Analysis of GPT-6 Service Failures  🔗 **Source:** [Analysis of GPT-6 Service Failures](https://share.snipd.com/episode/5e0b1c5d-551b-41bf-a98b-0eeb14eddc56) ## Highlights - **GPT-6 Astra Targets Computer-Using AGI** - GPT-6 Astra is positioned as a major step toward AGI, especially in computer use, coding, science, and cybersecurity. - OpenAI emphasized 98.6% on ARC AGI 3, perfect ExploitBench performance, and additional safety testing before release. Transcript: Jaeden Schafer So the GPT-6 release is obviously exciting. There's obviously a ton of hype on it. And it's definitely trending over on X. It's I think one of the top trending things today. Apparently, they trained this on over 1000 GPUs. They did it at the Texas Stargate site. Brockman called it a generational leap towards artificial general intelligence. And the model is doing really, really well on a bunch of different benchmarks. Specifically, it got 98.6% on the ARC AGI 3. And it got a perfect 100% on exploit bench. I kind of I think I covered that yesterday. Basically, it's really good at hacking. It's good at coding science and computer use computer use being one of the big things that they've highlighted. And I think one of the winning categories for AI models today, the things that I personally care about are basically computer use, because today I'm using both Claude and chat, Claude Co-work and chat GPT work. And I'm just telling them to go do things on my computer. So that computer use is basically everything. If it's not good at that, it's not something that I'm going to try. I've heard that GrokBot might be good at this. I haven't given it a lot of a lot of work. I'm never usually like the first early adopter. I'll wait till it's getting a little bit of traction. And I know it can handle things. I feel like I've been burnt. I think a lot of people might be in that same boat where we feel a little burnt, trying things for the first time with a lot of over promising and under delivering. But apparently this is rolling out to a bunch of selected organizations that have quote unquote, daybreak access. That's specifically, I think, for the model that's really good at hacking, then subscribers are going to get it. And Sam Altman had a tweet where he said, we hope it will begin to enable a new generation of entrepreneurship, scientific discovery and building. We believe it is the best model in the world for computer use, professional work, science, coding, cybersecurity, and more. I mean, obviously talking up his own book. He also said, it took us some extra time to ensure that we could meet the safety and alignment standards required for this capability level, but we think you'll find it worth the wait. ([Time 0:02:20](https://share.snipd.com/snip/2c8523d1-2df9-4765-9729-f6dfe3b4684f)) - **GPT-6 Turns Voice Commands Into Real Projects** - A launch demonstration showed voice-controlled GPT-6 transforming a drawing into a rocket, building it in Blender, and creating a printable 3D file. - Other examples included making a video game, scheduling tennis, and ordering a previously purchased meal. Transcript: Jaeden Schafer I actually watched a demo video, which I was really impressed with, where they essentially have a person sitting down in a room, they're pointing at a screen and they're like, in this One particular use case, they're like, draw like a draw yellow circle, it goes in and draws a yellow circle. And then the it says like, and turn that into like a rocket ship, it turns it into a rocket ship. Anyways, they have a whole bunch of different, you know, use cases that they're, they're showing what it's able to do. I think there was this comes from a very famous demo in 1979, where someone was showing off a computer, and it could kind of follow his his finger while he was pointing the remote at a screen. And so they kind of replayed the same is the same kind of demo idea. So honestly, I think it's a good ad. And in this one, they had the guy just like talking to chat to be voice, which by the way, I think is a big unlock. If you haven't used it already, you don't need to sit there typing, you can just have voice enabled and you can talk to it and it can take feedback and make changes live while you're going. So in any case, then a person in chair looking at ginormous projector screen in front of them, he's like draw a yellow circle, it draws a yellow circle, then he's like, okay, turn that Into the window of a rocket ship and makes it into the window of a rocket ship. He says like, okay, go open blender and turn that rocket ship into a 3d rendering of a rocket ship. It goes and does that. And they have a whole bunch of different people doing different things of some celebrities and stuff. You know, people saying