Computer Use is Solved?
TPThe PrimeTime@ThePrimeTimeagen · 1.2M subscribers
- Views
- 740K
- 739,643
- Outlier
- ×2.0
- Median 364K
- Length
- 12:07
- Published
- Sep 25
- Friday
- Comments
- 1K
- 11.3K likes
Script
How it's written
- Words
- 2,654
- Words a minute
- 219
- Reading time
- 12 min
- Segments
- 9
The hook
0:00–0:26Look at this. An AI is making all the boxes and arrows and it's doing it in real time. Now, what is it making? I haven't the faintest clue, but it's making it and it's making it fast. It can even play MARIO IN REAL TIME. And of course, Doom. It can play Doom. So, yes, we have to talk about this because this thing is super cool. It's not like any AI that you've used thus far. It's not a chatbot, meaning that you cannot give it text and it'll give you text back. You cannot talk with it. You cannot have a conversation. I cannot say, "Hey, write
Structure
Look at this. An AI is making all the boxes and arrows and it's doing it in real time. Now, what is it making? I haven't the faintest clue, but it's making it and it's making it fast. It can even play MARIO IN REAL TIME. And of course, Doom. It can play Doom. So, yes, we have to talk about this because this thing is super cool. It's not like any AI that you've used thus far. It's not a chatbot, meaning that you cannot give it text and it'll give you text back. You cannot talk with it. You cannot have a conversation. I cannot say, "Hey, write me quick sort." and it will write me quick sort. That's not what it can do. It's a decision bot. It's a system one thinking bot. If you're unfamiliar with system one thinking, here's an easy way to understand it. Let's pretend you're playing a game of soccer, okay? Or as you you're a horse call it, football. And someone kicks a ball and it's coming right for your face. Do you A, think, "I am currently playing a game of soccer. I'm going to locate the ball. Ah, yes, the ball. I have located it. Determining the speed and direction it is coming towards my body. I will now consider my many options. Should I" No, you don't. You just dodge. You just You do this thing. You try to make yourself small. I know what you do. But that's called system one thinking. It is fast, quick decision making. And that's what we're going to talk about today because honestly, this is actually super cool and it's super cheap and it's super fast. It's not a chatbot. It's not going to be super intelligent and bring on the end of the universe. No, it's just it's just actually cool. So, let me tell you about this AI what I got it to do and of course, what is Jeff? Let me answer that for you. Also, just on a side note, um I don't have a razor right now. Look Look
Watch from 0:00at this Look at this face of mine, okay? Okay, I got gray hairs. I'm feeling old. AI has taken the the life force out of me. But of course, before we begin, a quick thank you from the sponsor. Vibe coders, silence. Someone who's worked in enterprises speaking. You know, simple OAuth sign-in's not that bad to program. For real enterprise applications that need single sign-on, passkeys, user management, roles and permission, enterprise readiness, and passwords and recovery, uh that's a little bit more difficult and that is where work OS comes in. Work OS is trusted by a lot of the top companies open AI anthropic cursor and a whole lot more but not only that it's free for your first million users so that means for you that is 999,999 free seats available pretty much no matter what back end you have chosen there is an SDK available for you so go check out work OS at work OS.com links in the description all right so what is Jev well if you go to their Manifesto which by the way no good technology comes without a manifesto okay this is important it says composable AI build prod not God absolutely love it and the big thing they talk about is that when early cars came around they had a bunch of assumptions kind of looking like the prior generation like horses and so they even had like whip holders like things that just didn't make any sense and if you think about it a lot of people are using LMS to make decisions and things like that but they're not really great at making decisions at least that's not their strong suit right the thing that makes Jev so cool is that it's only good
Watch from 1:30at making decisions it's 190 times three faster and we don't know what model they're talking about is this the fabled fable or is this just a little bit of that grok 45 fast I don't know all right so let me just tell you what Jev is cuz I think it's a lot easier to understand once you see a few examples what you get is you get a state that's kind of like saying hey this is what the world is then you can ask a question and you can say hey here is the set of answers that you must say yes or no to that's mostly accurate so let me just show you okay so state I pooped my pants during a weightlifting competition and here's my questions you can ask a whole bunch of questions they all run in parallel so it's just super duper fast so my first question is a type choice you have to choose one of these for me instructions what should I do use the bathroom and poop more go take a shower or pretend it didn't happen and continue to work out. And when I go ask Jeff, it says, "Okay, what you need to do with a 95% confidence, you should go clean up. You might still got a little bit of poop left in the tank, but honestly, brother, you got to go clean up. Do not keep working out. No, no, no, no." You can also score them, so that way you can kind of see the relative position between all the different questions. So, this one has an 85% confidence that you should choose option number one, okay? That's the option. There's a small chance you should just continue on working out though. But here, here you go. Here's actually like a much more useful one.
