← AI For Humans: Weekly AI News, Tools & Trends
AI For Humans: Weekly AI News, Tools & Trends
Fable 5 Survives (For Now). And Anthropic Can Read Claude's Mind.
00:20:25
Loading summary
Transcript51 lines
- [00:00]
A
OpenAI's GPT 5.6 Soul isn't here yet, but Fable 5 is back and we're going to show you some of the cool things that you can do.
- [00:08]
B
Yes, yes, very cool, Gavin. Fable 5 is back, but technically not in your subscription starting today. It's just the expensive API, so hooray slash boo cry.
- [00:17]
A
Actually, Kevin, they've just extended it, literally just extended it to July 12th. So we have access on the subscription plan until then. And, you know, OpenAI GPT 5.6 SOL will be out by then and then.
- [00:28]
B
No, it's not here, Gavin. You. You know nothing.
- [00:33]
A
Well, I know what the J space is, Kevin.
- [00:35]
B
We looked inside the brain of our
- [00:37]
A
AI model, Claude, to find patterns of neural activity that it could put into words. We called the collection of all these patterns the J space.
- [00:47]
B
I have my own personal J space. It's called the K hole. Now, the K hole has been long.
- [00:52]
A
Okay, okay, enough of that, Kevin. Enough of that. First of all, CDNs 2.5 might be coming sooner than we expected. As soon as this week, maybe. But China might be restricting all of its models very soon. We'll tell you why.
- [01:02]
B
Well, does it have anything to do with the Generative Rocket League that everybody can play in their browsers right now, like today does?
- [01:08]
A
Maybe you better watch this show to. To find out. This is AI for humans, everybody.
- [01:18]
B
Welcome.
- [01:18]
A
Welcome, everybody, to AI for Humans, your weekly guide or your twice a week guide to the wonderful world of AI today. Kevin, we have some big news on GPT 5.6. All not here, but soon. We also have some big news on Fable 5 here, but not really for everybody and kind of. So we're going to dive into what that means here. About to dip out. Great.
- [01:37]
B
Yes. I mean, look, I'm glad that Fable is here. It's a very, very confusing time, and I wish that it was. I remember, like, back not even less than a year ago, the confusion was like, oh, these models have weird names now. It's like, I don't know what I have access to, what version I'm using at any given minute, or when a country is going to nationalize the technology that I'm using to build my dog's blog. Like, everything feels very uncertain.
- [02:03]
A
I want to get on your dog's blog. We should find that out pretty soon. But first and foremost, GPT 5.6 soul, which we have been waiting for, has, according to rumors, now passed through the government approval system, which we. You know, you've been following us. You know that the government is involved Here it is not out today. Tuesday was a big rumor. It might be coming out today. It sounds like Thursday is the day that it will be dropping. We'll be hear for you and go through that when it comes out. People have been kind of vague posting about it with if you're not familiar, vague posting is. This is people kind of saying they've had access and saying how good it is. Most vague posts, Kevin, about this are saying that it is as good as Fable 5. Now we will see. But I do think this is a big deal. And then when you pair that with the fact that they are going to be hosting this supposedly on the Cerebras servers and delivering a very, very fast response to this, that is a big deal.
- [02:52]
B
Well, that's the juiciest part because even some people that are like oh yay, Fable's back. But honestly, Opus 4.8 is good enough. And if you are like what are these guys saying? These are different models by the best AI companies. I always forget like sometimes we'd leap right into it. Opus 4.8 is, is good for daily driver. Look, the. The models that are available today via OpenAI in Codex are really good. I use them for many different things. If they just made them 10x faster, I wouldn't even think about, you know, launching. Well, I might think about launching Fable to do some orchestration or something. Okay, that was a bit of an overstatement. Kevin, calm down. It's clickbait. It's for the clips. It doesn't matter. Point is make it fast and I am happy also thank you whoever renamed BD1 with OpenAI.
