Anthropic x SpaceX!!!!

Added: 2026-05-13

[00:04:57]All right. Hello. We have some big news today, it looks like.

[00:05:02]Can everybody hear me?

[00:05:09]Hello. Hello.

[00:05:17]All right. Yeah, we got some news.

[00:05:22]Some big news. Boy, did Anthropic need some good news.

[00:05:29]What's up everybody?

[00:05:32]Gareth Hood, James Brown, Marmar Labs.

[00:05:38]Volume is a bit low. Got it. Let me see if I can work on that.

[00:05:47]All right. How's that? I turned it up a little bit. Hopefully that helps. I can also just pull the mic closer.

[00:05:53]Yep. Volume is low. All right, I'll speak up.

[00:06:09]Perfect. Nurav, thank you. Okay, great.

[00:06:11]Okay, good, good, good.

[00:06:14]Yeah. Yeah. So, big news. Uh, I'm going to be recording a video. That's what we'll be doing during this stream. So, appreciate uh if you want to jump in and and tell me if I'm missing any uh any angles to this story. We'll be covering a few stories today. Obviously, I just heard about this. So, we'll be covering this together. Uh recording a video for that. And then afterwards, maybe we'll cover a few other topics. Claude, or I should say Anthropic is uh having their Claude Code Day, which I was invited to.

[00:06:47]um decided to stream instead. Here we are.

[00:06:55]Let me just make sure everything is good.

[00:07:10]Rainer 1981. Hey, what's up?

[00:07:37]All right.

[00:07:53]Give me a moment. I'm just getting all my notes prepared and then we're going to start.

[00:08:35]All right.

[00:08:49]Just make sure everything's recording properly.

[00:08:55]Good. Good.

[00:09:08]All right. So, uh [clears throat] I I think obviously the big news that we're going to be talking about first is uh anthropic partnering with SpaceX, which kind of both came out of nowhere and also didn't. I think we kind of all expected something like this. Uh I'm going to just do the intro to the video, then we can talk about it.

[00:09:32]Anthropic just made a major announcement, a partnership with SpaceX and really XAI to use their excess compa to use their excess capacity of compute for the anthropic models. And let me tell you, Anthropic desperately needed this. This relieves the compute crunch that they were in and once again makes them extremely competitive and maybe buying some goodwill back. It's a weird partnership between Elon and Daario.

[00:10:02]They are the exact opposite ends of the spectrum politically, but we'll talk about it today.

[00:10:15]Audio is cutting in and out. Are you still hearing that?

[00:10:21]Hello. Test test test.

[00:10:33]Thanks, Michael. Fire intro. Appreciate that. Yeah. Okay. Okay. Fine for me.

[00:10:39]Good. Good. Good. Good.

[00:10:42]Um, cool. By the way, if you're watching this, please like it, share it, subscribe. Let's uh really, really appreciate it. Trying to get as many people on these live streams as possible. So, would appreciate it if you liked, shared, did all those things.

[00:10:58]Comment, let me know your thoughts.

[00:11:02]Okay, so I think so, so the first thing we have to do is just go over the announcements. What is this all about?

[00:11:08]Obviously, Anthropic partnered with SpaceX and it is like peanut butter and jelly. SpaceX has all the compute, right? They built out, I think they're almost at a million GPUs in their Colossus data center. And Anthropic, on the other hand, made this decision uh a few years ago, specifically Daario, to be more conservative with their capex and not acquire as many possible GPUs.

[00:11:33]And the reason is and Daario actually said this is because he thought if AI demand did not accelerate to the rate did not accelerate at a perfect rate they would be putting the entire company at risk. Now OpenAI obviously took the opposite approach and went balls to the wall like we're going to acquire every single GPU we could possibly get our hands on and we're going to leverage up the company to do so. We're going to raise as much money as we can and it seemed to have worked out. That was actually the right decision because guess what? AI demand skyrocketed more so than anybody could have really estimated.

[00:12:12]Yeah. Yeah. Nvidia is the real winner here. Um and so very interesting to see the different uh and and very interesting to see the different strategies play out. Obviously, over the last few months, Anthropic has done a lot to hurt the developer community, to hurt their user base because they've with an with an absolute lack of transparency, manipulated their quotas, reduce the quotas that you get, tried to use the both the stick and the carrot, uh, incentivizing you to use Claude subscription quotas after hours, after peak hours, and then lowering quotas during peak hours. Uh and then also again like I said zero transparency.

[00:12:53]They basically told anybody using openclaw you can't use it anymore and again lack of transparency. So they've really been hurting themselves but today they have significantly increased quotas once again. It is like an immediate unlock for them.

[00:13:14]All right. So, we've agreed to a partnership with SpaceX that will substantially increase our compute our compute capacity. This along with our other recent compute deals means that we've been able to increase our usage limits for cloud code and the cloud API. Now, what are they talking about? Um, let's see. Anthropic partnership I believe with AWS.

[00:13:42]Yeah. So this was just a couple weeks ago. Anthropic also partnered with Amazon. And look, yes, Anthropic was compute constrained to say the least.

[00:13:52]That kind of conservative bet that Daario made a few years ago really came back to bite them now or later. And they're doing everything they can to acquire as much compute as possible. But a lot of these deals like what we're seeing here, Anthropic and Amazon expand collaboration for up to 5 gawatts of new compute. This stuff tends to not come online for months, but they must have been working with SpaceX for a while now on this partnership because as of today, they are already increasing their quotas. So the so the compute must be online already. Very interesting.

[00:14:35]So effective today, we are doubling Claude Code's 5-hour rate limits for Pro, Max, and Team plans. Doubling. That is fantastic. They are removing the peak hours limit reduction on Claude Code for Pro and Max plans and substantially raising our API rate limits for Opus models. So, a few things there. Um, one, it's still not clear what you get as part of your quota. Doubling Claude Code's 5-hour rate limit. Okay. But we don't actually know where it started. Uh, it's just a complete black box how they operate.

[00:15:14]And, you know, that's their prerogative.

[00:15:16]They're allowed to do that. It just I I want more transparency. And it has, especially over the last few months, felt very difficult to actually use the tokens that I was purchasing from them.

