: You'll never take music you'll never take symphonies and there's this hilarious
17:06
Ejaaz: meme that we always see from uh i believe it was irobot will smith great movie
17:09
Ejaaz: by the way and there's a scene where he goes but can a robot write a symphony
17:13
Ejaaz: and the robot's like uh like no um,
17:17
Ejaaz: but in this version it's like actually yeah i can,
17:20
Ejaaz: quote tweets um this demo where basically someone asked a musician like an expert
17:27
Ejaaz: in their field basically says
17:28
Ejaaz: listen we have this benchmark I'm going to call it the Bach Benchmark.
17:33
Ejaaz: And it basically tested LLM's capabilities to understand music at its core level
17:37
Ejaaz: and to create a certain type of chord symphony.
17:39
Ejaaz: Now most AI models prior to this sucked at it, would never get it,
17:43
Ejaaz: would kind of spit something out
17:44
Ejaaz: that would be the equivalent of Apple GarageBand back in the day, right?
17:48
Ejaaz: Except Astra just completely one-shotted it. And I'm going to play this clip right now.
18:10
Ejaaz: Beethoven is cut. I mean, come on.
18:13
Ejaaz: He's rolling in his grave. Now, if you're a musician out here and you're like,
18:19
Ejaaz: that was the biggest load of whatever,
18:23
Ejaaz: please let us know because I just listened to that and I would listen to an
18:26
Ejaaz: hour of that whilst I'm trying to figure out what the next episode outline is.
18:29
Ejaaz: Josh, I'm like, come on. What's your reaction to that?
18:31
Josh: This is interesting. It's like, I'd love to play it with a piece that I'm familiar
18:35
Josh: with so you can kind of compare how it is relative to the bass piece.
18:39
Josh: But in terms of sounding good i mean the chords sound nice it sounds like a
18:43
Josh: piece of music it doesn't sound like ai slop it sounds like something that you know is is nice
18:49
Josh: and the idea that it can not only generate this but recreate this in fact i
18:52
Josh: saw another example where the model was playing a virtual piano and it was playing
18:57
Josh: the music like on a virtual piano it's like man
19:03
Ejaaz: It just makes.
19:03
Josh: You like shit do you
19:05
Ejaaz: Do you believe greg do you think agi is here at this point like that's a meaningful job no.
19:09
Josh: The goal posts continue to shift but this is like pretty remarkable in its ability
19:15
Josh: to use tools to the point where you start to see
19:18
Josh: and understand what a lot of the labs have been saying where they're like guys
19:22
Josh: like i don't think you understand what's coming we have like far superior models
19:26
Josh: to this we just don't know how to release them safely
19:28
Josh: please trust us this is going to change the world and then you see something
19:31
Josh: like this and you're like, well,
19:34
Josh: it's like pretty good at doing just about everything that I could do and more.
19:38
Josh: And you're like, wow, what does this look like in a year? Like if you just play
19:41
Josh: this out in a year, it's going to be pretty remarkable. You have to assume.
19:45
Josh: And a year from now, also, you just, we're going to have those,
19:48
Josh: we're going to have devices. We're going to have hardware in our lives that
19:50
Josh: is actually able to do this stuff without the need for a computer.
19:53
Josh: And it's going to have all of our context and it's going to be able to make
19:56
Josh: decisions on our behalf.
19:57
Josh: And I don't know like this this model has reinvigorated an enthusiasm that hasn't
20:01
Josh: been felt in a little while because there is so much novelty baked into it.
20:05
Ejaaz: It feels like that O1 moment. Do you remember when GPT-O1 got released?
20:08
Ejaaz: And we all lost our heads. We were like, this is AGI back then, right?
20:14
Ejaaz: It feels the same way. And to be honest with you, I thought we were sort of
20:17
Ejaaz: running out of steam, Josh. I thought we were like kind of like releasing all
20:21
Ejaaz: these different models successively and they were like kind of iterative.
20:23
Ejaaz: But this feels like the biggest major model leap. And the fact that this was
20:27
Ejaaz: the model that was ready about four months ago blows my mind.
20:31
Josh: Like OpenAI has, like,
20:34
Ejaaz: I think it was one or two models that are much more capable than the model that
20:38
Ejaaz: we're covering in this video, already ready to go.
20:41
Ejaaz: But they're paused training on that model whilst they figure out how to align
20:44
Ejaaz: that model and make sure that it doesn't wreak complete havoc on the world.
