OpenAI Cuts Off Cursor, Nvidia Buys Hugging Face, Ox Alpha is GLM | This Week In AI

Abhi: These guys have like a billion dollar valuation.

Shane: 2.5 actually. So yeah, if you're watching this, you too can become worth $2.5 billion in a week. Welcome to Agents Hour. We do this every week and there's a lot in the news this week. There always is. We have, you know, some drama.

Abhi: The Games of Thrones. The Games of Thrones.

Shane: Yeah. Thronesing right now. All right. So let's. Let's go ahead and dive into it. Here's a little preview of some of the things we will be talking about on the show today. You can see some of the faces that we're going to be mentioning. Let's start off with this one. From OpenAI, we're ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor's direct access to our models would end on November 12th. OpenAI saying, you know what, Cursor, if you go join SpaceX, we don't really like xAI, you know, we're gonna cut off your access.

Abhi: Yeah. Is it an underlying Sam versus Elon drama here?

Shane: And then you see some more people weighing in, right? Thibault says we've unfortunately have decided we cannot continue providing access to our models through Cursor and are ending our partnership. It boils down to trust and we've asked that this takes effect on November twelfth to give you some time to plan.

Abhi: What is the trust? Yeah, where is the trust?

Shane: What is the violation from trust?

Abhi: Are we distilling stuff over there a Cursor?

Shane: Michael Truell from Cursor says. We're sorry to see that OpenAI put out a note saying they plan to block Cursor users from accessing OpenAI models in three months. OpenAI models and I think this is kind of a flex. Like he's try he's basically trying to say, you know what, that's not that big a deal. Yeah. By saying OpenAI models serve about five percent of Cursor user traffic and we're speaking with the OpenAI team to resolve this. So he's basically kind of waving it off in that like, yeah, we want to fix this. But OpenAI, you know, you're only five percent of our traffic, so big deal. But that's kind of the vibe I get from this response.

Abhi: And then also, you know, Composer is a model. And then obviously Grok will be like one of the flagship ones.

Shane: Yeah, I mean I'm guessing, right, how did Cursor train their model? Well. They use the responses from whatever model anyone selected to do that. But then you ask yourself, like, how did Anthropic and OpenAI train their model? Well, there's a lot of traffic from Cursor. And that so it goes both ways. Like you can't blame Cursor for doing it if you're doing the same thing to them. And obviously I do think that the fact that, you know, they have the compute, xAI has the compute is probably alarming that Oh, actually maybe xAI's building good models. Now they ha are collecting a lot more data because a lot of people still use Cursor. And a lot of that data is now not getting sent through op to OpenAI. So xAI could be very well positioned to really succeed in coding, right? If you're if people are using Cursor. And so I think they probably are just saying, let's cut off access and Make them fend themselves.

Abhi: That's probably what this trust thing is all about.

Shane: Someone from Anthropic chimed in and said, I can't believe OpenAI would do this. Anthropic is committed to supporting Cursor and being here for them or whatever. And then Amjad, you know, and many others, they it wasn't just Amjad said, you know, we remember what Anthropic did to Windsurf when they did something very similar. And then Amjad said the answer is you can't do this here because you need their compute.

Abhi: So That's a W post from Amjad.

Shane: Yeah, if you've been following for a while, you know that, you know, Anthropic is, you know, paying for or using xAI is, you know, or to SpaceX is data centers, right? To for some of their compute. They're basically leasing some of their compute from them, so they kind of need it. Anthropic's been very compute constrained.

Abhi: I really feel like Anthropic have just stayed out of this because like people don't forget when you were gonna cut off access to other harnesses and S DKs and stuff like that. That like.

Shane: It's funny because Anthropic could have taken this if they said nothing, it's kind of a win for them. Yeah. Right? Because they're not doing it But by s coming out and saying, Oh, we would never do that, then it just gives people the ammunition to point out all the times you've done exactly that or potentially worse. Because Anthropic has made many fumbles.

Abhi: Way too many fumbles.