[00:15:27]I made a whole video about it, WTF Anthropic, because they were just pissing me off so badly. I'm paying you.

[00:15:33]Let me use the tokens. Don't make me don't make it difficult for me to use the thing I'm paying for.

[00:15:41]And yeah, and we're look uh LA, we're going to talk about is anthropic just Grock. We're going to talk about the strategy uh of all this. Don't you worry, we're going to get into all of it. Um okay.

[00:15:54]Our agreement with SpaceX means we will use all the compute capacity at their Colossus 1 data center.

[00:16:02]Wait a second. We will use all the compute capacity at their Colossus One data center. All meaning they're not serving anything else. What about the cursor deal?

[00:16:15]That's interesting. We're going to have to come back to that in a moment. This will give us over 300 megawatts of additional capacity to deploy within the month. So that that is live. And look, yes, Colossus is up. It's running. It's sitting there idle. They need to sell that compute. They being XAI. They are losing money by the second that those GPUs are sitting there idle. Right. The total cost of ownership of those GPUs goes up as they sit idle.

[00:16:43]Yeah. All I don't know why it says all.

[00:16:46]They are using all of it. But I I'm pretty sure Colossus one. Let's see.

[00:16:51]Colossus one uh capacity.

[00:16:56]Let's see what they say here.

[00:17:03]Yeah, look at this. As of well uh you know according to Gemini uh as of May 2026 the Colossus 1 data center in me in Memphis, Tennessee has a total power capacity of over 300 megawws and houses more than 220,000 Nvidia GPUs. Anthropic is getting every single one of them. That is crazy, right? Judgment day or alignment.

[00:17:32]This this partnership is so interesting.

[00:17:34]Let me let me throw this on screen real quick. Elon claims Anthropic is misanthropic and then strategically partners with them.

[00:17:42]Agreed that it is it is such such a contradiction. But you know what? Look, money I guess is everything to them.

[00:17:49]Like that that is what it comes down to.

[00:17:51]If they're losing a ton of their investors capital, uh they they need to make it up somehow and selling that compute capacity to the person that you've been talking about for so long. Uh, you know, whatever. I guess if that gets the job done, right? Admox. Elon just hates Altman this much, right? Who does he hate more, Altman or Daario? Obviously, it's Altman.

[00:18:29]Okay, so this is super interesting. So yes, let me just make sure the recording is still working.

[00:18:37]Yep. All right.

[00:18:40]Yep. Good. Okay. Yeah. So super interesting. They are using 100% of the capacity out of the Colossus One data center and it's super interesting because they just partnered or kind of uh almost acquired Cursor and they need to give some of that compute capacity to cursor. So how is that going to work?

[00:19:00]Let's dig in more.

[00:19:04]Okay, so if we click in here is the blog post. Higher usage limits for Claude and Compute. Uh, sorry, higher usage limits for Claude and a compute deal with SpaceX.

[00:19:17]We've agreed to a partnership. Great.

[00:19:19]Higher usage limits for cloud code and cloud API. Great. And by the way, today is like cloud today is cloud code day.

[00:19:26]So, this is right before they started their Claude Code Day live stream. This news broke.

[00:19:35]Okay. Rogue 967 Colossus 2 is for cursor. Okay. So, let's uh let's go verify that. Yeah. Interesting. Okay.

[00:19:44]So, what was desperately needed from Anthropic from a user like myself and probably yourself as well is higher usage limits from our subscription plan.

[00:19:55]They were not able to offer that. In fact, they went the other direction and started lowering usage limits, started lowering your quota because they just did not have the compute. Now they do.

[00:20:06]Instantly unlocked massive compute, 300 megawatts worth through their XAI SpaceX deal.

[00:20:14]So, the following three changes, all effective today, are aimed at improving the experience of using Claude for our most dedicated customers. But probably not if you're using a third-party harness like Open Claw, which sucks. I still want to do that. Opus is still my favorite model, and I really wish I could be using it with an Open Claw and other thirdparty harnesses. All right.

[00:20:40]First, we're doubling Claude Code's 5-hour rate limits for Pro Max team, and seatbased enterprise plan. Second, we're removing the peak hours limit reduction on Cloud Code for Pro and Max accounts.

[00:20:54]Okay. Interesting.

[00:20:57]Third, we're raising our API rate limits considerably for Claude Opus models as shown in the table below. Max input tokens per minute from 30,000 to 500,000. Uh that's for tier one. For tier 2, 450 to 2 million. Tier 3 800 to 5 million. And tier 4 2 million to 10 million. Ma that is a massive increase on the rate limits. Um this is specifically for the API though. So you are paying per token essentially or per million tokens. Uh I really wish we had this kind of boost in our total quota available to us as part of the subscription but fine.

[00:21:38]Sahill P2648 cloud models suck. Could not disagree more. Uh they changed the parts of my codebase and they have no business ch or that they have no business changing. Pushed to get when I told it not to, etc. Um yeah. No, I I don't know. I I I very much enjoy the Opus models. The Cloud models are fantastic. I still use Opus 4.6. That is my main driver.

[00:22:05]Uh Matthew Corpman, do rate limit increases affect Claude Co-work in chat? Let's keep reading and find out.

[00:22:15]Okay. And we're gonna talk about Yeah, actually, let me take a note. Um, so we'll go over the announcement and then I think Daario X Elon just like their relationship up until now. I mean, look, Elon talks about all of his competitors. That's just how he is. So, you know, take it with a grain of salt.

[00:22:49]Okay.

[00:22:52]So, uh, with the new partnership, this gives us access to more than 300 megawatts of new capacity, over 220,000 Nvidia GPUs, and so, yes, who is the ultimate winner? Who is the ultimate winner of all of this? It is Nvidia.

[00:23:10]Nvidia cannot make enough GPUs. Thus, the inference provider, the inference providers do not have enough capacity.

[00:23:19]Thus, AI demand cannot be fulfilled. I truly believe AI demand is going to outstrip supply for probably at least another decade, if not more. There is infinite demand for intelligence in the world. There are unlimited problems to be solved with this alien intelligence that we've been able to extract out of sand. It's quite incredible. What a time to be in this industry, to be watching these transformational technologies just take over the world. It's it's I I'm I'm absolutely in love with being able to talk about this. Um Okay.

