Latest / In the Loop AI / THIS WEEK IN AI: Real Anthropic Revenue, Meta Muse Spark, OpenAI's Next Model

Transcript

THIS WEEK IN AI: Real Anthropic Revenue, Meta Muse Spark, OpenAI's Next Model · In the Loop AI

0:00
Josh: This week, Meta released MuseSpark, its new flagship AI model built from scratch
0:04
Josh: by Alexander Wang's Super Intelligence Lab.
0:07
Josh: And noteworthy for the very first time, it's not open source.
0:10
Josh: They closed source the entire model, but the model isn't what caught our attention, it's distribution.
0:15
Josh: MuseSpark is rolling out across Facebook, Instagram, WhatsApp,
0:19
Josh: and the Rayvan Metaglasses to over 3 billion users.
0:22
Josh: That's more than every other AI company on Earth has. This is a really big deal.
0:27
Josh: Meta's been calling us personal superintelligence. You log in with your Facebook
0:30
Josh: or your Instagram account, and then the model pulls from your social graph that
0:33
Josh: it has based on the past, what, decade that you've been using these applications.
0:37
Josh: So no other model in the world has the type of information that this new meta
0:41
Josh: AI model is going to have. Now, here's where it gets interesting.
0:44
Josh: Meta is also developing another research project in parallel called Tribe V2,
0:48
Josh: something that almost nobody's talking about, but it's trained on brain scans from human beings.
0:53
Josh: And it understands how you react to certain imagery in certain videos so that
0:57
Josh: it knows when a video or a piece of content is going to light up part of your
1:01
Josh: brain that makes you interested and engaged.
1:03
Josh: Now, they're claiming to do both of these separately and in parallel,
1:06
Josh: but you have to ask the question, what happens when these things merge together?
1:09
Josh: What happens when Meta knows everything about you, they know everything that's
1:13
Josh: going on in your brain, and they're able to serve the absolute best content possible.
1:16
Josh: They become the 21st century drug dealers through this artificial intelligence.
1:20
Josh: And maybe that's a doomer take.
1:22
Josh: I know, Ejaz, you have a very different take on this, perhaps.
1:24
Josh: We're going to talk about that and a lot of other things on today's episode,
1:27
Josh: including Anthropik's not-so-real annual recurring revenue. We have some news
1:31
Josh: around OpenAI and then some cool robotic arms that we actually had the guest on a few months ago.
1:35
Josh: He is back with the final products. There's a lot going on this week,
1:38
Josh: but first, Meta, Ejaz. Meta's a pretty big story this week.
1:41
Ejaaz: They released MuseSpark, which is the first Frontier AI model that they've released
1:45
Ejaaz: in over a year now, which is a crazy amount of time in the AI world. Now, I'm going to
1:52
Ejaaz: I'm thinking about the story arc of Meta, and it's hard to be over-enthusiastic about this model.
1:58
Ejaaz: Last year, Zuck burned around $75 billion on AI CapEx.
2:03
Ejaaz: He then fired 600 of his AI staff, and then hired 150 more staff,
2:08
Ejaaz: spending $25 billion to do so.
2:11
Ejaaz: $15 billion, of which he used to hire one guy, Alex Wang, who leads Meta Super
2:16
Ejaaz: Intelligence Labs, who helped build this model. So I'm expecting big things at this point.
2:21
Ejaaz: Unfortunately, the model underperforms Claude Opus 4.6 and GPT 5.4 across the
2:28
Ejaaz: most important benchmarks, which is coding and reasoning.
2:31
Ejaaz: So it's easy to be bearish about this model, but I dug into it more.
2:34
Ejaaz: And what I realized is that's not what Zuck or Meta is going after at all.
2:38
Ejaaz: They're going after what I'm dubbing personal AGI.
2:42
Ejaaz: Personal AGI is basically super intelligent AI that's trained on your own personal
2:47
Ejaaz: data. It'll become an AI assistant that basically lives and breathes you.
2:51
Ejaaz: So how is Meta going to be able to pull this off versus all the other competitors
2:55
Ejaaz: who are leagues ahead of them?
2:57
Ejaaz: Well, they have something that none of them have, which is the personal data
3:00
Ejaaz: and very incriminating data of 3.4 billion daily active users.
3:05
Ejaaz: So not only do they have a crap ton of data, they also have the data being refreshed every single day.
3:11
Ejaaz: Now, I'm personally not a user of Facebook, but I use WhatsApp.
3:14
Ejaaz: I use Instagram every single day.
3:17
Ejaaz: So they have already collected a bunch of data around what type of conversations
3:21
Ejaaz: I'm having, what type of media that I'm liking, what kind of memes that I enjoy.
3:26
Ejaaz: And they can use this to engineer the perfectly crafted AI assistant that no
3:30
Ejaaz: other lab can do. And this is what they're doing.
3:33
Ejaaz: So the benchmarks that do matter, if you look at this table over here,
3:36
Ejaaz: they excel or this model excels in visual reasoning and multimodality.
3:40
Ejaaz: So it can look at a picture and understand it and perceive it better than Opus 4.6 and ChatGPT 5.4.
3:47
Ejaaz: And so if they're going to end up building a model that ultimately can feed
3:50
Ejaaz: you a better algorithm or feed you the best content, this is going to be the
3:53
Ejaaz: model that enables that. And that's what they're doing.
3:55
Josh: Yeah, the personal AGI, it's a good way to put it. I don't think this needs
3:58
Josh: to be everything for everyone. This just needs to be better than nothing.
