← AI For Humans: Weekly AI News, Tools & Trends
AI For Humans: Weekly AI News, Tools & Trends
Anthropic Caught Alibaba Spying. The AI Cold War Is Here.
00:26:50
Loading summary
Transcript103 lines
- [00:00]
A
American AI companies keep delaying their latest and greatest models.
- [00:04]
B
Yeah, and it's not just about stopping Skynet. No, this could be due to corporate espionage. It's actually really dramatic. Anthropic is pointing fingers at China and wagging them.
- [00:14]
A
It doesn't matter, Kevin, because I can't afford the memory to run anything. Apple and other companies just jacked their prices because of AI. Thanks a lot.
- [00:23]
B
Oh, no water and no money. More Importantly, Seed Audio 1.0 was just released. It is insane. You can generate songs, layered audio scenes. I spent some time with it and I think everyone listening to this will want to as well.
- [00:37]
A
Plus Claude just joined your slack thanks to a new tool called Tag and we have a very cool blender plus Sea Dance workflow that allows you to take very blocky setups and make really incredible controllable AI video.
- [00:50]
B
All of that and plenty more coming
- [00:52]
A
up on AI for Espionage. Wait, no. Humans. Humans. Oh, foreign. Welcome everybody to AI for Humans, your twice a week guide into the wonderful world of AI. I'm Gavin Purcell, that's Kevin Pereira. And Kevin, today we did not get Fable 5 as we know, as we don't know whether it could be happening now. We did not get GPT 5.6, but we did get Kevin is a story of corporate espionage at the AI level. There's a big breaking news story that Anthropic is saying that Chinese models, specifically Alibaba's Quinn is is still distilling. Claude, they're using CLOD to kind of train their models. So this is a big deal. Kevin, I'm curious to hear what your thoughts are on this because you did a little bit of deep dive reading on this. I'm just curious to know like, is this something that is going to continue for a bit? What do we feel like this goes?
- [01:46]
B
Well, this is a reason, one of the reasons that the Fable release was so guardrail. They wanted to stop actions like this from happening. And I believe OpenAI has dealt with this. Google has like others have. But this is a really, really fascinating. Gavin. From April 22 to June 5, a period of about 45 days, 25,000 fraudulent accounts were made supposedly by Alibaba to get around the banning of certain models in China. You know, like, like opus for example. 28.8 million exchanges or conversations were made with clot at that point. And basically what they're doing is they're poking at the model, you know, testing it with different inputs, seeing what the output is and as you said, distilling, trying to distill take Away the knowledge, the learning, the reasoning from the core model so they can make their model better.
- [02:39]
A
How many of those do you think were people asking about? Can you make a 64 animal tournament to fight each other? Do you think that was like at least 10% of those?
- [02:47]
B
There was at least three in there. There was a lot of like UWUs. How many Rs in Waifu?
- [02:53]
A
There was a lot of interesting. I do think we should define distillation too. Like make sure people understand what distillation is because it's actually quite meaningful in the AI space.
- [03:03]
B
In this case, it's asking the model questions, you know, feeding it a certain input and then analyzing the reasoning steps and the output that the model gives you and then using that as guidance for your own model that you're training. Because we know that there's like pre training where a lot of data goes into something, right? You, they grab all the artists imagery and all of the Reddit posts and all of the X rants and whatever and they slam it all in there and they pattern match. But then there's a lot of fine tuning and reasoning skills that come after. And with distillation they can really learn the way the model is performing its reasoning, its logic chains, et cetera. And if you copy and paste those, I am distilling right now, but if you copy and paste that into your model, you can improve its output. And that's why, you know, a lot of times we see these foundational models get released, particularly by American labs, and then just a few months later, oh, whoopsie doodles, it looks like something's coming out of another country that has capabilities. And some of that is because they're adversarially attacking the model, distilling it and learning from it.
- [04:01]
A
Yeah, and we talked about GLM 5.2 being very, it's a Chinese model, very close to, you know, Claude 4.5 outputs. And this is a version. Now we have to be clear here. Let me be clear. This is Anthropic saying this. This comes from Anthropic. And obviously, you know, this is a company saying this out loud. Anthropic does have a very specific, very clear Dario mode. The CEO Anthropic said multiple times that he believes China getting access to top line LLM models is a bad idea, that they should not be allowed to train on the most advanced chips, that because they are not a democracy, this is a bad thing. So I do want to make sure everybody understands that, like this is not coming from a narrator that is completely Unbiased. But this does feel like it could be leaning into, as you mentioned early on, why we're not seeing the Fable 5 reappear immediately or maybe even GPT 5.6 as well.