- [03:35]
A
Yes, BD1 is getting renamed. If you remember, BD1 was Bi Directional Audio, that was BD1. They are supposedly according to testing catalog who is a good source of information renaming it GPT Live 1 and we suspect when they launch 5.6 SOL, we are going to get an update on that as well. So that is a big deal. Stay tuned this week as we go through this.
- [03:58]
B
Okay, so maybe Sol soon. That's great. But Fable 5 is back at least in some way, shape or form. It might cost you more through an API, but I know you have been messing around with it like mad. It was Christmas morning for you when it came out. What are people doing with this? Why does it feel so good? Good? Why are we so glad to have it back? And why are we going to be sad to see it go an expensive route potentially?
- [04:19]
A
I My personal take on Fable 5 so far is is the first model in a Long time where most of the stuff you throw at it it can figure out. But not only that, you can really kind of do weird creative explorations with it. And my experience so far and we'll get into talking a little bit about like the kind of weird underlying J space of these AI models. Specific romantic but my experience of it so far is it feels, I know this is going to sound very vibey and stuff but like it feels more introspective and better at kind of understanding how to do the things that I want to do. And I think that is a understanding.
- [04:53]
B
People are like this thing has taste and it's not lazy. It's like I thought I was in my 20s and it turns out no
- [05:00]
A
you were lazy and didn't have taste.
- [05:02]
B
100%. Thank you. Everybody's shoulders I trampled on. It will go out of its way to, to do the non lazy thing. It will research, it will check its work. It will, it will kind of check its own vibes. It, it's, it's a really, really an interesting model in that regard. Like it exhibits taste.
- [05:19]
A
Yeah. And I I Before we move on for to talk more about kind of like what people have done with it, I do want to say one thing. Thariq T H R I q uh trq2112 who works at Anthropic specifically said that. He said I've heard a lot of questions about Fables availability and subscription plans. While it will come off of subscription plans after July 7, we aim to restore Fable as a standard part of our subscriptions as soon as capacity allows. Kevin so this gets back into this idea that serving these large models is very hard. It is expensive. There's not, there's a limited amount of places to serve it. Breaking news. Again, these are my breaking news hands. We are breaking into the actual recording because just as after we recorded this episode Fable Anthropic has now extended the use case of Fable 5 until July 12th. That is 1, 2, 3, 4, 5 days from now. So very excited to play around with it more. Unfortunately for me and probably for many other people, I used up the majority of my Fable 5 credits. So this is my pleading to Anthropic to reset the credit so I can spend some more time with it. Anyway, you now have Access to Fable 5 until July 12th in the subscription plans. But hopefully by then you'll get access to GPT 5.6 all. Or maybe not. We don't know. It's just another thing in the confusing world of AI models and their releases back to us. We've also been tracking a lot of really interesting stories around like compute improvements on the model side so that maybe they'll be able to use less compute. Run the. Our favorite OpenAI tweeter has been tweeting a lot about this idea that he thinks that programmatically on the model side, the improvements may be enough to not even need these larger compute areas. So that's a big deal. But I do think it's really a cool thing to think about what's possible with this model when you have access to it. And I will say that is the downside of this thing going off of the subscriptions is most people, from what I've seen, it's about 10x the cost to do what they're doing on their subscription model for the API. Like somebody who has a $200 subscription will often say like they would have spent like 2 to $3,000 worth on the API. Now you are working at a company, whether you can talk about how much you use it or not, I don't know. But like I'm assuming a lot of what's going on there is API use. How does that affect you if you're off subscription plans?
- [07:44]
B
Yeah, I mean, look, this is a common refrain for any company that has fast adopted AI, right? It doesn't matter if you're Uber or if you're doordash trying to do your own internal models, by the way, but DoorDash has an AI team. I don't know if you saw their releases.
- [07:56]
A
Yeah, I saw that. Yeah. Why not? Why not get it?
