Latest / In the Loop AI / THIS WEEK IN AI: Grok is Back, ChatGPT 5.6, Meta's New Models, China's Threats

Transcript

THIS WEEK IN AI: Grok is Back, ChatGPT 5.6, Meta's New Models, China's Threats · In the Loop AI

0:00
Ejaaz: I've been waiting so long to say this. Elon is officially back.
0:04
Ejaaz: The first major AI model released from SpaceX AI since they went public,
0:08
Ejaaz: since the $60 billion cursor acquisition is officially live.
0:12
Ejaaz: It's called Grok 4.5 and it's pretty damn good.
0:15
Ejaaz: It ranks around three after Fable 5 and GPT 5.5, but it's really good at coding
0:21
Ejaaz: and it's astonishingly cheap. It's around half the cost of Opus 4.8 and it's
0:26
Ejaaz: their quickest model yet.
0:28
Ejaaz: Elon only got to enjoy this for about a day or maybe even a couple of hours
0:33
Ejaaz: because GPT 5.6 was officially released from OpenAI. And if you're listening
0:37
Ejaaz: to this, you probably have access to it right now, a general release.
0:40
Ejaaz: It's their best model yet, extremely good energetic reasoning,
0:43
Ejaaz: but not quite fable worthy. But they are training GPT 6, which is going to be
0:49
Ejaaz: released supposedly a month from now. If you're releasing models,
0:52
Ejaaz: you're probably thinking that's amazing for the AI world.
0:54
Ejaaz: But this comes right at a time where our main adversary in China has been releasing
0:59
Ejaaz: so many models that are open source and everyone gets to use it.
1:01
Ejaaz: They're super cheap, but they might be considering cutting off the West from using their models.
1:06
Ejaaz: Meanwhile, Meta, who spent around $25 billion trying to build the best model
1:11
Ejaaz: and have seemingly failed, just released two new models there,
1:14
Ejaaz: Metamuse Image and Metamuse Video Models. We're going to get into all of this
1:17
Ejaaz: and more on today's episode.
1:19
Josh: There's a lot going on. It feels like nothing happens for a week or two,
1:22
Josh: and then five different releases happen all in one day. And that's what just
1:25
Josh: happened. We got new JATGPT. We have a new voice interface, which I love that we'll get into.
1:30
Josh: But I think we do have to start with Grok because I mean, like you said, Elon is back.
1:33
Josh: The SpaceX AI team, which is now officially formalized, they have a logo,
1:37
Josh: they have their official Twitter handle on X.
1:40
Josh: This is their first release in a long time and the results are pretty remarkable.
1:44
Josh: They're kind of classifying this as Opus class on a relative basis compared to Claude Opus 4.8.
1:49
Josh: And it seems to have actually accomplished that on just about all of the benchmarks.
1:52
Josh: In fact, it even exceeds it slightly in some benchmarks for a fraction of the
1:57
Josh: cost. And I think this is one of the more interesting things about this model
2:00
Josh: is just how efficient it is for how much intelligence you get out.
2:03
Josh: And I know that the XAI team or the SpaceX AI team, they've always been really
2:07
Josh: insistent on maximizing the.
2:09
Josh: Amount of intelligence per watt of compute. And it seems like this is the first
2:13
Josh: sign of that actually working.
2:15
Josh: For the same amount of power and electricity, they're able to generate tokens
2:18
Josh: that have more intelligence per token than just about anyone else.
2:21
Josh: Like you mentioned, Ejaz, the new Grok model is being trained.
2:25
Josh: It is currently cooking in the oven. They have all the GPUs humming.
2:28
Josh: And when that comes out, I really believe that they're going to have a shot at the top.