like, go schedule me a tennis appointment. And right, they got all these different use cases. The blender one I did like because he turned it into a 3d rocket ship. And then he turned it into a video game where he explained what the video game was doing. They have like lawyers and a bunch of other people using it. They're trying to show a bunch of cool use cases for the video game one while it's turning making this video game for him. He's like, hey, also go, you know, order food from that restaurant I ordered from last week. And then it's like, okay, looking at what you ordered last week, and it's going to get in it. They're obviously using like the the DoorDash plugin and all that kind of stuff to get it. And then he said, okay, take that rocket ship video game that I made. It like shows him playing his own video game, which obviously is not very realistic because this is a boring use case. You can make really exciting video games. This one's pretty boring. But then he's like, hey, go make a 3D, you know, make the right file that I can make a 3D print of that rocket ship. And it does it. So kind of cool, right? Shows a bunch of cool use cases. ([Time 0:04:16](https://share.snipd.com/snip/9fae3392-1aab-4a10-bf0f-05a3f0853aac)) - **Agentic AI Wins Through Low Latency** - Low latency may matter more than benchmark scores for agents that operate computers interactively. - Jaeden Schafer found GPT-6 Work powerful but noted browser glitches that sometimes required restarting the computer or software. Transcript: Jaeden Schafer This thing's very powerful. The thing that a lot of people are talking about with this and this entire demo I've just explained is the low latency instantaneous work. So as you talk, it's doing things very quickly. If you've used the chat GPT co-work, you know, this thing is very powerful. I think it's on par with, with, with Claude co-work or maybe a little bit better in a few areas. I've had, it had a couple of glitches with using my browser that Claude doesn't always have so I'm not sure what the pros and cons of of that are usually if I restart my computer and restart Software it works so a couple glitches perhaps if I'm being 100% honest but otherwise a like really powerful piece of technology I use it I would use it more often if it didn't have that One particular glitch. A lot of people are tweeting about what they're actually doing with it. Some of the things that I've been excited about is someone had like, they told it to go to Blender and make like an entire 3D replica of like a world and have all of these like characters That it pulled in and they're all talking to each other and they have to like, make it like a, you know, it's like kind of like a video game, but they're like making their own world and they Have to work together to build like houses and stuff and they're like doing it and talking and it's like pretty impressive some of the things that it can do in addition i saw some people Well one of the big funny ones is bojan tungas said where were you when agi was released to the public you know it's you know hyping up that this is the this is AGI ([Time 0:06:24](https://share.snipd.com/snip/98994975-1038-4370-97a8-f78f36f689ce)) - **AI Outages Expose Infrastructure Concentration** - Near-simultaneous outages at ChatGPT, Claude, and Grok exposed shared infrastructure dependencies rather than proving a conspiracy. - A Memphis compute-center outage reportedly affected both Grok and Anthropic, highlighting the risks of concentrated capacity. Transcript: Jaeden Schafer Now, I think what is an important element of all of this is the fact that all of the big models, I guess switching gears slightly to a new story, all of the big models, Chai Chippity, Claude And Grok all went down within 90 minutes of each other today at the same time. So on the one hand, we have this big, huge release from OpenAI and then everything kind of crashed. Now, I think with the OpenAI crash in particular, maybe it was just a coincidental timing that, you know, it crashed and went down because they have their big model release. Perhaps that pushed a lot of traffic over to Anthropic. And then what I think happened because we had SpaceX come up with a tweet a couple hours ago where they said, we are sorry for the issues you may have experienced with Grok following the Outage at our Memphis Compute Center this morning. We'd like to also apologize to our impacted compute partners, aka Anthropic, who pays them almost a billion dollars a month to use that same Memphis facility. They said all systems have now been restored and are functioning normally. So I think it could kind of be like this domino effect where there's so much usage that if ChatGPT goes down, perhaps because