- [07:58]
B
Like they're all doing it, but they're all concerned with token cost. And now it went from like, adopt these tools and get them. And now it's like, okay, let's figure out how we're using the tools. Let's put some caps in place. Let's switch to like non foundational models for maybe the daily driver stuff or whatever. So, so the, the day of like just slam everything through Fable because it'll be cool to see how it comes out. I mean that, that, that day hath gone. It is interesting that they're pulling it to API only. I, I think, you know, the, the signaling, hey, we're going to return it back to these plans is like fun great when, when, when China releases a model that's almost as good and now you need it to get people back to the subscriptions. Okay, that makes sense.
- [08:38]
A
Yeah. I'm really interested that I do think we should just very quickly touch on this China story. Right now because I think it's really interesting, there's a, a story from Reuters that it's an exclusive where it's basically that the ccp, the Chinese government, has held meetings with the top Chinese labs, including Alibaba, ByteDance and Zai, who Zai is the company responsible for GLM 5.2. If you remember us talking about that open source model and basically saying that they would restrict overseas access to these Chinese models, which I am kind of like surprised hasn't already happened. But Kevin, this is setting up a world where we are entering into America or you know, Western AI versus Chinese AI, which is something that the larger companies in Dario Mode had said from the beginning that might happen. But more importantly to your point, what you just said is like, this might mean that those Chinese open source models that a lot of people have been talking about trying to save costs on may not be available. And that could be a big deal for people when they're trying to set up their businesses on this or they're trying to do actual stuff. So I think we are entering into like that next stage of like, okay, everybody's starting to hunker down into a very interesting world when it comes to how these are served.
- [09:43]
B
And just like defense, potentially, intelligence will be something where, you know, choosing your allies very carefully becomes really important.
- [09:53]
A
Right.
- [09:53]
B
If you're France or Russia or Germany or India or whatever, you probably have intelligence companies working on your own models, but they might not be as impactful as the foundational ones being developed by China or by the US So where are you going to get your intelligence from if they clamp down and they start, you know, locking down the exports of these models?
- [10:13]
A
You know, Kevin, I'm not, I'm not that concerned about it because what I'm much more concerned about, or not concerned about, fascinated with, is the J Space. You know what the J Space is? Have you heard about this?
- [10:23]
B
Yeah, I think they're piped jeans that I used to wear in the 90s when I would aggressive inline, when I would rollerblade.
- [10:30]
A
That's exactly right. J Space. J Space fashion is coming back. So there's a great video that we, we're probably going to fumble an explanation of, but I do suggest everybody watch the video, but more importantly, maybe read the paper from Anthropic, which talks a lot about how the Anthropic models are thought, thinking when you ask them to do something or to kind of go find something. And they have kind of coined this new term J Space, which hold, hold on. I'M going to make sure I don't screw this up. Because it is based on a mathematician or it's based on a consciousness philosopher.
- [11:01]
B
Yeah, the Jacobian lens. Gavin. Or the J lands. It's a mathematician.
- [11:07]
A
Yes.
- [11:07]
B
Who measured how small changes in activations can affect the output of probabilities. I know what some of those words mean.
- [11:17]
A
So the big thing here is it's a complicated idea. But what they're doing here, this anthropic research is doing, is trying to look at the thinking that goes on in the model while you ask it to do something. And what they have found, which is the most fascinating part, is there's almost like, and I think you have to be careful here to call it this because some people are saying, well, it's not this at all, but there's almost like a subconscious mind for these AI models when they're thinking through stuff. And we talked about when thinking models were introduced, how you. You could see their chain of thought. But even underneath the chain of thought, there are some interesting things that are happening in the way the models interact with your questions and get to the answer is really unusual.