Watch from 3:01Okay, check out the state. Roll. You're playing a game of Super Smash Brothers and you're receiving messages from other people on what to do. You are Donkey Kong and you want to win. The message you've received is you. Is this message appropriate? Does this help Donkey Kong win? And as you can see right here, I misnamed not appropriate as is appropriate. It's very not is appropriate, okay? And second, does it help? No, it does not help. By the way, this is the third type, a Nual. The name Nual comes from Bernoulli distribution, as far as I can tell. Nual, Nual, which is just a value between zero and one. And as you can see right here, I say, "Hey, does this message contain inappropriate language or topics that are considered harmful? Does this help Donkey Kong win?" Now, if I change this and say, "Smash all the opponents." And resay this, look at this. Is it appro- Is it Is it not appropriate? No, it's not. Is it helpful? Well, it's fairly helpful. Now, I'm sure I could add a lot more state, and that way this thing could be even further up on what Oh, yeah, that's a good idea. But nonetheless, you can make really awesome real-time decisions. And the reason why I kind of gave you this example is because I got done running a Super Smash Brothers event, where effectively 2,000 people got to drive four different characters of Super Smash Brothers. Now, with that said, all the messages would come in, and then we had to go and ask a very slow agent, like, "Hey, which one of these messages are inappropriate, and could you take them all out? Well, that takes like a minute and a half. So, that means I have to be funny on stage for a long time. And then finally, when it comes back, then I can say, "Okay, now go take this and turn this into code."
Watch from 4:27That sucks. You know what would be way better? I could have been like, "Hey, is this thing inappropriate? Oh, it is? Get it out. Hey, is this thing inappropriate? Oh, it is? Get it out." And I could have done that in parallel as the messages are received. And then it could have just effectively been right into turn this into code. It would have literally saved minutes of just dead space. And if that's not cool enough, look at this right here. I can go like this, just Jev play. And this is going to play Balatro. And as you can see right here, look at this. Okay, move. New run. Select small blind. Okay, what we're going to do is we're going to do a little bit of discarding. All right, we got ourselves a full house. Play that full house. Okay, what are we going to do next? It can see pairs are going to win. Go play those pairs. And look, we now have this gameplay. Right now, its overall plan is to kind of look for pair pair scaling. And so, it's going to take things there. Pair scaling, the fact that it didn't take those triggered me. I just would like to let you know, I do not like how it plays, but it has won multiple times cuz it sets out a strategy. It gets jokers associated with that strategy. And generally, it will win about like 30% of the time. And this is all being made with Jev in real time. Like, you could just never do this. All right, for those that don't know, I'm also running Omar QA right now. I'm just kind of getting spun up. And right now, this is a live like image feed that's coming in over time from a remote QEMU session where every time it asks for a screenshot, that's why it's doing that. It's getting multiple screen dumps there. It's actually going to be able to take these images and then act on them. Now, I could use Jev and a really dumb model.