they have a big model that got released, maybe it was more popular, Whatever happened, maybe it was a big marketing stunt by Sam Altman to like make ChatGPT go down right after they launched a big model just to say that's how many people were using it. And I know you are rolling your eyes. But like, I remember distinctly when he distinctly being the word when Sam Altman put a tweet out, think it was maybe around chat GPT 5 when it came out or maybe it was chat GPT 4.5 and he Put a tweet out where he's like there's a distinct smell when you when silicone is melting or something after they like released a new model to say like their servers were like melting And obviously it's just a marketing thing and it's like funny or whatever but so anyways I haven't I haven't forgot that and wouldn't put it past Sam Altman to crash them a little bit on Launch. Okay, I don't actually think that happened. But basically, we're having this domino effect where one model probably went down. And then we had probably a lot of traffic go to Anthropic, potentially that took down the Memphis data center. I'm not sure. But one way or another, the Memphis data center went down, it took out Anthropic and Grok, a couple things, right? If you look at a lot of the comments on this, people are like, that just proves that like you can't bet against SpaceX because all the infrastructure runs on them. Okay, I also think it proves that like putting all of our eggs in one infrastructural basket is probably not a great idea. I wish probably diversify the data center portfolio pool here. But it is true. It's really hard to get new data centers built. SpaceX has taken a very aggressive approach, right? They're spinning up new data centers really quickly. And they're, you know, sometimes they'll like this Memphis one in particular, they spun the whole thing up, they got a bunch of gas turbines. I mean, they did some really incredible things to get this whole thing spun up. They like rented all the gas turbines in America brought them all the all the mobile cooling in America, they rented it all, brought it all to one place and spun up this whole data center. And everyone said they're kind of crazy. And you could conceptually say this cost them a lot of money. Maybe it was a big waste. They're getting a lot of lawsuits because of all the gas, yada, yada. But at the same time, they turn around and are selling that compute back to Anthropoc for a billion dollars a month. So undoubtedly, they're making some money off of this thing. But I do think we need some diversification in the market. ([Time 0:07:46](https://share.snipd.com/snip/28e278f0-cfc7-4dde-a80e-5d1e39f58607)) - **OpenAI Sacrifices Cursor Revenue Over Ownership Conflict** - OpenAI ended Cursor access after SpaceX acquired it, despite a partnership projected to generate more than $1 billion annually. - The decision may cost enterprise users because companies often standardize developer workflows around Cursor. Transcript: Jaeden Schafer Okay, OpenAI, I can't get away from the OpenAI story without talking about something that happened a couple days ago. But I just so much that so much juicy drama came out of it, which is that OpenAI walked away from a billion dollar cursor deal. The reason why I say it's a billion dollar cursor deal is because this particular partnership having OpenAI on cursor was projected to bring over a billion dollars annually. But rather than servicing that, rather than letting Cursor have OpenAI models on there, they decided to cut contact with the company because obviously they don't like Elon Musk and He owns SpaceX, so they don't want to help them out. Getting into the actual tweets, it came from OpenAI who had a tweet that they said, entering we're ending our partnership with cursor following its acquisition by spacex under our Proposal cursors direct access to our model would end on november 12th i think a lot of people talked about this and then they're like we know the the people most affected by this decision Are the developers who rely on opening eye models in cursor we care about their experience in this transition and we're ready to go above and beyond to support them blah blah blah I mean Basically they're like look come over and use our platforms and if I'm being honest I don't think it's as bad for open AI as a lot of people are making it out to be they're like look they're Going to lose a billion dollars but I think there is a percentage of people that maybe do like open AI models I'm one of them I like I really do use open AI like their model they have a couple More glitches than Claude I'm going to honest. But I do like their models and I get more done. I actually think you might get a little bit more usage out of OpenAI