Loading summary
Transcript1 lines
- [00:00]
A
Thinking Machines by Miriam Moradi, who famously left OpenAI after she was the CEO when Sam Altman got kicked out, have released their first open weight AI model called Inkling. OpenAI has built something called GPT Red. It's an LLM super hacker and they built this to harden their own models. AI Box has shipped an MPC server that brings 80 different AI models into Claude Chatgpt and Gemini. AWS is committing $1 billion to a forward deployed engineering organization. OpenAI and Anthropic are both scaling theirs. Apple Intelligence has been cleared for a China launch with Alibaba's Quen. And Meta plans to sell and resell their AI compute following what SpaceX is doing and kind of copying their playbook. Let's kick this off with Thinking Machines. This is a company that I'm really rooting for. They've raised, you know, over a billion dollars, but they haven't come out with anything super groundbreaking. But what I will say when it comes to a lot of these models, they're so expensive and they take so much time that um, when you think of something like even Anthropic, it felt like OpenAI had just run away from them and they were never going to catch up. And little by little they, they pick their lane, they make their product better and they're able to get market share. Until the point where Anthropic has now exploded and is making more money than OpenAI had surpassed them in revenue. I think we might see a lot of that same strategy played out by a lot of different AI companies if they don't kind of shrivel up and die or get sold off for parts or, you know, an aqua hire or something like that. And so Thinking Machines is one of these companies that I'm really excited about and, and I hope that they really go places. But this is run by Miriam Murati, who is famously the CEO of OpenAI when Sam Altman left and she has created this thing called Inkly and it's an open weight AI model, it has 975 billion parameters and companies can download this and customize it themselves so you don't have to go and pay for API access to OpenAI or Anthropic. You can actually just go and fine tune this all on your own. You can fine tune on your own data. And the bet that they want people to take is that that is going to be better than a one size fits all mod that OpenAI or Anthropic or Gemini or any of these other big labs are selling Inkling was trained on 45 trillion tokens of text, image and audio and video. They did it in about nine months, which is way faster than OpenAI's five year timeline or Anthropic's three years. In a test with Bridgewater Associates, the financial model, trained on the hedge fund's own expertise, scored 84.7% on financial reasoning benchmarks at roughly 114 the cost of the top models from Anthropic and OpenAI. So we're seeing some massive improvements when you're taking this kind of these kind of models and you're fine tuning it on your own data. The model activates only 41 billion of its 975 billion parameters per task, and users can dial up thinking efforts to trade speed for accuracy. So I'm rooting for them, but time will tell how well this model does. We just learned that OpenAI built something called GPT Red. This is an AI model trained to attack their own systems and, and they use it to catch security flaws before release. So the model cut successful attacks on GPT 5.6 from over 90% down to 23%, which basically makes it one of OpenAI's most secure releases yet. One particularly interesting attack that it was able to discover is called a novel fake chain of thought, which basically is tricking a model into making up fake reasoning steps. So it's kind of similar to convincing someone that you, you know, one plus one equals three and then, you know, saying that I already checked the math, that's what equals, and if it equals that, then therefore, and you, you know, go trick them on the next thing. So they found that when they gave the same task as a human, Red teamer who tested GPT5 in 2025, GPT Red found more effective attacks than the humans. They also got it to go and hack like this third party vending machine agent, which is kind of funny. It does have a bunch of limitations. So it's not very good at back and forth conversation attacks and, and it's not very good at exploiting images that are embedded into malicious text. The idea behind this is actually functionally working is that GPT Red is automating how all of the security holes are found because it puts an attacker model against a defender model. And they put them in these kind of simulated real world environments and they're battling it out. OpenAI is not releasing this externally, so they're not giving this to other people to test. Which is interesting, right, because we had the whole anthropic Mythos model, which was really good at security exploits and they gave it out to all of the top labs and said, hey, like go harden all your security with this. So OpenAI not giving this out externally beyond just using it for themselves. And I mean it's got a huge, massive drop in the success rate of a lot of these exploits. I think it shows AI powered red teaming can be just as effective or more effective than humans. I mean you can just brute force way more tests than humans and it's trained off of what humans are doing, but at this point it's doing better than what a lot of the humans are doing. AI Box, which full disclosure is my own startup, has just released an MCP server and we allowing people to plug 80 different AI models straight into Claude, ChatGPT, Gemini or Cursor. Basically it lets you use any model inside of whatever assistant you already are used to using. Personally, I use Claude all day long, although I'm kind of switching over to ChatGPT with their new ChatGPT app. It's really amazing. But both, either way, you get the AI Box MCP and it allows you to access all of the images that OpenAI can generate. It allows you, if you're on Claude or ChatGPT to generate all of the videos that Google VO3 can make. And no matter