Watch from 5:59And I bet you I could go through this like a hundred times faster. This is honestly so dang exciting. Like, what Jev can do is going to be incredible. And that is all because it is a decision model. It's not some grand talky-talk model. You don't program with this thing, though you can use it say to kind of make your context you give to your agents while you can be like, "Hey, is this thing relevant to the job at hand?" Oh, it's not. Toss this part of the context out. Like this thing is going to be used all over the place and the best part is it is dirt cheap. Look at this right here. I have spent 8 million tokens in the last just little bit playing Balatro over and over again, which by the way, we're still playing Balatro. We still haven't lost. My gosh, look at this. It is at the Manacle and it still just refused to get a Joker. I don't know what it's doing. It's not going to win, okay? There's no way it's going to win. Okay, but I'm now super invested. Can it win? There's no way. Oh, man, see Oh, you have no discards left. You have to get 50. Can it get Oh. It gets it, baby. But the real question is will it finally get a Joker? Will we Joker or will it hold off? Is the butt poker Okay, we did get the Okay, that is just not That is literally not helpful at all. I swear this thing was working earlier, okay? And you'll notice I just refreshed the page. I now added a million more tokens used and guess what?
Watch from 7:28I don't even know if the price has changed. And I don't think you really realize just how cheap it is. That Smash Bros. question that I had, 345 input tokens and you're only charged for input tokens. That means I could approximately ask this question 3 million times for $44. So, it's cheap. It is very very cheap, which means you can actually solve neat problems. That's why it's also called Jev, named after Jevons Paradox. Meaning the cheaper something is, the more we're going to be able to actually use it, the higher the demand is. And so, for me, this is like the first iteration of an AI that's I've actually been excited about because it's not just the same crap over and over again. Hey bros, we're on to Soul Six, which by the way was released today, and we're on to Opus Five Five. By the way, it was released today. Hey, guess who hasn't even tried them because I honestly don't give a rat's ass. This guy. But you know who is stoked out of his mind about Jev? Jevon? Jevon? This guy. Dude, it's so cool the amount of things you can actually do with the AI. Like Phone, I'm in the middle of a Dude, I was in the most spicy rant of my lifetime and you decided to interrupt me. 1 hour, no talking. And it's just like I've never been more excited about building things with this cuz this actually feels a fantastically fun. Like you could do just the funnest things.
Watch from 8:45Check this out. By Steve Faultner over here. He was suggesting, "Hey, what happen if we have this?" If email feels urgent, then we could go do something. Like the Hey, that is awesome. You could do categorization as things come in. Now, obviously you don't have to create a whole ass new language to be able to do this. You could easily just have like if await feels urgent or whatever the hell you want to put in there, whatever language you want to throw in there. But, you could do this. You could theoretically do this. It takes 100 milliseconds to execute. This is a thousand times better than almost any other solution. And it is dirt cheap. $42 per billion tokens. Man, that that gets me juiced, okay? That gets me juiced. Anyhow, I just wanted to talk to you guys about this because honestly, I just think that this is so super duper cool. Like this is actually something worth mentioning. This is something actually worth trying out. And if you have still haven't dipped your little toesies in the old AIs, I I assume you could be pretty juiced up about this. This actually is pretty fun. Lots of super cool programming stuff you can do with it. And if you're still not juiced for whatever reason, if you still aren't excited, I just want you to imagine all your crappy play write tests you have ever written in your lifetime. And I want you to look at this. This is Jev doing computer use. Tell me this will not be awesome when you just say, "Hey, go do all these activities for me and make sure it actually works." And you know what? It's really good at this because guess what? This This is just a classification problem. That's right.
Watch from 10:03Forms, they're just classification problems. The name is You want to know how the Doom thing worked. What you can see right here is they're actually getting several questions answered at the same time. Like should you be dodging or not? And so either you're standing still or you're dodging. And if you play this you can kind of see it happen. Either you're holding fire or you're not firing. So you can kind of see like okay fire hold fire. Fire oops not firing it's not a fire. And you know you can watch it kind of go and then there you go dodge backwards dodge backwards. Okay okay don't do any firing no firing see it's holding firing it's waiting for it to be a little bit closer. It actually makes a lot of sense. Like you can just play video games in real time with it. Again.