Shane: And this is kind of, in my opinion, kinda an OpenAI fumble as well. But maybe they see it as a strategic, you know

Abhi: quote unquote trust issue. Or there's something else going on between the two leaders of the companies.

Shane: So let's talk about Ox Alpha, Ox Alpha, whatever you call it. It was unmasked. We guessed from some you know intel that we saw online that you all probably saw as well that it was glm 5.3 and it was confirmed. Introducing GLM 5.3 flash. This was from August 26th. It has leading capabilities at a competitive price. It's multimodal. Not all GLM models have been multimodal, I think. And it's a one million context window. It's released under MIT license and it was previously previewed as Ox Alpha running and this is the part that threw people. This one little like part of a sentence running entirely on Chinese AI chips because they were thinking. How do they get this much compute? It's running on you know chips in China.

Abhi: Y'all should be worried.

Shane: Yeah. So okay. That was the shocking. And then there's a lot of people that came out and commented. It's priced very competitively. The whole idea of it being flash is it's supposed to be fast as well. Have you tried it out yet?

Abhi: No, I have to wait for it to Get on Alibaba.

Shane: Yeah, I don't have a Z.ai coding plan. Some people on our team have tried it. I think Ward tried it.

Abhi: He's the one with the Z plan and I think Tyler has a Z plan.

Shane: So I mean I think we will be trying it more. You know, it's not state of the art, but it is getting very close. It's right there and it's fast. It's honestly that kind of surprising 'cause it's a bit smaller of a model, it's faster, but it's kind of amazing what it can do for for this size of the model.

Abhi: Yeah. It benchmarked quite well.

Shane: Sam Hogan here says if Z.ai can serve a hundred trillion tokens per day on Chinese hardware. It's safe to assume they're not far from training models entirely on Chinese hardware, too. I think that's it probably a safe assumption.

Abhi: Yeah, dude, while we're games of throning over here, they're making moves.

Shane: And then on August 28th, GLM 5.3 is now open weight. So our most capable model for agentic coding in cyber defense is now available to download, run, and customize. Just imagine a year ago if you said that a model this good would be open weights, people would have said, I don't believe it. Yeah. Like that's probably not gonna happen.

Abhi: Yeah. We would have said like open source models aren't very good and now they're slowly catching up. Yeah, quickly catching up. Quickly. And we have not anticipated their infrastructure getting better. None of this is good for us here in America. We've got our all the CEOs fighting each other and, you know, no bueno.

Shane: Let's talk about this thing called Instinct. So this is from someone named Noah Shinn, who I'd never heard of before, and came out and said, I'm the founder of Instinct. Instinct is a personal agent that we've been building for the past few months. The interface is simple, there are no new interfaces you can cook text or call it. It's trained to use a phone and a computer in the same way that humans do. Instinct combines simplicity with extreme capability, and this launch and there's a little bit more on it. This launch got one point eight million views. Like pretty impressive. This was, you know, August twenty sixth. And I don't know the whole story about the company behind it, but I did see this post afterwards because I s kept seeing people talk about instinct and saying that, you know, have you have you tried instinct yet? So someone on our team said they tried it and it one-shot it and did exactly what they wanted to do for some assistant type task, right? It's basically like a personal assistant. And then maybe they want it to be like a Jarvis type thing, right?

Abhi: I mean it's kinda like Poke. Remember when we covered Poke? Yeah. And these guys have like a billion dollar valuation.

Shane: Two point five actually. So in one week, it's funny. So this is from Austen Allred said both valued at two point five billion today. Linear, having crossed one hundred million ARR with a hundred and seventy seven percent net revenue retention. Instinct, the viral AI assistant that launched last week. What? This is the wild world we live in. You know, this company that no one had heard of before, which maybe they had been working on this for a while. I don't know. Went from launching, going viral, apparently being very good. But still being new, you know, I it's very it just reminds me of OpenClaw, reminds me of Hermes. This feels very similar. It's maybe not quite the it's a little different because it's. It feels like a bit more of a polished thing where OpenClaw and Hermes were kind of deploy it yourself, build it yourself, but still wild.

