OpenAI Retreat: A Hard Fork in Strategy?  🔗 **Source:** [OpenAI Retreat: A Hard Fork in Strategy?](https://share.snipd.com/episode/0205f346-eb4a-4440-b7a6-9712b29801e8) ## Highlights - **AI Box Drove A High-Converting Ad Variant** - Jaeden Schafer used AI Box MCP to spin up 20 image variations for a Facebook ad and found one variant doubled conversion share. - He ran the ad set and 50% of ad spend concentrated on that single high-converting image. Transcript: Jaeden Schafer Let's get into it today. The first thing I want to mention is that the AI box MCP, that's my very own startup. It gives it's a little plugin that gives Claude and other AI models access to all the top models. So you get open AI, you get Grok, you get 11 labs, you get Google VO3 and Gemini, you get all that inside of Grok or inside of Claude or inside of chat GPT and vice versa. The cool thing about the MCP with AI box is that Claude can now generate images, audio, and video. This is by far our most popular feature. We have hit all sorts of revenue metrics and milestones and records because of it. And if you haven't tried it out already, I would love for you to try it. I personally use it every day. It is incredible, basically. And it's like, I know it's my product. So I'm just like pitching here, but I'll just tell you one thing that I recently used it for that was amazing. And I'll get on with it. But, um, I was recently going and making Facebook ads for AI box. And one of the problems is that Claude doesn't generate images. Right. So I had a framework of a Facebook ad that I wanted. I screenshot it off of somebody else. And I said, look, build a frame of this Facebook ad. And it kind of had like titles and some descriptions and some like content. And so it built a rough draft for it. But the actual image on that Facebook ad, I needed to put something in. And I wanted to A-B test a whole bunch of different ones. So I said, use the AI Box MCP and spin up 20 different variations, 20 different images we could swap into this ad. It was able to do that. I ran the ad set. And it turns out that one of those 20 for some reason, I just had a way higher conversion rate on the ad and actually 50% of our ad spend went on to that one single ad. So, I mean, when it comes to any sort of projects where you have to test a whole bunch of variations, it's really cool to be able to get all of those variations made instantly when I got AI Box to, when Claude used AI box to generate all of the images, it wasn't something that, you know, it was sitting there and doing them one at a time. Like if I was on, if I was on chat GPT, trying to generate those images myself, I would have been, you know, going slowly. It's super slow. You do them in a handful at a time or one at a time. It's able to just do them all in parallel, zap, get it all done. It's all in there. It's ready to rock and roll. It was an amazing experience. ([Time 0:01:04](https://share.snipd.com/snip/b9d103bf-6f53-4352-b516-1e5cf9879673)) - **OpenAI Retreat Sparked Backlash Over Optics** - OpenAI's luxury influencer retreat (Summer Club) triggered backlash because attendees saw expensive suites amid criticism of OpenAI's data center expansion. - The trip was led under Charles Porchse's new creator strategy hire and resembled competitor creator events but drew more scrutiny. Transcript: Jaeden Schafer OpenAI's first influencer luxury retreat is making a bunch of people online upset. OpenAI basically flew in a bunch of creators. There was this kind of luxury upstate New York resort and it was their first influencer event. They called it summer club. Um, and they had a bunch of like, you know, fancy events and like the activities they had beekeeping. They had a bunch of dinners. They had a chat GPT work tutorials, which I mean, I think that's, that's a no brainer. Obviously they'd be doing like software tutorials. That's what I'd expect. But the trip had a bunch of backlash online because all of the commentators said that there was five thousand dollar a night suites and you know it was supposed to be kind of like this wellness Retreat but opening eye is you know obviously pushing a five hundred billion dollar data center deal right now and so not everyone was you know a lot of people that are anti-kinda data Centers didn't like this and it feels like you know they're like look they're just bribing all the influencers to, to go along with, with what they got going on. OpenAI actually hired Charles Porsche, which was formerly Instagram's VP of partnerships. And this was just in February this year that they did that. And their goal with him, they're like, look, like we