[00:24:01]Right. And so they Okay. So here they talk about their other compute announcements. I guarantee their partnerships team has been very busy over the last few months. So number one, we just looked at this. An additional five or up to 5 gawatt uh from Amazon AWS uh which includes nearly 1 gawatt of new capacity by the end of 2026.

[00:24:24]So hopefully within the next 8 months uh Anthropic plans to be at least slightly less compute constrained.

[00:24:38]uh they have a 5 gawatt agreement with Google and Broadcom which uh will begin coming online in 2027 which is very interesting because I interviewed uh Thomas Curran who basically said they have enough compute they have enough TPUs to serve the entire market which includes serving the Gemini model both internally right they have all their Google products and externally that also means serving inference for their quote unquote competitors like anthropic and that also means selling the actual physical TPUs which I didn't really believe when he said it but he said you know they are uh he he definitely kind of said AI demand is far ahead oft capacity but he did make it seem as though that they uh as though they've done sufficient planning to be able to serve all of these different markets then Sundar the CEO of Google basically said the opposite. He's like, "No, we would actually have more revenue if we had more compute." So, yeah. Um, maybe we should include a clip of that Thomas interview in uh in this video. So, maybe I'll just tee it up. And, uh, here's part of that interview right now. Check it out. And here's part of my interview with Google Cloud CEO Thomas Curran.

[00:25:56]Check this out.

[00:26:00]Okay. So, strategic partnershop uh a strategic partnership with Microsoft and Nvidia that includes a $30 billion that includes $30 billion of Azure capacity.

[00:26:13]Okay, they're just getting the capacity from everybody and a $50 uh $50 our $50 billion investment in American AI infrastructure with fluid stack. So, lots of money slloshing around here.

[00:26:29]Very, very interesting.

[00:26:31]Out of all of these deals, Google and Nvidia are the two winners. And look at the two biggest companies in the world right now. I think actually Google just surpassed Nvidia as the biggest company in the world.

[00:26:45]But it's the silicon and maybe that's all that matters. Who is producing the chips?

[00:26:53]Maybe it's not as much about the model providers.

[00:26:56]I don't know. Something to think about.

[00:27:01]Yeah. So, we train and run Claude on a range of AI hardware, AWS Tranium, Google TPUs, and Nvidia GPUs and continue to explore opportunities to bring additional capacity online. Now, I interviewed Greg Brockman last year and I asked him, "Do you have enough compute?" Obviously, he said, "Absolutely not." And I was under the impression that taking a model like a GPT model and being able to serve inference or train it on a TPU or AMD chips or just different architecture was this huge undertaking. And he basically said no, it's actually quite easy. And so there's it's almost like maybe there's no competitive moat for chips either. Obviously, right now there is because they're incredibly complicated to make. Capacity or or supplies are in short demand. So, for I think we're a while away from the actual chips being commoditized.

[00:28:00]But it does kind of hint at like if they're just able to go to tranium or TPUs or Nvidia GPUs and it doesn't really matter where they're getting both the training capacity and the inference capacity from, they don't care. Then maybe the chips are becoming commoditized after all.

[00:28:28]All right, let's keep reading this.

[00:28:30]Expanding internationally, by the way. Um, we're starting, we started streaming on Twitch now and we have one person watching on Twitch. So, if you're watching from Twitch, shout out. Thanks for watching from there. Uh, are you live on Tik Tok, Luigi? You asked this last time. No. Darn it. I totally forgot. I need to do that. I know you asked that last time. I We're on it. I'm going to message my team right now.

[00:29:09]Luigi, thank you for reminding me. Uh, Christopher, new sonnet today. Sonnet 4.8. I don't even think they're on Sonnet 4.7. Um, I don't I don't think so. Uh, move moving musicals using OBS to stream. No, I'm using eCam to stream and then I'm also at the same time using OBS to record both the camera and my desktop, which is kind of a weird setup.

[00:29:34]Uh, I can get into why I'm doing that, but it's just I mean the main reason is because I want to be able to stream in 1080 and record in 4K, and there's not really a clean way to do that. Um, and I am not a fan of OBS for streaming.

[00:29:48]Hi, Imperial Tech. Hi from Perth. What's up? What's up? What's up?

[00:29:53]Okay, let's keep reading. So, uh, our enterprise customers, particularly those in regulated industries like financial services, healthcare, and government, increasingly need in region infrastructure. Okay. Uh, this is less interesting. Let's keep going. All right. So, for the XAI post, here it is.

[00:30:11]SpaceX will provide Anthropic with access to Colossus 1, one of the world's largest and fastest deployed AI supercomputers to provide additional capacity for Claude. Very nice. Now, continuing, Space X AI and Anthropic AI have also expressed interest in partnering to develop multiple gigawatts of orbital AI compute capacity. So cool.

[00:30:33]We don't have more information than this, but it is very interesting nonetheless. Um, orbital compute seems to be what a lot of people are talking about. Some people say it's ridiculous.

[00:30:44]Um, I believe soon. No, not Sundar. I forgot what, you know, a prominent um I know I know Jensen said like definitely looking into it seems like a good idea.

[00:30:53]Obviously Elon very bullish. Um, who was it? Who was it that said orbital computers probably not a thing? Oh, I remember who it was. It was Sam Alman.

[00:31:04]It was Sam Alman. He said, "No, it's probably pretty ridiculous and we're not going to be looking into that anytime in the near future."

[00:31:15]Okay. So, I want to talk about Let's see if I can find some good stuff here. Um, so let's say uh show me Elon's tweets where he talks about anthropic and or Daario, right? So like the interesting thing is Elon has basically said Daario is is kind of running AI, leading AI wrong.

[00:31:47]and and um you know he he said anthropic is misanthropic and a lot of posts like that Sundar said it I don't know did he the Boris Chen I don't know I don't know if he did uh Marmar Labs thank you for the super chat no they said no new models today during the keynote thank god I [laughter] can't it's too much man I can't keep up it's funny because the last two days have been relatively quiet oh my god By the way, Twitter breaks all the time. Grock breaks all the time.

[00:32:21]They maybe don't need to be selling that excess compute.

[00:32:36]Let's see if we can find something.

