Latent Space AI

Exploring the Shutdown of Atlas Today

12 min
Jul 10, 20268 days ago
Listen to Episode
Summary

The episode covers major AI industry developments including OpenAI shutting down Atlas browser, Meta launching MuseSpark 1.1 coding model, Sunrun's distributed AI compute initiative, SK Hynix's record IPO, and Fiji Simone stepping down from OpenAI due to chronic illness. The host also discusses a Chinese competitor cloning his AIbox startup.

Insights
  • OpenAI is consolidating products by moving Atlas features into ChatGPT app and Chrome extension, mirroring Claude's successful multi-platform strategy
  • Distributed residential AI compute via solar-powered homes could circumvent 70% public opposition to centralized data center construction
  • High bandwidth memory production is a critical bottleneck with only three global manufacturers (Samsung, SK Hynix, Micron), creating massive valuation opportunities
  • Meta's aggressive internet scraping and low-cost model releases signal serious competitive intent in the agentic AI market
  • Regulatory approval delays (40-60% rejection rate) are driving alternative infrastructure models like Sunrun's distributed compute approach
Trends
Consolidation of AI tools into unified desktop apps and browser extensions as the dominant UX patternDistributed computing infrastructure as solution to data center regulatory bottlenecksAggressive competition in low-cost coding models (Claude Haiku, GPT-4 mini, MuseSpark) driving commoditizationHigh bandwidth memory becoming strategic chokepoint in AI hardware supply chainChinese companies cloning Western AI SaaS products with minimal differentiationFederal government incentivizing domestic semiconductor manufacturing through subsidies and negotiationMark Zuckerberg's return to public communication signaling Meta's renewed AI ambitionsAgentic AI models with computer use and multi-step task capabilities becoming table stakes
Companies
OpenAI
Shutting down Atlas browser nine months after launch, consolidating features into ChatGPT app and Chrome extension
Meta
Launched MuseSpark 1.1 coding model priced to compete with Claude Haiku and GPT models for enterprise automation
Anthropic
Claude product mentioned as competitive benchmark; will potentially purchase compute capacity from Sunrun
Sunrun
Paying customers to host AI compute nodes in homes powered by solar/batteries; completed pilot, planning to resell to...
SK Hynix
Raised $26.5 billion in largest foreign IPO in U.S. history; manufactures high bandwidth memory for AI accelerators
Samsung
One of three global manufacturers of high bandwidth memory; subject of U.S. Commerce negotiations for domestic chip f...
Micron
Idaho-based high bandwidth memory manufacturer; pledged $250 billion U.S. investment in response to SK Hynix IPO
Google
Gemini side panel will compete with new ChatGPT Chrome extension for web page analysis and task completion
Perplexity
AI-native browser positioned to potentially gain market share as OpenAI and Claude consolidate into apps
NVIDIA
AI accelerators powered by high bandwidth memory produced by SK Hynix, Samsung, and Micron
AMD
AI accelerators powered by high bandwidth memory produced by SK Hynix, Samsung, and Micron
AIbox365.com
Chinese competitor cloning host's AIbox.ai startup with identical landing page, pricing, and redirects to aimirrorhub...
People
Fiji Simone
Stepping down to part-time advisor after three-month medical leave due to chronic illness (POTS) relapse
Greg Brockman
Expected to take over responsibilities previously held by Fiji Simone following her medical leave
Mark Zuckerberg
Broke three-year silence on X to promote MuseSpark 1.1 model; first post since July 2023
Howard Lutnick
Negotiating with SK Hynix and Samsung to build new chip factories in the United States
Kevin O'Leary
Mentioned as source claiming China funds campaigns opposing data center construction
Quotes
"They basically just got Jin Yang'd where I mean, he pitches, you know, Uber, but it's Chinese Airbnb, but it's Chinese. We now have AI box, but it's Chinese and everything else is the same except that it now points to their website"
HostMid-episode
"I'm proud to say that even with the ability to go and clone our site, we still beat them on pricing. They are $9. We're $8.99. So we're one cent cheaper."
HostMid-episode
"today we're releasing MuseSpark 1.1 a strong agentic encoding model at a very low price it's available through our new meta model API"
Mark ZuckerbergMid-episode
"70% of Americans oppose new data center construction. You know, they're worried about like pollution and noise and power drain."
HostMid-episode
"At the end of the day, nothing is more important than taking care of yourself, your family, your help. There are much bigger things than AI and technology."