2:32
Josh: And we talked about this way in the past, maybe like two or three months ago, about how the spacex ai
2:38
Josh: opportunities really just in their ability to build out infrastructure quickly
2:43
Josh: and as a result they actually had too much infrastructure for their use cases
2:47
Josh: they were selling it off to amazon they were selling it off to anthropic because
2:50
Josh: they just had access and it created this huge business model for them
2:52
Josh: but now they actually have a use case for this they're starting to take these
2:55
Josh: gpus they're starting to train the models and the results are actually great
2:58
Josh: and it's like okay wow they were they're acquired for a couple months but now
3:01
Josh: here they are right back on the frontier and you have to imagine
3:04
Josh: that the next model that's being trained on however many trillions of parameters
3:07
Josh: is going to be right up there at least very close to the frontier for what i
3:10
Josh: would imagine is a very very competitive price.
3:13
Ejaaz: I think i think the most hilarious part about this is this is probably elon's
3:16
Ejaaz: most humble model launch ever he was uh super up front saying you know this
3:20
Ejaaz: is opus 4.8 class grade model it's not as good as fable but we are training
3:25
Ejaaz: a much larger model speaking of size
3:27
Ejaaz: The size of this model, GROC 4.5, is only around 1.4 trillion parameter models.
3:32
Ejaaz: Now, if you compare that to Fable 5, that's around 10 to 15 trillion parameters in size.
3:37
Ejaaz: So the fact that they were able to achieve number three on some of these coding
3:41
Ejaaz: benchmarks, like the one that we're showing on your screen right now,
3:43
Ejaaz: DeepSwee Bench, is pretty impressive. And you can only imagine what a model,
3:47
Ejaaz: maybe five times the size of this, could probably be quite competitive.
3:51
Ejaaz: And this brings me to the absolute Hail Mary move that Elon and SpaceX AI pulled
3:56
Ejaaz: off a few months ago, or actually less than a month ago, which was the $60 billion acquisition of Cursor.
4:02
Ejaaz: Now, if you're wondering... Did that pay off? Yeah, it paid off massively.
4:04
Ejaaz: Like, just to bring people up to speed here, at the time of their public IPO,
4:11
Ejaaz: It was pretty much a foregone conclusion that SpaceX had fallen behind in the race.
4:15
Ejaaz: They hadn't released a meaningful model. Grok 4.2 kind of sucked.
4:19
Ejaaz: And OpenAI and Anthropic seem to be just running away with it.
4:22
Ejaaz: With 4.5, they have officially caught up and it's because they've been using
4:27
Ejaaz: this moat, this data moat that Cursor has.
4:29
Ejaaz: For those of you who don't know, for those of you who haven't used Cursor,
4:32
Ejaaz: Cursor is this amazing platform that kind of like pulls prompts across many
4:36
Ejaaz: different models and routes it to the right model at the right time to do the
4:39
Ejaaz: right types of tasks. It's also extremely cost efficient.
4:43
Ejaaz: Now that kind of data is really valuable towards training new models.
4:46
Ejaaz: It's not just about how much compute you put into train a model.
4:49
Ejaaz: It is about how you figure out how that model thinks. And Cursor's data mode
4:53
Ejaaz: is really useful for that. So that $60 billion acquisition seems to have paid off pretty damn well.
4:59
Ejaaz: Now, if you're wondering whether I should switch to GROC 4.5,
5:02
Ejaaz: if you're using your LLMs to do Google searches and do basic research, I would say no.
5:08
Ejaaz: But if you are a software engineer, if you're trying to build a new app,
5:11
Ejaaz: I would say absolutely yes, give it a go, because this model was designed to build.
5:16
Ejaaz: Some of the demos, which we'll show shortly, has been pretty awesome.
5:20
Ejaaz: It builds apps from scratch, builds simulators from scratch,
5:22
Ejaaz: and it's super cheap and fast to use.
5:27
Ejaaz: I it's around 17 times cheaper than opus 4.8 puttas it's pretty awesome
5:32
Josh: Yeah i was looking at the the sw bench um benchmarks that they offer and it
5:37
Josh: showed that the average is solved in about 16 000 tokens versus opus which is
5:41
Josh: solved in 67 000 tokens so it's about a 4.2 times gap
5:46
Josh: on efficiency which is like really impressive if you're paying for these tokens
5:50
Josh: and you want something more efficient this is a model at least to try to put
5:53
Josh: it in your toolkit to test it out to see where it walls.