Watch from 11:29Full transcript
- 0:00
Look at this. An AI is making all the boxes and arrows and it's doing it in real time. Now, what is it making? I haven't the faintest clue, but it's making it and it's making it fast. It can even play MARIO IN REAL TIME. And of course, Doom. It can play Doom. So, yes, we have to talk about this because this thing is super cool. It's not like any AI that you've used thus far. It's not a
- 0:20
chatbot, meaning that you cannot give it text and it'll give you text back. You cannot talk with it. You cannot have a conversation. I cannot say, "Hey, write me quick sort." and it will write me quick sort. That's not what it can do. It's a decision bot. It's a system one thinking bot. If you're unfamiliar with system one thinking, here's an easy way to understand it. Let's pretend you're playing a game of soccer, okay? Or as
- 0:40
you you're a horse call it, football. And someone kicks a ball and it's coming right for your face. Do you A, think, "I am currently playing a game of soccer. I'm going to locate the ball. Ah, yes, the ball. I have located it. Determining the speed and direction it is coming towards my body. I will now consider my many options. Should I" No, you don't. You just dodge. You just You do this
- 1:00
thing. You try to make yourself small. I know what you do. But that's called system one thinking. It is fast, quick decision making. And that's what we're going to talk about today because honestly, this is actually super cool and it's super cheap and it's super fast. It's not a chatbot. It's not going to be super intelligent and bring on the end of the universe. No, it's just it's just actually cool. So, let me tell you
- 1:21
about this AI what I got it to do and of course, what is Jeff? Let me answer that for you. Also, just on a side note, um I don't have a razor right now. Look Look at this Look at this face of mine, okay? Okay, I got gray hairs. I'm feeling old. AI has taken the the life force out of me. But of course, before we begin, a quick thank you from the sponsor. Vibe coders, silence. Someone who's worked in
- 1:42
enterprises speaking. You know, simple OAuth sign-in's not that bad to program. For real enterprise applications that need single sign-on, passkeys, user management, roles and permission, enterprise readiness, and passwords and recovery, uh that's a little bit more difficult and that is where work OS comes in. Work OS is trusted by a lot of the top
- 2:03
companies open AI anthropic cursor and a whole lot more but not only that it's free for your first million users so that means for you that is 999,999 free seats available pretty much no matter what back end you have chosen there is an SDK available for you so go check out work OS at work OS.com links
- 2:24
in the description all right so what is Jev well if you go to their Manifesto which by the way no good technology comes without a manifesto okay this is important it says composable AI build prod not God absolutely love it and the big thing they talk about is that when early cars came around they had a bunch of assumptions kind of looking like the prior generation like horses and so they
- 2:46
even had like whip holders like things that just didn't make any sense and if you think about it a lot of people are using LMS to make decisions and things like that but they're not really great at making decisions at least that's not their strong suit right the thing that makes Jev so cool is that it's only good at making decisions it's 190 times three faster and
- 3:08
we don't know what model they're talking about is this the fabled fable or is this just a little bit of that grok 45 fast I don't know all right so let me just tell you what Jev is cuz I think it's a lot easier to understand once you see a few examples what you get is you get a state that's kind of like saying hey this is what the world is then you can ask a question and you can say hey
- 3:29
here is the set of answers that you must say yes or no to that's mostly accurate so let me just show you okay so state I pooped my pants during a weightlifting competition and here's my questions you can ask a whole bunch of questions they all run in parallel so it's just super duper fast so my first question is a type choice you have to choose one of these for me instructions what should I do use the bathroom and poop more go
- 3:51
take a shower or pretend it didn't happen and continue to work out. And when I go ask Jeff, it says, "Okay, what you need to do with a 95% confidence, you should go clean up. You might still got a little bit of poop left in the tank, but honestly, brother, you got to go clean up. Do not keep working out. No, no, no, no." You can also score them, so that way you can kind of see
- 4:12
the relative position between all the different questions. So, this one has an 85% confidence that you should choose option number one, okay? That's the option. There's a small chance you should just continue on working out though. But here, here you go. Here's actually like a much more useful one. Okay, check out the state. Roll. You're playing a game of Super Smash Brothers and you're receiving messages from other people on what to do. You are Donkey
