Latest / In the Loop AI / THIS WEEK IN AI: Microsoft Build, NVIDIA at Computex, New Glenn Explosion

Transcript

THIS WEEK IN AI: Microsoft Build, NVIDIA at Computex, New Glenn Explosion · In the Loop AI

0:00
Josh: We had a pretty big and exciting week in the world of AI and frontier technology.
0:03
Josh: Lots of things happening. New AI hardware devices unveiled, rocket ships exploding.
0:07
Josh: I want to start with the Microsoft event. Microsoft held their build event.
0:11
Josh: Microsoft notoriously has been in the AI race only via companionship to OpenAI.
0:17
Josh: They are the largest investor, largest owner of OpenAI.
0:19
Josh: And that has been the connection. This week, they severed even more ties with
0:23
Josh: OpenAI as they decided to build their own new hardware.
0:27
Josh: Now, they released seven new AI models. they released two new AI devices and
0:31
Josh: to hear that Microsoft is building an AI hardware device was a little shocking.
0:36
Josh: So I went on this kind of like race to try to figure out more information to
0:40
Josh: understand how it worked and what I found is these two devices that we're seeing
0:42
Josh: on screen. We have a desktop device that looks kind of like an Amazon Echo Home
0:48
Josh: Hub and apparently it's very smart.
0:50
Josh: You can converse with it. It works with their new AI models and the second device
0:53
Josh: is this small looking key card that's something you wear around your neck.
0:56
Josh: It's very smart. It has a display built into it. And it's used for a lot of
1:00
Josh: kind of practical in the job type use cases.
1:04
Josh: What I found out that disappointed me is that all of this was a concept.
1:09
Josh: They marketed this as this amazing product that everyone was going to be able
1:13
Josh: to use showcasing all of the use cases.
1:15
Josh: But the reality is, is that they're never actually going to ship these.
1:18
Josh: In fact, this is just a concept for other companies to go off and get inspired
1:23
Josh: to then build their own versions of.
1:25
Josh: So here I was thinking, oh my god wait microsoft is making a comeback they have
1:28
Josh: seven new models they're releasing new hardware and i think the reality is is
1:32
Josh: that microsoft still kind of feels like they're in the same place they were
1:35
Josh: prior to this build event
1:37
Ejaaz: Yeah i've watched this video twice now and
1:39
Ejaaz: i still don't know what these devices do um i feel like i've been transported
1:44
Ejaaz: back a decade ago and amazon's just announced the echo again also why are we
1:47
Ejaaz: wearing lanyards as like futuristic ai tech now i got hopeful when i saw this
1:53
Ejaaz: announcement josh but i almost immediately killed that hope for one specific reason.
1:58
Ejaaz: I don't know anyone that is conceivably using Microsoft's AI products to a larger extent, right?
2:05
Ejaaz: So Copilot is like their flagship AI assistant, and they've admitted that their
2:09
Ejaaz: Copilot product is way behind the likes of ChatGPT and Claude itself.
2:14
Ejaaz: So if I already know that people aren't really using the Microsoft AI ecosystem
2:19
Ejaaz: at the software level, why on earth would they use it at the hardware level?
2:23
Ejaaz: Now, I dug in to a little bit about what this thing is supposedly meant to do.
2:27
Ejaaz: And it's supposedly meant to be like an AI agent kind of like physical surface area.
2:33
Ejaaz: So what I mean by that is it's meant to listen to the things that you do.
2:36
Ejaaz: It's meant to see the things that you write or speak about and then it's meant
2:39
Ejaaz: to kind of like ingest this data and improve your own personal agent,
2:43
Ejaaz: your own personal model. There's one issue though.
2:46
Ejaaz: Microsoft doesn't even have a functioning V1 of an AI agent that can do all of this for you.
2:50
Ejaaz: They kind of brand it as Copilot but it's really ChatGPT and it's an older model
2:54
Ejaaz: of ChatGPT and it's less effective than ChatGPT. So why wouldn't I just use ChatGPT?
2:59
Ejaaz: Now, understandably, Microsoft recently has been trying to move away from their
3:03
Ejaaz: partnership with OpenAI.
3:04
Ejaaz: There was a bit of tension between Sam Altman and Satni and Odella.
3:08
Ejaaz: And they're trying to kind of like sever that partnership and start becoming
3:10
Ejaaz: a more multimodal operation. They've started using Claude for a lot of their
3:14
Ejaaz: products and other types of models.
3:16
Ejaaz: This is a step into the device side of things.
3:18
Ejaaz: This wasn't exciting to me. And it maddens me so much that this is just a concept
3:22
Ejaaz: that apparently they don't plan to build at all.
3:25
Ejaaz: What did excite me was something that we covered earlier on in an earlier episode this week,
3:31
Ejaaz: which was their new Microsoft Surface AI engineered laptop, which is meant to
3:36
Ejaaz: run like one of Nvidia's latest chips, which allows you to kind of like run models locally.
3:40
Ejaaz: But this doesn't really inspire me.
3:43
Josh: Well, something that might inspire some people if the hardware isn't doing it
3:47
Josh: is the software because Microsoft isn't just sitting there being dormant.