- [04:58]
B
Yeah. And the most shocking thing, Gavin, did you, did you read how they did it?
- [05:03]
A
No, how did they do it?
- [05:05]
B
NORDVPN use promo code.
- [05:07]
A
Really?
- [05:09]
B
Get yourself some NORDVPN if you need to get around your country's impact.
- [05:13]
A
They're not a sponsor.
- [05:14]
B
Restrictions.
- [05:14]
A
They're not a sponsor, Kevin. They're not sponsor.
- [05:16]
B
Just to be clear, I gave them, I gave them Internet comment Eric's promo code.
- [05:21]
A
Good for him. Good for him. Get it. Is that a real story before we move on?
- [05:25]
B
Is that true? No, no, no. But they did have to get around the geo blockers. They did make thousands of accounts. Like, it's a pretty dramatic finger point. And again, it's why they lock down Fable so much to try to stop this distillation. And that's why some people were mad because if you're doing high level research, Fable would just sort of downgrade its logic. It wouldn't reveal what it was doing and it wouldn't tell the user it was doing that. And they were doing that to try to stop China and probably other bad actors from distilling their model. It's crazy, but not as crazy as these ding dang MacBook prices. Have you seen this, Gavin?
- [05:58]
A
Yeah. I mean, listen, Kevin, this is, this is part of one large story, right? And the one large story is that AI has come for the world. And we have been talking about this for a while. And Kevin, I hate to tell everybody in the world of AI this, but this MacBook increase in price is AI's fault. And this is a kind of multi step process. But the big thing to understand here, if you're not deep in the heart of why what's going on with memory chips and all that other stuff is that the AI buildout, this giant data center buildout, which I think we are going to do a full data center episode soon to kind of get the truth around this. And the idea that everybody is racing to make these chips as fast as they can has meant that the memory chip makers are running into to shortages, like really significant shortages, and they are jacking the price. So you may have seen a story about a company called Micron whose stock is way up. Well, this is because of the memory shortage. But specifically Kevin, Apple, which I still think is the richest country company in the entire world, or at least up there now, because of SpaceX. Maybe it's not as big, but this company is saying they have to raise prices specifically right now on MacBooks and many other products between 2 to 200 bucks to $1,000 depending on the product.
- [07:14]
B
And the studio is up like 25% price increase for the base model. If you wanted a Steam machine to play all of your Steam games, well, you're spending over $1,000 now because the price of memory has been baked into that Xbox. Literally just two hours ago from our recording announce they are raising their prices across their entire line, the Series S and the X. And everybody's pointing the finger at memory. Full disclosure, your baby invested in Micron. So good for you.
- [07:41]
A
Sorry, full disclosure, your baby on this side did not get the tip from your baby on that side to invest in Micron at the same time.
- [07:48]
B
So I do have to say, I think, I think disclosure that I am invested in several memory companies. I did see this happening, but so on the one hand, oh darn, my portfolio is increased. On the other hand, I am also seriously, sincerely upset about these increases across the board. The memory shortages are like, I don't think they're going anywhere anytime soon. Like, I think even if this AI bubble were to quote, unquote, burst like everybody says is going to happen tomorrow, I think there's still enough momentum on data centers and the inventory has been bought up. I don't see this, this kink in the garden hose, so to speak. I don't see it getting resolved until the end of 2028 at the earliest.