5:55
Josh: And hey, if you don't like this one, there's some better news for you because Elon says that,
6:00
Josh: spacex ai will ship a completely new from scratch foundation model every single
6:05
Josh: month through the end of 2026 which is
6:08
Josh: pretty unbelievable like that's one of those where i'll believe it when i see
6:10
Josh: it because like a foundation model from scratch to build every single month
6:14
Josh: seems like a lot they're gonna have.
6:16
Ejaaz: To be training a lot of these in parallel he has around 500 000 gpus which
6:19
Josh: Is so crazy it is the largest cluster.
6:21
Ejaaz: But he he has enough gpus to feasibly train these models i think it's it might be true yeah
6:26
Josh: Yeah so i mean if you think of it he's got this giant bakery and he's got like
6:29
Josh: a couple different ovens going at once all cooking various size
6:32
Josh: desserts those desserts coming in the form of grok so i'm really excited that
6:36
Josh: they're back i'm excited to see how they evolve over the next few months it
6:39
Josh: seems like now that they've launched the first one
6:41
Josh: the next ones are going to come in pretty rapid iteration and we will have our
6:44
Josh: finger on the pulse but that is not the only model launch in fact that is certainly
6:48
Josh: not the best model launch we had this week because,
6:50
Josh: gpt 5.6 is here finally it was kind of here it was here last week or two weeks
6:55
Josh: ago whenever they were released it but it unfortunately got a little blocker from the government,
6:59
Josh: they've cleared things up with the government they said okay your model is good
7:02
Josh: to go so now it is good to go and if you were listening to this podcast,
7:05
Josh: chances are you can go and use it right now we are a little early to the news
7:09
Josh: it hasn't quite released publicly for the world at the time of recording so
7:11
Josh: we don't have any interesting demos
7:13
Josh: we will probably record a lot of cool demos showcasing all the capabilities
7:17
Josh: model at some point in the near future but for now we have the news that it's
7:20
Josh: out and this is all three they have gpt 5.6 soul terra and luna all broadly
7:25
Josh: available, and they are very much at the frontier.
7:28
Josh: Their costs are, I think, $5 per million input, $30 per million output,
7:33
Josh: so cheaper than Fable, and seems like it's a pretty good model.
7:38
Josh: The results on X seem like people are really enjoying this so far.
7:42
Ejaaz: Yeah, so I was scouring across all the early tested reviews from Theo,
7:47
Ejaaz: from a bunch of other folks.
7:48
Ejaaz: And one thing stood out pretty massively to me, which is OpenAI seems to have
7:53
Ejaaz: solved the toughest problem that they've or the toughest critique that they've
7:56
Ejaaz: been facing, which is taste.
7:58
Ejaaz: So when you use ChatGPT, I don't know if you've used it recently,
8:01
Ejaaz: Josh, it still seems pretty sycophantic. It still sounds like I'm talking to
8:06
Ejaaz: an AI bot versus a human that can actually help me or like a companion that
8:10
Ejaaz: can like help me complete my task.
8:12
Ejaaz: With GPT 5.6, it seems that the feedback is overwhelmingly in favor of taste. It sounds better to use.
8:20
Ejaaz: It's more of a friend to kind of consult with.
8:22
Ejaaz: And that has swayed people to switch from Fable 5 to GPT 5.6.
8:28
Ejaaz: Now, it's important to mention that it isn't really a direct competitor to GPT
8:33
Ejaaz: 5.6, according to these reviews.
8:34
Ejaaz: Apparently, it still lacks when it comes to certain complex tasks at coding,
8:40
Ejaaz: when it comes to certain complex tasks at reasoning, but it is much, much better.
8:44
Ejaaz: It's a big jump from GPT 5.5. Now, if you're wondering what the fable competitor
8:48
Ejaaz: from OpenAI is, it's called GPT 6.
8:51
Ejaaz: And according to Greg Brockman, the president of OpenAI, he said that it's aiming
8:55
Ejaaz: to come out in under a month, potentially.