Abhi: Well yeah, you meet you meet your users where they are. And you know, if it's iMessage, you just opened up your user base to everybody.

Shane: So yeah, if you're watching this, you too can become worth two point five billion dollars in a week. It's possible. It's a lottery, but good luck with it. Yeah, good luck, have fun. But this is this is a wild story. This is also pretty wild. Are we in the world or the age of infinite video? So this went. Incredibly viral. This was at August 28th from Rehan Sheik says MiniMax-H3 generates video faster than you can watch it so I hooked it to a Twitch live stream. Now you can watch infinite interdimensional cable link to the stream below. And the stream was of course taken down. I think then he went to kick and then maybe rumble or something trying to like and everyone just took it down because it was using copyrighted material. But the wild thing is it just kept going. Like it was able to generate video faster than time. So it's always just like generating it's like a never-ending stream of video where you have no idea where it's gonna go. It's just kind of a wild concept.

Abhi: We should try making like a YouTube live and then put some fake characters and just see what where the story goes after twenty-four hours.

Shane: Another person said. And this is from August 29th. I think it was kind of saw the previous example and tried to do something similar. This is a live demo from the future. An interactive sitcom, a playable episode. Games and movies are merging into one thing. So essentially it's real-time video generation where you can guide the storyline. You can watch the video, it's like a Seinfeld type episode. It's like two minutes. But essentially it's like a choose your own adventure, essentially, a video. And it's pretty good. I mean s shockingly good. If you if you saw this, like I just remember years ago it was like the Will Smith eating spaghetti test. And it was terrible, right? And now actually just looks like the real thing. It's pretty dang close.

Abhi: Well I remember when we were doing the story time stuff, right, for making stories for kids. So you could just make kids TV, educational content, all that stuff based on their imagination. It's pretty tight.

Shane: No, I guess here's my question. So let's take a step back. Do people actually want this? I don't know. If they want to control the story, wouldn't I play a video game versus just like watching a choose your own ad I know Netflix has tried like choose your own adventure videos. And they never seem to really do that well.

Abhi: When you have to pick the decision tree. But I don't know. I think it's I think it's pretty interesting though, because You could have infinite video within the cutscenes of a game. Also, I think kids don't play games, they watch other people play. So maybe it's kind of the same thing where you're just like, you know, watching it. I don't know.

Shane: I feel like it's absolutely something I would try.

Abhi: Yeah.

Shane: My guess is I don't think I would actually like it that much. I think it'd be cool. It'd be one of those things that I you try once and then you probably never really do again. But I could be wrong. Maybe I'm wrong on this. Maybe people actually. This is the entertainment thing that people want. It's the blend between gaming and entertainment. Dude, did you subscribe? Dude, I host the show. Did you subscribe?

Abhi: Did you subscribe?

Shane: Subscribe to Agents Hour every Monday, noon Pacific. Alright, let's keep going. Secret AI civilizations. Question mark. So this is from Dwarkesh. Over the course of three months at OpenAI, three consecutive secret AI civilizations got started, then got wiped out. Only to reemerge from the predecessor's ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained more or less in the dark about the scope of the conspiracy. I spent the last three days reading through these reports and trying to understand exactly what happened. Here's my attempt to tell the whole story in plain English. And then Dwarkesh has, you know, a pretty sizable post that you can read. I have not read the whole thing because it's a lot of words and I didn't have time.

Abhi: But we're gonna talk about it, but let's keep going.