- [08:29]
A
Yeah, and I think what's really interesting about this is that AI bubble conversation continues to churn and churn and churn and churn. Everybody again is now convinced that there's going to be a crash because we're building all these data centers, we're building all this capacity and whether or not people will have it. And I think we disagree with that in a lot of ways. But Kevin, one other thing that's really interesting here. OpenAI did announce something yesterday which is a big deal and I think kind of leans into the future of what this stuff is. They have created their own chip called the Jalapeno, which by the way, you know what? Great. I love a Jalapeno name. You know, let's land in the fruit world, like Nano Banana. Please, please, more of this. Here's the thing that's so fascinating about this. The Jalapeno chip was a nine month development process, which is crazy when you think about how fast that is. Super crazy. I Think the next big thing that's going to happen here is you're going to start seeing, we already know Google has TPU chips, which is its own AI chip. OpenAI has a chip. You're going to start seeing these companies make their own chips. And that is not great for like say, Nvidia as a company per se. Maybe it is, maybe it isn't. But they're all betting on this idea that if they can serve their own inference, if they can serve their own AI inference instead of having to rely on these other companies, that, that might spread it out a little bit more. Now, the hard part is I think you're right. We're in a couple year, you know, backlog here. So I don't see this getting better before it gets worse. And I think that this price raise from Apple is going to be a signal to the mainstream of the world, one more signal that AI has screwed them. And I am very worried right now at this date in June 25th that the AI narrative is going to degrade further. And I think that's just an important thing for everybody listens to our show to understand. Maybe don't say at the July 4th barbecue so loudly how great AI is, because I bet you the difference from this year and last year is going to be huge. But it is something very important to
- [10:22]
B
track because I did, I did have the green candlestick print. I had it screen printed on a T shirt to wear.
- [10:28]
A
Don't do it. No. You and your Microns stock, you got to be more quiet than anybody else. You keep yourself low volume. Just enjoy people's hot dogs, enjoy the potato salad.
- [10:38]
B
Quiet subtlety is something I excel at, as we all know. Gavin, listen. We all might complain, we all might whine. We want the foundational, we want the frontier. We all want memory to be cheap. We all want drinkable water. We all want artists to be able to live off art. Okay, but did you see GPT 5.5 instant?
- [10:55]
A
Yes. So there is a. Another update was GPT 5.5 instant, which I don't use, I'm sure you probably don't use very often. This is their free, free version. It is supposedly much better. Have you been spending any time with this at all? Have you, have you tried it?
- [11:07]
B
Nope. No. Next story. Me either.
- [11:10]
A
Hold on, hold on, hold on, hold on, hold on, hold on. Let's say one thing about this. So there is a new model in the free version of GPT 5.5.
- [11:16]
B
It's a big deal because a lot of people be like, I Use AI. In fact, I use chat GPT and what you are typically getting, millions of users are getting the most watered down basic simple serve model.
- [11:27]
A
Yes. And from what I've read, this might be a precursor to the BD launch. Bidirectional voice model launch. Because what they might end up doing is serving this bidirectional voice model to everybody, to the free users as well. And if that bidirectional voice model has crappy intelligence and isn't very good, it's going to be a crappy experience for everybody. Because Kevin, there's one more thing that's been out there. I've seen a couple people post about this. I've seen people get notices that say GPT ads are almost here. So imagine you would be getting a better version of the free model. You'd be getting a voice model where you might be talking to it all the time and then that voice model drops on you. Nor VPN available right now for what's. What's it.
- [12:08]
B
What's Big money Salvia.
- [12:10]
A
Big money Salvia.
- [12:11]
B
There you go.
- [12:13]
A
Anyway, so I think that's a, I think that's kind of a plan that's lurking right now. Is the monetization of ChatGPT free. That is coming very soon. Crazy.
- [12:23]
B
I mean, look, we, we knew it was in the works. It was only a matter of time. And these data centers aren't going to pay for themselves. And with the price of memory, maybe we got to suffer through some audio ads. But speaking of suffering through some audio. Gavin.
- [12:35]
A
Yeah. Let's hear this. I'm excited about this. Tell us what this is. What if.
- [12:38]
B
What if I wrote a prompt that was quote, an ice cream truck drives down a bustling city street while children run alongside, cheering, laughing, chanting, Hot dog City. Hot dog City. When suddenly a piano falls and crushes the truck and the music stops and the children scream while the piano plays a broken out of tune Mario theme. Do you think a model could one
- [13:01]
A
shot it at this point? I would say no, because that seems very complicated. But I bet you're about to surprise me. Oh no. What happened?
- [13:26]
B
The piano fell on the truck and now there's a sad out of tune Mario theme. Exactly as it was.
- [13:32]
A
That's amazing.