8:58
Ejaaz: So again, we're looking at a new paradigm, a new era right now where the top
9:04
Ejaaz: model labs are basically releasing a new model every single month.
9:09
Ejaaz: Now, if you compare that to what that looked like,
9:12
Ejaaz: I don't know, a year ago, we had to wait months. We had to wait months.
9:16
Ejaaz: There were draughts between the news cycles, if you remember, Josh.
9:19
Josh: I don't know how we did it. It was so long between each model launch.
9:22
Josh: It was literally multiple months.
9:24
Ejaaz: It's so funny because I remember we had a lot of time to test these different
9:27
Ejaaz: models. Remember we attributed a personality to each of these different models?
9:31
Josh: Now they're kind.
9:32
Ejaaz: Of converging across this one uniform kind of face and voice,
9:37
Ejaaz: and they're kind of flitting between each other.
9:39
Ejaaz: I honestly am It is the best time alive if you are a builder,
9:43
Ejaaz: if you are someone that's interested in using these models and testing them
9:47
Ejaaz: out, you now get access to all of this stuff for a $20 a month subscription.
9:50
Ejaaz: So I'm super excited to actually get my hands on this and use it.
9:53
Ejaaz: If you're listening to this right now, I probably already am doing this,
9:56
Ejaaz: but you can expect an episode as soon as next week to show you what you can do with it.
10:01
Josh: And this is also great news for people who aren't developers,
10:03
Josh: who aren't builders, people who just want to use the model.
10:05
Josh: Not only is it more intelligent but also you have a new voice model today as
10:10
Josh: well which is really exciting so the voice model is something that we did have
10:14
Josh: time to play around with i was testing it last night i was chatting with it and it's um
10:18
Josh: pretty good i love the launch video that they have we're showing it can we play
10:21
Josh: a clip yeah yeah please play some audio it's very funny.
10:24
Ejaaz: Let's see let's see
10:26
Ejaaz: Today, we are announcing the all-new chat GPT Voice, powered by GPT Live 1,
10:33
Ejaaz: a full duplex conversational partner that is the most powerful voice model ever built.
10:39
Josh: Can you simplify that?
10:41
Ejaaz: Sure can. Can you explain what a full duplex voice model is?
10:45
Ejaaz: Yeah. Imagine a normal call with a friend.
10:47
Ejaaz: You can listen and talk at the same time. That's full duplex.
10:52
Josh: There you go. And that's the news. It's like you can now listen and talk at
10:56
Josh: the same time and you could talk over this voice in a way that you never were able to before.
11:00
Josh: And I will say this isn't novel technology. GPT launched this kind of as like
11:04
Josh: a research preview for more developers a month or two ago where it's kind of
11:08
Josh: live intelligence. But now it's rolled into the actual application and it's awesome.
11:12
Josh: You can yell at it. You can argue at it. You could have it think about things
11:15
Josh: while you are talking about something else.
11:17
Josh: I was doing a little demo where I was asking it to do some research about something.
11:21
Josh: And then we continued the conversation in parallel while it was researching in the background.
11:24
Josh: And then once it had the correct answer, it came back and it said,
11:27
Josh: hey, by the way, that thing you asked me, here's the answer.
11:29
Josh: And it felt very natural in a way that I've never felt before.
11:32
Josh: And watching this demo and getting the opportunity to try it,
11:36
Josh: I noticed that there were some new visual elements as well.
11:39
Josh: So when I asked like, hey, what's the weather in New York City?
11:41
Josh: I wanted to see how fast I could pull that up. It was close to instant.
11:44
Josh: And then it shows this like really pretty card on screen that gives you just
11:47
Josh: brief overview of all the weather that's going on.
11:49
Josh: And I'm thinking to myself, this is the first version of whatever their new
11:54
Josh: operating system is going to look like for this hardware device.
11:57
Josh: We have all the rumors in the world about what it's going to look like,
11:59
Josh: what it's going to feel like, and we know that voice is going to be at the core of it.