3:50
Josh: They are working on their own AI models. They're working on large language models
3:53
Josh: that I have to assume are somewhat derived from ChatGPT models because they
3:58
Josh: have that proprietary license. They must be leaning on it to some extent.
4:01
Josh: And what we have here is seven new models that they're working on in parallel.
4:05
Josh: And initially I was a little frustrated by this too. I'm like,
4:08
Josh: why the hell are they working on seven models at once? But in the context of
4:12
Josh: this image that we're seeing on screen, it makes a lot of sense.
4:14
Josh: The models are supposed to be working on specific modalities all together.
4:19
Josh: So we have an image model. We have a large one and a flash one for smaller.
4:24
Josh: We have a transcription model. We have a reasoning model. We have two voice
4:27
Josh: models and a coding model. So they're building all of these in parallel.
4:30
Josh: I have to imagine they're going to be building the orchestrator or at least
4:34
Josh: using Copilot as the kind of orchestration harness engine for these models,
4:38
Josh: knowing what to serve them.
4:39
Josh: And I think it's at least a start.
4:42
Josh: Microsoft has done pretty much nothing in the world of AI.
4:45
Josh: I'm not sure anybody uses Copilot who isn't forced to buy their business.
4:49
Josh: And I'm hopeful that these modalities, these models that they're working on
4:53
Josh: will actually yield some sort of impressive results from Microsoft.
4:56
Josh: Otherwise, I'm getting close to putting Microsoft in the Apple category of just like did not finish.
5:02
Josh: Like they just like we're not able to actually execute able to deliver anything
5:05
Josh: valuable in this ai model race maybe this changes things who knows but
5:11
Josh: After watching this, I kind of didn't feel that way.
5:14
Josh: It was like, okay, the hardware that you designed that was really cool looking,
5:17
Josh: it's just a concept. The software, well, it's great on paper.
5:21
Josh: These haven't been built. These haven't been deployed. It would have been awesome
5:23
Josh: if we could actually play with them today.
5:25
Josh: That is not the case. So there's still a lot of work that needs to happen for
5:28
Josh: Microsoft, but that was the build event.
5:30
Ejaaz: Yeah, I mean, the best thing that Microsoft has done in the entire AI race was
5:34
Ejaaz: invest in their biggest competitor, OpenAI. When it comes to actually building
5:39
Ejaaz: models, it's not exactly their specialty, but I applaud them for giving it a go.
5:42
Ejaaz: Now, one might think that they're using OpenAI's IP. They have OpenAI's IP for
5:48
Ejaaz: ChatGPT because they're one of the largest investors in that until I think it's
5:51
Ejaaz: 2030 or 2032. So they can technically use this IP to create their own models.
5:56
Ejaaz: But for one of these models specifically, which is the one I'm circling right
6:00
Ejaaz: now, MAI Thinking 1, that is their main reasoning model.
6:05
Ejaaz: It's their main orchestrator model that orchestrates all these other transcription,
6:09
Ejaaz: image, video, and audio models.
6:12
Ejaaz: And it hasn't been distilled or using OpenAI IP at all.
6:17
Ejaaz: It's been built from the ground up, which is pretty ambitious for Microsoft.
6:22
Ejaaz: Now, you might be wondering, okay, what are some of the advantages that it's unlocked?
6:25
Ejaaz: Well, collectively, across all of these different models, it uses 60% fewer
6:30
Ejaaz: tokens than your Frontier models.
6:32
Ejaaz: But the models that it can kind of compare with and that it's on par with are
6:36
Ejaaz: like three generations behind the frontier models that we're seeing from Anthropic and OpenAI right now.
6:41
Ejaaz: So it's a good attempt, but it's got me thinking like,
6:44
Ejaaz: I don't know why Microsoft is focused on this specifically when they should
6:48
Ejaaz: be kind of like doubling down on like the AI agent harness and trying to build
6:53
Ejaaz: a better AI experience across the humongous distribution that they have.
6:57
Ejaaz: They have like hundreds of millions of co-pilot users across the world.
7:01
Ejaaz: Like they should be able to kind of like monetize that somehow.
7:03
Ejaaz: And I think they're kind of wasting that time here.
7:07
Ejaaz: There was a bunch of Microsoft announcements, and one of them was announced
7:10
Ejaaz: in partnership with NVIDIA, and it was with their Surface laptop.
7:16
Ejaaz: And NVIDIA has been kind of like on a streak this week. So there's this conference
7:19
Ejaaz: that's taking place in Taiwan right now called Computex.
7:24
Ejaaz: And I'm kind of calling it NVIDIA GTC, which is their main conference 2.0.
7:30
Ejaaz: It's basically Jensen's show, and he's taken the stage. Everyone's kind of like
7:34
Ejaaz: crowded out the entire rooms and he's made a bunch of big announcements for NVIDIA.
7:38
Ejaaz: One of the main ones is Vera Rubin NVL72, which is the kind of like it's the
7:44
Ejaaz: next hottest generation that has actually been built for GPUs.
7:49
Ejaaz: Is finally coming online. And there's a few partners that are bringing it online.