4:02
Josh: Because the reality is that a lot of people still aren't using these AI tools
4:05
Josh: in their day-to-day life.
4:06
Josh: What I found interesting is the examples that they showcase.
4:09
Josh: Now, we don't actually know what it looks like applied to these social media
4:12
Josh: networks. We're not sure what that integration is going to look like.
4:14
Josh: But they showed a few examples in the blog post that they shared,
4:17
Josh: which is it has the visual intelligence.
4:19
Josh: When you point your camera at something like a refrigerator,
4:21
Josh: say, it knows what's in there. it can calculate the amount of protein in that
4:25
Josh: block of cheese or the amount of calories in that hamburger that's sitting there.
4:29
Josh: And I think that's really interesting.
4:31
Josh: They had another example of someone who was doing yoga and you could actually
4:34
Josh: analyze the positions of the arms and the legs and the body and you could see
4:39
Josh: if they were doing it properly.
4:40
Josh: So the visual element of it, I think the native multimodality is something that's
4:44
Josh: pretty noteworthy and interesting.
4:45
Josh: Yeah, here's the example of the food. Here's the yoga pose.
4:48
Josh: I'm not sure what that looks like fully integrated into...
4:52
Josh: Our social media experiences today, outside of what could be just like a much stronger algorithm.
4:57
Josh: But they did release meta.ai, which is a website that you can go to to actually
5:00
Josh: try this out yourself and run some tests.
5:02
Josh: So I would encourage anyone who's curious, who wants early access to the model
5:05
Josh: before it hits your Instagram feeds or WhatsApp to go actually on the website
5:09
Josh: and check it out for yourself. It's available for anyone with an account.
5:12
Ejaaz: Yeah. And it's been a labor of love for Zuck. I think he famously said on an
5:15
Ejaaz: interview about a month ago that he is willing to spend the entire wallet on
5:22
Ejaaz: AI CapEx and training the best AI model versus risking losing out on this race entirely.
5:27
Ejaaz: He would rather make an incredibly expensive mistake than produce a subtale model.
5:31
Ejaaz: They spent the last nine months pre-training this model. Fun fact,
5:34
Ejaaz: typically pre-training takes
5:36
Ejaaz: around a couple of months and then you kind of go into post-training.
5:39
Ejaaz: So they've invested a lot of time, money and heart into this.
5:42
Ejaaz: So that's why I'm kind of like, I'm kind of split as to why this model isn't
5:45
Ejaaz: as good enough. They delayed it.
5:47
Ejaaz: They delayed its own launch about a month as well. But the CapEx spend keeps on going.
5:53
Ejaaz: They today announced a partnership with CoreWeave to the tune of $21 billion.
5:58
Ejaaz: So they're just pushing ahead with all the data center type stuff.
6:01
Ejaaz: But you mentioned a model earlier on, Josh, which kind of went viral a couple of weeks ago.
6:06
Ejaaz: And this is the scary part of this entire story. I would love to think that
6:10
Ejaaz: Meta is in my best interests and wants me to become a better productive human.
6:16
Ejaaz: But stuff like this, this new model, Tribe V2, which basically reads your brain
6:20
Ejaaz: signals and predicts what type of content is going to stimulate different parts
6:25
Ejaaz: of your brain and make you more engaged in certain types of content, doesn't give me hope.
6:29
Ejaaz: Because presumably, they're going to use this model to understand what types
6:33
Ejaaz: of content stimulate different people's appetites for videos,
6:38
Ejaaz: reels, memes, or whatever that might be.
6:40
Ejaaz: And then AI generate this exact amount of content that they can feed you in
6:44
Ejaaz: the timeline, because they already have the distribution, right?
6:45
Ejaaz: They already have the social media platforms, they can create and perfect the
6:48
Ejaaz: content and own that without relying on creators.
6:51
Ejaaz: They now have a recursive dopamine loop that they can just trap you in.
6:54
Ejaaz: Doesn't give me hope. Maybe that's a do my take.
6:56
Josh: Well, I'm curious what the stated intention of this model is, right?
7:01
Josh: It's like, why are you trying to understand how the brain reacts to certain
7:05
Josh: type of content? And sure, it's interesting.
7:07
Josh: I know they're working on their metaglasses and they have that like wrist strap
7:10
Josh: that's kind of like a neural interface that detects your hand movements and
7:13
Josh: understand this information is probably more helpful to their future hardware products.
7:16
Josh: But also, I mean, the clear and obvious use case
7:19
Josh: for this is understanding what types of impulses trigger
7:23
Josh: when you watch a specific type of content and then funneling that
7:26
Josh: into your feed on a daily basis and luke who
7:29
Josh: is behind the scenes he was mentioning earlier that in a way
7:32
Josh: they become like the this high-end drug dealer that can feed
7:35
Josh: you like some dopamine here or some cortisol here and they know exactly
7:37
Josh: how a video is going to hit your brain and in fact they can optimize those
7:41
Josh: videos through a very tight feedback loop to improve them to a point in which
7:44
Josh: it can guarantee that the part of your brain that they want to fire will fire
7:48
Josh: and that level of deep understanding not only from the data and preferences
7:52
Josh: they've collected over the last decade but also now understanding the biological
7:55
Josh: brain and truly at a deep level how it works.
7:58
Josh: I can see the good use cases for it, but man, it gets scary quick because they
8:02
Josh: have their Vibes app. I don't know if anyone remembers it.
8:05
Josh: I'm not sure anyone's even used it, but it's the short form scroller that is
8:09
Josh: AI generated content only. And this is like injecting that with steroids.