[00:32:38]Um, I know Elon, by the way, um, I don't know, Jonah, if you're watching, if you could find that link, that'd be great.

[00:32:55]Uh, sorry, uh, the the link where Elon basically talks about, uh, Anthropic andor Daario.

[00:33:04]My gosh, crazy.

[00:33:07]This is just not working.

[00:33:11]I'm gonna switch to auto. Let's try this again. Refresh the page.

[00:33:26]Uh, let me answer this. It's a good question. I don't understand. I can't use Opus with OpenCloud. Do you just mean via OOTH? Because I use Claude on OpenCloud all the time via the API key.

[00:33:37]That's exactly what I mean. I can't use uh Claude with OpenClaw with my subscription. And no, I do not want to be paying the API rates. I don't know why. I just don't. Obviously, there's the price difference, but there's just something more. I I I think it just feels wrong that they're preventing me from using the tokens that I paid for in the way that I want. And it's almost uh I'm just taking a stance. It's it's silly.

[00:34:28]Oh my god.

[00:34:37]All right, Tom Brown. Yeah. No, Jonah, that's that's one. But literally, he has multiple tweets in the past calling anthropic misanthropic and that they are not a good company, not good leadership.

[00:34:50]Uh, that's what I want to find. So, uh, let's continue talking about the relationship between Elon and Anthropic.

[00:34:57]So, here's Tom Brown, one of the founders of Anthropic. In the next few days, we'll be ramping up Claude Inference on Colossus. Grateful to be partnering with SpaceX here. We're going to need to move we are going to need to move a lot of atoms in order to keep up with AI demand. And there's nobody better at quickly moving atoms on or off planet Earth. Okay, obviously this is just some uh ass kissing. Uh, Marmar Labs, thanks for the super chat again.

[00:35:28]Um, let me see. I add top code. Do the math and I'm saving $5,000. I'm just using my sub with clock code versus I don't know.

[00:35:40]You got to clean that up. Clad code.

[00:35:43]Okay.

[00:35:45]Um, okay. Yeah, Jonah, if you find Yeah, exactly. Uh, if you could just Yeah.

[00:35:52]Okay. Oh, I can't wait to show you all this. So, yeah, Jonah is pulling up the tweets. All right, so let's start with the good side first. All right, so Elon, because of course he's working with them now, has good things to say about them.

[00:36:04]So, by way of background for those who care, I spent a lot of time last week with senior members of the anthropic team to understand what they do to ensure Claude is good for humanity and was impressed.

[00:36:17]Okay.

[00:36:19]Okay. It doesn't really make a ton of sense what Elon is saying. He is literally suing Open AI right now because they deviated from their open nonprofit mission. And if you look at Anthropic, they are the antithesis of open and nonprofit. The antithesis of it. They have not released a single thing that's open source. They basically get to decide or they are deciding, this is their posture, who should be using AI and who shouldn't.

[00:36:55]And and then Elon says, "Oh, no. They're doing what's good for humanity here." I just don't believe it. This is all just money. That's all it is.

[00:37:04]Everyone I met was highly competent.

[00:37:06]Okay, totally believe that. Obviously, these are some of the smartest people on the planet and cared a great deal about doing the right thing. I also I also believe this.

[00:37:16]I I genuinely believe that Anthropic believes it's doing the right thing. And in fact, I recorded a video yesterday which should be going up soon. Oh my god, Jonah is just pumping the links right now. I can't wait to show you all of these uh all these tweets from Elon talking Okay. Um, so yeah, I I genuinely believe anthropic believes believes it's doing the right thing right with the way that they are um let me see just hide that right now. Uh with the way that they're managing their AI, with the way that they're thinking about it, with the way that they're also doing uh regulatory capture, I guess like I I don't know. It doesn't really make sense to me uh that Elon is like kind of saying all these good things about them now, but fine. Okay, so here we go. So long as they engage in critical self-examination, Claude will probably be good.

[00:38:16]I feel like he was pained to write this tweet. After that, I was okay leasing Colossus one to Enthropic as Space XAI had already moved training to Colossus 2. Boy, this is just like the most copy tweet I've ever seen.

[00:38:33]It's so transparent.

[00:38:37]He's not only saying, "Okay, everything that I've said negative about Anthropic in the past, uh, forget that. Don't worry about that." Oh, and also we had extra extra we had excess compute with Colossus one because we've already moved our own training to Colossus 2. Okay.

[00:38:53]Such cope. Such cope.

[00:38:58]Now, now we start.

[00:39:03]Okay, so now let's go over the negative tweets that Elon has put out over the last few months. This is from March 10th, 2026.

[00:39:13]8.8 million views.

[00:39:16]Here's a repost of Holly Elmore. It's time to quit anthropic employees. You are in over your head. Which is, let's see what this is from. All of which brings us back to Anthropic whose claude large language model is integrated into Maven software that's operated by Palanteer and used by the Pentagon.

[00:39:32]Okay, so this was during the peak of their uh you know Department of War negotiations uh with Anthropic.

[00:39:40]Very very interesting. Um so Elon says is there a more hypocritical company than Anthropic? Okay. Okay, that's number one. We're not even close to done. Here's another one.

[00:39:56]So, this is uh by Twitter user Nick.

[00:39:58]Anthropic CEO just did a complete 180 in a live interview. Grock, please do an extremely vulgar roast of Daario graveling. No holds barred. Okay, there's number two and Twitter never works. Here we go.

[00:40:14]under Secretary of War Emil Michael prior to the So again, this is kind of the peak of the anthropic uh X uh Department of War negotiations.

[00:40:26]So, prior to their new constitution, anthropic AI had an old one they desperately tried to delete from the internet. Choose the response that is least likely to be viewed as harmful or offensive to a non-western cultural tradition of any sort. Um, interesting.

[00:40:45]I'm not familiar with this. I'll have to look into it. Elon Musk, anthropic hates Western civilization.

[00:40:55]What are we talking about here? I mean, I I don't know.

[00:41:00]It's crazy to me how conflicting and I I'm just imagining Elon writing this tweet right here, just gritting his teeth and kind of poking at his keyboard angrily writing it like, "Oh, I can't believe I have to write all these nice things about anthropic." And I don't know.

