Latest / In the Loop AI / THIS WEEK IN AI: Kimi K3, OpenAI's Alexa, Grok Stealing Code, Thinking Machines

Transcript

THIS WEEK IN AI: Kimi K3, OpenAI's Alexa, Grok Stealing Code, Thinking Machines · In the Loop AI

0:00
Ejaaz: For years, the battle between open source and centralized AI models has been very one-sided.
0:06
Ejaaz: Anthropic open air has always dominated the field until this week where Moonshot
0:11
Ejaaz: Labs, an AI lab out of China, has finally released their next model,
0:16
Ejaaz: Kimi K3, and it is Fable 5 worthy.
0:19
Ejaaz: Now, the craziest part about this is Chinese and open source models in general
0:23
Ejaaz: have been behind American frontier labs. They haven't been able to close the
0:27
Ejaaz: gap. Dario recently said that that gap is roughly around eight months,
0:31
Ejaaz: having closed from about a year.
0:33
Ejaaz: Today, with Kimi K3, that model has closed it to today.
0:37
Ejaaz: It is as good at one-shotting visual prompts as Fable 5 is. It's amazing at
0:42
Ejaaz: reasoning in general. And we're going to get into a bunch of cool demos that
0:45
Ejaaz: we see on our screen today.
0:47
Ejaaz: Other big news in open source, Thinking Machine Labs, which,
0:50
Ejaaz: funnily enough, raised the largest seed round ever, $2 billion,
0:54
Ejaaz: headed by ex-OpenAI CTO Miro Murati, released their first model called Inkling.
1:00
Ejaaz: It's small, yet pretty mighty. We're going to get into that.
1:02
Ejaaz: And then OpenAI themselves have revealed some secretive details about their
1:07
Ejaaz: new hardware device. If you've watched the show at all, you know Joshua and
1:10
Ejaaz: I are obsessed with this.
1:11
Ejaaz: We're excited to get into the details and show you what is coming soon.
1:14
Josh: Yeah, so to start, we have a new king on the block, and that comes in the form
1:17
Josh: of Kimi K3. I mean, this very much feels like...
1:21
Josh: I wouldn't say the DeepSeek moment, but something somewhat familiar in the sense
1:24
Josh: that this is a really novel, leading edge, open source model coming out of the Kimmy team.
1:30
Josh: And it was paired with this really lovely launch video that actually.
1:34
Josh: I was shocked that it was coming from this lab because this looks like something
1:37
Josh: Apple would release as a product launch video.
1:40
Josh: It was beautiful. The sound design was amazing. The visuals were great.
1:43
Josh: And it's funny, they launched this before they even officially launched the
1:45
Josh: model. In fact, at the time of recording, they haven't even officially stated
1:49
Josh: that the model is live but
1:50
Josh: it's available it is available in the model picker for you to go and test it
1:53
Josh: now the model comes in the form of about a 2.8 trillion parameter model which
1:58
Josh: if you listen to the episode yesterday featuring grok that's about double what
2:01
Josh: grok 4.5 was in terms of parameter count
2:04
Josh: based on the early reviews that i've seen a lot of people have been using this
2:06
Josh: model kind of early it seems like this is pretty incredible um it seems like
2:11
Josh: people have been able to do really impressive long horizon tasks they've been able to do really
2:15
Josh: impressive coding generation gaming generation 3d generation and a lot of people are comparing this to
2:21
Josh: close to a fable worthy model and early testers they put it above gpt 5.6 actually
2:26
Josh: on coding and almost fable worthy at coding which puts it in a really solid
2:30
Josh: spot because the pricing of this thing is going to be very low and the quality
2:34
Josh: of it is incredibly high and to
2:36
Josh: our point yesterday talking about the two frontiers there is the
2:40
Josh: intelligence frontier and there is the cost Frontier, this is placing a new
2:44
Josh: point on that Pareto curve of cost and intelligence at a part of the curve that
2:48
Josh: I don't think anyone's ever made to before.
2:50
Josh: This is like very much a frontier model in terms of cost per intelligence.
2:53
Josh: It's really impressive.
2:55
Ejaaz: I think this tweet that I have over here summarizes it the best.
2:59
Ejaaz: He goes, the more I test K3, the more it feels like another DeepSeek R1 moment.
3:04
Ejaaz: It is often Fable level, maybe a little worse, but consistently better than 5.6.
3:09
Ejaaz: And that's a recurring theme across a lot of different takes from early testers,
3:15
Ejaaz: that it's not quite as good as Fable. It's as good as Fable when it comes to
3:19
Ejaaz: certain things, but not everything.
3:21
Ejaaz: But it is certainly better or on par with GPT 5.6, which is a bold claim,
3:25
Ejaaz: obviously, because OpenAI has invested billions and billions of dollars into
3:28
Ejaaz: their thing. Now, enough of us claiming theory. Let's talk about some actual examples.
3:34
Ejaaz: So on your screen right now, you're looking at the promotional video from the
3:37
Ejaaz: Moonshot Labs themselves.
3:39
Ejaaz: Now, what I'm going to shift to right now is Kimi K3 one-shotting the video entirely from scratch.
3:45
Josh: Whoa. And building it from scratch. Oh, that's really cool.
3:48
Ejaaz: Exactly. This is one single prompt. It was fed the promotional video that Moonshot
3:53
Ejaaz: Labs themselves created from scratch using humans and all that kind of stuff.
