Loading summary
Transcript22 lines
- [00:00]
A
Welcome to the podcast. Today I have some exciting news, and aside from the fact that Meta is unwinding their $2 billion Menace acquisition, I have some even more exciting personal news, and that is that last week, my wife and I welcomed our newest member of the family. We had a healthy baby boy. His name is Clay, and it is absolutely exciting in our household. So if you're wondering why my podcast posting hasn't been super consistent over the last week, and maybe for the next week or two, it might be a little spotty. I'll do my best, but I'm spending a lot of time snuggling that guy. Currently, him and mom are sleeping, so I figured this would be a good time to get a podcast episode out. But everything went incredibly well with the birth. Everyone is healthy and happy and doing great and growing. He's pooping lots, so we're pretty happy. He's now five days old. This year I turned 30 and this is our fourth kid, so we definitely have our hands full, but we feel super blessed. I always tell everyone to have as many kids as you can as soon as possible. They are the best thing about life.
- [00:59]
B
Okay, let's talk about what's going on in the news.
- [01:00]
A
I mentioned meta unwinding this $2 billion Manus acquisition. This is because Beijing told them that they had to divest it, which by this point we've all seen the headlines that Anthropics Fable 5 got banned over the weekend and pulled, so you no longer get access to that.
- [01:15]
B
But it turns out that an Amazon
- [01:17]
A
security paper is what triggered the White House banning this. We'll get into that. Apple's Craig Fred Griski says that the new Siri is going to refuse romantic roleplay.
- [01:25]
B
So we're going to talk a little
- [01:26]
A
bit about where Siri is going in the future and.
- [01:28]
B
And what it can and can't be used for.
- [01:30]
A
Jeff Bezos's Prometheus company has just raised $14 billion at a $41 billion valuation for physical AI. And there's some interesting things being built there. And the White House says they suspect a China linked group accessed Anthropic's Mythos model. So this isn't even their Fable 5. This is their Mythos model. And so there's some other drama falling out from there.
- [01:50]
B
The first thing we want to get
- [01:51]
A
into is the Meta story. This is absolutely crazy. I mean, this was a $2 billion acquisition.
- [01:57]
B
And what was interesting, just for like two seconds of context on the story,
- [02:01]
A
basically, it was started in China.
- [02:03]
B
It had a completely different name. It was like Butterfly or something like that. And then they moved it to Singapore. And I believe the CEO and the CTO are still, were still stuck in China, although they moved the whole team to Singapore and then changed the name and it kind of scaled. Became incredibly popular. Meta acquired it because they can acquire a company out of Singapore, but I think because the CEO and the CTO were still like, locked down in China, like they were banned leaving, this is kind of how they got them to unwind the deal, by sort of holding them hostage. It's interesting because in this particular case, you know, they're. When they first came out with the news telling Meta to unwind the acquisition, Meta said, you know, we believe we have legally acquired it. We're, you know, we're allowed to do this. We're going to, you know, discuss this in court with them. And, you know, Meta is a company that is banned in China, like Instagram, you Instagram, Facebook, all WhatsApp, all of their different social media platforms are banned in China because China censors, you know,
- [02:59]
A
the Internet and stuff.
- [03:00]
B
And so it's interesting that they still are unwinding this deal, even though, you know, Beijing doesn't really have that much. You would assume they don't have that much sway over Meta.
- [03:09]
A
Right?