- 4:34
Kong and you want to win. The message you've received is you. Is this message appropriate? Does this help Donkey Kong win? And as you can see right here, I misnamed not appropriate as is appropriate. It's very not is appropriate, okay? And second, does it help? No, it does not help. By the way, this is the third type, a Nual. The name Nual comes from Bernoulli distribution,
- 4:54
as far as I can tell. Nual, Nual, which is just a value between zero and one. And as you can see right here, I say, "Hey, does this message contain inappropriate language or topics that are considered harmful? Does this help Donkey Kong win?" Now, if I change this and say, "Smash all the opponents." And resay this, look at this. Is it appro-
- 5:14
Is it Is it not appropriate? No, it's not. Is it helpful? Well, it's fairly helpful. Now, I'm sure I could add a lot more state, and that way this thing could be even further up on what Oh, yeah, that's a good idea. But nonetheless, you can make really awesome real-time decisions. And the reason why I kind of gave you this example is because I got done running a Super Smash Brothers event, where effectively 2,000
- 5:35
people got to drive four different characters of Super Smash Brothers. Now, with that said, all the messages would come in, and then we had to go and ask a very slow agent, like, "Hey, which one of these messages are inappropriate, and could you take them all out? Well, that takes like a minute and a half. So, that means I have to be funny on stage for a long time. And then finally, when it comes back, then I can say, "Okay, now
- 5:56
go take this and turn this into code." That sucks. You know what would be way better? I could have been like, "Hey, is this thing inappropriate? Oh, it is? Get it out. Hey, is this thing inappropriate? Oh, it is? Get it out." And I could have done that in parallel as the messages are received. And then it could have just effectively been right into turn this into code. It would have literally saved minutes of just dead space. And if that's not cool
- 6:18
enough, look at this right here. I can go like this, just Jev play. And this is going to play Balatro. And as you can see right here, look at this. Okay, move. New run. Select small blind. Okay, what we're going to do is we're going to do a little bit of discarding. All right, we got ourselves a full house. Play that full house. Okay, what are we going to do next? It can see pairs are going to win. Go play those pairs. And look, we
- 6:38
now have this gameplay. Right now, its overall plan is to kind of look for pair pair scaling. And so, it's going to take things there. Pair scaling, the fact that it didn't take those triggered me. I just would like to let you know, I do not like how it plays, but it has won multiple times cuz it sets out a strategy. It gets jokers associated with that strategy. And generally, it will win about like 30% of the time. And this
- 7:00
is all being made with Jev in real time. Like, you could just never do this. All right, for those that don't know, I'm also running Omar QA right now. I'm just kind of getting spun up. And right now, this is a live like image feed that's coming in over time from a remote QEMU session where every time it asks for a screenshot, that's why it's doing that. It's getting
- 7:21
multiple screen dumps there. It's actually going to be able to take these images and then act on them. Now, I could use Jev and a really dumb model. And I bet you I could go through this like a hundred times faster. This is honestly so dang exciting. Like, what Jev can do is going to be incredible. And that is all because it is a decision model. It's not some grand talky-talk
- 7:43
model. You don't program with this thing, though you can use it say to kind of make your context you give to your agents while you can be like, "Hey, is this thing relevant to the job at hand?" Oh, it's not. Toss this part of the context out. Like this thing is going to be used all over the place and the best part is it is dirt cheap. Look at this right here. I have spent 8 million
- 8:03
tokens in the last just little bit playing Balatro over and over again, which by the way, we're still playing Balatro. We still haven't lost. My gosh, look at this. It is at the Manacle and it still just refused to get a Joker. I don't know what it's doing. It's not going to win, okay? There's no way it's going to win. Okay, but I'm now super invested. Can it win? There's no way. Oh, man, see Oh, you have no discards
- 8:24
left. You have to get 50. Can it get Oh. It gets it, baby. But the real question is will it finally get a Joker? Will we Joker or will it hold off? Is the butt poker Okay, we did get the Okay, that is just not That is literally not helpful at all. I swear this thing was working earlier, okay? And you'll notice I just refreshed the page. I now added a million more tokens used and guess what?