8:13
Josh: And I could see a world in which this content gets very good very quick.
8:16
Josh: And is it a good or bad thing?
8:17
Josh: That is TBD. You can argue that I would prefer to get good ads that are personalized
8:22
Josh: for me that actually sell me things that I'm interested in versus nonsense.
8:25
Josh: But at the same time, it's not my decision because they will have more understanding
8:30
Josh: of how my brain works than even I will. And that's a little unnerving.
8:33
Ejaaz: I'm going to argue on the side of Skynet. It's a bad thing.
8:37
Ejaaz: Prove me wrong, Suck, if you're watching this or anyone from Meta. I would love to know.
8:41
Ejaaz: Tell me the counter argument. But the good news is, even if they are going after
8:45
Ejaaz: your brain and hijacking your dopamine circuits, you might be able to get paid for it.
8:50
Ejaaz: This was a story that broke a few weeks ago where this kid, or I think she's
8:55
Ejaaz: like a young adult now, basically won a court case against Meta and YouTube
9:01
Ejaaz: where she got paid $3 million.
9:03
Ejaaz: For the effects of social media, specifically Meta's platforms and YouTube,
9:08
Ejaaz: had on her depression and personality, her core years that formed her adolescence.
9:14
Ejaaz: She got a three million paycheck.
9:15
Ejaaz: I'm wondering if I could do this because I definitely suffered from a lot of this growing up.
9:19
Josh: Yeah, what a come up. I mean, it sets a somewhat dangerous precedent,
9:22
Josh: right? It's like I spend a lot of time on my computer and my mental health,
9:25
Josh: sure, it could be improved.
9:26
Josh: Am I eligible for a $3 million raise? And if not me,
9:30
Josh: I certainly know some other people nearby that definitely qualify
9:33
Josh: for this if that's the parameter that counts so three
9:36
Josh: million dollar paycheck i don't know not a bad deal dangerous precedent
9:39
Josh: if people are going to be able to start claiming that um as they go because
9:41
Josh: i mean again this company has it's done a lot of great things but social media
9:45
Josh: as a whole has generated a lot of damage specifically around um the meta-owned
9:49
Josh: companies and we'll see what happens meta has been disappointing they continue to be disappointing
9:57
Josh: But I am hopeful that a founder-run company led by someone like Zuck can figure
10:02
Josh: out a way to really turn this into something special and meaningful and positively impactful.
10:07
Josh: So we'll see. I think that is the meta story.
10:10
Josh: Now we have another story that has left me a little unnerved,
10:15
Josh: which is the annual run rate revenue story between Anthropic and OpenAI and
10:19
Josh: how they're actually counting their revenue.
10:21
Josh: Because the headline is that what? Anthropic just went from $19 billion to $30
10:25
Josh: billion in annual revenue in like a month.
10:28
Ejaaz: Yeah. So the breaking news here was that Anthropic had signed a multi-billion
10:33
Ejaaz: dollar deal with Google to basically use their GPUs to train Claude.
10:36
Ejaaz: But there was a hidden nugget in this article, which revealed that Anthropic's
10:41
Ejaaz: revenue run rate, their ARR, has officially hit $30 billion.
10:45
Ejaaz: Now, for context here, at the end of last year, it was $9 billion.
10:50
Ejaaz: At the start of the year, it had just about hit around $12 billion.
10:54
Ejaaz: Last month, it hit $19 billion. In a single month, it has gained $11 billion.
11:01
Ejaaz: Now, that's the result of Claude Opus being amazing and the rumored Mythos model,
11:05
Ejaaz: which is now confirmed, we did a whole episode about this, being as good as people claimed.
11:10
Ejaaz: So people are obviously buying Claude subscriptions and running their revenue rate up to $30 billion.
11:15
Ejaaz: I thought this was amazing until we had a conversation this morning,
11:18
Ejaaz: Josh, where you said that this is fake news?
11:20
Josh: This is nonsense. The accounting is technically GAAP compliant,
11:25
Josh: but the way that they get there is so vastly different that you really can't
11:28
Josh: compare the two companies together.
11:30
Josh: I mean, this example is showing Anthropic is at $30 billion of annual revenue
11:34
Josh: and OpenAI is at $24 billion.
11:37
Josh: But then explain to me why OpenAI is valued at twice the current valuation of
11:40
Josh: Anthropic. It's an accounting problem.
11:43
Josh: Now, open ai has a deal with microsoft where open ai shares 20 of its revenue
11:46
Josh: with microsoft in the financial statements it counts those sales before that
11:51
Josh: deduction but for like azure cloud customers buying open ai models open ai only
11:55
Josh: books 20 of that cut as the revenue
11:58
Josh: And Anthropic, they have a deal with AWS, Google, and Microsoft,
12:01
Josh: and all three of those cloud providers resell cloud to their customers.
12:05
Josh: Anthropic books the entire thing as revenue, and then marks off the 80% cut
12:10
Josh: as a line item in marketing expenses.
12:13
Josh: And that is such a huge difference. I mean, both of those are technically GAAP
12:17
Josh: compliant, but they produce very different top line numbers.
12:20
Josh: And I think this is really important. And a lot of people are missing this,
12:22
Josh: That when they see Anthropic projecting a $30 billion annual recurring revenue,
12:28
Josh: that's only projecting out what they've currently done for the last four weeks.
12:32
Josh: And it's counting all revenue, including the amount that they're going to have
12:37
Josh: to give back to AWS, Google Cloud, and Microsoft Azure.