7:53
Ejaaz: You've got Dell that we mentioned earlier this week, we put out an episode,
7:56
Ejaaz: definitely go check that out, because Dell's actually a sleeper hit for AI investing.
8:00
Ejaaz: CoreWeave and a number of other providers have helped stand up the very first Vero Rubin stack.
8:07
Ejaaz: It looks incredibly complex and quite impressive from an engineering feat.
8:12
Ejaaz: I don't know if you've seen a picture of this, Josh, but like there are pipes
8:15
Ejaaz: and wires teeming out of this thing.
8:17
Ejaaz: And it's actually bigger than me, if I sit on, I'm around like 6.3.
8:21
Ejaaz: It's like a humongous type thing, incredibly impressive feat of networking architecture.
8:25
Ejaaz: That is one of like the major announcements, like Jensen's kind of like leaning
8:29
Ejaaz: into a lot of new hardware stacks.
8:31
Ejaaz: Now, if you've watched an episode that we put out about a month ago,
8:34
Ejaaz: we described why Jensen's kind of going all in on chips, if you didn't think
8:38
Ejaaz: he already was doing this, by recreating
8:41
Ejaaz: 15 of the core components that is required to build the best GPU.
8:46
Ejaaz: Now, typically, GPU providers aren't recommended to do that because if something
8:49
Ejaaz: goes wrong, you end up needing to delay the launch of your next model.
8:52
Ejaaz: Not only did Jensen go for it, but he pulled it off. And so now we see these new GPUs come to life.
8:58
Ejaaz: He's also entering a new market, which is the CPU market.
9:03
Ejaaz: Typically, Jensen likes to play in his lane. He kind of like sits at his own level of the stack.
9:08
Ejaaz: He owns the picks and shovels and he doesn't want to get involved in anything else.
9:11
Ejaaz: But he's peeked over at Intel and AMD and he's like, damn, these guys are making
9:15
Ejaaz: a bunch of money from CPUs.
9:17
Ejaaz: And in a future world where AI agents are everywhere, we're going to need CPUs
9:21
Ejaaz: to orchestrate these guys.
9:22
Ejaaz: And so he announced the launch of NVIDIA CPUs, which currently is on track for
9:26
Ejaaz: a $20 billion revenue run rate this year. It's pretty insane.
9:30
Josh: This is the first time I've ever heard of Computex as an event.
9:33
Josh: I didn't really even know that this existed, but it's basically the CES version in Taiwan.
9:38
Josh: Like the consumer electronic show it happens every january in the united states
9:42
Josh: it's where all of the cool new technology and hardware gets
9:46
Josh: Announced with computex it felt like
9:49
Josh: the gaming version almost of this and it's weird to
9:52
Josh: say that because sure a lot of this is ai focused but a
9:55
Josh: lot of the offerings a lot of the hardware was based on it's
9:58
Josh: funny there's there's a lot of crossover it turns out between ai local
10:02
Josh: inference running and gaming and that crossover is the
10:04
Josh: gpu and famously nvidia stopped producing the
10:07
Josh: higher end gpus that everyone was using in their computers to
10:10
Josh: run local inference but also to run video games
10:13
Josh: at very high fidelity what's happening here is through
10:16
Josh: the release of the rtx spark which is pretty much the flagship announcement
10:20
Josh: from this event that's the new super computer laptop macbook competitor it's
10:26
Josh: now bringing the power back to the consumer not only for gaming but also for
10:30
Josh: that edge ai compute and we have pricing for it which i thought was pretty interesting
10:34
Josh: and competitive the n1 systems featuring this rtx spark they are going to cost $1,800.
10:40
Josh: And that is for a very high end GPU that you can run locally on your machine
10:44
Josh: to run edge inference if you want to try small models or to actually run very
10:49
Josh: high fidelity video games.
10:50
Josh: The higher end of the system starts at $2,900. That includes the N1X system.
10:55
Josh: And that can run like pretty beefy language models locally on your machine.
11:00
Josh: It has 128 gigabytes of memory. It has, I think, 6,144 CUDA cores,
11:06
Josh: which for those who are unfamiliar with CUDA cores, it's a lot.
11:08
Josh: You can just do a lot of AI stuff on them.
11:10
Josh: And it's really impressive. The other announcement that I really loved is this
11:13
Josh: thing called a NUC, an N-U-K.
11:15
Josh: It's like they create these small computers. If you've ever ran an Ethereum
11:18
Josh: validator or a crypto validator, odds are it was run locally on a little device
11:22
Josh: called a NUC. They've built these...
11:24
Josh: Single purpose machines now with these new graphical
11:27
Josh: cards built into it so now you could have a desktop machine that isn't
11:31
Josh: really a laptop per se but it's meant specifically for
11:33
Josh: one purpose and that purpose is running local edge inference on your device
11:37
Josh: running video games the highest fidelity these are really powerful machines
11:41
Josh: and i have to imagine this is beginning to or i guess this is a continuation
11:45
Josh: of the trend that we've been seeing of companies pushing these consumers There's
11:49
Josh: two edge devices for this type of compute. It was very cool to see.
11:53
Ejaaz: Yeah, and it wouldn't be a Jensen Huang conference if he didn't bless one of
11:58
Ejaaz: the next companies that are going to be potentially a trillion dollar in valuation.
