Latest / In the Loop AI / THIS WEEK IN AI: Chat GPT-5.5 Beats Claude Mythos, SpaceX Cursor Rumors, Google's New TPU's

Transcript

THIS WEEK IN AI: Chat GPT-5.5 Beats Claude Mythos, SpaceX Cursor Rumors, Google's New TPU's · In the Loop AI

0:00
Ejaaz: The most powerful model in the world is here right now. In fact,
0:03
Ejaaz: it's so good that it beats Claude mythos.
0:06
Ejaaz: OpenAI just released ChatGPT 5.5 and it crushes Claude on every single benchmark.
0:12
Ejaaz: It's the new number one coding model. It can do 20 hour tasks that expert software
0:17
Ejaaz: engineers sometimes can't do.
0:19
Ejaaz: It's already discovered groundbreaking solutions in maths and frontier sciences such as genetics.
0:24
Ejaaz: And it's cheaper than GPT 5.4. This is the result of two years worth of frontier
0:30
Ejaaz: research released in this one single model.
0:32
Ejaaz: In fact, it's so good that an NVIDIA engineer said, and I quote,
0:36
Ejaaz: losing access to GPT 5.5 feels like I've had a limb amputated.
0:40
Josh: I think a lot of people are going to compare this to Opus 4.7,
0:43
Josh: and that's fair, but I really think the true comparison is to Mythos because
0:48
Josh: Sam Elman recently, he just posted something as the model was coming out that
0:51
Josh: felt very much like a jab at Mythos.
0:53
Josh: And we're going to get into the benchmarks comparing them, many of which will
0:56
Josh: actually beat the Claude model. But what I find most interesting about this
0:59
Josh: post is the second paragraph where he says, we believe in democratization.
1:03
Josh: And he mentioned specifically, we have been tracking cybersecurity as a preparedness
1:07
Josh: category for a long time and have built mitigations we believe in that enable
1:12
Josh: us to make capable models broadly available.
1:14
Josh: So this is very much a dig at Mythos, which is, as we all know,
1:17
Josh: privately available, only gated to the companies that are given allowance to it.
1:22
Josh: ChatGPT and OpenAI are like, hey, we're going to give you the powerful cybersecurity.
1:25
Josh: We're just going to bake in the precautions into the model so that everyone could have it.
1:29
Josh: And it ends by saying it's this really sweet thing. It's like we love you and
1:32
Josh: we want you to win. We believe in everyone having access to this intelligence.
1:35
Josh: And I really respect that. And I think it's an awesome way to set the precedence
1:38
Josh: for what the next generation of these models is going to look like.
1:41
Josh: But before we go any further, let's talk about the model itself. It's out right now.
1:44
Josh: If you have a chat GPT membership, you can go and use it, go and play with it.
1:48
Josh: EJS, what's the TLDR? What are the high-level things that everyone should know?
1:51
Josh: What's most new and noteworthy about GPT 5.5?
1:54
Ejaaz: Okay, so inspired by your mythos comparison, the first question that pops into
1:58
Ejaaz: my head is I use Claude Opus 4.7 every single day. So I'm like,
2:02
Ejaaz: is it better than this? Like, should I be switching back to ChatGPT right now?
2:06
Ejaaz: The answer might be yes. So if we look at the benchmark score right here,
2:10
Ejaaz: GPT 5.5 on the left over here absolutely crushes all the standard benchmarks
2:15
Ejaaz: that these frontier models are weighted against.
2:17
Ejaaz: And if you look on the right over here, Claude Opus 4.7, it either doesn't even
2:21
Ejaaz: measure in a particular category, or it's completely beaten by GPT 5.5.
2:26
Ejaaz: In fact, the only stat that GPT 5.5 doesn't beat Opus 4.7 in is something called
2:31
Ejaaz: Software Engineering Benchmark Verified Pro or something like that.
2:35
Ejaaz: It's like the pro software coding situation.
2:39
Ejaaz: But there's a footnote at the bottom of this blog where OpenAI states,
2:42
Ejaaz: Anthropic has publicly said that they might have gamed that particular benchmark
2:46
Ejaaz: and they need to be re-evaluated.
2:47
Ejaaz: So we might have a complete clean sweep for 5.5 as we see today.
2:52
Ejaaz: So it's an incredibly powerful model.
2:54
Ejaaz: But a question that popped to my head is, does it actually beat Mythos?
2:58
Ejaaz: And we have a direct comparison right here.
3:00
Josh: Yeah, so it shows that it does across some benchmarks. Now, again,
3:03
Josh: these benchmarks are pretty fuzzy.
3:04
Josh: We don't know which ones are gamed to do what. But there is a world in which
3:08
Josh: GPT 5.5 will outperform Mythos on some things, which ones we're not entirely sure.
3:13
Josh: I think as we kind of figure out ways to describe GPT 5.5, it seems as if it's
3:18
Josh: their first attempt at making a model built for autonomy instead of answers.
3:23
Josh: I think a lot of the benchmarks that they're working on is in agent decoding,
3:26
Josh: things like it handles tasks that are 20 hours long. We'll get into that.
3:30
Josh: It's doing 85% of OpenAI's internal work already.
3:33
Josh: And it also helped rewrite the infrastructure that
3:36
Josh: built it there was this amazing quote in the blog post it said open ai
3:39
Josh: says 5.5 itself helped optimize the stack
3:42
Josh: that serves it codex analyzed weeks of production traffic and
3:45
Josh: wrote custom heuristics for load balancing that boosted token generation speed
3:49
Josh: by over 20 so they're using the model to actually build the model and make it
3:53
Josh: maximally efficient based on the data that it's collected from users like us
3:57
Josh: who are interacting with the model on a daily basis so it's very smart it's
4:00
Josh: very clever it's not just there to give you answers.