12:40
Josh: And that's like a huge accounting difference. That is a big problem that I don't
12:44
Josh: think a lot of people are taking note of.
12:46
Ejaaz: I have so many thoughts on this. Number one, how is this compliant and legal?
12:50
Ejaaz: Why are they allowed to do that?
12:52
Ejaaz: Number two, this is financial accounting crime. They shouldn't be able to do that.
12:57
Ejaaz: But something isn't adding up for me, which is one thing that I see a lot of
13:01
Ejaaz: these articles comparing is how much revenue they're making,
13:05
Ejaaz: but also how much they're burning.
13:06
Ejaaz: So for example, with OpenAI, they're making, let's say, $25 billion this year,
13:11
Ejaaz: but they're also burning $25 billion dollars.
13:14
Ejaaz: And so if Anthropic is moving the line item for their cost center to just say
13:19
Ejaaz: marketing, shouldn't their burn rate still be higher than what is being reported on?
13:24
Ejaaz: Like something doesn't make sense. Like, is the Financial Times just wrong and
13:27
Ejaaz: they're missing this line item completely? Or is Anthropic genuinely just not
13:32
Ejaaz: burning as much still and they're still on track to making a profit by 2028?
13:36
Ejaaz: Because that's what all the projections have it at, right?
13:38
Ejaaz: They're going to be making a profit or turning a profit much sooner than OpenAI
13:42
Ejaaz: is who are spending way too much on capex so there's something i'm missing though.
13:45
Josh: Yeah i wonder if they're not counting it as burn because it's it's technically
13:50
Josh: not like they they are earning money on it just 20 of what they are actually
13:56
Josh: putting on the accounting book so like they're not actually burning cash they're
14:00
Josh: just accounting for more cash than they're going to right but the
14:03
Ejaaz: 80 is being put on the marketing expenses book right well i'm not misunderstanding that.
14:07
Josh: Uh, that one we're going to have to talk to an accountant about. I'm not sure.
14:11
Josh: I do know that we are comparing apples to oranges though. And we look at open
14:14
Josh: AI and Anthropics run rate. And I think this is something that,
14:16
Josh: um, an IPO is going to fix.
14:18
Josh: Once we have all of these documents publicly stated again, a lot of this is insider reporting.
14:22
Josh: You need a subscription just to access this document that we're showing on the screen.
14:26
Josh: It is this like really messy, uh, article also in the article or really messy accounting.
14:32
Josh: And also in the article, they mentioned how they count their revenue,
14:35
Josh: which is just projecting their prior four weeks out for a multiple of 13 to
14:40
Josh: account for all 52 weeks.
14:41
Josh: So there's a lot of handwaving going on in order to make these numbers go up into the right.
14:46
Josh: I just think it's important when you look at these headlines to really take
14:49
Josh: it with a grain of salt and understand that Anthropic actually didn't gain $11
14:53
Josh: billion in revenue in like a week. That's not happening.
14:56
Josh: They're just doing these funny accounting methods to make things look like they are going great.
15:01
Josh: And they are going great, just perhaps not as great as people perceive.
15:04
Ejaaz: And it might be getting a lot greater for OpenAI.
15:08
Ejaaz: I mean, talking of like the IPO rumor mill and like boosting valuations,
15:11
Ejaaz: it's all marketing, right? It's all storytelling.
15:13
Ejaaz: And there was a story that Axios broke this morning before we started recording
15:18
Ejaaz: that OpenAI plans to release their next model. Now they're dubbing it SPUD.
15:23
Ejaaz: Some people are calling it GPT 5.5. Some people are calling it GPT 6.
15:27
Ejaaz: But apparently it's going to be so good that they have to do a limited release
15:31
Ejaaz: that they can't release it publicly because it's too good or too dangerous.
15:35
Ejaaz: Now, if that sounds similar, it's because that's exactly what Anthropic did
15:40
Ejaaz: this week, announcing Claude Mythos, their next AGI-like model,
15:45
Ejaaz: which they're not releasing publicly because it's a cybersecurity risk.
15:47
Ejaaz: It discovered 1,000-plus security vulnerabilities.
15:50
Ejaaz: And what was interesting about that entire news cycle earlier this week is someone
15:55
Ejaaz: commented and said, well, if it's that good and if it's that expensive,
15:59
Ejaaz: we're probably not gonna get to use this model for our next another couple of months, basically.
16:03
Ejaaz: And OpenAI's head of model training, Thibault, basically said,
16:07
Ejaaz: Uh, I wouldn't count on that. He said, um, dot, dot, dot, which implies that
16:12
Ejaaz: OpenAI is going to release a very similarly capable model very soon.
16:15
Ejaaz: So that's what, uh, that's the news that basically Axios broke.
16:19
Ejaaz: But I just want to point out, I want to check myself here that it might also be fake news.
16:24
Ejaaz: Um, this post from Dan Shipper apparently spoke to someone within OpenAI and
16:28
Ejaaz: he basically said his, his contact basically said that we're talking about a
16:32
Ejaaz: different model here that is, uh, hyper-focused on cybersecurity specifically.
16:36
Ejaaz: Typically, and it'll be a separate release to GPT-6.
16:38
Ejaaz: So at this point, I have no idea what's going on, but I know that OpenAI's valuation
16:41
Ejaaz: for their IPO has probably gone up in the space of time that this has happened.
16:45
Josh: Yeah, I can't imagine they don't already have a model that is close to Mythos, if not already there.