12:02
Ejaaz: And he said it very explicitly on stage.
12:04
Ejaaz: The next trillion dollar company is Marvel.
12:07
Ejaaz: Now, I got to pull up the ticker. Yeah, yeah, please.
12:10
Ejaaz: MRVL and please limit it to the last five days of price action because after
12:15
Ejaaz: he said those words, the next day the stock boomed up. I think it was like 100
12:19
Ejaaz: to 150 percent, something obscene like that.
12:22
Ejaaz: And if you're wondering why Jensen brought the CEO and founder of Marvel,
12:26
Ejaaz: if you're wondering what the hell is Marvel doing, I must say it has got nothing
12:29
Ejaaz: to do with Iron Man or any of the cinematography around that.
12:33
Ejaaz: It is a much more boring kind of company, which is focused on networking architecture
12:38
Ejaaz: for all these GPUs and data centers.
12:41
Ejaaz: So a problem that a lot of AI labs have is when they're scaling their data centers.
12:46
Ejaaz: They get all these brand new expensive GPUs, but you need to be able to connect
12:50
Ejaaz: these GPUs together and get them to communicate with them at lightning speed as quick as they can.
12:56
Ejaaz: Sometimes these GPUs overheat, so you need to make sure you manage the electricity
13:00
Ejaaz: and make sure that the data is transferred at the right amounts of times and
13:03
Ejaaz: the right quantities overnight versus day.
13:06
Ejaaz: There's a ton of complex networking architecture that is required.
13:09
Ejaaz: There's wiring, pipelines, all this kind of stuff, Marvel as a company,
13:14
Ejaaz: as one of the leaders that specialize in this.
13:16
Ejaaz: And Jensen Huang placed a $2 billion check via NVIDIA into them. I think it was about,
13:22
Ejaaz: in March, potentially. And then he made this announcement basically saying like,
13:27
Ejaaz: you know, Marvel is the ones that are going to basically allow NVIDIA to kind
13:29
Ejaaz: of like build out its moat as a data center kind of like expert.
13:35
Ejaaz: And their stock went up a massive amount of time. And like he's pointing at
13:39
Ejaaz: one very important thing, which I want to mention.
13:42
Ejaaz: The next bottleneck constraint. So we've spoken about GPUs, compute.
13:46
Ejaaz: We've spoken about memory and architecture for memory a lot here.
13:49
Ejaaz: And those stocks have gone up. he's saying the next bottleneck is going to be
13:53
Ejaaz: on power networking specifically and that's what Marvel specializes in they
13:57
Ejaaz: specialize in like optical fiber networks and a ton of other stuff there so
14:00
Ejaaz: that's going to be an interesting one to view but moving on,
14:05
Ejaaz: there's been a massive explosion. I saw this all over my feed, Josh, over the weekend.
14:12
Ejaaz: And it's actually pretty intimidating to an extent, but this is Blue Origin's
14:17
Ejaaz: New Glenn rocket, which was being tested at Cape Canaveral and unfortunately exploded massively.
14:24
Ejaaz: What is going on here? Why did this happen?
14:27
Josh: Yeah, it's pretty intense. I actually had a post about this.
14:30
Josh: This is one of my largest ever 5 million people saw.
14:32
Josh: And it was because the comparison to this explosion is that it is roughly 20%
14:37
Josh: of the energy of the Hiroshima atomic bomb that was released on the launch pad of Blue Origin.
14:43
Josh: It is a horribly tragic happening for the space program, for the space program
14:48
Josh: as a whole, but also Blue Origin as a company.
14:50
Josh: The pad, unfortunately, is the only pad that is capable of launching the New
14:55
Josh: Glenn rockets to Earth, and now it's gone.
14:57
Josh: It cost them over a billion dollars to make, and SpaceX needed well over seven
15:02
Josh: months to rebuild theirs after they had a similar issue.
15:05
Josh: So this is a pretty large setback and it comes back against the problem that
15:10
Josh: Amazon needs, I think, the number of 1,618 satellites up in orbit by July 30th
15:16
Josh: to maintain its license with the FCC.
15:18
Josh: It currently has about 300. So the rocket that was supposed to really help with
15:23
Josh: this contract is now gone and the pad that was supposed to be launching it from is gone with it.
15:29
Josh: Now, they did have a response that I saw that seemed somewhat reassuring.
15:33
Josh: It was from someone within the company. I think he's the, oh,
15:37
Josh: actually the CEO of Blue Origin. It's funny.
15:38
Ejaaz: I think Jeff Bezos is the CEO.
15:40
Josh: Turns out it's this other guy named Dave Limp. And Dave Limp said that now that
15:44
Josh: we've had access to the pad and integration facility, we could share a good bit of news.
15:47
Josh: The propellant farm, oxygen, liquid hydrogen, and LNG tanks,
15:52
Josh: liquid natural gas, are all in good shape.