4:02
Josh: It's there to think deeply and actually solve problems for you in a way that
4:06
Josh: I think Mythos and a lot of these other frontier models are kind of pivoting towards now.
4:11
Ejaaz: The great thing about this model release is it reveals a few things that OpenAI
4:15
Ejaaz: has as an advantage against, say, a frontier lab like Anthropic.
4:18
Ejaaz: It's clear looking at these benchmarks compared to Mythos, which,
4:22
Ejaaz: by the way, the entire world is spiraling because of this model,
4:25
Ejaaz: because it's going to have the cybersecurity ability to take over any kind of government system.
4:30
Ejaaz: This model is pretty close, and Sam is going to be releasing this publicly,
4:34
Ejaaz: or OpenAI is going to be releasing it publicly for everyone to use.
4:36
Ejaaz: So a question that pops to my head is, does this mean that it's a matter of
4:41
Ejaaz: compute, and OpenAI just simply has more of them?
4:44
Ejaaz: Certainly, if you compare Sam Altman's ability to acquire compute and spend
4:48
Ejaaz: all these trillions of dollars to acquire it versus Anthropic,
4:51
Ejaaz: Anthropic has been extremely conservative, and now they're struggling.
4:54
Ejaaz: They recently signed a $5 billion deal with Amazon, which we'll get to later
4:58
Ejaaz: on. But the point is, this is a tale of two stories.
5:01
Ejaaz: Either OpenAI has enough compute and they're about to leapfrog Claude because of that.
5:05
Ejaaz: And they're proving that through this model that is a very good answer to Mythos.
5:10
Ejaaz: Or, and this is the alternative side, Anthropic's Mythos model is just plainly better than 5.5.
5:15
Ejaaz: And these benchmarks are actually verified, which is technically kind of true
5:19
Ejaaz: because I don't know how official these things are.
5:21
Ejaaz: These are just through tests that a small set of users have done.
5:24
Ejaaz: So it's a game of both. I'm sure Anthropic is watching this and thinking,
5:27
Ejaaz: hmm, maybe we should roll out Mythos, but they don't have to compute.
5:30
Josh: Yeah, they don't have the inference. In fact, speaking of the inference,
5:33
Josh: Sam actually made a post saying that he's really...
5:36
Josh: Excellent work by the inference team to serve this model so efficiently he
5:39
Josh: wanted to really highlight the fact that to a significant degree they've
5:43
Josh: become an ai inference company now and i think that's a
5:45
Josh: really big difference than what was previously stated like anthropic has really
5:48
Josh: tough time serving compute and we see that and even if they had mythos available
5:51
Josh: in a way that was safe they can't serve it open ai can and we see it reflected
5:55
Josh: in pricing because i mean we have some pricing for this model right and it seems
5:59
Josh: as if it's roughly at par with 4.7 if not slightly better?
6:03
Ejaaz: Slightly. It's slightly more expensive, but not by much.
6:06
Ejaaz: So for every million tokens input, it's both the same for Anthropic Opus 4.7 and GPT 5.5.
6:13
Ejaaz: It's $5 in, but the output is $30 for 5.5 per million tokens and $25 per million tokens for 4.7.
6:22
Ejaaz: So it's a little more expensive, but here's where you actually have more of
6:27
Ejaaz: a bargain using the more expensive model 5.5. It is cheaper than GPT 5.4,
6:33
Ejaaz: and it uses tokens way more efficiently to think.
6:37
Ejaaz: So what does that mean if you are an enterprise that wants to plug in this AI
6:41
Ejaaz: model and not worry about it and just have it power your entire profit engine?
6:45
Ejaaz: Well, you end up using less tokens, so you hit your rate limits in a much slower
6:50
Ejaaz: rate, which means that you end up getting more bang for your buck as long as
6:54
Ejaaz: you use the model like 24-7 or you use it effectively.
6:57
Ejaaz: Way if you are just kind of out there using 5.5 to like ask questions that you
7:01
Ejaaz: should maybe be asking google this is probably not the model for you but otherwise super powerful one
7:07
Josh: Yeah. And if these prices don't mean anything to you, that's fine.
7:09
Josh: As long as you have a $20 a month subscription.
7:11
Josh: In fact, this is going to be available to freezers fairly soon, I believe.
7:15
Josh: But anyone who is a subscriber has access to this. You don't need to use the
7:18
Josh: API. There's nothing fancy.
7:19
Josh: You open up your app on your phone, you go to the web browser,
7:21
Josh: it's there, it's available, ready to go.
7:23
Josh: Now, there's a few interesting things that you can do with this model that haven't
7:26
Josh: previously been possible.
7:28
Josh: And although we don't quite have access to it just yet, we're recording this
7:31
Josh: right as the model got launched.
7:32
Josh: We do have a blog post from OpenAI themselves who are showcasing a few demos
7:37
Josh: so again take these with a grain of salt these are straight from open ai but
7:40
Josh: they are seemingly pretty impressive and pretty noteworthy as to what they're
7:43
Josh: capable of doing starting with this space mission application which is um pretty
7:47
Josh: cool and very reminiscent of the moon mission that we just had yeah.
7:51
Ejaaz: Um so if you guys don't know um josh has a secret he has many secrets on this
7:55
Ejaaz: show one is he's a massive space fan and when he's not hanging out with me he's
8:00
Ejaaz: doing uh space simulations uh on whatever he can do right well okay maybe maybe
8:04
Ejaaz: be part of that is a bit of a lie.
8:06
Ejaaz: But with this new app that we're seeing in front of us right now,
8:09
Ejaaz: this was completely vibe coded using 5.5.