what platform you're using, you can access what 11 labs can do with audio. I spent basically the entire day today getting ready for a big Facebook campaign that we're launching here at AI Box. And in the past, doing Facebook campaigns usually meant hiring a Facebook ads team. It meant creating tons of creative, which just takes a lot of time. And it meant, you know, spending a ton of time on landing pages With Claude and with the AI Box MCP that we built into it, I actually got almost all of the. So basically what I did is I went and recorded a bunch of just selfies of me talking about the product on my phone. I dropped those into a folder and I had Claude go and edit all of them. And then I was able to say, hey, I need to create a bunch of dynamic so like image generated ads. I gave it the style that I wanted to have these different ideas. Some of them it looks like kind of like breaking news images. Some of them they actually look like it's like me doing a FaceTime call and there's like a text message popping up on the screen and it's some sort of text message about AI Box. So anyways, I had all these different creative ideas that I went and got from a bunch of people and, and to Go and you know, create a template on Canva and then switch out tons of different variations of the title and the backgrounds and stuff like that takes a lot of time. And before Claude couldn't do it because it couldn't do any of the image generation. But now that AI box is embedded inside of Claude and it can do the image generation. One in particular that I did is I was showing, I was showing the ability to generate images inside of Claude and I needed that generated as an ad. So you had to have like basically the design of a Claude interface and then you had to be able to show images generated and you know, have like AI box logos and stuff. So what I, what I did is I was like, hey, go, go create this mock up inside of Claude. But for the image part, just go use AI box to generate the image and pull that inside of. It's like the image inside of the image that it's generating is able to do that in a couple seconds. And I said, this is awesome. Go think of like 10 variations or like 10 different use cases of my product. And you know, and in this case it's the product is showing that you can create images. This is kind of a meta example. I know, but I was like, go think of like 10 different variations. So it's like, okay, well you could use for like UGC content, you could use it for like coming up with product images, for logos, for newsletter banners, whatever. It came up with all of these ideas and inside of the image ad that it's creating, it would go to AI box anytime it had to generate one of the actual graphics inside of its image. So like the, you know, the graphic of the logo or the graphic of the UGC person, whatever the actual image was, it would hit AI box, it would generate the image, it would pull inside of its own thing. So I didn't have to do anything. I literally just said, go. You know, we have the concept, we have the template that I like. Go create 20 variations, think of a bunch of good use cases and use AI box for all the images. And the cool thing is when, when Claude was making this specific kind of like UI mockup, because it's UI and it's not just like I went to, to chat GBT and got it to generate like one singular image which isn't perfect and the text might be funky and someone might be a little bit off or whatever, right? If you're just doing an image because Claude was kind of just creating it with code and then pulling an image into it, I said, okay, this is awesome. Now go put this actual graphic onto our website and animate it. So like where the chat bubble is, I'm like, have someone typing that out. Have the bubble up here, have a loading screen that makes the image pop up. So it's so cool because Claude is able to build these things and because we're just pulling the image in with AI box, it's not just a picture, but it's a full element that can be animated for the landing page. Then it can be turned into an ad for the Facebook ad and you could actually have the animated as a video ad anyways. So many possibilities. But if you're doing anything with Claude, which doesn't have audio, video or image capabilities, just go get the AI box MCP, it's like 7,99amonth I think, or 899amonth and you get access to over 80 different AI models. And anything you need while you're using Claude, it can go and grab that and pull it in and save you so much time. No more back and forth. So go check it out. There's a link in the description to AI box, AI, slash MCP if you want to get started with that. AWS is committing $1 billion to a new team of engineers who are going to be embedded inside of customers or in their customers companies. And they're going to be building custom AI agents tailored to each business and they're then going to hand off the working system when the project ends. There's been a ton of these companies being spun up by Anthropic and OpenAI. They're spending, they're basically making these like forward deployed engineers. So you send the engineer into your customer's company, they build some sort of automation workflow. And I think it's kind of like OpenAI and Anthropic's way of saying, hey, look, we know you guys want AI, but you also like don't know how to use it. We are the experts. We'll just send some people in there to build this thing for you. They're directly copying the playbook that OpenAI spent $4 billion on and anthropic spent $1.5 billion on so far. And they both have like, in those cases, OpenAI, Anthropic both partnered with some private equity firms to actually fund the staff for all of this FDE teams. Anthropic is doing it entirely with internal Amazon resources instead. And I'm going to be honest, this might be an interesting strategy if I was any of these big AI companies and you Know, like let's say Meta for example. And you were thinking about doing layoffs. Well, maybe instead of Meta doing layoffs, they