HostEnd of episode
Full Transcript
OpenAI is shutting down their Atlas browser nine months after it launched, and they're folding the features into ChatGPT, into a native app, and into a Chrome app. Meta is launching MuseSpark 1.1, and this is directly competing with Claude and GPT 5.6 on coding. Sunrun is going to pay customers to host AI compute nodes in their home, so you can turn your home into an AI data center. and Sky Hinnix has raised $26.5 billion in the largest ever foreign IPO in the US market. Fiji Simone is stepping down as OpenAI's application chief because of chronic illness. The craziest thing just happened to me yesterday. If you listen to the podcast, you know that for the last couple months, I've been talking trash about a bunch of Chinese AI companies on my podcast, but they have finally proved me wrong. So yesterday I discovered the incredibly innovative AIbox365.com. So essentially, I was getting Claude to do SEO for me, asking it, you know, a bunch of different tips and strategies for AIbox.ai and my own startups SEO. And it identified for me, it said, look, there's a red flag because there's another startup called AIbox365.com. It does exactly what you do. So it's a company, they went and they registered our exact name. They added 365. They cloned our landing page and our pricing. It's like the same pricing. But the only difference is if you click the sign up button on their website. it doesn't sign you up right there. It actually redirects you to a completely different site called aimirrorhub.com. And that one has a bunch of Chinese characters on it. I think the site is actually like translated from Chinese into English just automatically on Google, how you know how Google does that. If you've seen Silicon Valley, you know, they basically just got Jin Yang'd where I mean, he pitches, you know, Uber, but it's Chinese Airbnb, but it's Chinese. We now have AI box, but it's Chinese and everything else is the same except that it now points to their website, just basically poaching off of our SEO. So how do we stack up against this incredibly innovative Chinese competitor? Well, I'm proud to say that even with the ability to go and clone our site, we still beat them on pricing. They are $9. We're $8.99. So we're one cent cheaper. They have 10 models on their platform. We have over 80 different models on our platform. And if you go and get an annual subscription from them, it's $109, where our cheapest annual subscription is $86.99. So we're beating them on annual pricing, we're one cent cheaper on monthly, and we're not going to steal your data. In addition to the fact that we have an MCP, so you can take any of those image, audio, video models and stick them straight into Claude or ChatGPT or any AI tool you use, you can get all the other AI tools built into it as well. So if you're interested, it's trusted by over 3000 teams. If you're looking for every model in one place, do not go to our Chinese knockoff competitor. Go to AI box dot AI I leave a link in the show notes OpenAI is shutting down Atlas This is a browser that they launched nine months ago and they moving all of the features from that into a Chrome extension And they also upgrading their chat GPT desktop app instead I mean this is basically what Claude already has with the Claude app that has like all of their different Claude code and Claude co-work and even the chat all inside of one app. It looks like OpenAI is trying to consolidate and do that same thing. And I even saw Claude today launch the fact that they have a browser built into that. So I wouldn't be surprised if chat GPT did something similar and had some sort of browser built in eventually to the chat GPT app. They want to consolidate. They don't want multiple apps. And Claude also has a Chrome extension. And so it looks like OpenAI is basically following Claude's playbook. It works really well. I use both the Chrome extension and the app on my desktop. And sometimes they trap back and forth and they use each other. But anyways, this is, I think, the way to go. And OpenAI is copying that. The new chat GPT Chrome extension is going to compete directly with Google's Gemini side panel. And I think it's going to basically answer any questions that you have about a web, any specific web page, so I can look at the content on that page, it can handle multi step tasks, you're not leaving the browser. The other cool thing that they do is they're building a cloud browser. So it's going to run on their own servers. And basically, it lets chat GPT agents complete tasks for you while your desktop app is staying responsive. There's a couple different cool things here. One is like, I love that Claude Cowork, for example, uses the files on my computer. Or, you know, let's say Claude, the Claude Chrome extension it can use and like take care, take control of my screen. But a lot of times I just want something done and I don't want it to mess up with my screen or what's happening on my computer. I just want it to go happen in the back end. So that looks like what OpenAI is building here. But this is now leaving room for the browser company and for perplexity that both have their own AI native browsers to potentially take more control of the industry. Meta has just released Muse Spark 1.1. This is a coding model and it's designed to build features and manage workflows inside of enterprise software without any human intervention. At 1.25 per million input tokens, it's basically priced directly to compete with Claude Haiku 4.5 and GPT 5.6 Luna. So these are kind of the dominant cheap coding models. So it's not, you know, the best. It's basically trying to compete with the cheap coding models. The thing that I love about this is that Mark Zuckerberg has broken his three year silence on X to promote this. This is his first post since July 2023 back when I think threads dropped and when threads dropped it's kind of funny because he was making a post about threads and it was you know cloning x specifically what he said about this he said today we're releasing MuseSpark 1.1 a strong agentic encoding model at a very low price it's available through our new meta model API and meta API MuseSpark 1.1 is the strongest at agentic performance tool use and computer use it does well on long running tasks with a million token context window and can delegate execution to sub agents running in parallel and is trained to use computer interfaces on desktop mobile and browser the meta model API allows developers to build using Muse spark for the first time our focus is on delivering strong agentic and multimodal models at a very low cost more to come soon honestly I pretty stoked that we getting Mark Zuckerberg tweeting again But also I think this is a good model I excited to