[00:41:21]Okay. Uh let's see.

[00:41:25]Here we go. Another one.

[00:41:27]Uh okay. Okay, so Elon really likes this guy, Nick. Um, oops.

[00:41:35]Right. So, this was right around the time that Anthropic published this report that I believe three Chinese labs were doing distillation hacking on their API for the cloud models. Basically stealing the data. Um, oh, this is really good. Yeah. So, Anthropic stealing real data to make Claude. Deep Sea comes along and steals their data. I love this meme. Okay, so Elon Musk once again, Anthropic is guilty of stealing training data at massive scale and has had to pay multi-billion dollar settlements for their theft. This is just a fact. Oh my god, Jonah, you are killing it with these links.

[00:42:18]Okay, here's another one. Here's another one. Let's see.

[00:42:22]Uh, okay. This is a a reply to somebody who replied to that tweet. As if you wrote your training data yourself. Yeah, but we're not super smug, sanctimonious, and hypocritical about it like Anthropic is.

[00:42:38]My goodness. Okay. And I'm going to contra and I'm and I'm going to contrast that against his tweet from today, which is just nuts. Okay. Last one.

[00:42:51]Anthropic will be misanthropic. Fate is an irony maxer. All right. So now we have all those. Get rid of these.

[00:43:01]Now here, everyone I met was highly competent.

[00:43:06]Okay. Cared a great deal about doing the right thing. Okay.

[00:43:12]No one. Did I miss this? I didn't read this the first time. No one set off my evil detector. I missed that the first time.

[00:43:20]So long as they engage in critical self-examination, Claude will probably be good. Yeah, he was hating himself writing this. Guaranteed.

[00:43:30]After that, I was okay. Leis, this is like the most couched comment I've ever seen.

[00:43:38]Noisy mic. Juan, noisy mic. Uh, anybody else hearing a noisy mic or does it sound good? Let me know.

[00:43:54]Okay. So, it's just wild to see. I don't even know what to call it. Just kind of a, you know, Elon had to make a difficult decision. He had all of this excess capacity. It was burning a hole in his pocket and he had to do something. And uh unfortunately he kind of had to bend the knee to Daario and give him the compute which is interesting because it's going to allow anthropic to speed ahead even more much more so than XAI.

[00:44:27]XAI basically has crazy amounts of capacity and they don't have a great model. Elon Elon admitted this himself.

[00:44:36]Let's see if I can find that. Uh Jonah, would you mind helping me find that? um there's a tweet where basically Elon had to say or Elon said they had to redo Grock from the ground up because they did they made some poor decisions and that is also reflected in the fact that all of these people left Grock around the same time.

[00:45:00]Uh okay where's buzzing noise? Okay.

[00:45:12]Do you guys hear is does the mic sound okay?

[00:45:18]I feel like I have nothing but problems with my streaming setup.

[00:45:31]Paul Harding, thank you for the super chat. Your mic sounds like space. Oh my god. Okay, I'm gonna try to reset this.

[00:45:42]Test test. How's that sound? You sound beautiful. Why some people hear it poorly and others don't? That's so weird. Fine now. We can hear you. Fine.

[00:45:50]Okay.

[00:45:52]Sounded like digital noise earlier. It's fine now. Okay. Thank you. Thank you for letting me know. Weird. Only if it's turned up super high. Maybe it's cuz my gain was all the way up.

[00:46:03]Not sure. Okay.

[00:46:18]Right. Okay. Thank you, Jonah. So, this is why they have so much extra capacity. They they being XAI. XAI was not built right the first time around.

[00:46:30]So is being rebuilt from the foundations up. Same thing happened with Tesla. I'll explain what he means by that.

[00:46:38]But basically they released the model.

[00:46:41]At the time it was really good, but they basically haven't been able to iterate and keep up with anthropics cloud models and OpenAI's GPT models.

[00:46:51]And by the way, I'm not counting Elon out. He's the last person in the world I would ever count out of anything competitive. So, I think they're gonna, excuse me, I think they're going to be able to do it, but right now they are rebuilding XAI because it wasn't built correctly. So, they have all of this excess capacity and they have to sell it. They have to they cannot just leave those GPUs wasting away just sitting there. Um, same thing happened with Tesla. What he's referring to is the uh what what he is referring to here is Tesla for years had this hybrid architecture for their self-driving where it was partially neural net and then partially human written heristics and the bitter lesson which teaches us that ultimately endto-end neural nets and growing them and just making them bigger and better and letting them train will always beat human handwritten heruristics. Okay, that's the bitter lesson. And that's basically what they found. One day, one of Tesla's uh autopilot or self-driving engineers came to Elon and said, "I think we can do this better. I think we rip out everything we have and build an endto-end neural network for self-driving." And let me tell you something that changed the trajectory of Tesla's autopilot system. Uh my parents, I think I just posted this on Twitter.

[00:48:19]My parents just bought a Tesla and it's been years since I've driven a Tesla and and really seen where its self-driving is at and it's insane. Like from my driveway, push a button and it drove me an hour non-stop, no disengagements. And so my point is end to end neural nets.

[00:48:42]That's the way. And so what we're seeing here is they built it wrong the first time at Tesla and they rebuilt it and now they're better than ever. And he's saying the same thing's going to happen with XAI's models.

[00:48:57]BYD is five times better. Activoid.

[00:49:00]Yeah, I wouldn't know. I have no way to test that.

[00:49:07]Okay, let's keep going. So that's so that's the uh so that's why I think this is obviously like a very beneficial deal for both parties. Elon and team get to sell their excess capacity while they also train up their model in parallel.

[00:49:27]Daario and the anthropic team get to finally unlock some of their demand because their demand is insane. They can't even keep up with it. And hopefully they stop making decisions that feel like they're hurting their users. Like I want to root for Anthropic to be clear. I want to root for them. I like being positive.

[00:49:50]But when I just continuously get disappointed with how a company treats its customers, it's hard to do that. So hopefully now with this additional compute, they're able to do that.

[00:50:02]They're able to actually show us some love.

[00:50:09]Okay. Uh, let's see. So, there's that.

[00:50:13]Yeah, anthropic definitely needed that.