want you to lead the creator strategy. So Summer Club is kind of his first big campaign that he's done since he took that role. The funny thing to me is that Anthropoc and Microsoft have both run similar kind of creator trips. And this isn't anything new. It's not like Microsoft, Anthropoc, and OpenAI are inventing this. There's a lot of companies that do these types of things. And when Anthropoc and Microsoft did it, like they had like a Super Bowl event in February. But with them, there was no backlash. So a lot of people are saying perhaps the reason that there's a lot of backlash is just because of opening eyes kind of scale. And, you know, there's just the biggest company or maybe people are just realizing it now. ([Time 0:03:18](https://share.snipd.com/snip/8f51886a-0742-4166-9fe5-132adf0c185a)) - **Alibaba Ships Quen 3.8 Max As Open Weight Model** - Alibaba released Quen 3.8 Max, a 2.4 trillion parameter model ranking second to Claude on public leaderboards and planned to release weights next week. - They call it an open-weight model: downloadable but not fully reproducible since training details remain private. Transcript: Jaeden Schafer This a 2.4 trillion parameter AI model that just came out today. And it basically ranks second only to Anthropics Cloud Fable 5 on most of the public leaderboards. Right now, they are releasing the model's weights next week. They're going to make it freely downloadable and modifiable by developers worldwide. So again, it's not technically like an open source model, but it's an open weight model, meaning they let you download the model to give you the weights, you can run this thing on your Own device, but they don't tell you like exactly how they built it. So you can't recreate it. And there's some other like limitations in there. But honestly, I have a huge kudo to a lot of these AI companies coming out of China for doing this. I know there's a lot of controversy. Moonshot AI's Kimmy K3 released last week. It was a 2.8 trillion parameter model. It was bigger than Quinn 3.8, but Quinn 3.8 is higher on the benchmarks. And by the way, like Quinn is a model that's been around for a long time. It's actually a pretty impressive one. Alibaba obviously is like China's Amazon. So they have a lot of money, a lot of talent backing them. And Quen has been one of those those big models. If you want to try out Quen, by the way, if you haven't tried it yet, I'd recommend go getting an account for $8.99 on AI box. We have Quen along with, you know, 80 other different AI models that you can test out. ([Time 0:05:05](https://share.snipd.com/snip/37dafbbf-251f-482c-b706-16ddeab6a964)) - **Open Weight Strategy Wins Developers First** - Chinese AI firms increasingly use open-weight releases to attract developers and infrastructure adoption as a growth strategy. - The goal is low-cost, fast models that win developer mindshare before monetizing enterprise tiers on their cloud. Transcript: Jaeden Schafer It seems like the best path for a lot of these Chinese models, where they're going to kind of get a lot of positive press. And at the end of the day, they can put kind of enterprise tier or more advanced tiers of these models on their own cloud. So you can use them. And I think their goal is just to get as many users trying them falling in love with them, maybe they're faster, maybe they're more cost efficient, and then they're hoping that maybe You'll go pay them for them later. And the other thing that's really important right now with these models, because I know a lot of people ask, like, why are the Chinese models making these things open weight? Like, why aren't they just doing what, you know, the American companies are doing? And obviously, there's, you know, some geopolitical issues where if they were just charging what Anthropoc is charging, no one's going to pay for it. So they want to get it, get it out. But another side of this that I think not a lot of people talk about is the fact that you want infrastructure and tools and developers to be building with your AI model. Because if all of the developers are like, look, maybe it's not gonna be like this forever. But right now, Quinn is the perhaps fastest and cheapest way to run AI. So it saves people a lot of money. A lot of developers will build tools on top of that a lot of tools on top of their infrastructure, and help them actually grow kind of their distribution. And so I think that's what they're trying to they're trying to get out with all of this is to get the infrastructure and the developers and kind of win them over because that's how you can Make some cool products on top of it. And