16:53
Josh: And if they don't, then they're, I'm sure, just weeks behind actually having something like that.
16:57
Josh: What I found interesting about this story is it says that they're working explicitly
17:01
Josh: on a cyber product, which implies that it's for cybersecurity.
17:06
Josh: And with Claude Mythos, they weren't actually training it on cyber at all.
17:09
Josh: It was just a downstream effect of really powerful code.
17:12
Josh: And when I think about the pivot that OpenAI has had recently from retail to
17:16
Josh: enterprise, a lot of that focus has been around code.
17:19
Josh: So I wonder if they're just building a really strong coding model and this is
17:21
Josh: a downstream effect, or if they're genuinely training something explicitly on
17:25
Josh: cybersecurity. And I have to imagine, if it's trained explicitly on cybersecurity...
17:29
Josh: It probably will become better than mythos and then
17:31
Josh: you have to ask the question what happens three months from now
17:34
Josh: when these tools actually do become available and also who's going
17:37
Josh: to decide when to release them if open ai is taking the clod route or
17:40
Josh: the anthropic route and they're keeping it private well now suddenly
17:43
Josh: we have everyone's worst nightmare where the labs are kind of all working together
17:47
Josh: they all have the most powerful stuff and there is no counterbalance to whatever
17:53
Josh: they decide to do with it and they in a way become those king makers and when
17:57
Josh: you think about like the Department of War contract that we had this whole episode about is a big mess.
18:01
Josh: Anthropic kind of has all the leverage. And if OpenAI moves with them,
18:05
Josh: they have the ability to crack nation state software at will.
18:10
Josh: And that seems really powerful. And they are now the gatekeepers.
18:13
Josh: And I don't know, Chris, a lot of interesting questions as we move into this
18:16
Josh: next paradigm of Blackwell models that are unbelievably powerful.
18:20
Ejaaz: But as we mentioned earlier, Project Glasswing, Anthropic mentioned their breaking
18:24
Ejaaz: tier model, Claude Mythos, which is coming out. We did an entire episode about
18:28
Ejaaz: this earlier this week. Definitely go and check this out.
18:31
Ejaaz: Now, I want to end this segment on a bit of chart crime, Josh,
18:35
Ejaaz: because now that you've told me about the fake revenue numbers,
18:37
Ejaaz: I can't look at this Wall Street Journal analysis and think,
18:40
Ejaaz: is Anthropik's numbers fake? On the left here, we see yearly AI model training costs.
18:45
Ejaaz: And it basically shows that OpenAI is spending way too much money.
18:48
Ejaaz: And Anthropik is spending a fraction of that.
18:50
Ejaaz: But now I'm realizing that that's probably in projection with their revenue.
18:53
Ejaaz: And if you look at the bottom left over here, it shows that they're kind of the same.
18:57
Ejaaz: So I think there is some chart crime or county crime happening with Anthropic,
19:01
Ejaaz: and I want that to be talked about more.
19:03
Ejaaz: But it's not all rosy with OpenAI.
19:06
Ejaaz: They did have the UK Stargate project go down, right?
19:09
Josh: The whole Project Stargate thing is really just a huge disappointment.
19:13
Josh: It was supposed to be this giant grand buildout domestically, internationally.
19:16
Josh: I think they also did this in Saudi Arabia, somewhere in the Middle East.
19:20
Josh: They were planning to build one. They didn't do it in the UK.
19:23
Josh: They're not doing it in the U S they're not doing it. Project Stargate was really
19:27
Josh: just as big, like raw, raw project that was initiated with the government.
19:31
Josh: Elon famously said the post the day that it was announced, you don't have the
19:34
Josh: money to do this. Turns out he was right. No one actually wants to foot the bill for this.
19:39
Josh: Logistically, it's very technical and challenging. And it turns out it's just
19:42
Josh: really hard to build things at scale,
19:44
Josh: particularly internationally in places like Europe and the UK,
19:47
Josh: where there's a lot of regulatory issues, both environmental and energy related,
19:52
Josh: that just make doing these things incredibly difficult.
19:54
Josh: And we're seeing it here domestically. We have a story a little bit later about
19:57
Josh: AI data centers and how they're just having a really tough time getting them
20:00
Josh: online. And this is probably one of the future trends that we're going to look
20:03
Josh: out for is just the idea that building these data centers is hard.
20:08
Josh: And not always just because of the technicalities, but also because of the regulation
20:11
Josh: and legislation associated with this and the public sentiment. It's not looking good.
20:16
Josh: There is certainly a rift in the world right now between people who want to
20:20
Josh: build these things and people who do not.
20:22
Ejaaz: Shifting gears slightly, I wish Anthropix Claude Mythos was the main headline,
20:27
Ejaaz: but they're still shipping other products somehow.
20:30
Ejaaz: Maybe they're using an AGI-like model to do so, right? They announced a new
20:33
Ejaaz: product called Claude Manage Agents, which is basically, think of AWS,
20:38
Ejaaz: but for spinning up AI agents.
20:39
Ejaaz: It's a platform that allows you to design and architect an agent through a single
20:44
Ejaaz: prompt or a couple of prompts.
20:45
Ejaaz: You can amend the memory, a bit of its own custom design, and then launch it,
20:48
Ejaaz: which may not sound novel, right?
20:50
Ejaaz: You could create AI agents before, but there's a distinct difference here.
20:54
Ejaaz: Typically, when you create an AI agent, you can't scale it to your millions of users.