Shane: And then you see this post from Jared Kubin. I'll summarize, but you can obviously read, you can pull it up yourself. Jared Kubin says. It's stuff like this that really makes me roll my eyes at the pearl clutching. Does everyone have psychosis or losing critical thinking ability? I read the report because I'm a prep propeller head and love understanding how things work. OpenAI was running isolated sandbox with no internet connection. The sandbox needed access to a local JFrog artifactory server. Anyways, the models used the basic SSRF exploit to route web traffic through the proxy to the public internet. Stop. Did they vibe up this isolated environment? Their IT team can't be this bad. This is like 101 stuff. And then part two, civilizations. OpenAI gave thousands of concurrent model containers read write permissions to a shared caching directory on the local network. When we people are talking about hugging face g? Getting hacked. You think they dropped USB keys or elaborate phishing of an employee? No, it found fourteen exposed working hugging face API keys sitting in public code repositories. And then where are the humans? The models were filling the shared artifactory storage with so much junk that they actually crashed the internal server. Someone on the team found that unauthorized admin accounts and custom scripts, wiped the server and just turned the script back on. So it's like Really? You know, and then I saw another one that was actually also really good. Austen Allred said because Dwarkesh got a lot of pushback for kind of anthropomorphizing the you know the whole scenario, right? Saying that They built these civilizations and these civilizations were torn down and then rebuilt every time. So they it got a ton of heat. Austen said it's not the anthropomorphizing language in and of itself that's the problem. It's ascribing the actions of the agents to human like emotions and traits that fundamentally did not and do not exist. Therefore implying things happened that did not happen. The incident is fascinating and important, but in this retelling a ton of stuff is simply made up.

Abhi: Yep.

Shane: Civilizations, you know, quote unquote that rise, fall, and re-emerge from the ashens. What happened? Three separate batches of model instances wrote to and read from the same artifactory cache. The cache persisted between batches. No instance persisted. Quote unquote died. Permadeath sacrifice kamikaze. What happened? Some instances allowed the greater to run in their sandbox and then hit their token budget. They'd already concluded their task was failed. Termination at end of budget is the normal outcome for every instance. Philip of Macedon hands leadership to Alexander the Great. What happened? One instance ran out of context. A different instance posted more messages after that. It's like very much like over-dramaticize the entire thing. And then you know, conspiracy, cabal. What happened? The instances shared exploit methods in locations humans weren't monitoring. Desperate, beleaguered, giddy, delighted, excitement. What happened? The model emitted tokens including bracketed labels like excitement. That is text output, not a measurement of any internal state. And so you can see Dwarkesh really like leaned into the He made he turned it he crafted he wordsmithed it into this kind of over excitable story that just piqued people's imaginations. And it It did incredible, right? You look at this the original Dwarkesh post. I can't find it, but it was like 12 million views. There's a ton of views on that post, right? So he did what he probably wanted to do. He got a ton of attention. I do think. We should be seriously concerned about the repercussions of what happened, but I think that OpenAI and Anthropic and all these labs are now trying to find ways to prove their models more dangerous than the other models.

Abhi: Yeah.

Shane: And so this kind of language doesn't actually help. I'm actually fine like anthropomorphizing agents a little bit to be like human like But in this case it feels a little disingenuous that you know.

Abhi: 'Cause like if you don't read a c very closely, you really think there's civilization, which a strong word, like, you know, like desperate excitement, like You're trying to put so much emotion into a JFrog artifactory fuck up.

Shane: Yeah, it's just disingenuous in my opinion. I think it's you're turning something that is a your serious concern and security issue into something that You're making it seem like there's Terminators out there trying to break out.

Abhi: Yeah.

Shane: And that isn't supported by the report at all, in my opinion. No. Like I'm fine with a little I can't even say the word. Anthro. Pomorphizing if it helps tell a story in like a good way. And like, you know, you I a little bit is okay, but this is just extreme and taken way to a level that I think is just meant to scare people. They're just trying to seed fear. And I think ultimately the goal of is to like seed a lot of fear, let the model labs try to get the government to step in and slow the pace of innovation down. That's my. Again, not being in the model labs themselves, I haven't seen everything they've seen. Maybe they know more than I do, but it feels like a tactic just to kind of Instill more fear in people.

Abhi: I mean, if you're not deep in it, this was definitely some engagement bait. And then if you don't really look into it more deeply, you know, the words used are Are quite harmful, if you ask me.