- [03:09]
B
Like, they can't really tell Meta what to do when they, you know, they can't blackmail them, like, hey, we're going to ban your other apps. Like, they don't have as much leverage. So it's interesting to me that they're still unwinding this deal. I'm assuming it probably has something to do with kind of holding the other co founders hostage over there and probably some other things. The co founders are in preliminary talks to raise about a billion dollars to buy the company back from Meta. So I guess this is the way that they want to do it is, you know, Meta bought this for $2 billion, but I think China's like, okay, fine, look, we'll, we'll, we'll basically let them raise money elsewhere and pay back Meta. So I'm curious to see how that goes. Although from what I'm seeing, they're in preliminary talks to raise a billion dollars and Meta bought it for 2 billion. So I don't know if I wonder if Meta's gonna just straight up lose money or how this whole thing will work. Meta has cut Manus off from their internal systems. They're blocking employees from using Manus tools. Now. What's interesting is I actually have a friend that works in advertising and he works in a marketing firm that does Facebook ads. And he was telling me that, you know, he loves Manus. It does a lot of stuff inside of his company. It's been super useful. And so it's interesting to me that once this is getting rolled back, he's going to get cut off from that. Now what I will say is I feel like this actually isn't that terrible of a thing for Meta, and I think it is a pretty terrible thing for Manus, in my humble opinion, because Claude Cowork, you know, even my friend that works doing a lot of stuff inside of Facebook ads, he's like, yeah, like I use Manus for like, you know, these three things because it has access to these three files. But then I get. I just plug it into Cowork and it gives me like, more stuff and it actually does more of the doing for me. So what's interesting is I think Cowork is much more powerful. I think Meta could build a Manus clone or Cowork clone themselves. They could build that tool. I'm not sure if they will because I'm not sure what their strategy is. That's kind of just more general purpose. Whereas they kind of bought something that they could plug straight into Facebook ads and they kind of had a good use case there. So I'll be curious to see what they do. Although Manus was, you know, for kind of general purpose as well. So I'll be curious to see if Facebook rolls out their own version. Um, assuming they get all of their money back. I mean, it was a big headache, but I made them look cool for a minute. So assuming they get all of their money back, I think they could spend that money and, and build something similar that would be quite valuable to their company anyways. So I think, in my opinion, bad move on China's part. And I mean, I get it. They're like, look, we want to. We're kind of behind in the AI race. We want to be able to hold any AI companies that are breakouts inside of China and keep controlling them. But yeah, I'm not sure that's. It was the best move. But from China's side, I get it, right? It's. They feel like they'd lose power to. To the west by giving up their. One of their very few popular AI companies. Not few. I'm sure they have plenty of big AI companies that are used by Chinese people. Just haven't made a lot of breakouts in the broader market. Deep Seek isn't really in the. In the Headlines like it used to be. Okay, let's talk about Amazon. They had a security paper that triggered the White House's ban on Anthropic's Fable model. This is interesting because it's a counter to the narrat that I heard from a bunch of people, which is basically, the White House is mad Anthropic. So that's why they're kind of banning the Fable thing. I didn't really. Fable 5, anthropic's latest and greatest model. What's interesting about Fable 5, it is the Mythos model. So this is the model that, you know, they went around telling everyone this is like super dangerous, it can hack and breach everything in the world. And then they're like, hey, we put some guardrails on it and now we're releasing it to the public. Everyone can have it, but it just has guardrails on it. And there's a bunch of articles I read on AI chat daily.com that were basically saying if you ask it any biology related questions, it would, it would not answer them. And Anthropic said that's on purpose because it didn't want people to make bioweapons with them. So then it would forward you to Opus 4.8. If you ask any cybersecurity related questions, it would also not answer and forward you to Opus 4.8 for, you know, doing any sort of security audits because they don't want it to find any codes. So they had these guardrails, they released it to everybody. And apparently Amazon CEO Andy Jassy briefed the White House on a security report about Anthropic's Fable 5. And they, and basically because of that, they were able, they pulled, they banned Fable 5. And according to David Sacks, who's posted on Twitter, used to be the Aizar, he said, talking to people inside of the administration because he's no longer the Aizar, but I'm sure he's just got a lot of the connections there. He said from his understanding, Amazon came out with this paper, said they had, you know, figured out ways to get around the guardrails. They told Anthropic to fix them. Anthropic refused to fix them because they said it wasn't necessary. And then the US Government said, okay, well, in that case we'll ban it. I'm sure Fable 5 is going to come out again. Hopefully that will get fixed. David Sack said he's really perplexed because he loves Anthropic. He loves what they're, what they're Building and that he said the administration has a lot of respect for their technical capabilities of their models. And to be honest, it felt like the whole relationship there was kind of thawing out because they were briefing the, you know, different departments of the government. So it felt like it was kind of thawing out, which is what I personally would like to see more of. But he said he was perplexed that they didn't want to go and update those safety models. Maybe it was hard, maybe I'm not really sure what's going on, but this, this will be interesting. Anthropic disputed the jailbreak, apparently the one that Amazon came out with. And they said, they argued basically the same prompts and outputs work on GPT 5.5 and any other public models. So you know, kind of what they're saying and the reason why they didn't fix, quote, unquote fix is because they, they're like, look, yeah, you can get around some of our guardrails with some of those techniques, but you can use those same techniques to go quote, unquote, jailbreak, GPT 5.5. So why do we have to fix it?