see what they do with this And a lot of people have been pointing out that like just from internet traffic, meta is scraping like crazy. They're obviously building some big things. We don't know exactly what but I'm excited to have another competitor in the ring. A company called Sunrun is paying customers to host AI compute hardware in their homes, aka little mini data centers at your house. They're trying to basically spread the AI workload across 1.1 million residential solar and battery systems instead of building massive data centers. There's a couple interesting things right now. One of them, I think it basically gets them around a huge problem, which is that 70% of Americans oppose new data center construction. You know, they're worried about like pollution and noise and power drain. There's also a ton of like political campaigns going up. Some people have accused China of funding a lot of these different campaigns. Kevin O'Leary said that's the case in many instances. So being able to just, you know, not have to worry about getting one of these things approved because I saw a stat that it was like 40 to 60% of all data center projects don't get approved on the local level because of, you know, people pushing back or whatever. So if you could just go directly to consumers and if everyone had a little mini data center on the side of their house, I've seen people, you know, float like, hey, you get free electricity and other different benefits to it. So if that was the case, I guarantee tons of Americans would sign up for free electricity to host a data center on the side of their house. And then you wouldn't have any of this kind of regulatory burden as well. Sunrun has completed a proof of concept pilot and they're planning to resell the compute capacity to AI companies like Anthropic and OpenAI within the coming months. Home-based nodes like these little mini quote unquote data centers will essentially run partly on rooftop solar and also batteries. and then they're going to basically reduce grid draw compared to a single warehouse scale facility which is drawing hundreds of megawatts and I think this is kind of the concept that a lot of people have talked about with billions data centers Trump has talked about it a lot saying hey look if you're building the data centers you should go and build like your own energy generation plant next to it basically telling people they should build their own gas electric solar wind whatever you want like energy generation should be built into these so that it's not pulling off of the grid and if you do this the way that they're doing this now, they're going to bring their own batteries, solar, and it would just basically increase what is on the grid and not take from it. A lot of the top AI labs are trying to get these data center deals. It's basically a land grab right now. And Anthropic just signed a 20-year Kentucky facility lease, but the permits are a huge bottleneck and it hasn't actually solidified yet. SK Hinnix, I've been talking about them a ton on the podcast lately have just raised $26.5 billion. This is the largest ever foreign IPO that's happened in the United States. I've been talking about this as it's been coming up. But the thing that's interesting is that they actually beat Alibaba's 2014 record. And Wall Street is bidding really aggressively for exposure to this one particular chip maker because they make high bandwidth memory which basically powers NVIDIA AI accelerators and AMDs as well And what interesting is there only three players that make this particular high bandwidth memory Samsung is one of them SK Hinnix is another And there's one called Micron. I believe they're based in Idaho, but there's only three of these companies in the world. Their stock prices have absolutely been soaring. The companies are raising and making tons of money. It's absolutely eye watering. And so right now They've decided that they're going to go public and capitalize on this. They sold 1.779, so 177 basically, or 178 million shares at $149 each. It was priced 2.7% above their South Korean average, and there was demand for more than seven times the available supply. The stock opened 14% above the IPO price Friday. The ticker is SKHYV. It then started doing regular trading on Monday and under SKHY. U.S. Commerce Secretary Howard Lutnick said that he's already negotiating with SK Hennix and Samsung to both build new chip factors in the United States. We already have one in Idaho, but, you know, one company that's based in Idaho, but we want to bring everyone else in as well. I think basically this just shows that in the United States, there's a ton of appetite for AI infrastructure, even if it's not in the U.S., even if it's not based in the United States. It's like, I mean, essentially by investing in this IPO, you're investing in hardware companies based in Korea, which isn't a bad thing. Washington right now wants to rebuild a lot of the domestic memory manufacturing, which all of it has kind of been abandoned like year, you know, in the last 10 or 20 years. Micron fired back with a $250 billion US investment pledge. A big part of this, I think, is federal subsidies and talent. Fiji Simone is stepping down as OpenAI's number two executive, and she is moving to part-time advisor after a three month medical leave. So she had consolidated products and a lot of business operations under her since May of last year, but she had a really severe relapse of a chronic illness. And that led her to prioritize recovery ahead of opening eyes IPO that's happening next year. She was diagnosed with pots. And Greg Brockman, I think is gonna take over a lot of her responsibilities. The reason I wanted to bring this up because I know this is sort of just like a random thing to throw on the end of a podcast. I mean, it is the number two person opening eye. But my point being life is short. All of this and AI, everything happening in the markets, all of this, you know, technology, it's super exciting and interesting. But at the end of the day, nothing is more important than taking care of yourself, your family, your help. There are much bigger things than AI and technology. Even though this is what we talk about every day. I think she obviously realized that life is short. I'm really, you know, praying for a speedy recovery for her and for anyone that is suffering from illness. All right, guys, thank you so much for tuning into the podcast. If you enjoyed this episode, make sure to go check out the real AI box dot AI, not the fake one that is based in China. And also make sure to leave a review. If you haven't left a review already on the podcast, it helps the show out so much. I am eternally grateful and I hope you guys all have a great rest of your day.