12:02
Josh: So getting an opportunity to interface with a new voice, I could very clearly
12:06
Josh: see this working well with an AirPods-like device. It's like when I'm walking down the street,
12:10
Josh: You could chat with it very naturally. It talks back to you.
12:12
Josh: It has the context of everything around you. And it's really exciting to try.
12:16
Josh: So I would encourage everyone who has a ChatGPT account, give it a shot.
12:18
Josh: It's like pretty cool technology.
12:21
Ejaaz: This might be a hot take, but I think this is OpenAI's biggest release so far
12:24
Ejaaz: of this year. And it's because I think voice is just the way people are going
12:29
Ejaaz: to communicate with AI going forwards. I can't type as quick as I can think.
12:34
Ejaaz: And in fact, like, I can't remember the last time I typed towards my LLM.
12:38
Ejaaz: I use whisper AI or I use the microphone and I just kind of like speak at Claude
12:43
Ejaaz: or I speak at GPT and it comes back with very useful insights.
12:46
Ejaaz: It's a much more natural way of interacting.
12:49
Ejaaz: Now, if you listen to this and you're thinking,
12:51
Ejaaz: Why is this significant? Like I've been using voice the entire time.
12:55
Ejaaz: Well, I hate to inform you, but you might have been using a lesser lower grade model.
13:01
Ejaaz: Voice typically works on older models, not the cutting edge models.
13:06
Ejaaz: It takes a while to kind of like catch up and re-implement with like the GPT
13:09
Ejaaz: 5.5, GPT 5.6. For example, I doubt this is supported by 5.6.
13:14
Ejaaz: So you're already accessing a lower level of intelligence. With this new voice
13:18
Ejaaz: upgrade, you now get access to the higher set of things.
13:22
Ejaaz: It also finally implements a really useful feature that I enjoyed from Thinking
13:28
Ejaaz: Machine Labs. So Mira Murati, who was the former CTO of OpenAI,
13:33
Ejaaz: left to set up her own AI lab.
13:35
Ejaaz: She created this model called the interaction model under Thinking Machine Labs.
13:40
Ejaaz: And it does a lot of what is shown on this video right now, where you can kind of like interrupt it.
13:44
Ejaaz: You can kind of like talk to yourself and it knows it's not talking to you.
13:47
Ejaaz: Very natural human type interactions.
13:50
Ejaaz: We've finally got this feature in OpenAI. So now I'm thinking about her lab
13:53
Ejaaz: and what she's built and what her mode is.
13:55
Ejaaz: And I'm thinking, what's the point of using that when I can just do this on OpenAI?
13:58
Ejaaz: And again, it's this trend repeated constantly of these frontier AI labs that
14:04
Ejaaz: are basically looking over the fence, seeing what competitors are doing and
14:07
Ejaaz: saying, you know what, that's pretty cool.
14:09
Ejaaz: I'll get AI to build this in a couple of weeks and launch it ourselves. Well done from OpenAI.
14:14
Josh: Yeah, I like the point of it being the most important release it very much feels
14:16
Josh: like this is the critical release like this is the future way that we are going to interface with ai
14:21
Josh: and for them to really nail this and to make it a great product in the mobile
14:23
Josh: app is step one before it gets moved everywhere else and you could i wonder
14:27
Josh: i have a feeling that's probably why they used older people these are all grandmas they
14:30
Josh: three grandmas it's like okay as for everyone of all ages anyone could speak
14:34
Josh: anyone could have a conversation now you have an ai to converse with.
14:37
Ejaaz: My fiance's grandma uses this every day and I'm not joking because I looked
14:42
Ejaaz: at her tracking usage, three to four hours on average per day. She talks to it.
14:49
Ejaaz: As a friend, as a companion, as a consultant, as a gossip artist,
14:53
Ejaaz: when she's gossiping with her friends, it's insane.
14:57
Josh: That's unbelievable. People use this. That's high signal too,
14:59
Josh: coming from an older audience, because I don't think anyone's really tapped into that just yet.
15:03
Josh: So this is definitely a place to follow along closely. And my God,
15:07
Josh: we are halfway there to the end of the year when we are going to be getting
15:11
Josh: a preview of whatever this new hardware device is.