3:57
Ejaaz: And it only took 25 minutes to do.
4:00
Ejaaz: Now, rumor, the cost to create this thing. Josh, actually, take a guess.
4:04
Ejaaz: How much did it cost to create this video?
4:06
Josh: Oh, my God. One shot. probably nothing right because the cost per million tokens
4:10
Josh: is 95 cents i think so what would this take maybe five million tokens or say
4:15
Josh: five five to ten dollars no more than 10 bucks.
4:17
Ejaaz: It costs under 10 bucks, which is just insane.
4:20
Josh: Oh, that's so cool. It's insane.
4:22
Ejaaz: Now, like if you have used products from Anthropic or OpenAire that are relevant
4:29
Ejaaz: to artifacts, the visual diagrammatics that they help you create,
4:33
Ejaaz: it takes a while. It takes a couple of bits of prompting, sequential prompting.
4:37
Ejaaz: With Kimi, it seems that it could just one-shot this completely.
4:39
Ejaaz: And it's not just one example I have.
4:41
Ejaaz: I have quite a few. If you look at this, you're looking at a 3D simulation.
4:46
Ejaaz: Now, you would be excused if you thought that this was Fable 5 worthy,
4:50
Ejaaz: because if you rewind a few episodes ago where we covered all of Fable 5's cool
4:54
Ejaaz: demos and stuff, we showed you a Hogwarts demo, a bunch of other things like that.
4:58
Ejaaz: It was of the same quality and par. Kimi K3 can now compete on the same level.
5:03
Ejaaz: And it's not just for simulations. It's recreated Mario Kart from scratch.
5:07
Ejaaz: Now, granted, it looks like Minecraft, but, you know, what AI model doesn't do that?
5:11
Ejaaz: And you could imagine with a few more prompts and like visual graphics added,
5:14
Ejaaz: it could get to, you know, maybe not a AAA, but like a AA type game vibe.
5:20
Ejaaz: So the point is, it's a very impressive model and it's physically accurate.
5:24
Ejaaz: It gets the physics of how things work, the mechanics of wings that you're seeing
5:29
Ejaaz: On your screen right now, how it works. It nails it.
5:32
Ejaaz: Now, where it falls or where it's a little lackluster is general reasoning,
5:36
Ejaaz: writing and coding in general. Now, coding for visual prompts, it's very good.
5:41
Ejaaz: Coding for hard specifically developer tasks, Fable 5 is still good,
5:46
Ejaaz: but it matches the likes of GPT 5.6 Terra and Luna.
5:50
Ejaaz: Maybe not quite Sol, but Terra and Luna and certainly Sol at least on medium
5:54
Ejaaz: effort. So if you're looking for a comparison as to whether you should use this
5:57
Ejaaz: model, it's pretty up there and you should be surprised that they've been able
6:00
Ejaaz: to close the gap so quickly.
6:02
Ejaaz: The final thing I'll say is on cost, it is super cheap. It is under a dollar
6:06
Ejaaz: input tokens, and I think it is a couple dollars output tokens.
6:09
Ejaaz: Again, none of this is officially confirmed. Just yet we're getting like leaks
6:13
Ejaaz: as we're recording this episode.
6:15
Ejaaz: That would price it at the same value as sonnet five or sonnet the sonnet model
6:20
Ejaaz: from from claude themselves so we're not even talking about opus level here
6:23
Ejaaz: we're talking about like sonnet cost which is like cheapest chips
6:25
Josh: Yeah very impressive i mean it's funny as we cover this model i think about the open claw community
6:30
Josh: and people who use these long-running agentic tasks i don't know where they
6:34
Josh: are now i haven't seen many of the open clock community publicly in a while
6:37
Josh: but this is such a dream model for them where it is so cheap it is so capable
6:42
Josh: of these long-running agentic tasks it has all the visual understanding it seems
6:45
Josh: like a very good companion
6:46
Josh: and this seems like a legitimate kind of dent in the moat of these companies
6:51
Josh: that are kind of holding the frontier of that pareto curve in terms of cost
6:54
Josh: in terms of intelligence so it's a really impressive model coming out of china
6:57
Josh: i'd love to know how this came to be
7:00
Josh: what types of models were distilled if any kind of whose ip are you taking here
7:05
Josh: whose stolen gpus are you training this on listen.
7:07
Ejaaz: Josh that's actually important context right you you raise a good point because
7:11
Ejaaz: um moonshot labs the guys that created this model have been accused by not just
7:15
Ejaaz: Anthropite, but also OpenAI and Google for distilling their frontier models.
7:20
Ejaaz: So we have no proof whether they've done the same thing here,
7:23
Ejaaz: but maybe it can be assumed. I don't know.
7:25
Josh: I mean, yeah, I find it hard to believe that it's anything otherwise than they've
7:28
Josh: just distilled the frontier models and turned them into cheaper models that are more capable.