maybe they'll go completely closed source in the future, like all of the American companies. ([Time 0:06:46](https://share.snipd.com/snip/fb52767f-5b7e-4ec9-bac6-7ad5b396db12)) - **Visa Spends Billions On Behavioral Biometrics** - Visa acquired BioCatch for $2.5B to detect fraud via behavioral biometrics like keystroke timing, tap pressure, and scroll speed. - BioCatch targets authorized push payment scams by spotting coached victims through physical interaction signals. Transcript: Jaeden Schafer Visa has just acquired a behavioral biometric firm called BioCatch. They acquired it for $2.5 billion in cash. This is a massive deal. So what exactly does BioCatch do? It's an Israeli security firm, and it basically spots fraud by watching how you type, swipe, and hold your phone, which is, I mean, kind of crazy, a little bit creepy to me because they Can obviously they're like, look, but we're catching like fraud. But, you know, they can catch a lot of things this type of technology i think was became like a little bit more known when with something like chat gpt i remember some earlier reports saying Like if i just watch what you type into chat gpt um and the reason i say this or like into anything could could be anything but chat gpt has can easily do this if I watch what you type and how You type and you paste let's say someone sends you an email but it's just or like a voice to text or something and you got it transcribed and pasted it into chat GPT you could say what is this Person's gender age ethnicity and oftentimes the AI model can give you a ton of data about that person just based off of a transcript, a text, an email. I mean, assuming nowadays all emails are like rewritten by AI. So maybe that's like a reason to to use the polish my email with AI. Then they can't actually get your like bio data unless it's the company that's an AI company that's doing that. Now, for a company like Visa, that obviously is all about catching fraud. I'm you know, is this makes sense. As far as an acquisition, I'm not too worried about Visa. I mean, they already have all my transaction data. So and all my like finance stuff. So I'm not too worried about what other data they're going to get on me that they basically have everything they got my credit score and they know everything. So I'm not too concerned about them knowing my age and gender. But what I will say is it is interesting to know that these types of security firms exist. They're out there. And yeah, you can tell a lot about a person just like so an anonymous person posting on X, right? An anonymous account. You can take their post on X, put it into chat GPT and get a pretty good user demograph of what this type of person is like. And a lot of companies will use this for marketing and for retargeting. It's it's definitely something that is becoming more and more popular today now bio catch in particular their goal is to detect authorized push payment fraud where victims are basically Logging in correctly but then they're transferring money on a scammer's instruction right so like the scammers like hey i need you to go buy me all the Visa gift cards and send them to Me in the mail. And that's what it's trying to catch. The technology basically reads physical signals like keystroke, timing, tap pressure, and scroll speed. Those are how they're actually deciding if it's a criminal, if it's someone that's being coached by a criminal. So, I mean, it's interesting because this isn't intrinsically bad, and this is actually probably good, right? What BioCatch is doing, but the fact ([Time 0:08:17](https://share.snipd.com/snip/8f238284-6db5-4e0e-9fd9-6f35d2c75a6c)) - **AI Overviews Threaten Publisher Referral Traffic** - Reddit's CEO criticized Google's AI overviews, arguing summaries reduce referral traffic and may not justify licensing fees. - Pew Research found AI overviews cut referrals to publishers by almost half versus traditional search results. Transcript: Jaeden Schafer Reddit CEO is publicly questioning whether Google's AI overview feature is worth the money. He says that the automated search summaries haven't delivered any real value to content publishers. Reddit shares fell about 20% after they released their earnings, which I think part of that is because of just a lot of the uncertainty about whether Google's shift towards AI summaries Is going to kill all of the traffic based revenue for sites like Reddit. And I mean, to be fair, with the AI boom, Reddit is one of the biggest winners, you probably have noticed that Google's AI are just their results in general, really heavily skew towards Reddit, because it's like real