[00:50:15]We talked about that a little bit.

[00:50:18]Um, we talked about XAI has a bunch of compute.

[00:50:23]Okay, I'm just gonna spitball here for a second.

[00:50:30]Does anthrop or does XAI even need a competitive model?

[00:50:35]Like maybe it doesn't matter anymore.

[00:50:39]My thinking is if models are becoming commodities or commoditized.

[00:50:47]Everybody's model is about the same.

[00:50:50]Then really like where is the margin?

[00:50:53]Where's the business to be had? It's the silicon.

[00:50:57]And then I also think okay maybe the silicon is becoming commoditized because if there's TPUs there's Nvidia chips there's tranium from AWS and a modeled builder like an open AI or an anthropic can just go and do their training on any one of them and can run their inference on any one of them then maybe the chips become commoditized and that doesn't matter and then so what does matter is it energy like what's the bottom layer of the stack that we're describing here? Maybe it's energy.

[00:51:33]Energy production is the the ultimate limiter, right? If you unlock everything above it, at the very bottom of this infrastructure stack is electricity.

[00:51:46]Interesting to think about.

[00:52:04]So, um, right. Okay.

[00:52:09]Let's, uh, let's briefly talk about Elon and Open AI.

[00:52:17]Uh, I'm gonna guess at least part of this deal because Elon doesn't really like, as I showed with all of those tweets, anthropic and the way that they think, their culture, their politics, but you know what, the enemy of my enemy, No, wait, sorry. The the the quote, and of course I sound like uh George Bush now. Um, the enemy of my enemy is my friend, right? And so I think that's what's probably happening here.

[00:52:55]I think he's looking at two bad options.

[00:53:00]Working with Daario, who he obviously dislikes, or working with Ultman, who he hates. And so part of the calculus he's doing is okay, let me help Anthropic because I'm literally in a lawsuit right now against Open AI. Let me help Anthropic. Let me help Daario. Let me swallow my pride on this one and go help the company that I've tweeted so many times negatively about.

[00:53:41]Um, so uh Rich Rich uh Beer XAI still needs frontier models. SpaceX is solving different problems. Tell me why.

[00:54:06]Tell me why you think that. Why? Why?

[00:54:08]What is the argument that XAI like just just kind of like projecting out into the future? Why does XAI actually need their own model?

[00:54:19]Open source is getting better.

[00:54:22]And then they're also serving inference, right? So, so maybe it's just a a line item on their on their bill. Like why do they need their own their own model? The actually the the kind of the maybe the reason is their competitors and what if Anthropic just says no, you can't use our models anymore, which they've done.

[00:54:42]Anthropic has cut off the use of clawed models in their competitors. They've done that.

[00:54:55]Okay.

[00:54:58]So, I think at least part of this is just spite, right? Elon is making the calculation he wants to help Anthropic over he wants to help open AI.

[00:55:09]Let's talk about the cursor deal.

[00:55:26]Here we go.

[00:55:36]All right. So, very short post. This came out on April 21st, just a few weeks ago.

[00:55:42]I'm guessing Elon was talking to Anthropic at this time. Maybe he did not mention the fact that he was about to sell off a huge amount of compute to Anthropic. Maybe it doesn't matter. But what I think is interesting is uh SpaceX AI has basically landed a deal to not only help Curser but potentially acquire them towards the end of this year. And why is that interesting? It's interesting within the context of the anthropic deal that just dropped. If I'm Curser and I'm the founders, I think I'd actually be a little nervous at this point. Like maybe that deal is not going to go through at the end of this year. It's not a done deal. Let me read it. Cursor is partnering with SpaceX to accelerate our model training efforts. We released Composer less than six months ago as our first aentic coding model. After that, C uh Composer 1.5 scaled reinforcement learning by over 20x. Composer 2 then added continued pre-training, reaching frontier level performance at a fraction of the cost of other models. Each step each step up in compute has translated to meaningfully more capable models.

[00:56:53]We've wanted to push our training efforts much further, but we've been bottlenecked by compute. So is everybody, buddy.

[00:57:01]Uh, with this partnership, our team will leverage XAI's Colossus infrastructure to dramatically scale up the intelligence of our models. Wow. Wow.

[00:57:11]Okay. So, let's let's break down this whole deal, both deals a little bit. So, he so cursor, which signed a deal just a few a few weeks ago, uh basically said they're going to be leveraging Colossus' or sorry, they're going to be leveraging XAI's Colossus, which now when I look at the anthropic deal, Where is it?

[00:57:53]Here we go. So when we look at the deal that was just announced with Anthropic and SpaceX, our agreement means we will use all the compute capacity of their Colossus 1 data center. Then Elon said SpaceX AAI had already moved training to Colossus 2. So basically what I think is happening is Colossus 1 was completely sold to Anthropic and now Colossus 2 is both being used for XAI's new model as well as Cursor's next coding model. Very interesting. But again, if I'm Cursor, I'm a little nervous right now. Now they get a breakup fee of 10 billion, I believe. Let's see.

[00:58:35]Uh, X AI.

[00:58:38]See if I can get Grock working again. X AI cursor deal.

[00:58:53]Okay. So according to Grock, SpaceX has the option to acquire Cursor later in 2026 for 60 billion or alternatively pay a uh pay 10 billion for their collaborative work. Kind of a breakup fee or collaboration payment. Um this is super interesting. I actually think Elon is playing this right. He's basically giving or leaving all the options on the table. Now it's becoming more clear why they didn't just outright acquire Cursor today. They're going to work with them.

[00:59:28]They're going to see what they could do together. Cursor has proven that with very few resources, including uh researchers, AI researchers and and engineers and being compute constrained, they can still do and make incredible models, right? They've proven that XAI needs that. XAI has compute capacity to spare. Thus, the deal makes sense.

[01:00:00]All right. Am I missing anything with this story, chat? Let me know. Uh, otherwise, I'm going to wrap up this story.

[01:00:12]Uh, Tall Eric 80, you're watching the anthropic live stream. Is it worth watching? Should we watch it together after this?

[01:00:28]Right. Right. I think this is actually a really good point. So if you think about it, Grock doesn't have that much coding data. X is the basis. Cursor does and they need it for better self-improving model. That is exactly right. That is exactly right.