8:12
Ejaaz: And it's used to simulate a specific space mission.
8:15
Ejaaz: Now, if this looks very similar, it's because we just had a space mission for
8:18
Ejaaz: some we visited or went back to the moon in 53 years, pretty big deal.
8:22
Ejaaz: And we can see a pretty accurate simulation going on right here.
8:25
Ejaaz: So as you can see, there's various different toggles, the physics of the entire
8:28
Ejaaz: thing is very important.
8:30
Ejaaz: And that's another point I want to make about this model. it is being
8:33
Ejaaz: used for frontier research not just in ai but in
8:36
Ejaaz: mathematics in genetics like it made frontier progression on both of these fronts
8:40
Ejaaz: and so what we're showing here is this is a model that goes way beyond just
8:45
Ejaaz: text and telling you what could be it actually implements this into a lot of
8:49
Ejaaz: different things and understands the world around it which is extremely powerful
8:52
Ejaaz: but we have another one here we have a we have an earthquake tracker
8:56
Josh: For anyone who wants to make websites, it's so good at making websites.
8:59
Josh: And this appears to be one of the strong suits. In this case,
9:02
Josh: there's a few things to highlight on this Earthquake tracker.
9:04
Josh: One of them being that it's one, just like a pretty elegantly designed website.
9:07
Josh: But two, all of the graphics are interactive. You'll notice that they update
9:11
Josh: dynamically as you hover over them, as you click. It looks very clean.
9:14
Josh: I assume that it is pulling up-to-date information from an API somewhere that it set up.
9:19
Josh: It is just truly competent and capable of doing these kind of longer tail tasks
9:23
Josh: that are a bit more complicated than a static landing page,
9:25
Josh: but have dynamic data have the richness that
9:29
Josh: you would expect from a high-end high-quality polished website
9:32
Josh: except just built with an ai model from someone who doesn't
9:35
Josh: need to know anything about coding at all and then
9:38
Josh: for the gamers also there's another great example of a dungeon game
9:41
Josh: which is they're describing as a playable 3d dungeon arena prototype
9:44
Josh: built with codex and gpt models now
9:47
Josh: i think this is something novel to this setup where codex handles
9:50
Josh: the game architecture the combat systems the enemy encounters and then the character
9:54
Josh: models the character textures and animations those were created with third-party
9:58
Josh: asset generation tools using something like image gen 2.0 so this is also one
10:03
Josh: of the earlier signs where you can actually merge a lot of these tools together
10:06
Josh: to build something dynamic in a way that you previously couldn't have done before.
10:09
Ejaaz: Yet actually the quality of this game looks like something out of uh league
10:13
Ejaaz: of legends or something like that at least that's what it reminds me of like
10:15
Ejaaz: the these games are getting way more high def than i expected i know it's just
10:19
Ejaaz: like it's pretty basic for anyone that's watching this they can kind of like
10:22
Ejaaz: pick with a finer eye but it's cool but for those of you who prefer like the
10:25
Ejaaz: more traditional side of games
10:27
Ejaaz: this might be something that you can kind of vibe code in a
10:30
Ejaaz: couple of minutes now it may look basic but theoretically
10:33
Ejaaz: this is like a 3D spatially aware game and that's not something you could achieve
10:38
Ejaaz: at least very easily with previous models what I love about this as well is
10:42
Ejaaz: it's also they've also created or included the prompt for all of these things
10:46
Ejaaz: so this is something that you can try right now like look at this And the prompt
10:49
Ejaaz: is no more than like, what's it?
10:50
Ejaaz: One, two, three, four, like 12 lives. 12 lives, dude.
10:54
Ejaaz: And you can have like a fully functioning game. You can probably then add an
10:58
Ejaaz: extra step or extra prompt saying, hey, can you deploy this to Vercel? And-
11:01
Ejaaz: Send that to your friends. Now you can use, you have a game.
11:03
Ejaaz: You're a game creator. You're a game developer.
11:05
Ejaaz: So the applications for this model cannot be understated.
11:10
Ejaaz: I'm going to be very honest. I thought this model was going to be just an iterative
11:13
Ejaaz: upgrade. I didn't think it would get anywhere near Claude Mythos.
11:17
Ejaaz: Two stories have now revealed themselves, which is, one, it's the answer to Claude Mythos.
11:21
Ejaaz: And two, it's really damn good. I am now convinced that compute is everything,
11:26
Ejaaz: but not in the way that I thought it would be useful. I thought it would be
11:28
Ejaaz: largely for pre-training.
11:31
Ejaaz: But to Sam's tweet earlier on, and also in Greg Brockman, the president of OpenAI's
11:36
Ejaaz: recent interview, they're going all in on inference, test time compute,
11:40
Ejaaz: which just means that if you have more compute and if you have a good enough
11:43
Ejaaz: model, it can do the thing.
11:44
Ejaaz: This thing, like I said, built itself. It's a self-improving model. Very, very impressive.
11:48
Josh: It's good for solving hard problems. It's good for thinking for a long time.
11:52
Josh: In fact, they marketed it as a model that can now think for 20 hours coherently.