should build one of these forward deployed engineer teams, get a bunch of money and go and get, you know, Meta embedded into their customers, businesses and just deploy the engineers over there. Anyways, I'm not sure if this is Amazon getting around layoffs, but it is an interesting strategy. You can imagine because OpenAI and Anthropic definitely didn't have any extra engineers that to partner with people for this. AWS has a really big advantage, I would say, structurally, because when the project ends, the customer runs AI agents on AWS's cloud and that's basically just creating long term revenue. OpenAI, Anthropic, Make Money per token used. That's a lot smaller of a payoff per engagement. But the cloud is a really big win for aws. Also, I say all of this like aws is copying OpenAI and anthropic. Like they pioneered this, but this is actually something that Palantir has spent the last 20 years doing. And so I don't think this is, you know, something that's just brand new in the situation. It's, it's something that's been done in the past and it's, it's kind of something that's been successful. We'll see how big Amazon is able to scale this. Apple has just won regulatory approval to launch Apple Intelligence in China. They're partnering with Alibaba to power all of the AI features there. It's interesting, right, because Apple said, hey, look like they were going to have their own AI and Siri and yeah, that probably would have been a big pain for them with China. But then they said, look, we're just going to say anyone with an iPhone can go and use whatever AI model they choose. So in China that's going to be something from Alibaba and in America that's probably going to be Anthropic or OpenAI or whatever. The model that you probably pay for that you have premium of, you could probably plug that straight into Siri. I think this matters a lot because China is Apple's second largest market, $20.5 billion in SAL last quarter. And I think being able to close that gap helps keep Apple their number two smartphone position. There's rivals like Huawei. Alibaba's Coin model is going to handle text and image understanding and generation across all of the different, you know, iPad and Mac and Vision OS. And Apple's exploring a deal with Baidu Deepseek and ByteDance, I think they finally have settled on Alibaba after kind of talking to all of the other options. Alibaba's U.S. listed shares went up 6% when that was announced. This deal is basically going to give Quinn some built in distribution to the very new iPhone that will be sold in China. I'm going to be honest, I've tried some cool, some Quinn models and been really impressed. Their text to speech model in particular is really good. But kind of the bigger story for me on all of this is that for not just Apple, but for any of the Western companies, China is not going to prove any sort of Western AI service if there isn't a local model partner that's locked in and that is powering everything, everything. Meta is launching a cloud business called Meta Compute. They're going to be selling AI compute powered AI models to other companies and they're doing this directly to compete with Amazon, Google and Microsoft. This and all this is right after Meta committed a whopping $182.9 billion to AI infrastructure, which is basically the SpaceX strategy. They're turning all of their excess data center capacity into revenue. It's interesting, right, because Meta, I think hopes that, or would have hoped that their AI model was way more used and way more popular. But having all of the extra AI infrastructure is, I mean, basically a gold mine. It's money that they're not spending and now money that they're actually making. Similar to Grok, I think Meta is going to sell both raw computing capacity, basically just like Core Weave does, and they're also going to have access to their own AI models that they'll sell, including the recently launched Muse Spark model, which has gotten a lot of attention because it kind of finally pushes Meta to the forefront. The Ohio Data center that is described by Mark as a Manhattan sized project is expected to open this year and it's going to provide all the excess capacity that Meta needs to resell. And SpaceX basically showed this model works. XCI signed deals in May with anthropic, Google and reflection AI to buy compute time and they basically turned all of their idle capacity into immediate revenue. And people paid a lot of money. I think Anthropic is paying over a billion dollars a month for access to that. For companies with deep pockets, this seems to be a good strategy. If at any point Meta's AI models get super, super popular, right, they can go and use their own capacity, but if not, they're just going to make all of the money from all of the other AI companies. And I think at the end of the day, the demand for AI is not going to decrease. So it's a pretty smart bet. Guys, that was everything for the podcast today. Thank you so much for tuning in. If you enjoyed today's episode, make sure to go check out AI box AI like I mentioned, where you can get our MCP that gives you 80 different AI models. And more importantly, if you're using Claude, you get images, audio and video generated right inside of Claude. It's like actually magical to use. I've been using it all day. It takes two seconds to connect and it's only $8. So please, I beg you, if you have Claude, you have to try this out. It will just make your life so much easier. Also, if you want to get all of these different stories that I talk about on the podcast here straight into your inbox, go check out aichatdaily. Com. That's my website, my news site that is tied to this. I have deep dive articles on every story that I covered here and I also have links over there where you can go and read more about it. There's a subscribe tab at the top of that website where you can get this all as an email or you can go get the deep dives on the site. Guys, thank you so much for tuning in and I will catch you in the next episode.