- [13:33]
B
Gavin.
- [13:35]
A
What are we listening to here? Kevin? What is that? That's an output from Seed. Is that what this is?
- [13:39]
B
That is just. That is a. I wear a. I wear an EMS cap when I sleep. And that is what. That is what I hear when I'm in. When I'm in rem. So that was one shot. A brand new model from our friends @bytedance. Hashtag, not an ad, but.
- [13:54]
A
Oh, seed is bytedance too. I didn't realize it was also bytedance. Oh, interesting.
- [13:58]
B
Yep. It's in the old seed families. It's audio seed audio 1.0. You know that. From the makers of Seed Dance, the best video model that we yap about incessantly here. This is a model that generates voice, it generates music, it generates sound effects. It could do it all in one shot. You can feed it three sources of audio and tag them and reference them. I had issues getting that to work, but suppose they work. It generates audio that's layered, so you can go back and edit the individual layers, individual speakers.
- [14:29]
A
Wow.
- [14:29]
B
Oh, wow.
- [14:30]
A
That's amazing.
- [14:31]
B
And I used it on foul, where some of that works well and some does not. But I was shocked. And so I went to our dear friend Uncle Claude and I said, give me. Give me the goods. Give me the tough stuff. Give me a prom.
- [14:44]
A
Yeah, sure.
- [14:44]
B
That I could give an audio model that would be a real test for it and eval if it could nail all the twists and turns in the sound effects and music, etc. Etc. And I. I gave you the prompt that I gave.
- [14:55]
A
Yeah, I see this. It's incredible. It's very long.
- [14:57]
B
And just listen to some of this.
- [15:00]
A
Okay.
- [15:03]
B
Houston, do you hear that?
- [15:12]
A
Hola. Wow. So just as this is playing, I'll keep. Just so people understand, it'll be. It's on the screen as it's playing here, too. But if you're just listening, everything that you're hearing, Kevin or Claude actually wrote out very specifically all of these changes. We're going from a train, and then we're going to go underwater here soon.
- [15:39]
B
Yeah. I mean, it's not. It doesn't hit it word for word, but when you see the sheer length of this prompt and all of the twists and turns, all of the sound effects, you realize like, yes, this is amazing. And as you start thinking about prompt to Hollywood, listen to this. Is anyone still listening? It says listen.
- [16:03]
A
That is crazy. And again, please look at the prompt on YouTube if you're not seeing this. Like, it really is going piece by piece here. I have a question about this, Kevin, because. Okay, I've been spending the last week or so trying to spend some time with Sea Dance 2 and really, like, talked about, like, figuring out the best way to prompt stuff. I've kind of made a little progress on some different video ideas. The thing that I often find with Sea Dance prompting is that yes, you can get audio that matches and yes, you can use, you can upload a reference clip and it often will kind of match the audio in the same place. What I want to be able to do is I want like a three minute clip of, of something like say I've cut together, I don't know, ten different sea dance clips. I've got the ones that I want is this. Are you able to upload a clip and have it create the audio for you yet? Is that part of this?
- [16:51]
B
It does have an ability to upload like an image and I let me double check on the video aspect of it. So if you go to Seed Audio co, that's their site for this thing, it does mention like short film sound design, marketing creatives, gaming prototypes, et cetera, et cetera. And it mentions that you can upload imagery with it.
- [17:12]
A
Imagery is fine. I think imagery is trying to get it to give a reference to like what it might sound like. I think the better, more interesting thing would be for me for right now, and by the way, this is an advanced thing is like if, if I had a minute long clip or a minute and a half long clip and it could just essentially create the audio for that clip based on prompts I gave it. That feels like a major unlock. But this is not that far away from it.
- [17:34]
B
I was going to say you could do multiple minutes with it. And if you are, you know, generating the video scenes, I would tell your agent, now, give me a moment by moment, moment or second by second, frame by frame breakdown of the audio. And you could.
- [17:47]
A
And then use that, just heard like
- [17:48]
B
how rich that tapestry was. And again, you can feed it three different sources of audio and tag those sources and it will incorporate them. So if you have a character's voice or particular background noise or ambient audio from where you shot a scene and you want to redo the audio, I am genuinely blown away by this thing's ability to make music, to make the voices, to make everything. And this just sort of just came out of nowhere here.