11:56
Josh: Great which is almost a full day it can work
11:58
Josh: on a problem yeah and what you're noticing from this prompt that's on screen is
12:01
Josh: it doesn't take that much to get it going you don't need to
12:04
Josh: kind of spoon feed it all the way through anymore it can make
12:07
Josh: decisions on its own it can infer conclusions on what
12:10
Josh: you want just based on the the knowledge architecture that it
12:12
Josh: currently has it's amazingly impressive in fact one of the people who got access to
12:15
Josh: it early just posted on x that he's posting
12:18
Josh: live as his um prompt is seven hours
12:21
Josh: into his task it has been running for over seven hours
12:24
Josh: he said this has literally never happened before the models would maybe run
12:27
Josh: for 30 minutes or so wow or or if you
12:30
Josh: really shouted them after two to three hours but he's on seven plus hours i
12:34
Josh: think this is going to be fun for people with complicated things if you really
12:37
Josh: want to make a triple a feeling video game or a simulator or a really complex
12:42
Josh: website this is the model to try out and to use it with codex and see how all
12:46
Josh: these things kind of piece together it's really i mean
12:49
Josh: I wasn't, I didn't have my hopes very high based on the Opus 4.7 to 4.6 incremental improvement.
12:54
Josh: This seems like a very solid improvement over 5.4.
12:57
Ejaaz: Absolutely. And listen, if you are listening to this and you're like, listen, I'm not a gamer.
13:01
Ejaaz: I can't waste my time with that. I focus on more serious things.
13:05
Ejaaz: Well, for you serious people, if you're a manager at a top company or whatever
13:09
Ejaaz: that might be, this isn't just a toy or a model used for coders.
13:13
Ejaaz: A lot of the examples that we just gave are around coding.
13:15
Ejaaz: You can use this for just admin stuff or managerial work, like the capability
13:20
Ejaaz: of this model to think more strategically and long-term and understand the context
13:24
Ejaaz: of the tasks that you're working towards.
13:26
Ejaaz: Like we said earlier, for coding specifically, it can work on 20-hour-long expert tasks.
13:32
Ejaaz: That also applies for administrative stuff or things that are more generalized,
13:36
Ejaaz: white-collar worker work.
13:38
Ejaaz: And so in this example, Noam Brown says, I'm a manager at OpenAI,
13:41
Ejaaz: but I'm using this model to basically manage my entire team and make sure we're
13:45
Ejaaz: focused on the right things.
13:46
Ejaaz: And guess what? but the output of this team and this product has been pretty amazing.
13:49
Ejaaz: So all around really excellent work by the entire team and the inference team
13:54
Ejaaz: specifically, as Sam Altman says here.
13:56
Ejaaz: And yeah, I'm looking forward to using this thing. I don't have access to it right now.
13:59
Ejaaz: I've refreshed my account probably like five times at this point and it hasn't
14:04
Ejaaz: appeared. So maybe it's like a slow rollout.
14:06
Ejaaz: But if you're listening to this and you've tried it out, let us know what you're
14:08
Ejaaz: using it for. Let us know what amazes you. I really want to hear more.
14:11
Josh: Yeah, OpenAX had a pretty incredible week. And this comes on the back of their
14:14
Josh: new ImageGen model that they just released, which was also unbelievable.
14:17
Josh: If you haven't seen that episode, we just recorded it yesterday.
14:20
Josh: So I would go advise you to see because, oh my God, it is amazing.
14:23
Josh: We also recorded an episode on Apple's new CEO this week and what that means
14:27
Josh: for the company, as well as the hardware race and how this, I mean,
14:32
Josh: this model, Opus, no, not Opus, this is GPT.
14:34
Josh: GPT 5.5 is very much part of the AGI class of models that is built on Blackwell
14:40
Josh: chips. and we've recorded an entire episode all about that.
14:43
Josh: Very interesting, very fascinating. Also interesting and fascinating because
14:46
Josh: as always, this is the weekly roundup. We have a few other topics to talk about.
14:49
Josh: We have some news out of SpaceX, which is a pseudo acquisition.
14:53
Josh: Now they haven't quite acquired Cursor being the company in question,
14:56
Josh: but they have at least partnered with them with the option to buy Cursor for
15:01
Josh: either $60 billion or pay 10 billion for the right to actually work together.
15:05
Josh: This seems like a big deal.
15:07
Josh: This seems like, I mean, XAI, we could call it SpaceX, but SpaceX AI is taking
15:11
Josh: AI very seriously. They're currently behind.
15:13
Josh: They clearly don't want to be behind. This is a huge step and a huge kind of
15:17
Josh: trust of support in Cursor with this minimum of $10 billion into accelerating
15:21
Josh: their progress and trying to get themselves into this game.
15:24
Ejaaz: This is actually a genius deal, and there are a few stories why it makes that so. So let me explain.
15:30
Ejaaz: If you're SpaceX AI, which by the way is a ridiculous name now,
15:34
Ejaaz: like we'll just call them XAI, you are currently harboring...
15:39
Ejaaz: One to 1.5 million of the frontier GPUs, mainly NVIDIA, in a warehouse. There's one issue.
15:47
Ejaaz: You're not really utilizing all of it because XAI has had a bit of a slow start
15:51
Ejaaz: to training their models. What's a genius idea?
15:54
Ejaaz: If I rent those out to another company to train their own model,
15:59
Ejaaz: then we can make money from that. Okay, so that's win number one for SpaceX.
16:02
Ejaaz: But then they've thought of another thing which is huh grok
16:06
Ejaaz: isn't really good at coding and we are
16:09
Ejaaz: losing the race every single day we don't update our model
16:12
Ejaaz: at coding because anthropic and chat gpt
16:15
Ejaaz: 5.5 is completely running away with it so
16:18
Ejaaz: how did they leapfrog and get ahead they should acquire the company that is
16:23
Ejaaz: using their own gpus to train a frontier coding model so then the question becomes
16:27
Ejaaz: well who the hell is cursor what what's the mode that they have like why do
16:31
Ejaaz: they have a good shot of training a better coding model than Anthropic and GPT-505?
16:36
Ejaaz: Aren't those two companies way ahead? Well, the answer is not quite so.
16:40
Ejaaz: Cursor, for the longest time, was the number one platform and tool for people
16:44
Ejaaz: to use to do their Vibe coding. Why?