- [08:39]
A
Right?
- [08:40]
B
So it is a fair point, right? Is GPT 5.5 going to get banned or you know, get export controls put on them, et cetera. So it, you know, people could say it's one sided, but for a company that's so focused on being a safety company, like that's basically anthropics, you know, mantra, I would assume they would take a slightly different approach and, and try to work with people to, to fix it and not sweep it under the rug. Mostly because they just came out and did their big PR stunt saying, you know, Mythos is going to take over the world. They got a lot of people that aren't even that into AI concerned and know about that story. So I would think they would, you know, for nothing else, for that reason, take it a little bit more seriously. Okay. Apple's Craig Frederick said that the new Syria is Siri is going to refuse romantic roleplay. This is interesting because Apple has kind of been on the sidelines with a lot of this AI for quite a while and they're finally coming out with their updated Siri. And so we don't really know all of its capabilities, what it's going to be capable of. We know that you can plug ChatGPT, Gemini, Anthropic, different models into it to power it. But it's interesting because this is one of the first times I'm seeing they're making some you know, decisions on what it will do. It's not just like, look, plug these other models in and chat with them like you would chat with ChatGPT. It looks like they're actually setting restrictions and rules on the model and probably for, you know, maybe like safety reasons or all sorts of different reasons, but they have like their own controls that they're going to be putting on top of the model. OpenAI and Google's chat bot, According to the person over at Apple, Frederick said that they are tuned for sycophancy and user stickiness. Which, to be fair, if you ask ChatGPT a question, it's going to answer and then it's like, would you like me to do xyz, XYZ and xyz? Like it's trying to prod you on to keep chatting with it. It wants to, you know, maximum conversations. So that was his point and he framed Siri's role a lot more narrowly. He said its goal is to answer the question, complete the task, it's going to decline the emotional bond. And it's basically the opposite of how a lot of the main AI apps are working right now. OpenAI apparently has loosened some adult content restrictions on Chat GPT. I remember Sam Altman making a big announcement about that. He walked it back when they kind of axed the Sora video model. And I think the whole adult content was something that kind of got cut back as well. So I'm not sure where they are on that as well. Apple is trying to basically differentiate and they're going to put these new guardrails on that stops any of that kind of. And they're just saying, like, look, that's not the use case. They said maybe they were going to be more of the quote unquote, boring assistant, but it's just going to be useful privacy on device. And I mean, for me personally, I'm happy if there is a model like that that is useful for my kids. Like, my kids could go and talk to Siri and use it versus me worrying about, oh, shoot, like, if they're talking to ChatGPT, is it going to say something weird or something weird going to get triggered or, you know, who knows what rabbit hole it goes down. So I do see a lot of value to, to making some of those editorial decisions. Okay. Jeff Bezos, his project Prometheus, his company Prometheus just raised $12 billion at a $41 billion valuation. He's working in the physical AI space. This is their Series B that funding they've raised is $18 billion and they've done that in less than a year. It is building software to automate the design and manufacturing of really complex hardware. So they're doing things like jet engines. They're also doing drugs. And this is basically a category that a lot of investors right now are seen as more defensible than just software. Right. Anthropic is going, is just software, but it's incredible. And so they're trying to get into, you know, how do we actually manufacture some of these harder things, you know, Anthropic's helping you build an app. Well, they'll help you build a jet engine. So it is pretty interesting. JP Morgan Chase, Goldman Sachs, BlackRock, all of them joined in this round. They have about 150 employees in San Francisco, London and Zurich. And a huge chunk of the $12 billion is going to fund compute infrastructure.