15:54
Josh: This is good because those are very long lead times. So it seems like a lot
15:58
Josh: of the core components that were going to be causing a longer delay
16:03
Josh: aren't actually affected too badly there was a
16:06
Josh: support tower that if you watch the video that thing got destroyed
16:09
Josh: that's the strong arm that holds the rocket up totally got wiped
16:12
Josh: out he said well good news we were planning to get rid of that anyway so they're
16:16
Josh: very much making light of the situation they're trying to work their hardest
16:19
Josh: i just wanted to share this because it was one unbelievable explosion great
16:21
Josh: content to watch if you are seeing the video version of this episode but also
16:25
Josh: that it may not be the worst setback in the world and in terms of getting these
16:30
Josh: Blue Origin satellites into space.
16:31
Josh: Thankfully, they still do have SpaceX with the Falcon 9, Falcon Heavy,
16:34
Josh: and soon-to-be Starship. So a bit of a setback in the space race.
16:39
Josh: It's unfortunate. SpaceX really could benefit from some competition.
16:42
Josh: We don't want them going into their IPO as a complete and total monopoly.
16:45
Josh: But here we are. They're the only ones that can figure it out.
16:48
Josh: And it's funny, we took space for granted because SpaceX has just been landing
16:52
Josh: boosters back on barges in the middle of the ocean for like five plus years now.
16:56
Josh: And it turns Turns out space and rocket science really are like pretty difficult categories.
17:01
Josh: So wishing the team all the best. A bummer, but.
17:04
Josh: Alas, we will carry on.
17:06
Ejaaz: It's probably like the best timing for Elon Musk. He probably saw that and he
17:10
Ejaaz: was like, this is upsetting for the industry as a whole, but perfect for my
17:13
Ejaaz: IPO that's happening in probably next week.
17:16
Ejaaz: Moving on, there was this interesting stat that we pulled from Cloudflare,
17:21
Ejaaz: which basically said that agentic traffic is growing so fast that it now has
17:26
Ejaaz: surpassed human traffic.
17:28
Ejaaz: It is over 50% of internet traffic online.
17:31
Ejaaz: Now, I can't say that this surprises me, as much.
17:35
Ejaaz: I think the adoption and growth of people using AI LLMs and setting up their
17:39
Ejaaz: own agents, vibe coding, different kinds of things that kind of like run autonomously
17:43
Ejaaz: 24-7 has increased exponentially at this point.
17:47
Ejaaz: And I think this is only going to get more aggressive.
17:50
Ejaaz: We were having a discussion prior to starting this recording,
17:53
Ejaaz: and I was like pondering about a world where like, you know,
17:57
Ejaaz: future kids or generations of kids are largely going to
18:00
Ejaaz: interact with entities or beings online
18:04
Ejaaz: that don't have a physical body or if they do it
18:06
Ejaaz: might be robotic and we got into like a debate
18:09
Ejaaz: around like what that might potentially mean like from an educational stance
18:13
Ejaaz: right if you had a tutor that can kind of like personally tailor its lessons
18:16
Ejaaz: to you for this particular child that might be end that might end up being useful
18:21
Ejaaz: but for someone who is like engaging in entertainment or leisure with a type
18:26
Ejaaz: of ai personality it gets a little blurry as to whether this is good or not.
18:31
Ejaaz: And I guess some versions of this already do exist.
18:34
Ejaaz: Like I watch a lot of YouTube, I kind of follow my favorite creators online.
18:38
Ejaaz: I don't really know them. I
18:39
Ejaaz: just know them through the screen and through the content that I consume.
18:42
Ejaaz: So what if it means that it's AI generated completely?
18:45
Ejaaz: Anyway, I think we're going to see a lot of this going forwards.
18:48
Ejaaz: I think we're going to interact with a lot of businesses that are completely
18:51
Ejaaz: autonomously run by agents, but we would never know because we interact with
18:54
Ejaaz: them via text or they generate their own physical, sorry,
18:57
Ejaaz: digital AI avatars, but just an interesting stat and one that has gone under the radar, I think.
19:04
Josh: Yeah, well, there's been this concept of the dead internet theory for a very
19:07
Josh: long time, which basically says that a lot of the internet has ceased to be
19:11
Josh: human-centric and is now mostly composed of bot activity. And we've seen this before.
19:16
Josh: We've seen this through mostly an election cycle.
19:19
Josh: So you remember back in 2016, a lot of the propaganda that we were seeing on
19:23
Josh: social media aggregator sites that were being published to the people,
19:26
Josh: a lot of it was bot activity.
19:27
Josh: So we've had at least the appearance of bot activity, particularly on curated
19:32
Josh: algorithmic feeds for a very long time.
19:34
Josh: Well, Cloudflare sharing is basically the idea that not only is it greater than
19:39
Josh: 50% of the content that you see, but genuinely now 57% as of recording this
19:44
Josh: based on their charts is bot activity, provably bot activity.
19:48
Josh: So we've officially crossed the chasm. It's no longer a dead internet theory.