16:46
Ejaaz: Not only did they have access to Frontier coding models from Claude and ChatGPT,
16:50
Ejaaz: they also had something called an agent harness.
16:53
Ejaaz: Now, you'll notice in GPT-505, it's really good at coding because of something
16:58
Ejaaz: called agentic That is something that Cursor pretty much pioneered.
17:03
Ejaaz: It's basically the harness, the prompts, the environment that they mold the model,
17:07
Ejaaz: or rather that they mold around the model that makes it so good and intuitive
17:12
Ejaaz: and remembers the context across every single project, like menial things,
17:16
Ejaaz: like understanding your GitHub branches and working on separate flows at the same time.
17:20
Ejaaz: A lot of the top software engineers in the world right now use tools like Curse
17:25
Ejaaz: and Argentic Coding to be able to pull this off.
17:27
Ejaaz: So Elon Musk thought, hmm, if I give you the GPUs to train a better coding model,
17:32
Ejaaz: which gives you a better product, I should have the option to acquire you.
17:36
Ejaaz: In acquiring you, I can integrate you with Grok and Grok somehow becomes the
17:40
Ejaaz: number one coding model over the next year or so, depending on if this deal goes.
17:44
Ejaaz: And if the deal falls through and they create a really bad model,
17:48
Ejaaz: well, you pay me $10 billion for the service.
17:50
Ejaaz: Well i pay you not a bad deal not a bad deal
17:52
Josh: Yeah it seems like they're they're going to be continuing to
17:55
Josh: work with other companies to accelerate in places that
17:58
Josh: they're weak at currently because i mean they they're so strong at
18:01
Josh: building out the hardware and creating these huge data centers they need
18:04
Josh: someone who could take advantage of all those gpus hopefully this will
18:06
Josh: help serve that cause and that's not the only spacex news
18:09
Josh: this week the other is that they have officially filed an s1 which
18:12
Josh: for those who are not familiar it means they're going public it's officially official
18:15
Josh: 100 they will be going public this year if there
18:18
Josh: were any doubts please let them be relinquished here we
18:21
Josh: have it spacex will be going published the most interesting thing from
18:24
Josh: this was i think the share structure of
18:28
Josh: how they're going to be organizing this for daddy elon who's going to be getting
18:32
Josh: quite a big payday if he does well so we have on screen here just a series of
18:37
Josh: some of the financials i mean we know starlink as a business has been doing
18:39
Josh: unbelievable they have about 25 billion dollars in cash 92 billion assets 50
18:44
Josh: billion liabilities that's.
18:45
Ejaaz: Quite a lot of liabilities on this my god
18:47
Josh: They got a lot of debt man i don't know we'll see we'll see once they finally
18:51
Josh: publish everything i'm very excited for the first earnings report where you
18:54
Josh: really get a true peek behind the scenes of what's going on there but it looks
18:57
Josh: like it's going to be going public at a 1.75 trillion dollar valuation now in terms of pay structure.
19:02
Josh: Elon is posed to get 60 million shares
19:05
Josh: which is 11 tranches vesting in
19:09
Josh: 500 billion dollar market cap increments from 1.1 trillion to 6.6 trillion dollar
19:16
Josh: share price oh um so for those unfamiliar with the current ceiling i think it's
19:21
Josh: nvidia nvidia is what five trillion under five trillion close to five trillion
19:25
Josh: it's like 4.3 yeah okay so not even close they're like 20 away from five trillion.
19:29
Josh: SpaceX needs to be, what is that? Like 20 something percent more valuable than
19:33
Josh: the most valuable company in the world.
19:34
Josh: But if they do, Elon gets 60 million shares.
19:38
Josh: Now I haven't done the math on exactly how much that is.
19:41
Josh: But if we make some assumptions here, the total value at Vest looks like it
19:47
Josh: could be about a quarter of a trillion dollars.
19:49
Josh: So pretty good payday for Elon. I think the most important thing is that he's
19:52
Josh: getting a lot of control over this.
19:54
Josh: It seems as if he's going to have 40 something percent control of
19:58
Josh: the company which is really ultimately what was most important to
20:00
Josh: him as they went public so really exciting news i am hopeful that it happens
20:04
Josh: this june which we can expect and it's without a shadow of a doubt going to
20:08
Josh: be the largest ipo in history i think everyone's going to be talking about it
20:11
Josh: there is a new vehicle in which some people are investing in we're actually
20:14
Josh: going to have the founder on the show soon so keep an eye out for that one.
20:19
Josh: And yeah, the SpaceX news is very exciting.
20:21
Ejaaz: Now, in the world of AI hardware, many people think that NVIDIA has run away with the win.
20:27
Ejaaz: And you could argue that with a $4.300 market cap, not many people are competing,
20:31
Ejaaz: except that there is one company, Google.
20:34
Ejaaz: Now, you might be thinking, Google does all my search engines and stuff.
20:36
Ejaaz: Well, Google is the only vertically integrated Mag 7 company that is involved
20:43
Ejaaz: or has a frontier capability at every single layer of the AI stack.
20:47
Ejaaz: Now, right at the bottom are these things called Google TPUs,
20:51
Ejaaz: Tensor Processing Units.
20:52
Ejaaz: And they're their version of the GPU. In fact, fun fact,
20:56
Ejaaz: Google's Gemini models has never trained on an NVIDIA GPU.
21:00
Ejaaz: It's all been their own internal warehoused infrastructure. And they've been
21:04
Ejaaz: working on this thing for 10 years.
21:06
Ejaaz: Now, just today, or rather this week, they released their latest generation
21:10
Ejaaz: of TPUs, the TPU-8T and the TPU-8i.