7:33
Josh: It seems like this is kind of well known, that they have smuggled GPUs into
7:36
Josh: the country, that they have like kind of,
7:39
Josh: nudge the numbers on how much the training costs were on a relative basis where
7:43
Josh: they say it's a lot cheaper than it actually is
7:45
Josh: there's a lot of malpractice going on behind these models which is why it's
7:48
Josh: a little difficult to support them but man on just a pure open weight basis
7:53
Josh: for them to publish a model like this it's pretty impressive and i think a lot
7:56
Josh: of people are going to get a lot of use out of this so,
7:59
Josh: i am going to go and try this i'm going to check my model picker after this
8:01
Josh: just to test maybe we'll have a follow-up episode showing some of the demos
8:05
Josh: all the interesting things you could do with it but if you are watching this
8:07
Josh: chances are it is live it's available. Go try it out and let us know what you think.
8:11
Josh: That brings us to our second topic of the day. And this one I am particularly
8:15
Josh: excited about because, Ejaz, if you remember, 18 months ago,
8:18
Josh: after leaving OpenAI, the co-founder of OpenAI, Mur Murati, she left and she
8:24
Josh: went to go start a company named Thinking Machines.
8:27
Josh: She is one of many co-founders who have left the company, unfortunately,
8:30
Josh: for better or worse, and went off and tried to do her own thing.
8:33
Josh: During this process, she raised the largest seed round in venture capital history.
8:38
Josh: Now, seed rounds, generally speaking, I remember back in the day, a seed round of like,
8:42
Josh: a couple million dollars if you raise like five to ten million dollars.
8:44
Josh: That's a remarkable seed round. This was two billion dollars.
8:48
Josh: And for those not familiar, the seed round is the first money into the company.
8:52
Josh: You have an idea and you say, I'm going to go build this idea and an investor
8:54
Josh: gives you some money to do it.
8:55
Josh: She raised two billion dollars at a twelve billion dollar valuation right off
8:59
Josh: the bat from companies like Andreessen Horowitz, NVIDIA, AMD and Jane Street.
9:03
Josh: These are the largest ones in the game.
9:05
Josh: So for the last 18 or so months, we have been asking ourselves the question,
9:08
Josh: well, what are they actually working on?
9:10
Josh: Two billion dollars is a tremendous amount of money to go off and build something
9:14
Josh: and mira and the team are clearly very talented because they were one of the
9:17
Josh: day ones at open ai who were building the gpt
9:20
Josh: early gpt days and as of today we finally have an answer and it comes in the
9:24
Josh: form of this model named ship shipling inkling what is this what did they actually release.
9:30
Ejaaz: It to be fair it doesn't roll off the tongue but it's called inkling and it is a
9:34
Josh: Model that falls.
9:35
Ejaaz: Yeah yeah it's a model that falls just under a trillion parameters in size,
9:39
Ejaaz: so relatively small given the landscape today.
9:42
Ejaaz: But would it surprise you if I said it's a pretty unremarkable model that is
9:48
Ejaaz: remarkable in a very specific way that not many people see? So let me explain what I mean by that.
9:54
Ejaaz: It's trained on about 45 trillion tokens of text, image, video, and audio.
10:00
Ejaaz: Now, the reason why that is pretty cool is that it's completely omni-model.
10:05
Ejaaz: It's like a multi-modal model that you can send text to, you can send audio
10:11
Ejaaz: to, you can speak to it, doesn't interrupt you.
10:14
Ejaaz: It's that type of model that is very intuitive. It feels very human.
10:18
Ejaaz: And Mira Murati released a really interesting blog post earlier this week,
10:22
Ejaaz: which described her vision for Thinking Machine Labs.
10:24
Ejaaz: It's not to be the frontier LLM, it's to be the frontier companion for humans.
10:29
Ejaaz: And so she's building these models in a very specific way. So when we look at
10:33
Ejaaz: the model itself, it was trained in under nine months.
10:36
Ejaaz: It is quite good at reasoning. It is okay at coding.
10:40
Ejaaz: It is nowhere near Fable 5 worthy. But where it gets super powerful is if you take their model,
10:46
Ejaaz: And then you fine tune it for your own specific task department or team.
10:50
Ejaaz: So what that looks like is, and there's a live example of this actually happening
10:54
Ejaaz: right now, where you can take this model and let's say you operate like one
10:58
Ejaaz: of the world's top trading hedge funds, right?
11:01
Ejaaz: You could feed it data. You could run it locally, by the way,
11:04
Ejaaz: feed it data, all encrypted,
11:05
Ejaaz: train the model to understand how your team company operates,
11:09
Ejaaz: what your investment theses are, feed it maybe even proprietary information
11:13
Ejaaz: because, by the way, this model is completely open-weight.
11:16
Ejaaz: So, you know, open sources come back to America. Well done, Mira Morati.
11:20
Ejaaz: And you can hyper-focus this model on your daily trading activities if you are
11:24
Ejaaz: in that hedge fund position.
11:25
Ejaaz: Now, if you don't believe me, there is a live example of this actually happening
11:29
Ejaaz: from none other than thinking machines themselves.
11:31
Ejaaz: They're working reportedly pretty closely with Bridgewater Capital,
11:35
Ejaaz: which is one of the biggest funds in the entire world using their Tinker API to fine-tune models.
11:41
Ejaaz: So the long story short is this model isn't meant to be remarkable on its own.
11:45
Ejaaz: It's meant to be taken by you, whether you're a startup founder or a hobbyist
11:48
Ejaaz: or whatever that might be, and fine-tune it to do the specific tasks that you can do.
11:52
Ejaaz: If you end up doing that, you end up with a more powerful model that is less
11:56
Ejaaz: generalized and more specific than a classic anthropic or open AI model.
