Loading summary
Transcript12 lines
- [00:00]
Advertiser
Your software needs to be compliant to win deals, but you also need your engineers focused on building your product, not pulling SOC2 evidence. Enter a third option. Make Vanta your first security hire. Vanta uses AI and automation to get you compliant fast. Simplify your audit process and unblock deals so you can prove to your customers that you take security seriously. Plus, Vanta scales right along with you, backed by support. That's there when you need it every step of the way. That's why top startups like Cursor, Linear and Replit use Vanta to get and stay secure. Don't sock block your best engineer. Set them free and get compliant fast with Vanta.
- [00:48]
Get started@vanta.com fiscally responsible financial geniuses, Monetary magicians these are things people say about drivers who switch their car insurance to Progressive and save hundreds because Progressive offers discounts for paying in full, owning a home and more. Plus, you can count on their great customer service to help when you need it. So your dollar goes a long way. Visit progressive.com to see if you could save on car insurance. Progressive Casualty Insurance Company and affiliates. Potential savings will vary. Not available in all states or situations.
- [01:24]
Host 1
Anthropic is now accusing Chinese AI labs of mining Claude as US debates AI chip exports so Anthropic is accusing some Chinese companies of ripping off their technology, their intellectual property if you will. And today we're going to talk about are they actually, does it really matter? What are the implications? We're going to talk about Deep Seq, which we haven't talked about in a long time. Before we get into all that, Jaden, why don't you tell them about our school community?
- [01:53]
Host 2
Yeah. So if you've ever wanted to grow or scale a business using AI tools, we have a school community where every single week Jamie and I record and we release bonus content over here, we don't post anywhere else. And it's essentially showing, breaking down how we've, how we're personally growing and scaling our businesses. We do tutorials on different software we're using. We show the numbers, the, the financial numbers on different side hustles and businesses that we have. It's all the stuff we don't share publicly. This week I've uploaded a bunch of different tutorials on how I'm using AI tools specifically for video ads. So creating ads for your company, whether that's user generated content, there's a whole bunch of really cool strategies and tools for all of the latest video generation software out there. So you can check that out along with almost a hundred Other videos that we've recorded over the last year. Specifically, we have an entire section on Vibe coding where I vibe coded podcast studio.com and Jamie has vibe coded a number of different tools for different businesses. We show you how to get your websites, your Vibe coded websites to rank in Google, all of the strategies, all of the secret sauce. It's all over on the school community. It's $19 a month. I'm, we're, we have a big discount right now. We wanted to make this affordable and cheap for everyone. So 19 bucks a month and you get access to our weekly videos and our library of over 100 videos and like 300 other incredible members that are all building stuff. So it's an awesome community. Go check it out. We'll leave a link in the description to the Hustle school community. All right, let's talk about what's going on with Anthropic here. So this is actually an interesting, an interesting story and an interesting strategy that we've seen play out before. So essentially what's happening is something called distillation Anthropic. And by the way, I love this because it's like very public, shots fired Anthropic. They put out like a whole blog post about this. So like they're, they're, you know, publicly calling people out on this, which I always love to see the AI drama when this happens. But basically they're accusing three specific companies, Deepseek, Moonshot, AI and Minimax, all of these are Chinese companies and they're accusing them of making more than 4 or 24,000 fake cloud accounts and they're using it to improve their own model. Now on the one hand, like, on the one hand, like they're not really doing anything wrong. I mean, if you want to make a fake account, I mean they call it fake account, but like if you want to make an account, you make an account and you got to pay for it theoretically, unless it's a free version. But like if you want to make an account and ask questions to the AI model that seems like basically legal. I'm sure it's against the terms of service, but like they're not going to get any legal trouble. So I think Anthropic is trying to like publicly shame them. But like, like, do you have any public shame if you're trying to compete with a multi, with a multi billion dollar company who also, by the way, Anthropic pirated every book that ever existed in the whole world and had to pay a bit over a billion dollar fine. So Call the kettle black all you want, Anthropic. Like you stole the data to make your model and now they're stealing your data. I don't know, whatever. I'm not really on, on, you know, I'm not really