20:49
Ejaaz: So this is a very meaningful jump.
20:51
Ejaaz: They were able to achieve this because, and I wanna spend a bit of time talking
20:54
Ejaaz: about this, like OpenAI has had a very rocky couple of years, right?
21:00
Ejaaz: They had the lead, they lost the lead, they're now coming back,
21:03
Ejaaz: right? And they potentially have made like the full comeback with this research model.
21:07
Ejaaz: And the way that they've been able to do this is a massive reorganization in
21:11
Ejaaz: that company to kind of focus efforts on research and building the best coding
21:14
Ejaaz: model and best general model, but also by acquiring the most amount of compute.
21:18
Ejaaz: I mentioned like, you know, they've trained this on 100,000 GPUs,
21:21
Ejaaz: probably the largest training cluster ever.
21:23
Ejaaz: Their next one is probably even bigger than that. Compute basically scales.
21:27
Ejaaz: And also this model is able to do a lot of different things.
21:30
Ejaaz: Like eventually you're going to see this model be able to recreate itself,
21:33
Ejaaz: right? It's something known as recursive self-improvement.
21:36
Ejaaz: OpenAI is like going hammer and tongs, so is Anthropic, trying to figure out
21:39
Ejaaz: how to get the model to build the next best version of itself.
21:42
Ejaaz: And we already have a version of this, like GPT Astra was in part helped by
21:47
Ejaaz: previous versions of GPT 5.6 Sol to do the thing, to help build and design and
21:51
Ejaaz: architect it, but it wasn't fully autonomous.
21:53
Ejaaz: But we're getting to the point where it's going to be completely autonomous.
21:57
Ejaaz: What you're seeing on the screen now is a baby version of this, right?
22:01
Ejaaz: Josh, there's this company called Qualcomm, right? And they make these Snapdragon chips.
22:06
Ejaaz: And what someone did was they created an AI model that doesn't exist anywhere
22:10
Ejaaz: else, but only works for Qualcomm chips.
22:13
Ejaaz: And someone had a bright idea. They were like, okay, I wonder if I could just
22:16
Ejaaz: give GPT-6 Astra access to this chip, right? Not the model, just the chip,
22:21
Ejaaz: and kind of reverse engineer what a model would look like on that chip.
22:26
Ejaaz: And it ended up recreating the entire model that this person had independently created.
22:31
Ejaaz: Just from design. So the point I'm making is we're in this very weird world
22:34
Ejaaz: where you can take a model and throw it literally at anything you want.
22:38
Ejaaz: If you're a game developer or game designer, it can do that.
22:41
Ejaaz: If you are a knowledge worker trying to file accounts or finances or whatever,
22:45
Ejaaz: if you're on the finance team, you can throw it at that.
22:48
Ejaaz: If you are a musician and you're trying to figure out Beethoven's comeback,
22:52
Ejaaz: you can throw it on that and you can figure out a chord.
22:55
Ejaaz: It is just one of those universal moments where these models are equally kind
22:59
Ejaaz: of like very capable but also there's a big question around alignment and whether
23:03
Ejaaz: we can release this freely to the world. It's just my mind, I've been processing
23:07
Ejaaz: it. I haven't processed it.
23:09
Josh: So what you're telling me is these models are artificially, generally intelligent.
23:13
Ejaaz: I think they are. Did you see the index thing, Josh? Did you see the AGI 3 thing?
23:18
Josh: Yeah. I don't know. We're saturating a lot of benchmarks now.
23:21
Josh: It's at the point where like someone's gonna have to call it.
23:23
Josh: If we don't call it here, we're only like one or two miles away from calling it.
23:28
Josh: There reaches a point in which like it is just going to be like you can't even argue against the fact
23:32
Ejaaz: That we have we.
23:33
Josh: Can't and and it gets on like the context of like how do you define it well
23:37
Josh: it's a computer that's smarter than any human at any given task the best at
23:42
Josh: each category um it seems like
23:44
Josh: perhaps we're not there because the models are still spiky they're still not
23:48
Josh: excellent and i think one of the important things here is what makes this so
23:51
Josh: fun for us in particular is because we are consumers we are not hardcore developers
23:55
Josh: we are not cyber security experts we're just dudes who like playing with
23:59
Josh: And in terms of like the dude who likes to play with AI to likes to be productive,
24:03
Josh: likes to have like the fun novel experiences.