21:14
Ejaaz: Now, the TPU-8T, T stands for training or pre-training.
21:19
Ejaaz: It is highly optimized for the pre-training part of an AI model.
21:23
Ejaaz: So this is like the bulk, arguably the more expensive part of training a model.
21:27
Ejaaz: It's like teaching it like, hey, these are words.
21:29
Ejaaz: These are the general fundamental set of facts that you need to know before
21:32
Ejaaz: we can kind of like put you out into the world and present you to our users.
21:37
Ejaaz: TPU AI is specialized or hyper-specialized in inference specifically.
21:42
Ejaaz: Now, the important part about inference is it's being used for so many different things.
21:46
Ejaaz: Number one, it's to answer all your different prompts. Whenever you write a
21:49
Ejaaz: prompt and you submit it to an AI model, it is known as inference.
21:52
Ejaaz: It's getting inference. It needs to query the model and make sure it does the
21:55
Ejaaz: right types of thinking and gives you the right answer.
21:57
Ejaaz: But the other part of inference is post-training, where a lot of people train
22:01
Ejaaz: the model, and then they do more training after the fact by using it to help
22:06
Ejaaz: the model reason and think of other alternative facts before it presents you the actual answer.
22:11
Ejaaz: And that's what that second TPU is. Now, Google's TPUs have been used extensively.
22:15
Ejaaz: In fact, their largest customer is a little-known AI lab known as Anthropic,
22:19
Ejaaz: which currently runs 1.5 million TPUs. So the argument can be made that TPUs
22:25
Ejaaz: are largely responsible for Claude's and Opus' success.
22:28
Ejaaz: So very impressive all around, but there's some other facts about this, right?
22:31
Josh: Yeah, well, I love the dual architecture training setup that they have here being hyper specific.
22:37
Josh: I mean, the AT chip in particular, it's built to reduce frontier model development
22:40
Josh: cycles, they said, from months to weeks.
22:43
Josh: And then we have the AI, which is the reasoning engine, which is specifically
22:46
Josh: served for agentic use to deliver tokens really quick, as fast as possible.
22:51
Josh: And as we know, Anthropic is working closely with them. And also,
22:54
Josh: I mean, Google is making these for themselves.
22:55
Josh: So I think whoever is working with Google, whoever's kind of focused on these
23:00
Josh: accelerators, is probably in for a nice little windfall as it relates to increased
23:05
Josh: velocity of the training and also increased ability to distribute these models.
23:08
Josh: As we know, Anthropik is having a very difficult time with this.
23:11
Josh: Now, NVIDIA and Jensen are probably feeling a little shook.
23:14
Josh: They got to be feeling a little bit of pressure here. and it seems as if
23:17
Josh: that's why they're pushing to be open source because if you
23:20
Josh: are a in a closed source world where everyone is
23:23
Josh: making closed source models on their own architecture then the
23:26
Josh: nvidia edge very quickly disappears and i mean i'm looking at these chips in
23:29
Josh: hand they look beautiful they're ready to be they're taped out ready to be manufactured
23:32
Josh: and i i think you could start getting kind of excited about this new world of
23:36
Josh: accelerated hardware and we're seeing this happen again and again because amazon
23:40
Josh: just made another big investment in who else other than Anthropic.
23:46
Josh: And the deal, I think, is like this has to be close to a record deal.
23:49
Josh: They're owning a tremendous amount of this company now.
23:50
Ejaaz: Yep. So the news here is Amazon announced they're investing $5 billion into Anthropic.
23:57
Ejaaz: They've just raised $5 billion. Congrats.
24:00
Ejaaz: And so the reason why this is important is, well, there's a few reasons.
24:04
Ejaaz: Number one, Anthropic knows that they don't have enough compute.
24:07
Ejaaz: The argument could be made that's why Claude Mythos hasn't been rolled out.
24:10
Ejaaz: Well, hey, hey, presto, now you have $5 billion worth more of compute.
24:14
Ejaaz: Now, for those of you who didn't know, Amazon is a primary investor already in Anthropic.
24:19
Ejaaz: Before this announcement, they owned around 17% of Anthropic.
24:23
Ejaaz: After this announcement, it's closer to 20%. So we're talking about one company
24:28
Ejaaz: that's publicly tradable right now that owns a fifth.
24:32
Ejaaz: Is my math right? Yeah, a fifth of the world's leading
24:35
Ejaaz: AI lab, which is pretty crazy. Now, if we look into the stats of this,
24:40
Ejaaz: this is a five gigawatt deal, which is more than any single data center that's currently live.
24:46
Ejaaz: It's actually a multiple of five. I think SpaceX AI's Colossus 2 is the largest
24:52
Ejaaz: right now with their 1 million TB.
24:53
Ejaaz: So it's going to be 5X larger than the average data center that we're seeing
24:57
Ejaaz: right now for AI specifically.
24:58
Ejaaz: And they're aiming to get one gigawatt online by the end of the year.
25:01
Ejaaz: Now, Now, the reason why this is so good for both teams is Anthropic already
25:05
Ejaaz: has a close relationship with AWS and Amazon's cloud computing department.
25:09
Ejaaz: So spinning up more compute clusters is gonna be so easy for them.
25:13
Ejaaz: They have a working relationship.
25:14
Ejaaz: They're used to training cloud models on this, so it shouldn't be too hard to
25:17
Ejaaz: ramp this up. If you're Amazon,
25:20
Ejaaz: Welcome back. That $5 billion is going to come right back to you.
25:24
Ejaaz: So I don't know what kind of like circle economy this is, but it's back and
25:27
Ejaaz: it's very impressive for them.
25:29
Josh: Is it ironic that today Amazon hit an all-time high? No, maybe, maybe not.