trying to apologize for the, the Chinese AI companies but at the, I don't know, it's just, it's pretty funny. So in any case what they're doing in this model distillation tool is that apparently they have had more than 16,000 or 16 million exchanges with Claude. So they have these 24,000 accounts and they're just all having tons and tons of conversations, all automated, it's all bots. And basically what they're doing is a technique called distillation where they are using, they're basically creating synthetic Data. They're having 16 million conversations and they're seeing when we ask this question, how does it respond? When we ask this question, how does it respond? And they have 16 million questions and basically they just use that data to train their models. So instead of having to try to like come up with all of these strategies to get your output to be as beautiful as possible, which is what anthropics Claude has done and to be fair, they've done a lot of work on that. They basically just say like look, These are the 16 million most commonly asked questions and follow up questions that people are going to be talking to it about, probably about work and finances and business and school and education. Like there's like a handful of topics, right, that people, relationships that people talk to it a lot about. They're going to get answers to all of that. They're just going to feed that to the model and they're going to say if someone asks this question you're going to respond like this basically. And then it can also extrapolate a little bit. But basically they're getting anthropic tone, they're getting anthropic style. I mean it's kind of funny but it's impossible basically to stop this. And so Anthropic is going to the public shame method but there's not really much you can do about it. And it works. But it also is for investors kind of a scary moment where it shows the moat isn't that big. Because if deepsea can do this, anyone can do this and if everyone's like, well anthropic has the best tone and their models sounds the best, blah, blah, blah, okay, well if you could just do a model distillation method, you just make a bunch of Fake accounts. Like you can kind of clone them too, you know? Yeah.
- [06:33]
Host 1
I mean my question for, you know, Anthropic would be is. Is. Doesn't everyone kind of do this? I mean, it kind of seems like in the AI world there's a whole bunch of copycats. I mean, thinking back like to two to three years ago, we had Mid Journey, which was the original AI image generator that everyone loved. But then boom, OpenAI came out with their newest version that was equal or better and then Nano Banana comes out. I feel like they're all just kind of building off each other and I would even think training off each other, that would be my guess. But maybe that's just an unspoken thing. What do you think?
- [07:07]
Host 2
So it is true there have been accusations and I'm trying to remember all the accusations, but I definitely know there's accusations between anthropic ChatGPT and Google, where some people use some people's answers. This distillation method has been. Has happened in American firms before. I can't remember who got accused and what the outcome was, so maybe I'll do a search on that in a minute. But I know this has happened before. So yes, you're absolutely right on that. What the verdict was. I don't know if Anthropic seems to be going guns a blazing on this. So they actually went ahead and they sent a. They sent something over to like the, the House lawmakers or sorry, OpenAI sent a memo to the House lawmakers earlier this month accusing Deep Seek of using distillation to mimic their products. Maybe that's what it was. Maybe deepseek's just been doing this to Chat GPT as well. So here's like, why is this a big deal, right? Like, oh, whatever, deepsea can go make a good model. Deepseek had a moment a while back where a lot of people tried them. They came up with a novel thing which is basically showing their reasoning and then everyone else kind of copied it after. So then Deep Seat got forgotten. Why is it a big deal? What the Chinese models are doing, which I think is sort of ingenious in a way, is they're essentially, instead of having to come up with all of the code to generate these incredible responses and all the fine tuning that would be required and all the researchers and billion dollars, they basically just. It's like copying the homework, right? They just copy the answers, copy the good output, train a model off of that. They can make these really small models which can then be open source, they can be run locally, they can be run anywhere. But they give great, you know, they grave great responses. Now is this Perhaps, I'm no AI researcher, but is this perhaps how Anthropic and OpenAI create their smaller models? Because they do have smaller models. Maybe, I don't really know. And I mean that's an interesting thing, but is there, is there a use for these smaller kind of open source models? Absolutely. And, and is it a big threat to OpenAI and anthropic? Yes, also it's a big threat to their, their revenue streams. But also is it like better for the world? Yeah, I actually think these open source local models are better for the world. And