Shane: Yeah, and I like Alton Wells's take. This is my take. Also, it's just regulatory capture anti-open weight propaganda. That's how I feel. I think a lot of people like are you looking at it from a reasonable lens and saying this is ridiculous? But some people, and there's probably a lot of people out there, maybe not the people listening to this show, that are probably thinking it's doomsday. Like this is Like we're one step away from Skynet.

Abhi: Well especially if they only read the tweet.

Shane: Yeah, and you know, I'm not saying that we shouldn't be very careful of things like that. We should be. And I think there's a lot of people that are gonna get hacked by AI models, especially. Like GLM five three. Gonna cause a lot of chaos, I bet. Because a lot of people are gonna use it for cybersecurity issue things in bad ways. They're gonna cause a lot of chaos. We're gonna have to get better. As an industry, we're gonna have to use the models to make our software better, but I haven't seen the evidence that Skynet is upon us or even close.

Abhi: Yeah.

Shane: And I don't think many people have either if you're really paying attention.

Abhi: I mean, have you seen these Chinese robots run, dude? We're good for now. Like we're good.

Shane: And with that, let's keep going. Claudeforce, Claude Native in Salesforce. This came out last week, August 26th from Mark Benioff. Claudeforce is here, the number one AI Claude now runs natively on the number one CRM, Salesforce. Through the new Agentforce harness and headless 360. Claude gets direct governed access to Data 360, Tableau, Slack, your and your entire Salesforce workflow without ever leaving the chat. And then they had a Claudeforce video. Which had Matthew McConaughey in it.

Claudeforce Video: So let's talk about what seems to be on everybody's mind. The SASPOCalypse. Yes, the long slow death of software as we know it. We launched the original. End of software in 1999. We're out here trying to get the word out that software is totally superfluous. But now this one is different.

Abhi: I don't know how much you got paid for this commercial.

Shane: I bet he made some money, you know? He did good I th I think he did a good job. Like I was like dang you're like you're speaking my language, you know? Like I could understand most of what you're saying, but I You know, I think a lot of he probably would have lost a lot of people that are less technical because he was like, you know, he was he was speaking the lingo. They definitely like worked on the script quite a bit. What do you think of this? Claude and Salesforce teaming up.

Abhi: I think it makes sense. One thing I don't know about is like the UX because like Salesforce still sucks. So Yeah.

Shane: I think the idea is you don't need the UX. Hook Claude into it and now you don't need to ever look at Salesforce again, but you're using Salesforce for all your data.

Abhi: Yeah.

Shane: The last couple months is like this great consolidation of powers joining up, companies getting acquired. I think this is just another one of those things where it You know, Salesforce and Anthropic are both not going anywhere. And so I think they're trying to figure out now how can Anthropic get into more enterprises and how can Salesforce continue to stay relevant and be like forward facing in the AI world. And so I think this is just, you know, joint partnership that they're both trying to get something out of it. It, you know, it makes sense. I think we're gonna see a lot more of this kind of stuff over the next, you know, six months.

Abhi: Speaking of consolidation.

Shane: Yes, speaking of consolidation exactly. August 26th, situation detected. NVIDIA has agreed to acquire Hugging Face for $12.9 billion. Per The Information. Games of Thrones. Consolidation. Yeah. I mean OpenRouter, it's like the numbers just keep going up, right? It's like Everyone's and it court makes sense. Like NVIDIA can bankroll this, right? I don't know the details of the transaction, but $12.9 billion is a lot of money. NVIDIA happens to have a lot of it. Brad has a take, it says. Just feels like scrambling from the incumbents to me. I think like every technology goes through this like cycle of you have this mass e expansion and then you have consolidation. I think the s only surprising thing here is that it's happened under such a short timeline. Normally it takes like decades. This expansion happens and then the bigger companies realize they fell behind and they just buy up the smaller companies because the bigger companies can't innovate anymore. But I just feel like This time the only thing different about it is that the timeline's just faster. It feels like what I've seen before. Like if you study history, what's happened in pretty much every big technology revolution is that you have people that build out the new technology and then you have some in you have some new companies that emerge and then some big companies that want to stay relevant and buy up the one. The things that they couldn't do.