15:13
Josh: I am certain that it will be running some form of this voice mode.
15:17
Josh: Next up on the docket, Chinese AI labs are building their own chips while buying NVIDIAs.
15:23
Josh: Ejaz, as our resident Chinese expert, what the hell's going on?
15:29
Ejaaz: Hi, reporting from Beijing over here. Okay, so a narrative that we've consistently
15:34
Ejaaz: talked about on this show is China versus the US.
15:38
Ejaaz: If you are a new listener, China versus the US is probably the biggest air battle
15:42
Ejaaz: that's happening out there.
15:43
Ejaaz: And the biggest kind of question is, can China catch up to America's frontier
15:48
Ejaaz: of AI models? So far, they've been able to keep up pretty well,
15:51
Ejaaz: but they've lacked in one main thing, which is their chip architecture.
15:56
Ejaaz: They've had to buy and source and rely on NVIDIA chips.
16:00
Ejaaz: Until maybe the last couple of months where they announced that companies like
16:05
Ejaaz: Huawei are building their own AI chips and the government has mandated all Chinese
16:09
Ejaaz: AI labs to use Chinese chips.
16:12
Ejaaz: Okay, maybe there needs to be some kind of political resolve there.
16:15
Ejaaz: But now we got announcements for not one, but two AI labs, including the major
16:20
Ejaaz: AI lab from China, DeepSeek, that they're developing their own AI chip,
16:25
Ejaaz: according to a bunch of people that are familiar
16:27
Ejaaz: with the matter, in order to reduce its reliance on NVIDIA and Huawei chips alone.
16:32
Ejaaz: So even just relying less on Chinese manufacturers and focusing on their own AI chips.
16:37
Ejaaz: Now, if I had to guess why an AI lab would do this, it would be twofold.
16:41
Ejaaz: One, you don't want to rely on export controls and restrictions.
16:45
Ejaaz: You know, Trump or any other future US president might say, I don't want to
16:48
Ejaaz: sell you any more GPUs and your model is dead in the water.
16:51
Ejaaz: But number two, the most important one is, I think these AI labs are trying
16:55
Ejaaz: to create a vertically integrated stack. What I mean by that is it's not only
17:00
Ejaaz: good enough to create your own intelligent model right now.
17:02
Ejaaz: You kind of want to own the hardware and infrastructure that it runs on because
17:06
Ejaaz: then you can optimize it specifically for your model.
17:09
Ejaaz: So it could be cheaper, it could work much faster, and it could arguably be more intelligent.
17:14
Ejaaz: If you don't believe me, look at what OpenAI is doing. Well,
17:16
Ejaaz: look at what Meta is doing. They're working on their own chips.
17:18
Ejaaz: OpenAI is releasing their new jalapeno chip nine months from now.
17:22
Ejaaz: So if they end up pulling this off, I think we see a world, honestly,
17:25
Ejaaz: a year from now, where a lot of these labs create their own chips and they run
17:29
Ejaaz: it on their own infrastructure. Less reliance on NVIDIA.
17:32
Josh: I found it really interesting how they've been so quiet about which chips or
17:36
Josh: if they're even making chips at all.
17:37
Josh: And then within 24 hours or within even hours of each other,
17:40
Josh: we got the leaks from two places, both DeepSeek and Jipu, both building their own custom chips.
17:47
Josh: And I think it's this interesting coordinated effort that we're beginning to
17:50
Josh: see with this rotation, perhaps from open source to more closed source as they
17:56
Josh: feel like they're getting more powerful.
17:58
Josh: So as they get the capability to build these custom asic chips these custom inference chips and
18:04
Josh: i'll note that these aren't really pre-training chips these are custom built
18:07
Josh: accelerator chips similar to that episode that we filmed with etched about etched
18:11
Josh: uh the other day where these are very purpose-built chips built for optimizing
18:14
Josh: a specific stack and it's very clear that china is now in this race and they're trying to
18:19
Josh: get themselves as close to the frontier as possible and in a way they're basically
18:23
Josh: now just mirroring the u.s one-to-one it's like open ai has jalapeno.