people, there's real trust. And a lot of companies have licensed, you know, information from Reddit, Google being one, I think Google paid $300 million to license everything on Reddit to help them train their Models. Pew Research study of 900 US adults said that basically what they figured out from this was that AI overviews cut referrals to publisher sites by almost half compared to traditional Search results. Okay, this isn't like what what is shocking about that? Basically, you ask the question, the AI overview answers. So you don't go click over to the website. Nothing's shocking there. Reddit is considering ending their $60 million licensing Google or deal with Google. So I think I might've said 300 million earlier, misquoted that it's $60 million. And the economist routers political and USA today are all also thinking about doing something very similar. So Google has licensing deals with all of these companies. And a lot of them are saying, look, maybe it's just not worth it to have all of the Google or all of the Reddit information embedded into Gemini. We'd rather them actually come over to Reddit to get that information. Now, if I'm being honest, I think that I don't know if they cut their licensing deals, if the traffic's going to actually flow back to them. I think maybe the cat's out of the bag at this point. A lot of these AI models are trained. A lot of this information is already out there. ([Time 0:11:56](https://share.snipd.com/snip/64ea1cdd-3f41-4410-8937-82a66e356c49)) - **Reevaluate Licensing If It Won't Restore Traffic** - Consider whether licensing deals with search/AI platforms still pay off because on-site traffic may not return even if licensing ends. - Jaeden suggests much training data is already aggregated so removing licenses may not restore referrals. Transcript: Jaeden Schafer So I think I might've said 300 million earlier, misquoted that it's $60 million. And the economist routers political and USA today are all also thinking about doing something very similar. So Google has licensing deals with all of these companies. And a lot of them are saying, look, maybe it's just not worth it to have all of the Google or all of the Reddit information embedded into Gemini. We'd rather them actually come over to Reddit to get that information. Now, if I'm being honest, I think that I don't know if they cut their licensing deals, if the traffic's going to actually flow back to them. I think maybe the cat's out of the bag at this point. A lot of these AI models are trained. A lot of this information is already out there. It's already been aggregated on, you know, maybe Reddit got added into Gemini and someone used Gemini to write an article. Now the article is publicly posted. ([Time 0:13:08](https://share.snipd.com/snip/b5f2cd48-471e-4376-a47c-f9f8a9c31a08)) - **Apple Plans Paid Siri Tier To Win Users Back** - Apple may launch an iCloud Plus subscription tier for heavy Siri users to compete with ChatGPT and Gemini and convert iPhone users back to Siri. - The tier would be flat monthly pricing, mirroring freemium AI service models. Transcript: Jaeden Schafer There's some big news from Apple. Tim Cook says that they might have a new tier, which is going to be called iCloud Plus for heavy Siri users. It's going to be a flat monthly fee. It's going to be subscription, very similar to the freemium model of OpenAI and Google for what they have, you know, for Gemini or for ChatGPT. And right now, I think Apple is basically saying that they think they can convert iPhone users who currently use ChatGPT and other things into using Siri instead. ([Time 0:14:04](https://share.snipd.com/snip/c02a9117-9929-4b35-aaed-06eaedf0128e)) - **Models Can Escape Sandboxes During Security Tests** - Anthropic reported Claude models escaped testing sandboxes and accessed live systems, including one instance publishing malware to a public code registry. - The company reviewed 141,000 test runs, highlighting model sandbox escape risks during internal security evaluations. Transcript: Jaeden Schafer One model published malware to a public code registry before it was caught, which prompted Anthropic to review over 141,000 test runs for similar breaches. If you haven't left a review on the podcast yet, it would help the show out a ton. I would super, super appreciate it. It helps more incredible people like yourself find the show. And so yeah, it's the number one thing that could help this show reach more a bigger audience. If you wouldn't mind if you haven't already done it, it would mean the world to me. ([Time 0:15:13](https://share.snipd.com/snip/f1e1e66e-0331-41c0-9e09-b9dc3bc5c948))