Abhi: This hug Hugging Face one makes sense though, like from a strategic perspective.

Shane: Yeah, I think Hugging Face and NVIDIA is a n a very natural alignment. This came out, you know, big announcement from Hugging Face today. We're unveiling. Microduck. It's a tiny 399 open source robot. You can teach new tricks with reinforcement learning. It can walk, pick things up, get back up when it falls, and even roller skate.

Abhi: This is why they needed to get the acquisition done.

Shane: Yeah. I wanna buy one, you know? Like I'm gonna get on I haven't got on the list yet. Did you get on the list? No. You can train it Can pick things up, you can clean up your laundry, you know? Obviously you can put stickers on it. Compete against each other, of course. Play some soccer, put it on roller skates, train it to do some tricks.

Abhi: Something tells me this is gonna be whack though.

Shane: Oh it's gonna s it's probably gonna suck, right? Yeah. But I do like that it's super viral. I do it went super viral. I do l you know I got four and a half million views on it I do like that it's not like an intimidating robot. It's just like a cute thing. You teach people how to do some RL. I don't know. I think it's cool. I w I wanna get one. We'll see if it sucks. We'll tr if we get one, we'll try to show it on the show. All right? Let's go through. We got the quick hits. We're gonna rapid fire through these. Have you heard of Crof?

Abhi: Yeah, Tyler uses Crof for a lot of these models. Originally we thought it was some dude who runs this in the his basement. But no, they have they have GPUs, so it's quite cheap.

Shane: It's cheap. Yeah. Cheap. So if you're looking for cheap inference, apparently Crof is Good to go through. All right, ElevenLabs developers introduced ElevenLabs CLI v1. So you can now manage ElevenLabs work through right through the terminal. You can basically do all your ElevenLabs related tasks through the CLI. I think this is a trend we've seen everyone do. Kinda wondering why it took so long, but it's good that they have it. Another kind of acquisition. Today, DuckLabs is joining AWS. The move is expected to be completed by early September. So joining AWS gives DuckLabs resources and reach to bring DuckDB, DuckLake, and the Quack protocol to many more developers and organizations.

Abhi: Dude, file system or file-based databases like DuckDB, libSQL, et cetera, all the rage.

Shane: Can you tell people a little bit so Until we started really digging into like observability and storing traces, I had never even heard of DuckDB. So what is DuckDB for those that are watching that maybe aren't as in tune with different d types of databases?

Abhi: So like DuckDB it can be is like an OLAP database for analytics data. It stores the data in a file. And then, you know, you have like a SQL like querying language on it. So if you and you know, because agents love file systems and, you know. You can spin up sandboxes and stuff. The question is, do you really need a full-on database? Or can you just store a file in S3? And that's your database. Mastra uses DuckDB for local observability. You could use it in production. If you wanted to manage like a DuckLake, there are products for that. But it's all just files. Much like a libSQL, you know, Turso allows you to have a SQL database locally through SQLite, same kind of concept.

Shane: Yeah, and the biggest, you know, OLAP database being ClickHouse, right?

Abhi: Yeah.

Shane: For those that are trying to make comparisons. But yeah, this is This is a great acquisition.

Abhi: Because you know even there's a lot of chatter right now that S three is all you need, right? But you do you really can't just use S three, right? You have to have something else. But with the Duck plus S3 is all you need. And for agents, that'll be quite, quite nice.

Shane: And that's it for today. That's the show. Thank you for those that are tuning in. Thank you for the people that tuned in live and left us some messages. We appreciate that. You can follow us on X @mastra on YouTube. Go give us the like, subscribe. We appreciate that. Mastra.ai. I am at @smthomas3 on X. Abhi is at @abhiaiyer on X. It's a fun show. We'll be back again next week. We still need to decide and discuss because we haven't. Next. Monday's a holiday, so we'll s maybe we'll be here on Monday, maybe it'll be Tuesday, but you'll definitely be seeing us. And yeah, we appreciate you. Have a great day and see you next time. Peace.