12:00
Ejaaz: And it is way, way cheaper than if you use any of these other generalized models.
12:05
Ejaaz: And as we've spoken about on yesterday's episode and a lot of episodes recently,
12:08
Ejaaz: actually, cheaper models might actually end up being the models that are used
12:12
Ejaaz: more than the intelligent models themselves.
12:15
Josh: Yeah, I think this is a really interesting angle. And I'm excited to see that
12:17
Josh: this is the way that they're going. They're not trying to compete with the frontier
12:20
Josh: labs. In fact, to use their own words, they said, it's not the strongest overall
12:24
Josh: model available today, open or closed. So they're very much aware that they are not on the frontier.
12:28
Josh: The interesting thing is addressing this common concern that we see talked about a lot.
12:32
Josh: If you remember, there's this viral clip on CNBC a couple weeks ago with Palantir CEO Alex Karp.
12:37
Josh: And he was describing how close frontier tools are too expensive.
12:41
Josh: And they have very somewhat murky IP protection where you kind of feed these
12:44
Josh: models. your contacts, and then perhaps they go off and use them to train new models.
12:48
Josh: And what Thinking Machines is doing is basically building this infrastructure
12:51
Josh: for you to build and own the IP custom to you. And I think this is the interesting
12:56
Josh: thing where when you think about monetization for a company like this, they raised $2 billion.
12:59
Josh: Now they're just giving away their technology for free. That's the reality to
13:02
Josh: an extent because this model is open source. You can actually go and download
13:06
Josh: it and use it for your own, but they have this platform called Tinker.
13:09
Josh: That's meant to hyper optimize these models specific for your use case with
13:12
Josh: a much less amount of data than traditional models would.
13:16
Josh: So traditionally, these models need like terabytes of data in order to go through
13:20
Josh: this reinforcement learning process and to become hyper specifically good at a certain task.
13:24
Josh: This model and using this Tinker product, you're actually able to feed it a
13:27
Josh: lot less context to get a very specific outcome.
13:30
Josh: So if you want to train it to be a financial wizard, like Blackwater was perhaps
13:34
Josh: doing, or if you want to train to be a professional private chef,
13:37
Josh: if that's what you're doing.
13:38
Josh: It requires a lot less context and you could kind of mold it to the form that
13:42
Josh: you want specific to you and i think that's a really powerful thing this is
13:45
Josh: mostly focused on consumer and enterprise but very specific use cases for it and i think this is a,
13:52
Josh: novel angle in the sense that a lot of the frontier labs their general purpose
13:55
Josh: you go to them with any question you want and they will answer to the best of
13:58
Josh: their abilities generally better than everyone else but this lab will allow
14:01
Josh: you to do a very specific answer
14:03
Josh: with a very low cost using open weight models without any potential leaking
14:08
Josh: of any of your intellectual property.
14:09
Josh: And I think that's a pretty novel use case and really cool to see out of the
14:12
Josh: Thinking Machines team.
14:13
Ejaaz: I think an obvious question to come from this as well is how were they able
14:17
Ejaaz: to achieve this? Well, I'm pleased to say that I think the Chinese are getting
14:22
Ejaaz: a little taste of their own medicine.
14:24
Ejaaz: What I'm showing on the screen here is in order to train this model,
14:27
Ejaaz: thinking machines used DeepSeq's V3 architecture.
14:31
Ejaaz: Now, if you've heard of DeepSeq before, it comes from an AI lab in China,
14:35
Ejaaz: funnily enough, a hedge fund that built their Chinese model called DeepSeq.
14:40
Ejaaz: And it uses this architectural mixture of experts. I'm not going to bore you
14:43
Ejaaz: with the details, but basically, you can have a really large model and only
14:47
Ejaaz: utilize the necessary parts of the model when you're answering a prompt.
14:51
Ejaaz: It's way more efficient.
14:52
Ejaaz: It's way more cheaper. And it allows you to customize it way more.
14:54
Ejaaz: That's how they were able to pull this off.
14:56
Ejaaz: And it's kind of nice to see this reverse distillation. And I say that justingly
15:00
Ejaaz: because, obviously, all the Chinese models are open source.
15:02
Ejaaz: But moving on, Miro Morati is my favorite topic.
15:07
Ejaaz: Mira Marathi's former boss has been up to some good and some no good.
15:12
Ejaaz: If you watched our previous episode earlier this week, you'll know that Sam's
15:16
Ejaaz: in a bit of hot water right now.
15:19
Ejaaz: The big boys at Apple are suing Sam and OpenAI for stealing hardware trade secrets
15:25
Ejaaz: to build their new AI hardware device.
15:29
Ejaaz: And a question on Josh and I's minds was, what on earth is this device?
15:33
Ejaaz: Josh, Maybe we have some light now. Can you give us some details?
15:35
Josh: Oh my God. How many times have we spoken about this on the show?
15:38
Josh: We got to be approaching like double digits now. This is probably the 10th time
15:41
Josh: we've talked about this freaking device.
15:42
Josh: We still don't know what it looks like. I am so desperate to know.
15:44
Josh: I've never wanted to hold a product in my hand so bad in my entire life than
15:48
Josh: I do want to hold this Johnny Ive OpenAI collab device.
15:53
Josh: And it of course has not come without its troubles. They are currently experiencing
15:57
Josh: lawsuits. I'm wondering if those lawsuits have impacted the trajectory of this
16:00
Josh: suite and family of products that they're going to launch.