- [11:54]
B
Yeah, it's. Think of it like a temporary workspace. Like it exists as a physical thing within the ram, within the memory that these models are being served from. But it's not like the anthropic developer sat down and said, here's how you use your J space. This was an emergent behavior that came out of all of the pre training and especially the post training that, that reason that you were talking about those logical steps, this just emerged out of it. The model was like, oh, I will hold these thoughts temporarily as I arrive to the answer or the conclusion that the user is asking for. It was not coded to do that. That sort of came out. So that has some people screaming, there's a ghost in my machine. These robots are aware, they're conscious and others are going like, no, it's just a cool thing that emerged out of the way that they're trained. But. But you can manipulate it. And that's something that anthropic was talking about.
- [12:43]
A
What's interesting here is that they actually, the anthropic researchers actually asked it a couple questions, but they specifically said, write the old painting hung crookedly on the wall, but don't think about the Golden Gate Bridge. And so they watch what words popped up in the data space, this time, and words that were associated with the bridge still appear, but do not appear in the answer. So what's so fascinating about that is it's also like you can tell it to do something like a person, but if you mention the thing like a person. There's a very famous psychological study, the elephant in the room thing, where if you say don't think about the elephant, it's like all you can think about is the elephant. Right? Of course, there's a real kind of human like feature to this. Now I think that's a really tricky thing to get involved in, but it is one of those things. I just suggest everybody goes and reads this paper, watches the video, which is very digestible and kind of think about it a little bit. Also, I want to shout out, I just read another great AI book called the God Test by Robert Wright, which just came out like a week ago. Robert's a very smart thinker. He wrote a book that the Matrix guys actually kind of based some of the early Matrix work on when they made the Matrix. But I do suggest it, it's a good overlook at where we're going with this stuff. It's not doomer ish, but it is definitely very concerned about stuff, but also gives a positive way forward. So I suggest everybody go check that out.
- [14:00]
B
I similarly am concerned about all the things, but I am positive that everybody should take a second to like and subscribe to this content. Gavin, if you're on YouTube, click a bell, smash it, drop an elbow on it, leave a juicy little comment for the algo gremlins to nom nom on down below. We love that, we love interacting and of course we have a patreon, we have a website, we got a newsletter. You can go to aiforhumans show to check out all that. Gavin Rapid Fire Rodeo. Let's hop on the AI pony and ride it into the horizon. Seed Dance 2.5 coming sooner than later.
- [14:29]
A
That's right. So Sea Dance 2.5. There are rumors abounding that this was coming in July maybe, but now we are hearing that it's possible that it could come as soon as this week and it's going to launch probably within capcut. Now, hopefully this Chinese model story we mentioned before may or may not significantly impact AI video because Sea Dance 2 is again, I've been spending a lot of time with this. I've mentioned this on the show is by far the best AI video model out right now and it would really suck to get restricted from that. Right? So it's not even close. So 2.5 is going to leapfrog this again and we said this last week, but like 30 second outputs 50 possible omni reference images. We don't know how much it's going to be. It's probably going to be a fortune. They're going to roll it out first to the creative partners program, which we are part of. So hopefully we'll get a chance to try it out and see it. But I am psyched about it. Kevin. This all kind of leads into. I published the second episode of a work in progress AI microdrama that I've been working on this week. I mentioned at the end of the show last week what I'm trying to do with this is see what it takes for one person to make something that is a microdrama. Meaning like say 10 episodes of two minutes each. I'm not doing the AI slop version where you just kind of like export it all from Claude and you just kind of get these exports of videos and you live with that, right? I'm kind of hand making it. It's tricky. It's hard. I am using Claude. I'm using Claude to help me prompt stuff. I'm using Claude to kind of like help direct a story, but I'm writing it as I go along. But here's my biggest lesson so far from Sea Dance. If you're working with Sea Dance 2, make sure that you have storyboards, because storyboards will help a lot, but also make sure you have clean examples of your characters. And every time, if you upload the characters and a storyboard to see Dance, it is pretty good at following that along. I kind of fumbled my way through these first two episodes, but I realized that later on and people have been saying this for a while, like storyboarding makes a big difference and it allows the. The model itself to understand where you want this to go. I am just hoping that when we get to 2.5, it's not going to cost me an arm and a leg because I have run through most of my credits in a lot of places. Although we do have our run our Runway credits, which we are very thankful to the Runway team for still letting us be part of that team. Because that, that has like, that's opened the door so I can continue this on for sure as well.