- [12:31]
A
Right.
- [12:31]
B
It's so interesting when these companies are raising these massive rounds, it's not like they just raise them to increase salaries, although it seems like Meta was paying some pretty huge salaries a while back. They were on a bit of a tear there, but it feels like they're, you know, they're just raising all this money because they need to buy GPUs, they need to buy data centers, they need to buy all of the infrastructure. So Prometheus cited Google's $920 million monthly SpaceX bill as kind of a floor for some serious physical AI programs. Bezos is arguing that AI is going to create labor scarcity rather than mass job loss. He said he's, it's kind of a different narrative than a lot of his other peers who have forecasted a huge white collar displacement over the next few years. Project Prometheus, right now they have a $41 billion valuation. I think this is basically showing a lot of people believe a lot of capital is flowing into this area where you're focusing on hardware, you're working with data, you're working with manufacturing, there's regulatory moats. It's not just pure, pure software, you know, helping people build software. Play like you have something with Anthropic. The White House is, has suspected that a China linked group got access to Anthropic's Mythos model. Now this is kind of funny coming on the back of everything with Fable 5 getting banned. But Mythos, right? Very hyped, very scary cybersecurity model. Anthropic gave it to a bunch of trusted partners, they gave it to Amazon, they gave it to Microsoft, they gave it to Mozilla. They gave it to Apple. A lot of their quote unquote trusted partners, they said, hey look, we found security vulnerabilities and all your stuff, take this, run it on it and fix all your security vulnerabilities before releases to the public. So hackers aren't going to be able to get a lot of the legacy databases they give to a lot of open source projects that were used by a lot of people. Now apparently of all of the quote unquote trusted people they gave it to, some Chinese linked groups did get access to it, according to the White House, who now has export controls on Mythos's on the Mythos model. And because of this, according to a Semaphore report, the breach they said is basically a critical vulnerability where they're showing that a rival state with query access to this model can use Distillation to train smaller capable copies without needing the outright original weights. Now basically what this is saying with distillation is you can take a model like ChatGPT and you can ask it 100,000 questions and see how it responds to those questions. You could take all of your questions and its responses and train a model to answer like those hundred thousand questions and you can actually get, it's called distillation. But you can get a pretty good copy of that model and how it interacts. You give it as much real world knowledge outside of that as you can. But basically it's like a lot of the pre training this is how you. It's kind of crazy. It's like there's less IP for these AI companies because you can do this distillation method. Now whatever. A lot of people have been doing this. I think Elon even admitted that Grok did distillation based off of OpenAI. Now what's interesting is if a Chinese group got access to Mythos, the security one, and they were able to do this, they could technically create smaller, more cape like smaller capable hacking models that are good at different things or you know, all sorts of adversarial tactics. And even if Mythos gets shut down, if you got access and good get a hundred thousand queries in, you could kind of distill these smaller models. So there's a disc, a discord group that was previously talking about how they got access to Mythos for two weeks before Anthropic detected and shut down this breach. And I think that they were like on the website and anyways they figured out how to do some stuff and so Anthropic told Semaphore that the government did not mention China during their export control talks. So this is kind of contradicting a lot of the reported China motivation restrictions, and I think it creates a bit of a gap between what we know and what the policy is. So this will be something that has to be worked out. But I mean, if it really is a massive risk, you definitely don't want it to be leaked. Thank you everyone so much for tuning into the podcast today. If you enjoyed the episode, please leave a rating review wherever you get your podcasts. I'll try to keep up to date on my upload schedule, although I do obviously have the new baby, so lots of fun with him over there. I was up Claude coding between the hours of 1am and 3am last night while rocking him on a bouncy ball because he was a little fussy and was trying to let mom sleep a bit. So lots of fun over here, but I am getting a lot of cloud coding done. Anyways, hope you guys all have a fantastic rest of your day. Make sure to go check out AI box AI if you want to get access to all of the latest AI models in one place, I'll leave a link in the description. Catch you guys all in the next episode.