16:02
Josh: But we do seem to have somewhat of a close confirmation to what the very first
16:07
Josh: device in this family of devices will look like and when we'll actually be able
16:11
Josh: to get our hands on it now mark german who
16:13
Josh: some people might know as the apple leaker he is kind of the guy who you go
16:17
Josh: to when you want to know about leaks he has very reliable sources he is generally
16:21
Josh: right when he publishes these things for his publication i think he works with
16:24
Josh: bloomberg i'm pretty sure that's where he publishes all of these and
16:28
Josh: we have an idea of what this looks like and drumroll please it is coming in a form of a speaker.
16:33
Ejaaz: Which is interesting.
16:35
Josh: It is described as a screen-free smart speaker that will act as a new type of
16:41
Josh: home computer for the AI era.
16:43
Josh: So it's screenless, which is important to note, but it's certainly not senseless.
16:47
Josh: The idea is that it will have a camera plus these environmental sensors, probably like light.
16:51
Josh: It'll have microphones. It'll have a rechargeable battery, which I found interesting
16:55
Josh: because that means that this will be a portable device.
16:58
Josh: And then, Ijaz, I was asking you about this yesterday, is, does that mean that
17:02
Josh: it's going to be something that fits in my pocket? Like, is this the puck?
17:06
Josh: Or is this going to be, I think, of the Amazon Echo, like that little...
17:10
Josh: Little block that sits on your desk and you could kind of carry it from room
17:13
Josh: to room. I'm kind of now hypothesizing of what this is going to look like.
17:16
Ejaaz: Well, the secret is revealed in this tweet from Mark Gurman and none themselves.
17:22
Ejaaz: He says, as one of his first descriptors of this device, that it's a mobile device.
17:27
Ejaaz: And the idea, if you open up this article and read through it,
17:30
Ejaaz: is he talks about this device being somewhat portable.
17:32
Ejaaz: It's meant to live in your home. It's meant to ingest and listen to everything
17:36
Ejaaz: that is said, the conversations that you have, you can also speak to it.
17:39
Ejaaz: But the goal, the stated goal of what OpenAIR is trying to build here is,
17:44
Ejaaz: A companion, a companion that can sit by your side, that you can speak to,
17:48
Ejaaz: that sounds very human, that understands the work or tasks or goals,
17:51
Ejaaz: whatever you're trying to achieve for that day and over a set number of days,
17:55
Ejaaz: and it can work productively with you.
17:57
Ejaaz: Now, a big question for me is, how different is this from just opening up the
18:03
Ejaaz: ChatGPT app on my phone and speaking to it?
18:05
Ejaaz: And I think I've answered my own question, which is, it is quite different.
18:08
Ejaaz: There is a big delta between having a device that is omnipresent,
18:14
Ejaaz: that is ambient, that listens to everything you say, that can chime in when
18:17
Ejaaz: necessary, like a real human does, like a companion or an advisor or a mentor does,
18:22
Ejaaz: versus having something that is a little more archaic, manual,
18:25
Ejaaz: that you have to trigger when you think of it.
18:28
Ejaaz: And I think this kind of like jump going from kind of like manual devices where
18:34
Ejaaz: you have to click buttons, swipe and go to the specific app versus having this
18:38
Ejaaz: omnipresent artificial in terms.
18:41
Ejaaz: It's funny. It's like I'm describing like the Terminator plot right now.
18:43
Ejaaz: And I'm listening to myself and I'm like, oh, God, like maybe this isn't good for us.
18:47
Ejaaz: But I get that that's where we are moving to. Now, of course,
18:50
Ejaaz: all of this is really exciting to see.
18:52
Ejaaz: But if there is a massive Apple lawsuit looming over their heads,
18:56
Ejaaz: there are grounds for this potentially being delayed. Now, the target date is
19:01
Ejaaz: early 2027, specifically February. Sam wants to ship around 100 million units initially.
19:07
Ejaaz: Josh and I are probably going to buy up 50 million of those units.
19:10
Ejaaz: But the point is that this is going to be really in demand, but it might get
19:14
Ejaaz: delayed to later on in the year if Apple has grounds for them stealing secrets.
19:18
Josh: I really hope not. The reports look like it's going to come out at around $200 to $300.
19:23
Josh: And you have to imagine if they're building a hundred million of these,
19:25
Josh: a couple hundred million of these, it's going to be a fairly simple design.
19:28
Josh: There won't be a ton of moving parts. There won't be a lot of moving pieces.
19:31
Josh: This will kind of be a solid machine is my assumption.
19:35
Josh: If I know anything from my obsession of Johnny AI, his design.
19:39
Josh: Entire goal is to reverse the damage that has been done through the smartphone era in the sense that
19:45
Josh: smartphones have taken so much of our lives away from reality and they've placed
19:50
Josh: them into this digital world and i know he is on a mission hell-bent on reversing that
19:55
Josh: and part of that is removing the screen but the second part of that is trying
19:58
Josh: to make the experience more human and i think we got an early release version
20:01
Josh: and look at this with the new chat model from chat gpt
20:04
Josh: on how conversational it is how it feels very warm and friendly and i expect
20:08
Josh: this to very much be an extension of that
20:10
Josh: in a way it's kind of converging on the voice models and miramirati miramirati's
20:15
Josh: like tinkering model where it will kind of build up this context window about
20:18
Josh: you it will be able to be your personal assistant and allow you to remove yourself
20:23
Josh: away from the matrix that he has,
20:26
Josh: created such a large amount of through his hardware devices almost unintentionally
20:31
Josh: and to me that's super exciting and i cannot wait to get my hands on it in early
20:34
Josh: 2027 so we will be keeping our eyes peeled any more rumors you know we're going
20:39
Josh: to talk about it because we can't help but talk about it this is like the 10th
20:41
Josh: time but that is the update on the open ai device coming soon.