20:59
Ejaaz: Let's say if you're a Fortune 500 company, because you need to set up a bunch
21:02
Ejaaz: of other production and dev tooling environments in order to support that.
21:06
Ejaaz: That typically takes anywhere between three to six months.
21:09
Ejaaz: Now, I ran a bit of the math on this. Typically, an app development at scale
21:13
Ejaaz: for, say, a million users costs around $50,000, depending on the specific type
21:18
Ejaaz: of feature or product you want to launch.
21:19
Ejaaz: This reduces it down to $100. That's like a 500x reduction, and you can do it in under an hour.
21:26
Ejaaz: But there is a new cool thing that this product unlocks. Josh, can you guess it?
21:30
Josh: I'm going to guess the end of white call it work or am I being dramatic?
21:34
Ejaaz: Well, that's also that, but they figured out a new revenue run rate for them.
21:39
Josh: Of course they did. Of course they did.
21:42
Ejaaz: Right now. Now listen to this genius plan, right?
21:44
Ejaaz: Typically, every single AI model provider charges you based on the amount of tokens you use.
21:51
Ejaaz: Tokens in, tokens out, you pay a subscription, or you pay an API cost.
21:56
Ejaaz: Anthropic, for this product, is charging you for the amount of time that your
22:00
Ejaaz: agent takes to think of a solution.
22:03
Ejaaz: So the tokens it's using to think is now being charged to the tune of $0.08
22:09
Ejaaz: per call. Now, if you assume that, and I actually don't need to assume,
22:13
Ejaaz: they used a live example of Sentry processing 1 million bug reports.
22:17
Ejaaz: If each agent session is 10 minutes, that's 166,000 session hours,
22:22
Ejaaz: which turns out to be around $13,000 per run in fees.
22:27
Ejaaz: That all adds up massively. If you're a massive corporate enterprise where you
22:31
Ejaaz: distribute this platform to whatever, 50 plus teams, you end up making millions of dollars.
22:36
Ejaaz: Genius unlock anthropic well done i'm clapping for you also you're going after
22:40
Ejaaz: all the other startups out there you probably killed out a bunch of different
22:43
Ejaaz: agent harness startups that will value their billions of dollars well done.
22:47
Josh: If you use a computer you just have to assume that
22:50
Josh: you're months away from no longer needing to use a computer for anything you
22:53
Josh: don't want to do yeah um it will just be automated you
22:56
Josh: can have the computer watch your screen emulate your decision making processes
22:59
Josh: do all the clicking and thinking that you would need to do in order
23:02
Josh: to accomplish whatever you're trying to do and that's been
23:04
Josh: it's been a recent realization particularly with mythos about how close we are
23:07
Josh: to this reality and i have to imagine that all these features are being built
23:11
Josh: with that model and are therefore resulting in this incredibly fast iteration
23:15
Josh: loop that anthropic's having where every day we get some like groundbreaking
23:18
Josh: technological impact if you just push this out three months six months even 12 months
23:24
Josh: there's no way that you're going to need to use your computer for anything you
23:27
Josh: don't want to it It will just automate the entire process for you.
23:30
Josh: So Anthropic with another big win. Quad is just on fire. The rate of acceleration
23:35
Josh: is truly through the roof.
23:37
Ejaaz: Josh, are you a Black Mirror fan?
23:40
Josh: Oh, huge fan. Watched every episode.
23:42
Ejaaz: Huge fan. So what if I told you there is now a robotic lamp that you can buy
23:47
Ejaaz: that turns into pincers that can fold your clothes, make your bed,
23:52
Ejaaz: and maybe make you a cup of coffee?
23:54
Josh: You buy that? I say 100% chance it stares me through the heart when I'm sleeping.
23:57
Josh: But I'm going to hope that's not the case because we actually did have this
24:00
Josh: founder on the show a while back to talk about this product, which is now out?
24:04
Josh: Yes. Kind of. On a second, Cali, explain what's going on here.
24:07
Ejaaz: Okay. So what you're seeing on your screen is a lamp which basically extends
24:11
Ejaaz: into a robotic arm and it can do a bunch of chores for you.
24:15
Ejaaz: So what you're seeing on the screen is this lady is apparently putting her clean
24:20
Ejaaz: laundry on the front of her bed and her lamps are now activating.
24:24
Ejaaz: And now you can see their claws coming out and they can start folding your clothes
24:28
Ejaaz: and making your bed or starting your record player.
24:32
Ejaaz: And the point is robots are going to be pretty pervasive in human society.
24:38
Ejaaz: They may not necessarily look like humanoids. And that's the point that Sincere
24:41
Ejaaz: and Aaron Tan, the founder, is aiming at.
24:44
Ejaaz: Now, the last time that we spoke about this founder, he only had a mock-up.
24:48
Ejaaz: And to be honest, the lamp looked pretty scary. The pincers were much larger.
24:51
Ejaaz: It's nice to see that they've now got like a kind of curved metal piece over
24:55
Ejaaz: the pincers. So maybe they can't necessarily stab you. It's more colorful, it's more amenable.
25:00
Ejaaz: It looks kind of slow, but the good news is you can start ordering this right now.
25:04
Ejaaz: I signed up on the wait list and I got an email the other day saying,
25:06
Ejaaz: hey, you can now order this thing. I don't know when they're gonna start delivering
25:08
Ejaaz: it, but it might be something that I try out. Are you trying it out?
25:12
Josh: No, I will not be trying this out. Okay.
25:16
Josh: I love founders who are trying cool things. And I think building a narrow use
25:22
Josh: robot like this is so cool.