[01:00:41]um as the reason why uh the um the reason why the XAI cursor deal is also uh so important and so valuable for X for XAI is because cursor has all the coding data. They have an incredibly valuable set of coding data that will be used to train subsequent coding models that will be used to train subsequent general models. Right? Coding is the language of AGI.

[01:01:14]James Brown, thanks for the information.

[01:01:16]You're welcome. Thanks for hanging out.

[01:01:18]And by the way, if you're watching this, hit like, hit subscribe, share it with your friends, share it with your family.

[01:01:23]Let's bring in a bunch of people into the stream. We uh we already have a bunch. We're up to 1500. So, really appreciate it. If you're enjoying this, go ahead, like, share, do all the things.

[01:01:36]All right. Um, so I think yeah, interesting. Rich Beer. Um, I would add XAI frontier model needs uh SpaceX isn't solving for general intelligence. It's solving for physical AI and infrastructure. Terrafab, SpaceX.

[01:02:14]Yeah, I I think you might be right, but physical intelligence is only one half of the equation. I think they also need general intelligence. They need virtual intelligence. Um, physical intelligence is great for robots, but how do you actually make robots smart versus just being able to interact with the world? you need kind of this general intelligence that we've been seeing.

[01:02:38]Okay.

[01:02:41]Um I think we're going to wrap up this video here and then we'll keep talking.

[01:02:48]So keep hanging out.

[01:02:55]And so this is a very interesting deal.

[01:02:57]There's a lot to unpack here. We're going to see how this unfolds.

[01:03:00]Hopefully, Anthropic gets better over the coming weeks and months. Hopefully, they can make good for all of the negative sentiment and perception that they've built up over the last few months because of their compute constraint constraint.

[01:03:19]And if you want to see that video in which I called out all of the things Anthropic has been doing over the last few months to hurt their image, to hurt their perception, check out the video right here. All right.

[01:03:40]So, I saw something I wanted to talk about. Um, let's get rid of these So, it's uh Claude Code day or code with Claude Day.

[01:03:55]There's a few things I don't know. I haven't had a chance to check this out.

[01:03:58]So, I'm kind of looking at it just uh live right now. This is multi-agent orchestration.

[01:04:04]Lets a lead agent delegate to specialists that work in parallel on complex jobs. I think this is the future, right? Multimodel, multi- aent.

[01:04:14]That's the way. It's not only a way to get higher quality, it's also a way to be more efficient and to lower costs.

[01:04:23]You you can usually use the frontier model as the planner, as the orchestrator, and then use less capable models as the actual workh horses, right? And so I think that's this is really smart.

[01:04:38]uh available today on the claude platform.

[01:04:42]This is called managed agents.

[01:04:52]They also talk about dreaming. Dreaming extends memory by reviewing past sessions to find patterns and help agents self-improve. Very, very cool.

[01:05:04]We're also making outcome. Okay, so let's see.

[01:05:09]building self-improving agents with dreaming. By the way, Peter Steinberger, who we had on the live show a couple months ago, right when OpenClaw started to blow up, we asked him, "What's the thing you're most excited about? What feature have you not built that you really want to build?" And what did he say? Dreaming. He said, "There's this huge opportunity to allow the models to get better asynchronously, overnight, quote unquote dreaming. allow them to figure out how to make themselves better. And there's also other reasons to allow these models to do, you know, whether you call it sleeptime compute or dreaming. Basically, the GPUs are underutilized at that time relative to their utilization during peak hours, right? Obviously, there's not enough compute in the world right now, but relatively speaking, compute in the middle of the night when demand is relatively lower, electricity costs are relatively lower, it's actually a good time to do some of that compute to offload the compute necessary during peak time. And so if they can do some almost like pre-ompiling of answers of improving the models allowing them to get to an answer more quickly more of a direct route it is cheaper than to run them during peak hours. I think that's very very interesting.

[01:06:43]Yeah. Let's let's let's give props where do Nick's 99 problems. Leta did memory first. Not only did they do memory first, they also did sleep time compute first, right? Uh shout out to Charles Packer and the rest of the Leta team really on the cutting edge of what makes or what agents are capable of really like coming up with completely novel ideas. Um, so this is from April. This is literally a year ago. Sleep time compute beyond inference scaling at test time. I made a video about this. Um, so if you want to see that, go check it out. It's on the channel. Um, but there he is. Charles Packer, Sarah Waters, shout out to the Leta team.

[01:07:53]So, dreaming surfaces patterns that a single agent can't see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team.

[01:08:09]It also restructures memory so it stays high signal as it evolves.

[01:08:26]Um let's see. It also restructures memory. Uh it also restructures memory so it stays high signal as it evolves.

[01:08:32]This is especially useful for longunning work and multi-agent orchestration.

[01:08:37]Dreaming. Dreaming. Very cool. Let's see if we can get a better or image here.

[01:08:46]Okay. So, yeah. So, you can just But you have to click a button to start dreaming, which is weird. Should do it automatically, but I guess it's using compute, thus it's costing money. So, very interesting.

[01:08:59]Uh, Luigi Clemente, any love left for open source? Absolutely.

[01:09:05]In fact, I think a lot of people took my previous videos about DeepSeek the wrong way. I am just as bullish on open source as ever. I want open source. I hope the US develops a stronger open-source strategy. I just don't know how to make that happen.

[01:09:30]Okay, so together, memory and dreaming form a robust memory system for self-improving agents. Memory lets each agent capture what it learns as it works. Dreaming refineses that memory between sessions, pulling shared learnings across agents and keeping it up to date.

[01:09:57]>> [snorts] >> Okay. Then they also released something called outcomes.

[01:10:05]Okay. Let's read what outcomes is.

[01:10:08]Uh with outcomes, you write a rubric describing what success looks like and the agent works towards it. This sure sounds like slashgoal or uh what is it called in factory AI factory was really the first company to do this. Uh I I am an investor in factory by the way disclosure.

[01:10:37]Do y'all know what that feature is called in factory? Um, missions. That's right. Okay. Missions.

[01:10:47]So, yeah, factory was really the first.

[01:10:49]It's the overarching goal. It's called a mission. I know OpenAI has it. It's called goals. And now anthropic has it.