20:45
Ejaaz: I i will say that um apple is planning to at least announce their ai hardware
20:51
Ejaaz: devices plural by the end of this year so they might just be doing this whole
20:55
Ejaaz: lawsuit thing just to to front run them.
20:57
Ejaaz: And I know that Apple is working on, I believe, some kind of a disc pendant,
21:02
Ejaaz: AirPods with cameras, as well as some smart glasses. So I'm excited to see these
21:06
Ejaaz: two honestly go at it because the end benefit is us. We get to buy a bunch of
21:10
Ejaaz: cool gadgets and hopefully wear them every day.
21:13
Ejaaz: Moving on to the next topic.
21:16
Ejaaz: Elon Musk, surprise, surprise, is in a bit of trouble.
21:20
Ejaaz: Now, he has this product called Grok Build, which is basically their Claude Code competitor.
21:26
Ejaaz: It is Grok used specifically for coding and building different apps.
21:30
Ejaaz: And some people leaked that XAI's Grok Build CLI, so their API basically,
21:36
Ejaaz: or their front-end interface, was taking people's code,
21:40
Ejaaz: uploading it to private servers for SpaceX AI and then using that data to train new models.
21:46
Ejaaz: Now, if you've ever been in this model game for a while, personal data,
21:50
Ejaaz: privacy, and encryption are incredibly important.
21:53
Ejaaz: Anthropic and Open AI take this very, very seriously because,
21:55
Ejaaz: of course, you want to make sure that they're not taking your personal information
21:59
Ejaaz: that you're imparting to these AI models and using it for nefarious ways or
22:02
Ejaaz: taking advantage of them.
22:05
Ejaaz: Grok, or rather SpaceX AI and Elon were caught doing this And they had to do
22:10
Ejaaz: a lot of damage control, which ended up being, a few days later,
22:14
Ejaaz: them open sourcing the entire Grok build.
22:17
Ejaaz: They apologized and basically said, zero data retention policies are super important
22:22
Ejaaz: where we don't retain any of the data. And if we do, we will be upfront.
22:25
Ejaaz: You can turn this off in settings, but it sounded like it was on by default,
22:29
Ejaaz: which is the warning part.
22:30
Ejaaz: But in order to retaliate, they've said, we've open sourced this entire product,
22:34
Ejaaz: take it, fine tune it, do whatever you want with it, and see the code yourself
22:38
Ejaaz: and see that we're not taking any of your stuff.
22:42
Ejaaz: It's a bit of a roller coaster for them. I'm glad they're being able to do damage
22:45
Ejaaz: control, especially as like Elon is like hurling insults to his enemy at Sam
22:49
Ejaaz: and OpenAI, which arguably haven't done this yet.
22:53
Josh: I respect it. It feels like the early days of Facebook, the move fast and break
22:55
Josh: things where they're just going balls to the wall, full speed.
22:59
Josh: If they make a mistake, they correct it. And in fact, I feel like they really
23:01
Josh: even over-indexed on correcting it.
23:02
Josh: They open source everything and they had zero data retention policies.
23:06
Josh: They really went like fully locked down, almost Apple status,
23:09
Josh: where now there is zero data that is going to cross the line.
23:12
Josh: I think this is going to prove to be a very unique advantage.
23:15
Josh: When you think about the Apple edge, a lot of it has to do with guaranteed privacy of data.
23:19
Josh: And if Grok is going to be the only person who is going to provide that guarantee,
23:24
Josh: that is a huge benefit because remember, they have the cost per token down about
23:28
Josh: as low as it gets. They're on the frontier in terms of cost per intelligence.
23:31
Josh: And now if they have privacy baked in, that's a very compelling product that
23:34
Josh: I think is noteworthy to watch.
23:35
Josh: There is a second bit of news about elon and it is about elon not about his
23:40
Josh: company spacex not about tesla
23:42
Josh: but about elon musk the individual who just acquired a company himself for an
23:47
Josh: estimated one billion dollars which i found interesting i was like why is
23:51
Josh: elon the person acquiring this company
23:53
Josh: and i mean surprise surprise it's an energy company named apr energy.
23:57
Josh: And they are a Jacksonville fleet of trailer-mounted mobile gas and diesel turbines.
24:02
Josh: Why on earth would he want to buy this?
24:05
Josh: Well, it's because they have a gigawatt plus of capacity. And to buy a gigawatt
24:09
Josh: of capacity for a billion dollars, it turns out that's a pretty good deal.
24:13
Josh: And as we know, Grok has these Memphis data centers. They have Colossus 1,
24:18
Josh: they have Colossus 2, they need to spin them up.
24:20
Josh: And one of those big constraints has been power. And we've seen how much in-demand
24:25
Josh: power has been in the ability to deploy quickly through companies like Bloom
24:29
Josh: Energy. We've talked about Bloom Energy a lot on the show.