18:27
Josh: Now deep seek is going to build their own custom asic chip and they're doing
18:30
Josh: it the same place they're probably using some of the same suppliers i have to
18:33
Josh: imagine at least the same memory supplier so they are now at war in terms of
18:37
Josh: supply chain with the united states so now traditionally the u.s versus china has been
18:42
Josh: fought on a software front this is the first instances in which there's probably some
18:46
Josh: going to be some serious competition on the hardware front at least as it relates
18:50
Josh: to supply chain and manufacturing these things because the capacity to do this
18:54
Josh: is so limited so really interesting developments coming out of china and we'll
18:56
Josh: see what they can build with these now we know the process is very long to go from
19:00
Josh: chip architecture to taping out a chip to building a chip and actually producing
19:05
Josh: it at scale then plugging it in
19:07
Josh: and then training a model and then publishing a model so that takes a very long
19:11
Josh: time it's going to be a while until we get the first models that have actually
19:14
Josh: been trained on these chips but the seeds are being planted now for pretty,
19:18
Josh: i would say impressive future from china if they continue on this path.
19:21
Ejaaz: I i personally think it's a it's a bigger problem than people are making out
19:24
Ejaaz: like i remember two years ago when people were like oh china would never catch
19:28
Ejaaz: up and back then it was true they were like three years behind
19:31
Ejaaz: now that gap is only really six months and if you can project into the future
19:35
Ejaaz: let's say that they create a chip that's maybe not nvidia level but it's like
19:38
Ejaaz: 80 to 90 percent nvidia level
19:41
Ejaaz: you could probably train a fable 5 or mythos 5 grade model from that and the
19:45
Ejaaz: chinese His Chibu founder has literally said, we'll have this ready in a few
19:50
Ejaaz: months, definitely before the end of the year.
19:52
Ejaaz: Now, remember, China isn't the US. There may not be any strict restrictions
19:58
Ejaaz: on this. They may use it nefariously.
20:00
Ejaaz: They may use it maliciously. And these models are pretty capable of some dangerous stuff.
20:04
Ejaaz: So I think one of the biggest missteps that we've made, at least on the Western
20:09
Ejaaz: side of things, is imposing these export controls because it's meant that they
20:13
Ejaaz: have been forced to create their own chip architecture. and
20:16
Ejaaz: that's what China relied on a lot. I think if they, if we could have just sold
20:19
Ejaaz: them a few more NVIDIA GPUs, we would have made them more reliant on the architecture,
20:24
Ejaaz: but now it's going to become a heavily siloed thing.
20:26
Ejaaz: And I think, like you said, we're going to lose access to these open source
20:28
Ejaaz: models. I bet you they're going to close source a bunch of the future meter-square models.
20:32
Josh: Yeah. Well, I have another company for you that used to be open source and they've gone close source.
20:37
Josh: So maybe we have a peek into the future because that company comes in the form of meta.
20:41
Josh: Remember them when they were so hell bent on creating open source models that were going to
20:45
Josh: change the world be accessible to everyone well it turns out that didn't happen
20:48
Josh: uh the same way that the metaverse didn't happen but we still call them meta
20:52
Josh: sorry i i feel bad for roasting them but oh my god it's so easy to every single
20:55
Josh: time but anyways meta is back in the news because they have done something interesting and novel in fact
21:00
Josh: um well i will say i didn't actually realize this was the news ej has filmed
21:03
Josh: me right before we recorded i was like oh wow they did release something this
21:06
Josh: week and that comes in the form of muse image and muse spark you'll remember the Muse models.
21:12
Josh: Oh, I feel so bad roasting them. But like the last experience we had with Muse
21:15
Josh: models, that was when everyone's Instagram accounts got hacked.
21:17
Josh: So they're back at it again with a new model. This one now focused on visuals and videos.
21:22
Josh: What is different about this model than the previous one? Because I remember
21:26
Josh: we had that like AI slop TikTok feed thing that kind of came and went from meta.