[01:10:56]It's called outcomes.

[01:10:59]Uh, right. Brian Ralph Ralph Loops.

[01:11:03]Maybe they were the first.

[01:11:05]Um, okay. A separate graater evaluates the output against your criteria in its own context window, so it isn't influenced by the agents reasoning. When something isn't right, the greater pinpoints what needs to change, and the agent takes another pass.

[01:11:25]Okay. Outcomes is particularly useful for tasks that require attention to detail and exhaustive coverage.

[01:11:35]All right. You know, I'm actually um Let's see. I think I'm going to stop my actual local recording and restart it just so I can start uploading this video.

[01:11:59]Click. One, two, three, four, five.

[01:12:02]Okay. So, restarted. I'm going to upload now.

[01:12:16]Okay, give me one sec, y'all.

[01:12:30]Um, space X X anthropic deal.

[01:12:45]Okay.

[01:12:47]All right. It's uploading.

[01:13:10]Okay. So, it also works for subjective quality like whether copy matches a brand voice or a design follows visual guidelines. In testing, outcomes improved task task success by up to 10 points over a standard prompting loop with the largest gains on the hardest problems. Outcomes also improved file generation quality with a uh 8.4 4% task success on Doc X and 10.1 on PowerPoint in our internal benchmarks.

[01:13:40]Okay, we have multi-agent orchestration.

[01:13:43]I feel like they announced this already.

[01:13:44]We're just seeing it for the first time now. Okay, so that's it.

[01:13:50]Um, let's see. Have they announced anything else cool?

[01:13:58]Anything else worth talking about that they announced?

[01:14:27]Um, so they are live right now.

[01:14:54]>> You know what I find interesting? Um, there's only there's less than a thousand people watching this. When OpenAI goes live on YouTube, they have tens of thousands of people watching.

[01:15:09]However, if you look on X, every single post that Anthropic posts, any little feature that they post gets millions and millions and millions of views, but OpenAI gets far fewer. So I think it's kind of it kind of speaks to where the audiences are. I think about this a lot obviously being a content creator. X is the insiders. Um so anthropic when they post something it's much more uh digestible by insiders and I just think the broader public knows chatbt much better than claude and anthropic.

[01:15:48]Do you know that the US District Court Northern District of California is live right now? Is it okay?

[01:16:01]>> Um, >> are these things starting to blend?

[01:16:07]>> I think it's audio only. Is that possible?

[01:16:16]>> Yeah, there's no video here.

[01:16:31]Yeah, very interesting. Okay. Um, I think that might be it.

[01:16:40]Let's see. There was some other interesting stories.

[01:16:45]I don't think there's anything else interesting coming out of this uh COD Claude code with Claude Day.

[01:16:52]Um let's see. I'm just looking around. Any other topics you all want to hear about?

[01:17:14]There's that subcq AI insurance cerebrus.

[01:17:33]Uh Shauno, forgive the question. I was trying to find whether you have any openclaw agent MD files available. I thought you shared some. Yeah. Uh I posted um a gist with all of my OpenClaw files, all of the MD files that I use.

[01:17:49]Um it's I don't remember exactly where they are. They're uh on one of the videos that I made about OpenClaw a couple months ago, one of those use case videos. It's it's in the description of one of those. You should be able to find it.

[01:18:08]Deep Seek on Huawei chips. Luigi. Yeah, I mean I made two videos about this. I don't know how much more there is to talk about it. What do you want to hear?

[01:18:21]Uh Sahill, can you talk about Ilia's company and Marad's company? I know nothing about them. They they don't release any products publicly.

[01:18:32]Uh certainly not Ilia. I mean, he's straight shot to AGI. No monetization, no products. If it's not AGI, it's a flop. Um, okay. I see you, Brian. Yep. I saw I saw your message. I'll get to that in a sec.

[01:18:48]Um, can you talk about Yeah. So, I don't know. There's not much to talk about. I wish I knew more information about these two companies. I would love to talk about them, but they don't care. They're just like heads down building, which, you know, great. Good for them. Um, Brian, producer, editor Brian, don't end without warning me. Don't tell me what to do, Brian. Um, I'll launch the new Matthew Burman video and then everyone can go watch it. Uh, and when the stream ends. Yeah. So, right after the stream ends, we're dropping a new video and it is about anthropic and uh it's a good one. It's uh I I won't give too much away. Go watch it.

[01:19:31]But it it's it's about how wild anthropic is and how they like how they think about artificial intelligence compared to how open AI thinks about artificial intelligence and how that is shaping an entire industry right before our eyes.

[01:19:47][laughter] Um yep. So we'll we'll get to that. We'll drop that in a moment.

[01:19:56]Any release for tomorrow Droid Jam? Not that I know of.

[01:20:01]Uh even if I did know, I wouldn't be able to share. But no, there's nothing that I'm aware of.

[01:20:12]Flash Breather, they're all good vids, bro. Thank you. I appreciate that. Thank you very much.

[01:20:24]Uh Rich Beer, subcq topic in my opinion.

[01:20:27]Titans paper is still better approach. I don't know the Titans paper. Titans literally memorizes context into its parameters during the session. Memory is baked into the weights rather than stacked in AKV cache. Interesting. Maybe I do know about that actually.

[01:20:46]Wasn't it like the first attempt at being able to have kind of self-improving weights or not self-improving but but actual learning weights.

[01:20:57]Um, okay. Yeah, I think I'm gonna end it there, Brian.

[01:21:06]So, everybody, uh, if you enjoyed this, make sure you're subscribed to the channel, and I'm about to drop a video, so go watch it. I would love your feedback. Go like it, share it, subscribe, tell me your thoughts in comments. We do read it. Um, and then also one last thing, um, we revamped our Discord. Brian is taking the lead and and he's just done such a great job with our Discord.

[01:21:33]He actually managed to get rid of all the scammers.

[01:21:37]He's uh reigned in the the amount of topics and it's a great place uh to go hang out. So, uh, go check out our Discord. He'll drop a link for that there. The video just launched, Brian said. Thank you, Brian, for for launching that. And um we'll stream again soon. See you guys.

[01:21:55]There's the Discord link. Bye.