24:31
Josh: Their stock has gone basically vertical recently because they are one of the
24:34
Josh: few people who can modularly deploy energy to these data centers quickly.
24:38
Josh: This company that Elon bought now is able to stand up these projects in about
24:42
Josh: 15 to 30 days, which is really exciting because these grid interconnection queues
24:47
Josh: to tap into the mainline grid normally take a couple of years.
24:51
Josh: So this really shrinks that time to get energy to the data center by like a
24:55
Josh: order of magnitude and i think that's a really big deal so this is exciting
25:00
Josh: it was kind of under the radar it didn't get published
25:02
Josh: anywhere it was just kind of discovered through these filings that elon had
25:05
Josh: made the purchase and i think it's a testament to how fast they are moving they
25:08
Josh: need power elon's going out he said i'll get you a gigawatt myself he bought
25:12
Josh: it they're going to deploy it in 15 to 30 days they're going to bring way more gpus online
25:16
Josh: and that's a testament to the velocity in spacex ai so again really bullish
25:20
Josh: news like keep an eye out on them they're they're making their way up to the frontier very quickly.
25:25
Ejaaz: The problem Elon's solving with this is he has way too many GPUs to the tune
25:30
Ejaaz: of like 1 to 1.5 million of all the latest NVIDIA GPUs, but he has no means
25:35
Ejaaz: of bringing them online.
25:36
Ejaaz: The actual wait time to get your GPUs hooked up to the energy grid is around five years right now.
25:43
Ejaaz: And New York City, or rather New York State, this week became the first state
25:48
Ejaaz: to pause or ban data center development and it's following a trend of a few other states.
25:53
Ejaaz: If this continues to happen, workarounds like this, like acquiring companies
25:58
Ejaaz: like this to get portable gas turbines and powering your GPUs may be the only way to work around it.
26:04
Ejaaz: And the important part about this, why is this important at all,
26:06
Ejaaz: is you need compute at the end of the day to build the best models.
26:09
Ejaaz: We've seen Meta release MuseSpark 1.1, which isn't quite Frontier, but is pretty good.
26:14
Ejaaz: And you've seen Grok 4.5 be able to compete with the big boys unexpectedly.
26:19
Ejaaz: And the way they've been able to do this is because they've acquired so much
26:21
Ejaaz: compute, but none of your GPUs make sense. They're worthless pieces of metal
26:25
Ejaaz: if you can't power them up in the first place.
26:27
Ejaaz: One thing I will say is we've spoken about Bloom Energy a lot on this show.
26:32
Ejaaz: Leopold Aschenbrenner, who runs one of the biggest AI funds,
26:35
Ejaaz: 24-year-old kid, has made a big investment in Bloom Energy.
26:39
Ejaaz: This is their direct competitor. And Elon just went out and looked at Bloom
26:43
Ejaaz: Energy and thought, I don't want to have to deal with this. I'll just acquire
26:45
Ejaaz: the company themselves and integrate them into what I'm doing here at SpaceX AI.
26:49
Ejaaz: And I think it's an incredibly smart move. Now, moving on.
26:53
Ejaaz: Yeah, there is a popular product that I believe both of us like to use quite
26:58
Ejaaz: a lot called Whisperflow.
27:00
Ejaaz: And the way that that works is instead of typing prompts, it takes a lot of
27:04
Ejaaz: time, you can tap a button, hold it down on your laptop or on your computer.
27:09
Ejaaz: And you can speak into a microphone or speak to your computer and all your words are recorded.
27:15
Ejaaz: And it has been a major unlock for me. Voice to AI is just the way forwards
27:19
Ejaaz: and I honestly can't see anywhere else where it doesn't result in that.
27:23
Ejaaz: Now, this guy, Farza, released a new product from his company that is called
27:28
Ejaaz: Screen Aware Dictation.
27:30
Ejaaz: Now, one problem that I have when I hold down the button in Whisperflow and
27:34
Ejaaz: I speak into it and it records exactly everything that I say is,
27:38
Ejaaz: I'm also thinking in real time.
27:40
Ejaaz: I'm like, oh, I want to say something in a tone that is kind but not too assertive.
27:47
Ejaaz: And I want to talk about this topic. I don't know the definition of this topic,
27:51
Ejaaz: but I need to do some research on that. And if I was using Whisperflow,
27:55
Ejaaz: it would record everything I say verbatim and put it in that email.
27:59
Ejaaz: With this new feature, it listens to you, it understands when you don't understand
28:03
Ejaaz: something, it goes away, does the research, comes back and formulates the reply
28:08
Ejaaz: or the response that you want.
28:09
Ejaaz: And it's this massive jump up where you now have this feature which kind of
28:13
Ejaaz: like simultaneously listens to what you want to say exactly,
28:17
Ejaaz: also understands when you don't want to say something or what you don't know,
28:21
Ejaaz: goes away, does the research, pulls all the context from your personal files
28:24
Ejaaz: that it's connected to, and then comes up with the perfect response.
28:27
Ejaaz: You can literally say, hey, I want to create a Google slide deck that is to
28:31
Ejaaz: do with this document that I have on another screen or another tab.
28:34
Ejaaz: It'll go away, research that, come back, and put that into a very well-presented
28:39
Ejaaz: slide deck. So it's a really amazing product or feature. I've been using it a bit this morning.