19:52
Josh: It is a dead internet and we are living in a bot driven world. the
19:55
Josh: downstream implications of this are going to be interesting
19:58
Josh: to think about to discuss i'm sure this is a very
20:01
Josh: long conversation we're going to be having for many many years
20:04
Josh: going forward but it's official the flipping has happened as
20:07
Josh: confirmed by cloudflare one of the largest internet providers i
20:10
Josh: think the largest internet provider in the world so it's
20:14
Josh: funny too they were guiding they say in the post for the end of 2027 and then
20:18
Josh: they bumped their timeline up to early 2027 and now we're in mid-2026 and that
20:22
Josh: flip has happened so things are clearly accelerating faster than i think anyone
20:25
Josh: expected and what that means for the internet as a whole for the people the
20:30
Josh: human users of it is to be determined
20:34
Ejaaz: Now, one thing that we've spoken a lot about on this show is the idea of scaling laws.
20:40
Ejaaz: So that follows the rule where if you want to create a better AI model,
20:43
Ejaaz: if you want to train a better AI model, if you want to inference a better AI model,
20:46
Ejaaz: you need a lot more compute that comes in the form of GPUs, it comes in the
20:50
Ejaaz: form of energy or electricity to power these GPUs and to train these different models.
20:56
Ejaaz: Sarah Fryer, the CFO of OpenAI, is one of the few people probably in the industry
21:02
Ejaaz: that knows to the bones how true this statement is or how true this scaling law actually is upheld.
21:09
Ejaaz: And she had a very clear statement on her appearance on the All In conference
21:12
Ejaaz: where she said, compute through 2027 is completely sold out.
21:16
Ejaaz: And that is the clearest statement ever that the scaling laws are still kind of holding.
21:21
Ejaaz: We're still in the early parts of 2026. We've heard similar trends for AI memory
21:27
Ejaaz: being sold out through the end of 2027, but not really on compute.
21:30
Ejaaz: And we also know that compute data centers specifically is the current focus
21:35
Ejaaz: of every major AI CEO in terms of scaling. They're trying to like...
21:39
Ejaaz: Boots on the ground and build these things, hire all the technical employees
21:43
Ejaaz: that are required to set these things up.
21:44
Ejaaz: It is the most aggressive expenditure. Google just raised $85 billion.
21:48
Ejaaz: All these IPOs that we're talking about, Anthropic, OpenAI, and SpaceX is all
21:52
Ejaaz: to get more money to invest in AI CapEx. It's the number one focus.
21:55
Ejaaz: It's Krishna Rao, CFO of Anthropic, number one focus.
22:00
Ejaaz: So for Sarah Fry to also confirm this and say that she feels like,
22:04
Ejaaz: compute is going to be short for a long period of time through 2030 and 2032,
22:09
Ejaaz: is the clearest statement ever that OpenAI, Anthropic,
22:13
Ejaaz: and the likes of other top major AI labs aren't going to go away when it comes to AI capex.
22:19
Ejaaz: Now, one thing that this forces me to think is, or like challenge myself is,
22:23
Ejaaz: it probably means then that we are not limited by capital.
22:27
Ejaaz: We're not limited by kind of this AI bubble that is being inflated.
22:32
Ejaaz: We're being limited by physical atoms itself.
22:35
Ejaaz: Like we don't have enough, no matter how much money you have,
22:38
Ejaaz: we don't have enough human capital to actually build these things.
22:41
Ejaaz: And so we're constrained by that.
22:43
Ejaaz: The regulations determining what land to build on it takes a lot of time.
22:47
Ejaaz: And that backlog, in my opinion, is probably going to go on for a lot more.
22:50
Ejaaz: So we're probably going to see a lot more investments of this trend.
22:53
Josh: Sarah Fryer is incredible. I would encourage everyone to go and watch this. There's
22:56
Josh: a strong case to me that she should be running the company because man she
22:59
Josh: is incredibly well spoken and thoughtful and revealed a lot
23:02
Josh: of things that weren't obviously apparent to me one of them talking about the
23:06
Josh: time frames in which she deals on like i'm very much under the impression that
23:09
Josh: these ai labs are thinking maybe 12 to 24 months out because further than that
23:13
Josh: the supply chains are nearly impossible predict she was sharing that they think
23:17
Josh: like up to six years out and she's placing guidance for 2032,
23:22
Josh: which to me feels like this impossibly complicated task. So maybe she's crazy.
23:27
Josh: Maybe there's some kind of merit to these ideas, but it was a really fascinating
23:31
Josh: episode from the all-in guys just to understand
23:34
Josh: Kind of where OPAIs see themselves and how they plan to navigate through this
23:38
Josh: next five to six year period in the sense that, like you're saying,
23:41
Josh: EJS, there's just not enough compute.
23:43
Josh: So therefore, they're not worried about any slowing demand because the demand
23:47
Josh: is going to continue to outpace that compute over the next six years.
23:50
Josh: And so long as that demand outpaces the expense, these companies are going to become profitable.
23:54
Josh: They're going to continue to go up. And I mean, obviously, she's guiding for
23:57
Josh: this positive future, but it seems like they have the numbers to back it up.
24:00
Josh: A second thing she mentioned that I found very, very exciting was she talked
24:03
Josh: about my favorite, our favorite hardware device, the Johnny Ive design device.
24:07
Josh: She actually referenced it by name.
24:10
Josh: Johnny Ive was mentioned and I think it was earbud that was mentioned or earpiece.