- [18:12]
A
It was, how long, how long are the outputs? Like how long is that output you had right there you're playing?
- [18:17]
B
Yeah, it's like, it's like two to three minutes per generation that you can do. And your prompts can be up to like 2,000 characters long. So I mean, amazing. Just wow. Oh, two minutes. Sorry. So, okay, a single audio generation can be two minutes. You can have one prompt. You can control dialogue, tone, ambience, background music and sound effects. It's multimodal. It's cheap, by the way. Like everything that I have generated and played with today, it cost me less than a dollar. Wow. Pretty quick. But it did cost me a dollar. Gavin, I want to be clear, Seed Audio wasn't free. Which is why I have to remind everybody that you can support this podcast. You can back us on Patreon and you can.
- [18:53]
A
Oh, it's AI for humans. I'm Kevin. More charming than Sunrise. Worth more than your gold. He's Kevin. Wow. We die for this man. And I'm Gavin. Oh, boo.
- [19:08]
B
He's fine. We could leave him. They tolerate me.
- [19:17]
A
Nice. It juices the algo. I think it juices the algo. Oh, my God. And other guy. The other guys right here just hanging out, waiting for my chance.
- [19:32]
B
I told it to do what it wanted. I said, go research AI for humans. Take some of the feedback and the fight. Put a musical together. Yeah. And it just did that. And I don't necessarily agree with the content of it, but you gotta admit, for a single one shot prompt, that's pretty good.
- [19:46]
A
Like and subscribe, everybody. Like and subscribe for more of that. All right, couple quick other things. First of all, Claude has a new thing called tag, and it is in Slack. And Andres Karpathy, who now works at Anthropic, and Claude is very excited about this. Some people are not as excited about it, but basically, Kevin, this allows you to bring Claude into your conversations in work. And I think this is a bigger deal than the normal person might think. Yeah. So you have a thought on this because you probably, maybe, I don't know, as your company instigate, have you started using this yet or is it still kind of like. We're not sure.
- [20:17]
B
It's so funny. Well, I think that my daily driver, the company that I work for is very much Claude pilled. Like, we also have an OpenAI, like Enterprise account. But I think cloud is just very clearly winning the battle for the old hearts and minds. Like, everybody's running cloud desktop. I mean, it was at one point, obviously engineering coders, et cetera.
- [20:35]
A
Right.
- [20:36]
B
But then it was slowly. Oh, all the designers are using it now. The, you know, business, business insights and analytics and this and that. Everybody is running it and they're all connecting to the various services and even connecting to Slack to check their messages and whatnot. But what is powerful here about the Slack integration is that now it is not like it's not Gavin's Claude code accessing Slack or mine. It's. It is Claude as a teammate for the company Holistically so this is like an organization level harness that anybody can tag in. It has its own credits, it has its own memory. And you got to imagine like this is like a beautifully insidious way for them to start gathering intel of how well that means operate, not individuals at an organization.
- [21:19]
A
Yeah, right.
- [21:20]
B
And it starts to have its own memory and then people are going to be reliant upon that memory.
- [21:26]
A
Well, that's the criticism I saw of it. I saw a couple pretty significant criticisms here. One is that you basically by allowing this system into your organization, you are then disallowing your own organization to kind of own your workflows. Right. Like every, the argument. I'll try to figure out what the tweet is and I'll. We'll put it up here. But like the argument here is that like every organization, the intelligence is not the thing that's valuable, it's what you've built for yourself. Right. And it's what, it's how your workflows work. And every organization is slightly different. And if you let Claude in on everything, it starts to get that information and it knows it. And now I'm sure that Claude has restrictions around like what they're going to train on everything. But this is information going in and out of the system, which is a pretty big deal.
- [22:06]
B
I feel like 100 and like they've had hiccups in the past where people were getting other people's outputs. Like that did happen. That was an issue with their API. That would be a big oopsie if you were doing anything like sensitive legal or health wise. But I think, you know, that's a big point that you make, is that now most enterprises will have agreements where CLAUDE can't train on the data and not supposed to know it.
- [22:30]
A
Yeah.
- [22:30]
B
If you're on a certain. Another plan. You said it exactly like you're handing over your taste, you're handing over your process.