25:23
Josh: The design is awesome. It's a lamp that does robot things. And that's very cool.
25:28
Josh: When I look at the trailer here, and this isn't really showing many use cases,
25:31
Josh: one of which is the folding of laundry.
25:34
Josh: You kind of see the actuators there. They're a little like fidgety.
25:37
Josh: They're not quite smooth.
25:38
Josh: They're like, it doesn't seem that it's moving very quick. I can't imagine this many use cases.
25:42
Josh: Like we're looking at a robotic lamp that is being manually adjusted and turned.
25:46
Josh: Like why is that happening there's a lot of questions i have about what the
25:49
Josh: actual use cases of something like this are and how effective it is at those
25:52
Josh: use cases but again i love the idea that people are trying new things and trying
25:56
Josh: to build something unique with beautiful design that's actually effective inside
26:00
Josh: of a home and as soon as you get yours delivered i'm coming over and i'm trying
26:03
Josh: it out because i want to see i
26:05
Ejaaz: Might be dead so i don't know maybe you're recovering my corpse at that point
26:08
Ejaaz: josh but in other news space x ai is back and they've signed a massive partnership
26:14
Ejaaz: with Intel, American made fabricator of AI chips.
26:20
Ejaaz: And the question becomes, why on earth are they doing this? Well,
26:23
Ejaaz: there's a few different reasons.
26:25
Ejaaz: Number one, Elon Musk announced something called the TerraFab,
26:28
Ejaaz: which is pretty much the most ambitious AI chip manufacturing project that is
26:33
Ejaaz: ever going to be achieved if it does get achieved over the next, say, five to 10 years.
26:38
Ejaaz: The idea is to achieve one terawatts worth of compute, 80% of those AI chips
26:43
Ejaaz: are actually getting sent out to space to harvest the sun's energy to train
26:46
Ejaaz: AI models, presumably Grok.
26:48
Ejaaz: Now, if all of that sounds insane, we have a bunch of episodes that you can
26:51
Ejaaz: go and check out and we explain everything.
26:53
Ejaaz: But the point is, why Intel specifically?
26:56
Ejaaz: I think there's two main reasons. Number one, you need these AI chips to be
27:00
Ejaaz: American-made and American-manufactured.
27:03
Ejaaz: AI has become a huge geopolitical weapon and taiwan the threat of taiwan being
27:09
Ejaaz: taken over by china and tsmc being.
27:11
Ejaaz: Within Taiwan is a massive threat to the US AI production of AI models,
27:17
Ejaaz: GPUs, etc. NVIDIA realized heavily on TSMC.
27:20
Ejaaz: Intel is the closest American-made lab or manufacturing plant that we can get
27:24
Ejaaz: to building A-grade AI chips.
27:27
Ejaaz: But why is Elon signing up with Intel specifically?
27:29
Ejaaz: There was this little nugget that I saw Robert Scoble post about,
27:33
Ejaaz: which is there's this compound called gallium nitride, which basically makes
27:37
Ejaaz: these AI chips radiation-hardened, which, there you go,
27:41
Ejaaz: is going to make it perfectly suitable to launch these AI chips into space.
27:45
Ejaaz: So Elon's already thinking way in advance.
27:47
Ejaaz: Robert got into a bit of trouble because he published that Elon had liked this
27:50
Ejaaz: tweet, aka confirming that this was partially the reason why they did that.
27:54
Ejaaz: But yeah, this might be the next unlock for achieving the TerraFap. It's pretty cool.
27:59
Josh: It's one of the most ambitious projects I think any company is undertaking on earth.
28:02
Josh: There's no one who's really trying to offset this monopoly that exists on chip
28:07
Josh: fabrication and production.
28:09
Josh: And I think one of the most important things that's underrated about the TerraFab
28:11
Josh: is the fact that they have a separate staging facility separate from the TerraFab
28:15
Josh: that has all of the required pieces needed to make these chips.
28:18
Josh: It has the lithography, it has the masking, it has the packaging.
28:20
Josh: And what that allows you to do is iterate very quickly on the actual design
28:25
Josh: of these chips. A lot of times a chip gets submitted and then a year goes by
28:29
Josh: or even longer until you actually have the full thing completed.
28:31
Josh: This compresses that iteration cycle because it's all into one roof and allows
28:34
Josh: them to make chips that are far better very quick because they can do all the testing in one place.
28:39
Josh: And they can even do that prior to the TerraFab going fully online because it's
28:43
Josh: a small sample set of that.
28:44
Josh: The TerraFab is going to be hard. I'm sure there's going to be a lot of negative
28:47
Josh: press as they make mistakes and as things get delayed.
28:50
Josh: But the outcome of the terafab is so
28:53
Josh: profound that it's hard to imagine a world in which tesla
28:56
Josh: does accomplish this at scale or spacex ai does accomplish
28:59
Josh: this at scale and they are collectively not the most valuable company in the
29:03
Josh: world because that implies not only do they have the chips but they have the
29:05
Josh: robots they have the satellites they have the spaceships they have all the infrastructure
29:10
Josh: required for this next generation of embodied ai of space trained intelligence
29:15
Josh: and super super intelligence and I'm really bullish on it.
29:19
Josh: Intel's badass. I'm glad they're helping them, and I'm just so stoked for the
29:22
Josh: tariff app in general. This is going to be a fun one to follow over the next few years.
29:25
Ejaaz: Yeah. I have a huge Intel bag, so please, please, please keep signing all these partnerships. Please.