24:15
Josh: She said Johnny would kill her if she called it an earpiece,
24:18
Josh: kind of acknowledging the fact that this is actually something that goes in your ears.
24:22
Josh: The things that we have been seeing are somewhat real.
24:25
Josh: Jason, one of the people who hosts the show, he referenced the puck,
24:29
Josh: the earpiece. She did not
24:32
Josh: Rebuke she kind of was just like yep it's it's happening
24:34
Josh: it's amazing she's used it yes like
24:37
Josh: people internally have been using it it seems like this incredible device
24:40
Josh: and she confirmed that it will be announced before the
24:43
Josh: end of this year so we have confirmation at least on timelines happening
24:47
Josh: sometime i would assume around q4 where we'll get more information and
24:50
Josh: that is the holy grail of hardware i cannot wait to get those to try it out
24:54
Josh: just to see what it looks like so that was really exciting uh open ai also they
24:59
Josh: launched sites which is a cool new feature back to like practical ai use cases
25:04
Josh: that you can actually use um tell us about sites ejaz i know you've been looking
25:07
Josh: through how these kind of work so
25:09
Ejaaz: I'm a very visual learner and one thing that i interact a lot with when i'm
25:14
Ejaaz: creating kind of like research and agendas with claude is like can you create
25:17
Ejaaz: me like a visual artifact um this is basically open ai's attempt to kind of
25:22
Ejaaz: like not only compete, but excel for that product.
25:24
Ejaaz: So in their own words, with sites, Codex can turn your work,
25:28
Ejaaz: ideas and plans into an interactive website or app your team can explore, use and share with a URL.
25:34
Ejaaz: Now, why I enjoy this the most is typically when you create a visual artifact,
25:39
Ejaaz: it's static, it's like an image, it kind of you can kind of tell that it's created
25:43
Ejaaz: by an AI, because it uses the same font, it's structured in a very particular way.
25:46
Ejaaz: And what I typically do with that is I take that I take it to code code and
25:50
Ejaaz: I say, Hey, can you create an HTML version of this and enhance it in a B and C types of ways?
25:56
Ejaaz: This is a one button, one plugin, one feature fix that OpenAI is doing via codex.
26:02
Ejaaz: So it's, you know, they're creating this like super app and this is kind of
26:05
Ejaaz: like a step towards that direction.
26:06
Ejaaz: Well, when you ask for a visual artifact, it will now do this.
26:09
Ejaaz: And what I like about this is it's easily shareable.
26:12
Ejaaz: So like you can create your own webpage, you can share it with someone and it
26:15
Ejaaz: introduces just a new medium and way to kind of read and learn about something.
26:18
Ejaaz: Like I would love to scroll a custom website with graphics, infographics that
26:23
Ejaaz: are kind of like responsive to the way that I move my cursor,
26:25
Ejaaz: that kind of explains concepts very clearly to me, I would learn much faster
26:29
Ejaaz: that way versus just reading a block of text or looking at a static image.
26:33
Ejaaz: And I just think it's a really neat feature and I want to see more companies
26:36
Ejaaz: rolling out features like this.
26:38
Josh: Yeah, it seems like it's going out to business and enterprise price plans currently.
26:41
Josh: So if you have one of those, congratulations, you can go and try it out.
26:44
Josh: I assume all the other plans will be following shortly after. I'm excited for this.
26:48
Josh: So many people want to have their own websites. They've dreamed of making websites.
26:52
Josh: They've paid tens to hundreds of thousands of dollars to have other people build websites.
26:57
Josh: And now it's a matter of clicking a button and getting something that looks
27:00
Josh: really lovely and that is functional and that accomplishes all of your goals.
27:03
Josh: So I think in terms of people, just the average person, this is such a powerful
27:06
Josh: tool and use case to empower them to build these websites. So that's great.
27:11
Josh: We have two more topics quickly that we could probably breeze through.
27:14
Josh: The first is Grok launching with Composer 2.5.
27:18
Josh: I'm just reading the headline here, making it competitive product to Cloud Code.
27:23
Josh: When I think of Cloud Code, I think of, oh my gosh, this model is unbelievable.
27:26
Josh: People are actually doing real work.
27:28
Josh: Awesome. When I think of Grok, I think of, oh yeah, that's right.
27:32
Josh: That's the thing I used like six months ago and haven't touched since.
27:34
Josh: So what's changed over those six months?
27:36
Ejaaz: Yeah. So it's this one layer that sits on top of the model called the agent harness.
27:42
Ejaaz: Now, a few weeks ago, we covered XAI or SpaceX AI forming a large partnership
27:47
Ejaaz: with this company called Cursor.
27:49
Ejaaz: Cursor doesn't actually create their own AI model, but they have specialized
27:52
Ejaaz: remote known as the agent harness, which basically sits on top of any frontier
27:56
Ejaaz: AI model and tells it, you know, kind of how to work, where to direct prompts.
28:01
Ejaaz: It kind of like fine tunes it in a very specific way that actually includes
28:05
Ejaaz: a lot of information that is very valuable if you're able to own it as a company.
28:11
Ejaaz: Cursor has been able to do that.
28:12
Ejaaz: Now, in this partnership, Elon has basically been able to acquire the right
28:16
Ejaaz: to buy Cursor within 30 days of their IPO.