for the same intelligence level. And I really like ChatGPT work. So I think they'll have some percentage of people move over there if you really like it. But honestly, at the end of the day, I don't think a lot of people, maybe there's a percentage, but a lot of people don't have loyalty. You just switch back and forth between whatever model gives you the best bang for your buck, especially if you're a developer. So in this regard, I'd be curious. The one thing that I do think they're going to be losing out on is there's a ton of enterprise customers. Like I have a lot of friends that work at companies where they have licenses to cursor, right? Their company wants to be able to access all of the top AI models, not just one. So just go get a license to cursor. And so if opening eyes getting ripped out of cursor, like your company isn't approving you for a new thing, they're just like, just keep using cursor that we've always been using, you Have all of your workflow and stuff set up there. So they are going to lose like a big chunk of customers and users by doing this. ([Time 0:10:44](https://share.snipd.com/snip/c748288b-f820-4410-8aba-78de4961dd09)) - **Cursor Shrugs Off OpenAI’s Five Percent Share** - Cursor CEO Michael Truell responded that OpenAI models represented only about 5% of Cursor traffic. - His response framed the cutoff as manageable because Cursor already supported other models, especially Anthropic’s. Transcript: Jaeden Schafer This is Michael Truel. He's kind of a gangster if I'm being honest, because in building Cursor, like the OG team, all of them are multi, multi millionaires. Now, a lot of the early people that joined the company are like a hundred million, you know, some of them made a hundred million dollars. Some of them are making, you know, tens of millions of dollars on this whole sale to SpaceX. So he had an incredible outcome for his team. This is what he said. He said, we're sorry to see the OpenAI put out a note saying they plan to block cursor users from accessing OpenAI models in three months. OpenAI models serve about 5% of cursor user traffic. And we're speaking with the OpenAI team to resolve this. Cursor was one of the very first users of OpenAI. We've worked closely with their team for years, and we've trusted their platform to be neutral infrastructure for our business. It's kind of true. It's kind of shady just because you don't like Elon Musk. You can't like work with a company that's owned by a company that's owned by a company that he owns. Like, I mean, it's just like at some point, I don't know, like Sam Altman still tweeting this and OpenAI is still posting this on X. So they can post it on X, which Elon Musk owns, but they don't want to service their customers through Cursor. I don't know. I think it's kind of ridiculous. And I think that Michael threw a little bit of shade at them being like, OpenAI models serve only about 5% of the traffic on Cursor. So, you know, he's like, look, it's not the end of the world. We have lots of other models. Obviously, Anthropica is goaded in this regard. But yeah, so kind of funny. ([Time 0:13:05](https://share.snipd.com/snip/081c2153-4ee3-4065-9377-c88900c1139a)) - **Meta Reenters The Frontier With MuseSpark** - Meta’s MuseSpark 1.3 reached third place on Artificial Analysis’s intelligence index, driven mainly by agentic and scientific improvements. - Meta’s 95% discount for training-data sharing reflects how model providers exchange cheaper access for valuable user data. Transcript: Jaeden Schafer Meta is offering a 95% discount on MuseSpark for users who are going to share their training data. This is I mean, the new model that Meta came out with is genuinely impressive. I like Alexander Wang, he was tweeting about it on x. But there's a tweet also that came out from artificial analysis. That's the number one benchmark I use if you haven't if you don't know much about benchmarking. Artificial analysis publishes the results of a lot of different benchmarks, but they also have their own benchmark. Specifically, they'll have like a benchmark where you get free access to AI models, but it asks you two questions side by side and you pick the response you like the most. And then they basically make this benchmark where they're like, and it's blind. You don't know what model said what. You just ask a question, see two responses, pick your favorite one. And in that blind test, they're able to say, look, users prefer responses from this model over that model. So it's kind of a it's a cool blind test benchmark that I really do like. Anyways, artificial analysis said meta has released Muse Spark 1.3, their fourth Muse Spark model release in five months. Muse Muse Spark 1.33 max which is in limited preview