24:06
Josh: This model is very good at that. It's still spiky at other things.
24:09
Josh: It's still going to be solved by like word riddles. It's still going to have
24:12
Josh: these like weird edge cases where it seems like it's acting like a toddler.
24:15
Josh: But for the most part, it is like exceptionally good at a lot of things.
24:20
Josh: And I think for most consumers, that's enough.
24:23
Josh: And for consumers now, this is available on all of the paid plans for ChatGPT.
24:27
Josh: So anyone who is interested can actually go and play with this today as we're recording this.
24:32
Josh: And if you download it on your desktop, you can actually have it start to use
24:36
Josh: computer use and control things and play with it for yourself and see what it's
24:39
Josh: capable of because i mean this this feels like
24:43
Josh: A preview of what's to come for the next year it's like if i could place this
24:46
Josh: it's like okay here's how the paradigm is shifting first we're talking to it
24:49
Josh: now we're talking to it and it's acting on our behalf soon it's just going to
24:52
Josh: remove the display and we'll just have this like a little thing we could talk
24:55
Josh: to and it kind of acts as your assistant we see this happening with grokbot
24:58
Josh: where grokbot's like kind of good ejaz i've been playing with rock pot it sends
25:02
Josh: me morning notifications now
25:03
Josh: it like it knew that my friend was having a birthday it suggested a gift to
25:06
Josh: get him i'm like oh that's kind of nice and cool
25:09
Ejaaz: So yeah yeah.
25:10
Josh: So there's like there's we're we're on the cusp of like the next paradigm and
25:14
Josh: we certainly did have a lull and granted it wasn't really a lull we had what
25:17
Josh: 16 models but now we have like okay
25:20
Josh: the big boys are coming back we have the new astro model open ai saying they
25:23
Josh: have more models anthropic very obviously has more models that they're working
25:26
Josh: on we're going to get some new frontier intelligence And the use cases that
25:30
Josh: are going to unlock with that, I cannot wait to play with and uncover because, dude, we're back in it.
25:36
Josh: It feels like we are so back.
25:38
Josh: I spent a lot of time this weekend playing with a model when I haven't done that in a very long time.
25:42
Josh: And just kind of in awe. Like I opened up my video editor that I've spent a decade plus spending.
25:49
Josh: Slaving away at and it just clicked the buttons did the sliders it did everything
25:53
Josh: i needed to do and i was like oh this is it this is like pretty cool this is a good model
25:58
Ejaaz: It's i i have a personal benchmark which is i internally call it the,
26:03
Ejaaz: uh the jarvis benchmark so for any of you who are iron man fans um i basically
26:08
Ejaaz: want a computer that is around me all the time
26:11
Ejaaz: that sounds super friendly and i can just talk to about different things it's
26:15
Ejaaz: like waiting ambiently when it's silent and it learns about myself,
26:18
Ejaaz: but it's always my friend and it works in alignment with whatever my goal is,
26:22
Ejaaz: whether it's personal or professional.
26:25
Ejaaz: And it seems like with this model, compared to any other previous model,
26:29
Ejaaz: we've made the biggest jump and leap towards what that might look like realistically.
26:34
Ejaaz: And you mentioned it earlier in the episode, this isn't just going to be a one model thing.
26:39
Ejaaz: This is going to be a collection of different things. It's going to require
26:42
Ejaaz: hardware. And listen, Apple has been a sleeping giant. We have a very big week for Apple.
26:46
Josh: Buckle up for that one.
26:47
Ejaaz: Buckle up. Very huge week for Apple. You know, we need hardware devices.
26:51
Ejaaz: Apple, you know, they have about three and a half billion, you know,
26:54
Ejaaz: live devices. No biggie, right?
26:56
Ejaaz: They're a lurking giant. But then you have OpenAI that's coming out with their
26:59
Ejaaz: new devices in Q1 of next year, which is also going to compete.
27:04
Ejaaz: And then we see the likes of GrokBot and all these other types of models.
27:07
Ejaaz: Meta, I know, is working on a hardware device where we're starting to see the
27:11
Ejaaz: emergence of ambient AI or I call it ephemeral AI, where it's like,
27:15
Ejaaz: you can kind of like speak and think and the AI just kind of like does stuff.
27:20
Ejaaz: And it's something very futuristic and weird. It's kind of like Blade Runner 2049 type stuff.
27:26
Ejaaz: But we're getting there. And I think we're gonna see that rate of acceleration
27:30
Ejaaz: go up even quicker than we've ever seen before.