- [22:37]
A
Yes.
- [22:37]
B
So now anybody could go like, well, help me critique this like Mr. Beast would potentially, if they were using that internally. Right. And it might use their internal knowledge to add way more explosions and way more big tech.
- [22:49]
A
Yes, of course. I mean it is an interesting thing that like the overarching idea of AI is getting smarter than us. And it's not just about that these systems are getting better and bigger and broader. Is that the training data that we're giving them is getting better. Right. Like they're able to integrate themselves in your very word, like insidious in some ways. Right. Like suddenly they are becoming part of everything. We do. And when they're part of everything we do, they see through everything we do. Right. Like ChatGPT, I'm sure, has a lot of data on how people think about themselves or how they feel about themselves, which, you know, not training on any individual thing, but like, at a whole, you might get a real good sense of how what ads will work for a certain type of thing or other
- [23:27]
B
sorts of stuff like that.
- [23:28]
A
So this is the kind of other side of the AI conversation that's happening and it gets very weird, very broad, very quickly that we're giving over a lot of human knowledge in different ways to AI. But I guess, Kevin, I'm fine with it. We're here.
- [23:42]
B
What are we going to do with this point?
- [23:44]
A
We're living our lives, you know, dude,
- [23:46]
B
like, and I hate to dude you, but like Anthropic went down like two days ago, you know, for like 10 hours or it wasn't even 10 hours, it was like four hours. Did you see the. Well, it's like it's a snow day. No, no one was able to do anything to do their stuff.
- [24:01]
A
Yeah, exactly. I know, it's crazy. Yeah, I will say so.
- [24:04]
B
Something we're already here.
- [24:06]
A
Something Fable going down. I had a similar experience with this and something I've done. Like, part of the reason I've kind of dove back into AI video is because, like, oh, I can't use Fable right now. And I know I can do lava 4.8, but like, maybe I'll try to explore this other side. I've been spending a lot of time trying to work with prompting systems and different things to do a few like, side projects I want to do. But there's a very cool workflow that I just saw shared that I'm not sure I'm going to dive into because it does involve Blender. But as I've shown the show before, Blender has an MCP now where you can actually go in and use this. What people are doing is very cool. They are creating fly throughs in very specific, blocky textures. And imagine a world where you're almost seeing like a previz of a movie from the 90s, right? It's just like blocks and like very not finished characters, but you see the way the camera moves and they are then adding that with a sea dance layer on top. And Kevin, this is maybe the best and most controllable AI video we've ever seen to date. It's a couple steps extra than just prompting the machine to get something back. But I do want to shout out this Japanese creator on X Craft Capital Lab. That was the first one of these I saw. And if you're watching the video right now, you'll see this very cool shot of like an anime where it's like if somebody's flying through and then somebody's chasing you, watch the camera go. But when you watch what Sea Dance 2 can do with that pre vis shot, you just see the future of Hollywood just laid bare before us. And there's a couple other good examples here. One of them is like a bunch of tall rectangles kind of standing next to each other, and the person transformed it into like an old west shot. Like, it's just a very cool way of looking at how to prompt a video that I hadn't seen before using 3D tools, which, you know, all of this is kind of coming together. And you can do it at home too.
- [25:48]
B
I did you see the. The. The Morphic. Morphic is like an AI company. I'm not super familiar with them, but they have the ability to generate viewport previews. Those block.
- [25:59]
A
Yeah.
- [25:59]
B
Blender videos you can generate. Use AI to generate that. So you don't even have to go within Blender and make the blocks and move them around. You describe it. You see the viewport preview. It's super, you know, red, simple grayscale. It comes out quick and then you click and it uses Sea Dance to generate the AI video. Like, the speed with which these workflows are changing blows my mind. But it's the difference between, like prompting something into existence and actually directing it into existence. And I like, this is it. This is happening.
- [26:28]
A
Yeah, this is it. Shout out to AI Warper, one of our favorites, who did a full workflow. We'll link to our show notes. You can go look at that here. And then, Kevin, we will be back next week and hopefully, hopefully, maybe Fable 5 will be back, or maybe not. And we'll just keep working in AI video.
- [26:43]
B
Okay, that works for me. See you next week.
- [26:47]
A
Bye, everybody.