for meta partners across scores 62 on the artificial analysis intelligence index behind only claude fable 5.1 and claude opus 5 The variant the variant gains come primarily from improvements in agentic work and scientific capability okay so basically they're saying in this particular artificial analysis In the intelligence index, it's scoring number three, just slightly behind Claude Opus 5. And, you know, it's four points behind Claude Fable 5.1 max, which just came out. So I mean, we have the brand new model from Anthropic, which is it's like right behind there. And so they're really at the frontier. I've seen a lot of people talking about the fact that people kind of were betting on meta to fall behind. And it seems like they were able to pull through because they have a really incredible talent pool, say whatever you want about meta, they have incredible engineers, they obviously Have a lot of money coming in from Facebook ads and Instagram and all that kind of stuff. And then they have a lot of money they invested in compute. And they invested a lot of money into researchers for AI from the very beginning of this. So yeah, I'm not I'm not surprised that they're able to pull ahead of, you know, even Claude Fable 5. And they were able to pull ahead of GPT 5.6 and Grok 4.6 and Kimi K3 and GLM 5.3 and Gemini 3.8 flash. They're able to pull ahead of all of those. So great work to everyone over at Meta. And the other thing that I was mentioning earlier in this news story is that they're offering a 95% discount. Look, and that's if you're sharing your train, you're like, whatever you're talking about as training data, look, this is a great way to get more data. And lest you think Meta is the only person doing this. It's interesting, they're offering a 95% discount, I get a 100% discount on a lot of my on some different projects. I have different chats and API chats for OpenAI. They've been doing this for over a year. So I have certain things that I work on that I get free OpenAI access. And it's honestly a good way to go, especially if you're offering like freemium models for tools and stuff. ([Time 0:14:25](https://share.snipd.com/snip/0e64abb8-e771-4680-bf11-35c1240be2ee)) - **NVIDIA Buys Hugging Face Without Forcing Its Hardware** - NVIDIA’s acquisition of Hugging Face gives it influence over a major open-model community of more than 18 million users. - Hugging Face will remain platform-neutral, although NVIDIA integration will likely become especially seamless and prominent. Transcript: Jaeden Schafer NVIDIA is to acquire Hugging Faces. This has kind of been in the news for a day or two, but it looks like it's all official now. The rumors have officially happened. Jensen Huang posted on LinkedIn. He said, I'm excited to announce that NVIDIA has agreed to acquire Hugging Face. Over the past decade, the Hugging Face team has built something remarkable, a vibrant home for the open model developer community. More than 18 million developers, researchers, and creators use its platforms to share models, data sets, and applications, while companies around the world use it to discover, evaluate, And customize and deploy AI. Together, we will scale Hugging Face's platforms, strengthen its infrastructure, and expand access to AI for developers and institutions worldwide. This is a big, I mean, this is a great move for NVIDIA. They're getting a lot of goodwill in supporting a project like Hugging Face that has open source models. And they've been quite vocal. Jensen's been very vocal over the last couple of months when everyone's talking about regulating AI that he thinks open source models need to be publicly accessible and available. He kind of signed this big open letter on X and had a bunch of other big industry leaders get in on this. In his note, he said, Higging Face will remain an open platform for the entire AI ecosystem. Developers will continue to choose the models, frameworks, cloud, inference providers, and compute platforms that best fit their work. NVIDIA compute will not be required to build on or deploy through the platform. This is probably an anti-monopoly, you know, clause that got put in there. I mean, it's goodwill for everybody, but it's like, they're not going to force you to use NVIDIA. However, it's probably going to be the default is probably going to be encouraged. It's probably going to be easily linked and easily accessible and set up to kind of like, be like a very seamless integration. However, you don't have to use it. And I think at the end of the day, if they're going to buy it, like they got to get some value out of it, they're going to try to get some usage out of it. ([Time 0:17:21](https://share.snipd.com/snip/7287d1f1-be3e-4eba-b6f4-4d88cf7dba98))