28:44
Ejaaz: I think he got overloaded since he announced this literally under 25 uh 24 hours
28:49
Ejaaz: ago and it's got like 3 million views but um i don't know i just want to highlight
28:52
Ejaaz: this feature because i think it's it's really cool
28:54
Josh: Cool yeah i'm excited to give it a shot this feels like something that like
28:56
Josh: open ai can release tomorrow and kind of
28:59
Josh: knock off the the grid and call it a day i think this is a
29:02
Josh: thematically interesting project though because again everyone is really trying
29:06
Josh: to collect the context if you think about what thinking machines released they
29:09
Josh: are just collecting context so that they can give you a custom tailored experience
29:13
Josh: the open ai speaker that is probably going to be coming out its job is just
29:16
Josh: to collect context about you and to then customize your experience same thing
29:20
Josh: with this so much of the ai story now is kind of pivoting to context because
29:24
Josh: the more an ai understands about you understands about your preferences the
29:27
Josh: better it can serve you and this is another interesting example of it i saw
29:30
Josh: there's a few other kind of rapid fire points that we can get through the next
29:33
Josh: one being deep seek is preparing an ipo which is kind of crazy
29:37
Josh: um that's going to be interesting to talk about,
29:40
Josh: are there any interesting talking points with this ipo or is this kind of a
29:43
Josh: rumor you guys do you know.
29:44
Ejaaz: Um it's it's kind of an unofficial confirmed thing at this point a lot of the
29:49
Ejaaz: main investors have kind of been like referring to this ipo and referring to
29:53
Ejaaz: this new raise the reason why this is like compelling news is
29:56
Ejaaz: they literally raised i think 15 billion dollars a week and a half ago
30:01
Ejaaz: so as soon as that round closed they've raised yet another round that values them at like almost
30:07
Ejaaz: one and a half X higher than they were before.
30:09
Ejaaz: And now they're touting the IPO rumor, which you know, they're going to join
30:13
Ejaaz: cheapu and a bunch of other Chinese AI companies in that public race.
30:17
Ejaaz: Another interesting fact is the founder
30:21
Ejaaz: Liang Wenfeng of DeepSeek owns around, I think it's 94% of the company,
30:26
Ejaaz: which means that his net worth is roughly around $26 billion,
30:28
Ejaaz: making him one of the most wealthiest AI CEOs out there.
30:32
Ejaaz: This random guy from China who was a hedge quant fund trader literally about a year and a half ago.
30:37
Josh: It's insane. Good for him, man. He's doing well. And then with that,
30:41
Josh: I think that is pretty much all the news for this week. That's it.
30:44
Josh: Yeah, we got it all covered.
30:45
Josh: It's been an eventful week. There's a lot going on.
30:48
Josh: It's amazing. I'm always surprised at how every week we can sit down here and
30:51
Josh: somehow manage to find so many topics that are interesting to talk about
30:54
Josh: the just frontier and the world of ai and technology is moving so rapidly so
30:59
Josh: if you've made it this far through the episode thank you for watching if you've
31:01
Josh: listened to the three prior episodes this week which i encourage you to do so
31:05
Josh: we had some pretty interesting ones
31:06
Josh: one was about the gpt 5.6 release the other one was about open ai getting sued
31:11
Josh: by apple and the third being about um why the price war actually matters and
31:16
Josh: we referenced that episode quite a bit about how the new frontier is kind of
31:19
Josh: being battled on price instead of intelligence
31:22
Josh: so those were all really great episodes if you enjoy this one
31:24
Josh: please do not forget to share with your friends and if you did make it through
31:27
Josh: all of that congratulations you're caught up you can go touch grass enjoy your weekend
31:31
Josh: and yeah each has any final parting thoughts before we head out for the weekend.
31:34
Ejaaz: I want to know what people are doing on the weekend. Are you using any of these models?
31:40
Ejaaz: What are we tinkering with? You know, pun very much intended.
31:43
Ejaaz: We, I think 90% of our content this week has been about new models.
31:47
Ejaaz: There's OpenAI releasing literally three different models of three different modes.
31:51
Ejaaz: We've got all these new open source models and they're super cheap to use.
31:55
Ejaaz: So I'm wondering what you folks listening to this are tinkering with.
31:59
Ejaaz: What are you doing? What are you building? What are you doing with these models? Are they even useful?
32:02
Ejaaz: Are you still just kind of like ignoring all of this and using Fable 5 or GBT
32:06
Ejaaz: 5.6, let us know in the comments, DM us.
32:08
Ejaaz: And the final thing I'll say is we are on the lookout for sponsors.
32:12
Ejaaz: Limitless has spun out. We are officially on our own. We're keeping the lights on through our own.
32:16
Ejaaz: And we would love to partner with someone that has a product or service that
32:20
Ejaaz: they feel incredibly passionate about and want to share with our audience of
32:24
Ejaaz: users who love to use these different things. So if that sounds like you,
32:27
Ejaaz: or if there's a friend that sounds like that, please reach out to them, reach out to us.
32:30
Ejaaz: Our email is in the description below. You can DM us on X or wherever we are.
32:34
Ejaaz: And yeah, we'll see you on the next one. You're all caught up. Yeah.
32:37
Josh: Awesome. Well, have a great weekend, everyone. See you next time.