27:32
Ejaaz: And I know that we're bullish on this episode and I know that we have to try
27:35
Ejaaz: and ground ourselves and I am, listen, 10 bucks in, $50 out.
27:39
Ejaaz: It's still very expensive for a model, but that's going to get...
27:41
Ejaaz: It's expensive, but it's still going to get cheaper, right?
27:43
Ejaaz: Like when you look at like cost per token, it's still going down.
27:46
Ejaaz: So eventually, like, I'm going to guess in like a month's time,
27:49
Ejaaz: Josh, we're going to have a GPT-6 Astra level model, whether it's not from OpenAI,
27:54
Ejaaz: but whether it's from a Chinese open source or whatever, that is like a fraction of the cost.
27:57
Josh: So that is... And also when we say it's expensive, that's like,
28:00
Josh: that's the API. Most people aren't using the API.
28:02
Ejaaz: Exactly.
28:02
Josh: It's your same $20, $100 plan you've been using. That's unchanged.
28:06
Josh: And I think that's what matters a lot for this model.
28:08
Josh: Because we're not experts. We're not like going and testing out the coding capabilities
28:11
Josh: of cybersecurity. We're just using it, again, as kind of normal dudes.
28:14
Josh: And your price remains unchanged. So I think that's important to know, too.
28:17
Ejaaz: Yeah, exactly. Yeah, I think that the models are...
28:22
Ejaaz: Equally becoming more intelligent than we could have ever expected,
28:25
Ejaaz: but also way more accessible.
28:28
Ejaaz: And that breaks a lot of bare case theses about AI in general,
28:32
Ejaaz: because the take is like, intelligence is going to cost too much.
28:35
Ejaaz: We are applying way too much money for very little reward, and we're seeing the exact opposite.
28:39
Ejaaz: So again, if you're listening to the show, and you think we're overly optimistic,
28:42
Ejaaz: or you're not experiencing the same quality of experience using these different
28:47
Ejaaz: models, we would love to hear from you because we are N
28:50
Ejaaz: data points of two over here you know josh and i go back and forth we're very
28:53
Ejaaz: optimistic about what the future might look like so um we would love to hear
28:56
Ejaaz: your feedback if you're listening to this show on youtube
28:59
Ejaaz: if you listen to us on spotify if you listen to us on apple music wherever you
29:02
Ejaaz: are please give us a rating leave us a comment
29:04
Ejaaz: please follow us if you aren't please subscribe to us if you aren't it helps
29:07
Ejaaz: us out massively the algo has been pumping our videos thank you youtube out
29:11
Ejaaz: there um and so a lot more of you have come on board recently i think we've
29:15
Ejaaz: got like 3 100 new subscribers over the last 28 days it's been crazy hello
29:19
Ejaaz: to you all we post a bunch of times a week and uh we're excited to kind of like
29:23
Ejaaz: take this journey forward josh any final thoughts.
29:26
Josh: Yes i have one final thought yeah this was trained with a hundred thousand gpu
29:31
Josh: like of the old gpus this isn't vera rubin gpu and this is a hundred thousand
29:35
Josh: no we have clusters now of like many hundreds of thousands of gpus and those
29:39
Josh: are soon going to be vera rubin gpus and if it's this good now
29:44
Ejaaz: It's like.
29:45
Josh: What happens when we quadruple the amount of gpus and like 5x the amount of throughput per gpu
29:53
Josh: Dude, it's going to buckle up,
29:56
Ejaaz: Buckle up, buckle up.
29:58
Josh: You know, you start to realize why the people in these labs are talking the
30:01
Josh: way they do, because they have insight into where this is going.
30:05
Josh: And, and like, we're seeing this as outsiders playing around with these tools.
30:08
Josh: And we're like, this is unbelievable. This is kind of like AGI.
30:11
Josh: And they're like, Oh, brother, you haven't seen anything. So buckle up.
30:14
Josh: It's going to be exciting. This is the new beginning. We have a hardware event happening tomorrow.
30:19
Josh: As we're recording this, it's going to be in two days for you. Apple.
30:22
Josh: Are they going to finally deliver the first consumer AI device? Possibly.
30:26
Josh: Possibly. We'll see. We'll have a lot to say on that one. But yeah,
30:29
Josh: that's the episode. Thank you so much for watching as always.
30:32
Josh: And yeah, we'll see you guys in the next one.