25:33
Ejaaz: I'm holding stock. I got the stock.
25:34
Josh: Clearly, clearly they're doing something right. Amazon is a phenomenal company.
25:37
Josh: They're the largest shareholder in Anthropic.
25:39
Josh: It's hard not to be bullish on them. It's hard not to be bullish on the accelerated
25:42
Josh: computing stack. And I think that's probably what Jensen is getting nervous
25:45
Josh: about. That's why NVIDIA is pushing open source.
25:46
Josh: And the good news is, is he has some help. He has some assistance.
25:50
Josh: From the folks overseas in China who have been pumping out unbelievable models
25:54
Josh: all week long as it relates to Kimi and Quen, our Chinese favorites.
25:58
Josh: We have Kimi K2.6 and Quen 3.6. There's a lot of digits and numbers.
26:04
Josh: All you need to know is that the best open source models in the world didn't exist last week.
26:09
Josh: They now exist this week and they are better at pretty much everything,
26:12
Josh: but exceptional at coding.
26:14
Josh: In fact, word on the street is that some of these models are as good as GPT
26:18
Josh: 5.4 was and only a few points off of Claude.
26:21
Josh: I mean, these are pretty amazing open source models that, again,
26:23
Josh: are free to run locally on your machine if you have the machine capability of doing so.
26:28
Josh: That's a big, that's a big game changer.
26:30
Ejaaz: Okay, so typically the story we tell with these open source models is,
26:34
Ejaaz: wow, aren't they so amazing?
26:36
Ejaaz: Yeah, they're the good younger brother. They're not as good as the Frontier AI Labs.
26:40
Ejaaz: That completely changed this week. So Kimi K 2.6 is the latest model from a
26:45
Ejaaz: Chinese lab called Moonshot Labs. I believe it's Moonshot or Moonshot AI.
26:49
Ejaaz: And they released their model, which ends up being as good as coding or at coding.
26:54
Ejaaz: As opus 4.7 and it's 100 open source
26:57
Ejaaz: like you mentioned josh which means that maybe you could run
27:00
Ejaaz: this on a local device now the answer that you would typically get back from
27:03
Ejaaz: this is hey like listen it's uh it's too
27:05
Ejaaz: large to run on my laptop and that is true but with the latest quen model which
27:09
Ejaaz: is a 3.6 version you can run it as an 18 gigabyte sized model slightly quantized
27:16
Ejaaz: on your laptop today so the point that i want to make about these models isn't
27:20
Ejaaz: exactly the specifics but across all benchmarks,
27:23
Ejaaz: it's not as good as the Frontier AI Labs, but it's a few points.
27:27
Ejaaz: That difference and gap has closed massively over the last couple of months,
27:31
Ejaaz: which tells me two things.
27:33
Ejaaz: Number one, China has figured out some kind of groundbreaking way to train their
27:37
Ejaaz: models that they haven't told the West about, and they're going to keep it closed
27:40
Ejaaz: guard and eventually close source their model releases going forwards.
27:44
Ejaaz: And number two, they've figured out a new way to use inference to their benefit.
27:48
Ejaaz: Like one thing I'm going to highlight here is this new Kimi K 2.6 model can
27:51
Ejaaz: code continuously for 12 hours straight using 300 agents.
27:56
Ejaaz: So the unlock here isn't one model itself. It's spitting up 300 versions of
28:00
Ejaaz: itself and getting it to attack the problem.
28:02
Ejaaz: That's something Sam realized and what he's implementing in 5.5.
28:05
Ejaaz: That's something Opus 4.7 realized and is doing probably similarly with Mythos.
28:09
Ejaaz: So I have this question here, which is like, how do they have to try to do this?
28:12
Ejaaz: Well, I think every three months that there's a new open model that gets released,
28:16
Ejaaz: they're making these jumps because they're using these models to train themselves
28:19
Ejaaz: we proved that with kimmy k 2.5 there's too many two point whatevers um and
28:23
Ejaaz: the same thing is happening with quen it's just all around pretty amazing stuff um
28:27
Josh: Yeah chen is crushing okay so before we go we have two quick things to hit the
28:31
Josh: first being one that we missed last week which we need to touch on quickly anthropic
28:34
Josh: has a design tool now if you are a designer if you are interested in building
28:38
Josh: web pages videos graphics slideshows pitch decks any type of visual asset you're
28:43
Josh: claude now has an entire design suite built just
28:47
Josh: for this purpose it's called claude design it exists separately
28:50
Josh: you can access it through the desktop app or on your browser and
28:53
Josh: it basically allows you to build visual assets in a way that
28:56
Josh: you couldn't previously previously with claude you had artifacts an
28:59
Josh: artifact you could generate something dynamic it could kind of
29:02
Josh: build you a web page this takes it to a whole new level you could generate wireframes
29:05
Josh: if you want to try it to use less tokens you could fill it out and create properly
29:09
Josh: created prototypes that are actually clickable it's amazing the video we're
29:13
Josh: seeing on screen highlights a few of them unfortunately there was a big loser
29:16
Josh: in this because uh this sounds like a lot of what that little design company named figma does.
29:21
Ejaaz: Yeah the little company
29:22
Josh: Stock market did not love the reaction to that did it nope.
29:26
Ejaaz: Nope it is down almost 20 on the week um i actually tracked the stock price
29:31
Ejaaz: after the announcement was made so like it wasn't even readily available it
29:34
Ejaaz: was literally just the tweet 20 minutes after it was tweeted the stock was down
29:38
Ejaaz: six percent so So the point being,
29:40
Ejaaz: whether this is market speculation or not, like, listen,
29:43
Ejaaz: Claude Design isn't as good as Figma.