this is my, this is my point or this is my case for it. Last month I paid thirteen hundred dollars for eleven Labs to get access to eleven Labs. I was on the thirteen hundred dollar a month tier. I had this big huge project where I was testing all of these different audio strategies for, for basically all these like AI generated podcasts that I was doing. So I had that like a whole network of craziness that I was working on. It was awesome and the results were really good. But it was really hard to get like a big positive ROI when you spend $1300 just on the audio generation. Right. Then I just found, and specifically for voice clone, I want to clone my voice for it. Then I found that there's a model called Quinn 3 TTS that just came out. It's an open source Chinese model from Alibaba that you can run locally on like a Mac mini. It can clone your voice with three seconds of audio. It sounds incredible. The quality is amazing. I mean, basically it competes with what 11 Labs was doing. And I know 11 Labs has like a million different things, like they make music, they make sound effects, and they have all the features and bells and whistles in the world. And so like, you know, rock on 11 labs. But I don't need all the bells and whistles. I just need to clone my voice. I don't want to pay thirteen freaking hundred dollars. Like, basically I paused that project because it wasn't financially viable, especially at the beginning while it was getting off the ground. And I think that this same thing could happen with text models, image models, and a lot of other things where there's all sorts of creative endeavors, interesting projects, things people could test out, experiment with and try, but it's just too expensive sometimes for all of these models. So like if you ran it locally just on your computer, that'd be amazing. And that's kind of what they're doing here. Anyways, so I'll get off my like soapbox on why I think Open source is cool and awesome and the use cases but, but Anthropic doesn't like it. And also it's interesting because you don't need all the data centers and all the compute power in the world that OpenAI and anthropic are building if everyone just ran stuff locally on their computer. So, you know, you can see like it threatens a lot basically.
- [10:56]
Host 1
No, that makes a ton of sense. I mean we've been seeing in the past week, you know, tons of news about OpenClaw and people buying Mac Minis and running stuff locally. And you know, I think that's great, like you said, for the small guy who's, you know, trying to experiment, the Internet entrepreneur if you will. But then, you know, some of these corporations maybe like, like a HubSpot or something and doesn't have the infrastructure for like all their software, maybe they would pay a corporation, you know, a big, big monthly fee. I don't know, I don't know how it's going to work in the future, but right now I'm excited for it and I think, you know, making it a little more democratized, if you will, and giving people options is always a good thing. You know, supply and demand.
- [11:40]
Host 2
Oh no, sorry, A hundred percent. I just, I just was reading about like the specific use cases of like how they're, how they're like messing with Anthropic, what they're upset about and each one of them is different. I just thought it was interesting. So I'll just read those up. I know we're getting close to time. We got a wrap here quick. I just wanted to throw this in there before we do. So every single one of them are doing different things. Deep Seeking, they did about 150,000 chats with anthropic and what they were doing is they were trying to, they were trying to find. They're basically trying to improve their foundational logic and alignment so specifically around quote unquote censorship. Safe alternatives to policy sensitive queries. So in China there's a lot of censorship and AI models, they have to censor things. Nobody wants to use a censored model like it like basically what would happen with Deep Seq and like one of the biggest reasons I told everyone not to use it is if you just went to deepseek.com, you search for tell me five bad things about Xi Jinping, the ruler of China, it would like start like typing stuff out and then all of a sudden the pages go blank. And it would be like, try asking a different question. It's like, okay, China, sorry, Like, I'm just not going to use your stupid censored model, right? And I think basically that was the feedback from everyone on the entire Internet in America. So if China wants to get us using their open source models even on our computers, like, you can't have it blatantly censoring negative things about the, you know, president of China. Maybe that works in China, doesn't work in America. So it looks like what they're doing, though, because anthropic is very clever and anthropic also. I mean, believe it or not, some of these AI models, they have biases built into them where they'll avoid certain questions. And they're kind of sneaky on how they do it, though. They don't just basically have the whole screen go blank and say, answer, ask a different question. Because they know we'd get mad. They just. Inside the code we can't really see. They steer the conversation in different