29:30
Ejaaz: Now, the reason why Josh and I started this show, the reason why we do Limitless
29:35
Ejaaz: is we're very optimistic about the tech.
29:37
Ejaaz: Now, we know that there's a lot of doomerous takes, but the fact is we believe
29:41
Ejaaz: AI is going to change the world for good, and there are many different ways
29:44
Ejaaz: that we think that's going to happen, and we're going to track every single news story.
29:47
Ejaaz: Supports that. But there are obviously people in the world that don't believe that's the case.
29:52
Ejaaz: And unfortunately, this week, we had a pretty serious story where someone fired
29:56
Ejaaz: 13 shots into the home of an Indianapolis counselor with a note reading,
30:01
Ejaaz: no data centers left at the scene.
30:04
Ejaaz: Now, we don't know the exact motivations of this person, because I don't believe
30:08
Ejaaz: they've been caught just yet.
30:09
Ejaaz: But you can hypothesize what the takes are, which is stuff that we've covered
30:13
Ejaaz: on the show before, which is people are worried that data centers are going
30:16
Ejaaz: to empower AI models that are going to eventually replace them or take their job.
30:19
Ejaaz: They're worried about the energy costs. They're worried about the water consumption.
30:23
Ejaaz: Now, the issue with this is, number one, AI data centers take up less water
30:28
Ejaaz: than your average golf course.
30:30
Ejaaz: That's like in your neighborhood itself. We did a whole episode covering this.
30:34
Ejaaz: Number two, for the electricity charges that kind of increase for people's bills,
30:38
Ejaaz: a lot of governments and states have mandated that the AI labs responsible for
30:42
Ejaaz: this pay for that extra surplus so that it doesn't actually hit you.
30:46
Ejaaz: Also, we're working on different ways to deal with this electricity consumption,
30:49
Ejaaz: like launching GPUs into space. So it's just a sad and very concerning story to see.
30:54
Ejaaz: There's a growing contingency of people that are against data centers,
30:57
Ejaaz: and I understand the concerns, but this shouldn't be the way to deal with it.
31:01
Josh: This is dark. You know what's way worse than not having data centers?
31:04
Josh: Trying to getting mythos first and then using it to attack all of our infrastructure
31:08
Josh: and then using it to iterate and build even more powerful models that are even
31:12
Josh: more dangerous, more harmful, and then applying that back to us.
31:15
Josh: And I think the impact of that is far greater than the impact of some patch
31:20
Josh: of grass that is so detached from most towns getting turned into a data center.
31:24
Josh: And I think the moral dilemma here is that people are saying they want one thing
31:30
Josh: and then doing something else.
31:31
Josh: And it's disturbing to see the sheer size of the population that doesn't want
31:39
Josh: to move forward as it relates to AI and progress,
31:43
Josh: totally unaware of the fact that this moves on whether we're a participant or not. And...
31:49
Josh: As these things become more powerful, there's a lot more profound downstream
31:54
Josh: effects of not having these models in our core, having them like having the power on our side.
32:02
Josh: And I hope that this becomes more of a realization for a lot more people.
32:07
Josh: There's this clear divide happening now where it's like people who are using
32:10
Josh: these AI tools to further empower themselves and to do better work in their
32:13
Josh: lives or to handle more things in their personal lives. and then those who don't.
32:16
Josh: And that K-curve that's going to come out of this in the economy,
32:20
Josh: in society, just throughout our general day-to-day lives is going to become pretty wide.
32:26
Josh: And I hope a lot of people really reflect on
32:30
Josh: What it looks like to like actively be on the wrong side, be on the slowdown
32:36
Josh: side of history, be on the, what is it?
32:38
Josh: The deceleration side of history and like what the actual implications of that
32:41
Josh: are as we continue progress forward.
32:43
Josh: I don't know. It makes me sad, but hopefully it's something that will change over time.
32:50
Josh: And perhaps it's just a messaging thing. It's funny. A lot of people that hate
32:53
Josh: SpaceX were very excited about Artemis.
32:55
Josh: So perhaps like the goals are the same. We just need to package them differently.
33:00
Josh: In a way that's more digestible that these mobs can get behind.
33:04
Josh: And I don't know, maybe it's a messaging thing. Maybe it's a moral thing.
33:07
Josh: But that's where we'll leave you at the end of this week on the conclusion of the AI Roundup.
33:12
Josh: That's four episodes in a single week about all of the hottest topics.
33:15
Josh: If you missed anything, you can go back and watch them.
33:17
Josh: I think it's been a big week. I mean, we had a few huge models released.
33:20
Josh: We had Mythos. We had OpenAI versus Anthropic.
33:24
Josh: We had Gemma 4, which was very cool and a very powerful model.
33:28
Josh: But that wraps up everything for this week. Gijos, any final thoughts before
33:31
Josh: we let these lovely listeners go?
33:33
Ejaaz: Nope. Thank you guys for watching and listening. I'm curious if you guys have
33:38
Ejaaz: any thoughts on any of the topics that we've discussed today,
33:41
Ejaaz: or if there are any topics that you think we have missed or that you want to hear more of.
33:45
Ejaaz: We are trying to cover any and every breaking topic, as well as some novel analysis
33:48
Ejaaz: into the actual tools. It's one thing announcing the tools.
33:51
Ejaaz: It's another thing using them. We're going to be doing more demos in the future.
33:54
Ejaaz: But yeah, that's the end of the agenda for this week. And we will see you next
33:58
Ejaaz: week. Thank you guys so much for listening.
34:00
Josh: See you guys in the next one.