29:45
Ejaaz: They're working with a few of these different partners, such as Canva.
29:49
Ejaaz: Two weeks ago, one of Anthropic's former most execs left the board of Figma.
29:55
Ejaaz: And the rumors was because they were building a competitor. So it's pretty clear.
29:59
Ejaaz: Anthropic is going off to every single sector, whether you're a designer,
30:02
Ejaaz: a software engineer, a mathematician, a research scientist, doesn't matter.
30:05
Ejaaz: They're going off to everything because the model is applicable to everything.
30:08
Ejaaz: And I don't know what this means for certain modes that companies like Figma
30:11
Ejaaz: holds, but it's certainly going to affect the stock price.
30:14
Josh: Can you do me a favor and click the max button real quick for me just to show the chart?
30:17
Ejaaz: Oh!
30:19
Josh: Yeah, minus 86% since IPO for those who are not watching on screen.
30:23
Josh: It's been a pretty bad, rough run for Figma.
30:27
Ejaaz: We have to start naming Anthropic the stock killer, Josh. This is like every
30:31
Ejaaz: single tweet is tanking a stock.
30:33
Josh: No, it's tough. It's brutal. We had one last thing that you wanted to mention.
30:36
Josh: I know. We got to end on this strong. What do we have?
30:38
Ejaaz: How good is your accent or impersonation of your president, of our president, Josh?
30:45
Josh: Pretty horrible.
30:46
Ejaaz: Not good. Okay, well, then we're not going to attempt it.
30:47
Josh: I'd love to hear your British take on it, though, if you're feeling ambitious.
30:51
Ejaaz: Okay, so my British take on this is, this is, albeit hilarious and somewhat
30:56
Ejaaz: terrifying, that the President of the United States is saying this.
31:00
Ejaaz: He commented, okay, on the government's relationship with Anthropic.
31:04
Ejaaz: Now, if you're wondering why on earth he's commenting on it,
31:06
Ejaaz: they're going to be releasing this cold mythos model.
31:08
Ejaaz: It might be a security risk. It's probably good for the government to have access
31:11
Ejaaz: to this thing and prepare necessarily.
31:13
Ejaaz: The government has been having very important conversations with bankers and
31:16
Ejaaz: governments all around the world, just try and figure out, you know,
31:19
Ejaaz: how best to prepare for this.
31:20
Ejaaz: And after having an in-depth discussion with Dario Amore, which by the way,
31:25
Ejaaz: he blacklisted that CEO and Anthropic entirely from the government using it.
31:29
Ejaaz: He's now rekindling it and saying, maybe there's a deal on the line.
31:32
Ejaaz: He goes, and I quote, I'm not going to do the accent.
31:34
Ejaaz: We'll get along with Anthropic just fine. Trump said on CNBC.
31:38
Josh: We'll get along with Anthropic just fine. I think they can be of great use to us.
31:43
Josh: They're high IQ people. Very good. Very good. They tend to be on the left,
31:46
Josh: radical left, but We get along with them.
31:49
Josh: I don't know. That's all I got. But that is what he said.
31:51
Ejaaz: Were you practicing that? That was actually pretty good. I was practicing my head.
31:54
Josh: I was rehearsing.
31:55
Ejaaz: Damn. I closed my eyes whilst you were doing that, whilst I was laughing. Did it feel right?
31:59
Josh: It sounded like him? Good.
32:00
Ejaaz: It channeled his spirit. It was there. It was a good effort.
32:03
Ejaaz: But I believe that's it. That is the end of the roundup.
32:07
Ejaaz: Josh and I are recording this. FYI, it's 4 p.m. over here. Typically, we're morning birds.
32:12
Ejaaz: We deliver this in the morning, but we waited for the announcement of Spud GPT
32:17
Ejaaz: 5.5 just for you guys. and we're going to be bringing you the cutting edge news every single week.
32:22
Ejaaz: As Josh mentioned, we had three other amazing episodes that we filmed earlier
32:25
Ejaaz: this week. Definitely go check them out. They're all each 20 minute song.
32:28
Ejaaz: It's your commute to work.
32:29
Ejaaz: It's your gym session if you're not that active.
32:32
Ejaaz: Definitely go check it out and let us know what you think. But yeah,
32:35
Ejaaz: Josh, any final thoughts?
32:36
Josh: Call me crazy, but I like the afternoon recordings. I got good energy.
32:39
Josh: I'm like woken up. I'm 100% right now. I'm rocking and rolling.
32:42
Josh: I'm feeling good. So I don't know. Maybe we'll have to lean into this a little
32:45
Josh: bit more, but that's everything.
32:46
Josh: If you've made it this far, if you're still listening to this and you've heard
32:48
Josh: our other episodes, you're caught up. You're done for the week.
32:51
Josh: You can go touch grass. Enjoy your weekend.
32:52
Josh: There will be a lot more to talk about next weekend. But for now,
32:55
Josh: you have fully synchronized with all of the chaos happening on the frontier of AI and technology.
33:01
Josh: Thank you so much for watching. As always, we very much appreciate it.
33:03
Josh: If you enjoyed this episode or any of our previous episodes from this week,
33:06
Josh: don't forget to share them with a friend who you also might enjoy it, possibly.
33:09
Josh: We have a newsletter on Substack that goes live twice a week.
33:12
Josh: Just went live yesterday, going live again tomorrow.
33:14
Josh: The Friday issue is a recap of everything that happens this week,
33:17
Josh: which is always fun and exciting.
33:18
Josh: In fact, I'm gonna go write that as soon as we finish this episode.
33:21
Josh: So thank you all for watching.
33:22
Josh: As always, don't forget to subscribe, like, comment, all the good things,
33:25
Josh: and we will see you guys next week.