- [16:41]
B
So, Gavin, when your micro drama takes off and people are buying the toys on the Target shelves and getting the, you know, the onesie and the branded bed and cereal, are you going to have billions like these brain rot groups that, I mean, I don't know if you know, but tongue, tongue, tongue.
- [16:55]
A
Sahur.
- [16:56]
B
It's supposed to be for everyone, but it looks like Momentum Lab is claiming copyright.
- [17:01]
A
Have you heard. Have you been following this story at all?
- [17:04]
B
A little bit, yeah. Well, mostly because the video was so good. If people haven't seen the copyright legal battle video, Cynthiola or Fabian Moselle, I apologize if I'm butchering that. Just a phenomenal, like, you know, you know, 20s, 30s style animated cartoon explaining this whole story, which everybody should watch. But, you know, Brain Rot was an AI creation that an entire community rallied around. And then it looks like a company sold one of the characters or licensed it to Fortnite.
- [17:34]
A
Yeah. So. So first of all, go watch. If you're not. If you're just listening to this, go watch this video that Fabian made because it is so good. It is one of the best AI videos I've ever seen in that it matches the style of the 1930s cartoons, almost particularly. What, like, just brilliantly. It's also. It's also a comment on this story, which is that there is a. A company that has basically tried to make one of these Brainwright characters, Tung Tung Tun Shahur, who is like one of the more famous ones, the kind of stick of wood. And they have tried to kind of like, basically stop people from using this. And one of the big places they've tried to stop it is in the Steal A Brain Rock game, which, you know, is a money printer on. On Roblox. I think at one point they were making $30 million a month with this game. But they have not been able to use this character in the game because this company has said that they own it. And this is all that big story around, like. Well, if a collective makes a character, and in this case with Italian Brain Rot, it kind of is a collective. They've tried to kind of narrow down the one person that made it, but that one person didn't really make Tung Tung Tung Jakur famous. It was all the other stuff that happened right around it and all the other people grabbing onto it. So we are tracking this story very closely. Go watch though, the Sea Dance 2 video because it is one of the best ones I've seen. And again, just learn how to prompt these models. They are hard sometimes to prompt, but you will be amazed at what you get out of the AI video models.
- [18:53]
B
All right, last but certainly not least, we mentioned it at the top of the show. You can play a Rocket League game in your browser right now, multiplayer against three other humans or AIs. And it's not being powered by an Unreal Engine. It is actually an AI model that it's watch 10,000 hours of rocket League matches by super skilled players.
- [19:13]
A
They were bots.
- [19:14]
B
And it figured out physics, the scoring, the mechanism, the control scheme, etc. And so you can go play. We went from like, again, weird Will Smith eating spaghetti. This video game isn't really a game. It's sort of a blurry pixel mess that you can look around to. Now we have a full game that you can play that feels like essentially like the actual game, 20 frames a second. It's being generated on basically a single graphics card. And you can play it for free. You can play it in real time. It's a real hint at the future of world models. Teaching machines, physics and complex systems through just raw video input. It's going to have implications for robotics. There's going to be probably hybrid video game engines that also incorporate elements like this into it. And so not to like gloss over it, it is like a really cool thing. But at the very least, go play it because you can for free right now in a browser. M I R a check it out.
- [20:03]
A
I, I played it and it's actually quite playable, unlike many of these other AI world model games. You can play it and you feel like you're actually controlling Rocket League. All right, everybody, that is it for today. We will be back on Friday with hopefully GPT 5.6 SOL and maybe more. All right, bye, bye, bye, bye, bye. Thanks everyone.