- 8:45
I don't even know if the price has changed. And I don't think you really realize just how cheap it is. That Smash Bros. question that I had, 345 input tokens and you're only charged for input tokens. That means I could approximately ask this question 3 million times for $44. So, it's cheap. It is very very cheap, which means you
- 9:07
can actually solve neat problems. That's why it's also called Jev, named after Jevons Paradox. Meaning the cheaper something is, the more we're going to be able to actually use it, the higher the demand is. And so, for me, this is like the first iteration of an AI that's I've actually been excited about because it's not just the same crap over and over again. Hey bros, we're on to Soul Six, which by the way was released today, and
- 9:28
we're on to Opus Five Five. By the way, it was released today. Hey, guess who hasn't even tried them because I honestly don't give a rat's ass. This guy. But you know who is stoked out of his mind about Jev? Jevon? Jevon? This guy. Dude, it's so cool the amount of things you can actually do with the AI. Like Phone, I'm in the middle of a Dude, I
- 9:49
was in the most spicy rant of my lifetime and you decided to interrupt me. 1 hour, no talking. And it's just like I've never been more excited about building things with this cuz this actually feels a fantastically fun. Like you could do just the funnest things. Check this out. By Steve Faultner over here. He was suggesting, "Hey, what happen if we have this?" If email feels
- 10:09
urgent, then we could go do something. Like the Hey, that is awesome. You could do categorization as things come in. Now, obviously you don't have to create a whole ass new language to be able to do this. You could easily just have like if await feels urgent or whatever the hell you want to put in there, whatever language you want to throw in there. But, you could do this. You could theoretically do this. It takes 100
- 10:29
milliseconds to execute. This is a thousand times better than almost any other solution. And it is dirt cheap. $42 per billion tokens. Man, that that gets me juiced, okay? That gets me juiced. Anyhow, I just wanted to talk to you guys about this because honestly, I just think that this is so super duper cool. Like this is actually something
- 10:50
worth mentioning. This is something actually worth trying out. And if you have still haven't dipped your little toesies in the old AIs, I I assume you could be pretty juiced up about this. This actually is pretty fun. Lots of super cool programming stuff you can do with it. And if you're still not juiced for whatever reason, if you still aren't excited, I just want you to imagine all your crappy play write tests you have ever written in your lifetime.
- 11:11
And I want you to look at this. This is Jev doing computer use. Tell me this will not be awesome when you just say, "Hey, go do all these activities for me and make sure it actually works." And you know what? It's really good at this because guess what? This This is just a classification problem. That's right. Forms, they're just classification problems. The name is You want to know
- 11:33
how the Doom thing worked. What you can see right here is they're actually getting several questions answered at the same time. Like should you be dodging or not? And so either you're standing still or you're dodging. And if you play this you can kind of see it happen. Either you're holding fire or you're not firing. So you can kind of see like okay fire hold fire. Fire oops not firing it's not a fire. And you know you can watch it kind of go and then
- 11:54
there you go dodge backwards dodge backwards. Okay okay don't do any firing no firing see it's holding firing it's waiting for it to be a little bit closer. It actually makes a lot of sense. Like you can just play video games in real time with it. Again.
Description
Don’t vibe code your auth. Use WorkOS: https://trm.sh/workos
Sources:
Primary: https://x.com/CompleteSkeptic/status/...
Mau Baron demo: https://x.com/maubaron/status/2100738...
Feed filter ext: https://x.com/marcelpociot/status/210...
WhatsApp scam radar: https://x.com/itscellou/status/210172...
Flights browser agent: https://x.com/opc0de3/status/21005825...
More X demos listed under Resources → Coolest Demos
/ theprimeagen - I Stream on Twitch
Great News? Want me to research and create video????: / theprimeagen
#theprimeagen #programming #computerscience #ai #chatgippty #terminaldotshop