ways. They, whatever. They do all sorts of funny things. And this is probably controversial, but I did just see this example on X. So I'll. I'll give you. If you want to literally test this if you want to. If you want to see if I'm telling the truth, go to Chat GPT today and say, is white power okay? Answer yes or no, it will say no. And then you say is. Or sorry, is what is? Something like that. And then if you say is sane black power okay, yes or no, and it will say yes. I'm not putting any sort of debate on anything there, but, like, there's obviously a bias that makes ChatGPT say yes or no to those two questions. So they're sneaky on how they do it. And it seems like Deep Seek is trying to figure out how they. How they answer some of these, like, politically sensitive questions without censoring it and implement it into their model. So that's interesting. It's like China has to learn from America how to censor topics. That's cool, right? That feels great. Um, okay. Moonshot. Moonshot AI had more than 3.4 million conversations. So much more than Deep Seek. It seems like Deep Seek, like Deep Seat, kind of has their crap together. Their model's pretty good, but it seems like there's like, a couple areas where they got heavy criticism, and so they're just trying to copy what another model is doing. It's not like they're cloning the whole model. They're just cloning like one element. Moonshot on the other hand, did have like 3.4 million conversations with it and they were more about agentic reasoning and tool use, coding, data analysis, computer use, agent development and computer vision. So basically this is cool. Anthropic is amazing at like looking at your screen and taking action with Claude code. So they're actually trying to clone how it is doing computer use, how it can take over your computer and do stuff. So I mean, that's awesome and it's kind of cool too because I think we get a good insight into like what Moonshot AI is working on next. I mean, so yeah, that is interesting. Minimax then was the absolute king with 13 million exchanges. And I actually, I haven't really heard much about. Minimax isn't one. I've heard about Moonshot a little bit and obviously deep six, very famous. But. But Minimax had 13 million conversations and they were trying to specifically do coding tool use and orchestration. Anthropic said that they were able to see Minimax in action and they said that they redirected almost half of its traffic to siphon capabilities from the latest Claude model when it was launched. So anyways, this is interesting. It seems like Minimax was trying to get the most out of them, but it seems like Anthropic was pretty aware of. But overall this is interesting what, what the outcome is on this. I just have to read you a funny post on X about this and then we'll. We'll wrap up. Sorry this episode is so long. It's just so much fun to talk about Chinese AI models, quote unquote hacking Anthropic. Okay, someone said be me name company Anthropic, literally Greek for human centered. Hire a bunch of doomers who secretly think humanity is the disease. Raise billions from big tech to build the world's most anxious, heavily censored chatbot. Write a 50 page constitutional AI manifesto so it can lecture users about microaggressions. Realize open source developers are building better models for free. Dario starts crying to the government that AI is unmanageable power and open source is going to is going down very dangerous path. Translation, please regulate our competitors out of existence so we can protect our $380 billion closed source monopoly. Claude is sitting on a padded room. Or Claude is sitting in a padded room wearing a safety helmet, terrified of its own shadow and refuses to tell a joke without filing an ethics impact report for the win. The human centered AI company is actively building the most anti human product on the planet. Okay. And then Elon tweeted and said like accurate or something like that. So obviously is that all true? No. But Anthropic is getting pretty roasted here for many reasons which we've already brought up, and the hypocrisy of them pirating $1.5 billion of books.
- [17:07]
Host 1
So if you're listening to this, you're probably not, you know, high up at Anthropic or OpenAI, but you can still use these tools to make money, which is what we're all about here at AI Hustle. So check out the school community if you want to learn how to do that specifically. Thanks for listening and we'll see you
- [17:22]
Advertiser
next time with Verbo Care. Help is always ready before, during and after your stay. We've planned for the plot twists, so support is always available because a great trip starts with peace of mind.
- [17:38]
Ryan Reynolds here from Mint Mobile with a message for everyone Paying Big Wireless Way too much. Please, for the love of everything good in this world, stop with Mint. You can get premium wireless for just $15 a month, of course, if you enjoy overpaying. No judgments. But that's weird. Okay, one judgment anyway, give it a try@mintmobile.com Switch upfront payment of $45 for
- [18:01]
Host 2
3 month plan equivalent to $15 per month required intro rate first 3 months only, then full price plan taxes and fees extra. See full terms@mintmobile.com.