28:20
Ejaaz: And as part of this agreement, that means he can take Cursor's very valuable
28:24
Ejaaz: moat, their harness, slap it on Grok 4.3, which is the latest model that he's
28:29
Ejaaz: training right now that is, I believe, publicly available now as of this week.
28:33
Ejaaz: And hey, presto, you have a frontier coding model.
28:36
Ejaaz: That's how powerful the harness layer is over the model. It's not just how good
28:41
Ejaaz: your model is, it's how good your harness is as well.
28:43
Ejaaz: Like Dario and Sam have both confirmed this.
28:47
Ejaaz: And actually, if you slap Curse's harness right now on top of Claude Opus 4.8
28:51
Ejaaz: or GPT 5.5 high, you end up having a better coding model than those models stand on their own.
28:56
Ejaaz: So Elon has been able to buy his way to the frontier top.
29:00
Ejaaz: Composer 2.5 on its own, which is Cursor's own model that they wrapped a harness
29:06
Ejaaz: over, is comparable and on par with 4.8 and 5.5.
29:11
Ejaaz: So if you wrap it around Grok 4.3, I can imagine the same is the case.
29:14
Josh: That's pretty amazing. They have made some unbelievable progress.
29:17
Josh: And I think, again, it's a testament to never count Elon out.
29:20
Josh: They are making incredible progress.
29:22
Josh: And as we know, they have the training models. They have the hardware.
29:26
Josh: They've currently given it out to Anthropic. Elon restated publicly this week that
29:30
Josh: It is more of a shorter term deal is kind of the idea that they both have the
29:34
Josh: option to opt out at an inhuman moment.
29:36
Josh: And I think the terms were 90 days that he said or something like that.
29:39
Josh: But it is cool to see another competitor in the race. It's great to see another
29:43
Josh: harness that people are excited about. I mentioned there was two topics.
29:45
Josh: The last one is actually Google's $80 billion race that we covered yesterday in the episode.
29:50
Josh: So unless there's net new information, I think I'm going to tease it a little bit.
29:54
Josh: And then if you want to see the full thing, we could go back to our episode from yesterday. Yeah.
29:59
Josh: Yeah.
29:59
Ejaaz: And the only quick.
30:01
Josh: Oh, yeah, what do you have? Do you have any updates?
30:03
Ejaaz: Well, yeah, the only quick update is they doubled down and raised it to $85 billion.
30:08
Ejaaz: So there's an extra $5 billion that they raised at the last minute.
30:11
Ejaaz: And I just wanted to spend like 10 seconds saying a lot of people's takes on
30:16
Ejaaz: this raise is it's bearish that Google selling stock, they're not confident
30:20
Ejaaz: in the future of CapEx. I actually think it's the opposite.
30:23
Ejaaz: I think Google has $126 billion on their balance sheet, and they can use that
30:28
Ejaaz: to raise higher debt in the future. So they're being strategic.
30:31
Ejaaz: This is financial bookkeeping, creative financial bookkeeping to make sure that
30:35
Ejaaz: they have a healthy balance sheet in the future for when they need to raise even more money.
30:39
Ejaaz: So selling a bit of stocks now won't get punished if their stock value ends
30:43
Ejaaz: up going up more for the fewer shares that they end up having.
30:46
Ejaaz: So I think this is ultimately a good move. And I just want to go on record saying
30:49
Ejaaz: that for when it eventually is proven.
30:51
Josh: And there you have it so if you have made it to this part of
30:54
Josh: the episode congrats you're totally caught up on the week you can go touch
30:57
Josh: grass go enjoy your weekend you have been exposed and up
31:00
Josh: to date on everything as for the next few weeks buckle up
31:03
Josh: it's going to get exciting next week we have wwdc from
31:07
Josh: apple early in the week which is huge this is make or break for them if they
31:11
Josh: cannot deliver again apple we're gonna have to write them off to zero we need
31:14
Josh: the following week is seeming like it's going to be spacex ipo week that's going
31:19
Josh: to be huge biggest ipo in history and i'm sure along the way,
31:22
Josh: there will be many other releases.
31:24
Josh: So buckle up. Very excited. Ejaz, any final parting thoughts before we head into our weekend?
31:28
Ejaaz: No, that's it. I'm curious if any of you were out at any of the major Microsoft
31:32
Ejaaz: Build or Computex events.
31:34
Ejaaz: Let us know if there are any different takes that we missed.
31:37
Ejaaz: But aside from that, wherever you listen to us on YouTube, Spotify,
31:40
Ejaaz: Apple Music, please subscribe if you haven't.
31:42
Ejaaz: Please give us a rating if you haven't. Please leave us a comment if you haven't.
31:44
Ejaaz: All of these are incredibly valuable in helping us grow our channel.
31:47
Ejaaz: And we have grown massively.
31:48
Ejaaz: I think it's like 12,000 people over the last month, which is just absolutely
31:51
Ejaaz: insane. It's lovely to meet you guys.
31:53
Ejaaz: And yeah, we will see you on the next one next week.
31:56
Josh: Thanks so much.
31:56
Ejaaz: See you guys.