21:31
Josh: Is this kind of the next iteration on that model?
21:34
Ejaaz: It's not the next iteration, but it's the...
21:37
Ejaaz: It's the power infrastructure. It's the thing that will power all those future
21:41
Ejaaz: MetaBibes apps and stuff.
21:42
Ejaaz: So the quick TLDR is, this is their new image and video model,
21:48
Ejaaz: and it is a lot, lot better than their previous image and video models.
21:53
Ejaaz: Now, the bar admittedly was set pretty low back then.
21:57
Ejaaz: I would say this is where Grok Imagine currently is.
22:00
Ejaaz: When I look at this, when I look at Grok Imagine, they're relatively the kind
22:04
Ejaaz: of same grade of quality, which is impressive for Meta because this is a big
22:07
Ejaaz: jump from their previous model. Now, why I like this model specifically is how it works.
22:13
Ejaaz: There's been this recent trend across any LLM in OpenAI, Anthropic,
22:19
Ejaaz: with the recent CROC 4.5 model, which is called agentic reasoning or agentic thinking.
22:24
Ejaaz: What that means is they allow the model to, one, access more tools to do its
22:29
Ejaaz: thinking, and two, to think more. They give it more tokens. Meta has one of
22:33
Ejaaz: the largest data centers out there.
22:35
Ejaaz: So that's what they did with this model. So it's not necessarily an intelligent model,
22:40
Ejaaz: thinks more. It uses web search to look at potential images that might make
22:44
Ejaaz: sense for the request that you have.
22:46
Ejaaz: It then gets inspired from those images and it generates its own version and
22:50
Ejaaz: then it looks at it and it thinks, hmm, will Josh like this?
22:54
Ejaaz: If it's not good enough to Josh's expectations, I'm going to either edit this
22:57
Ejaaz: using my editing tools or I'm going to completely regenerate this from scratch.
23:01
Ejaaz: So you may have to wait a little bit longer, but you end up getting a much better output. And guess what?
23:06
Ejaaz: It's cheaper because the tokens that they serve from their own data centers are cheaper to serve.
23:10
Ejaaz: So all around a good engineering effort from Meta. I don't think they're anywhere
23:15
Ejaaz: near Fable 5 LLM worthy, but I actually don't think it matters.
23:18
Ejaaz: I think their main focus should be on niche things like this to do what they're
23:21
Ejaaz: good at, which is social media.
23:23
Josh: Yeah. I mean, it seems like it looks pretty good. This is certainly far from the frontier.
23:28
Josh: And I'm starting to realize now through all of these releases,
23:30
Josh: how difficult it is to compete at the frontier.
23:33
Josh: It's just like, I mean, I think of Google, I think of Meta, I think of,
23:36
Josh: spacex ai these are all incredible companies and they are putting all of their
23:40
Josh: weight behind getting to this frontier
23:42
Josh: and yet they're still having a difficult time of actually reaching it at the
23:45
Josh: same rate that some other companies like anthropic and open ai are doing that
23:48
Josh: said i mean this is a good model for what it is i love the idea that it has
23:51
Josh: reasoning i think with open ai and their most recent image model that they they published the,
23:57
Josh: novel and really exciting thing was that it had reasoning abilities baked in
24:00
Josh: it didn't just take your text prompt at face value and generate an image.
24:04
Josh: It actually thought about it.
24:05
Josh: It works through the physics. It had this chain of thought that you kind of
24:08
Josh: trace to get to a final image. And therefore the outputs were significantly better.
24:12
Josh: This is another version of that where it's very smart. So I imagine we get some
24:16
Josh: fun kind of like mobile based...
24:19
Josh: Optimized features with this where it can recognize details and images really
24:23
Josh: well it can perhaps like if you have
24:26
Josh: for example on this image there is something sitting on the table you could
24:28
Josh: swap out that element in particular i'm sure there's a lot of customization
24:32
Josh: being built around this reasoning ability to understand what it's looking at
24:35
Josh: and kind of reason through the images that you want
24:37
Josh: and yeah overall pretty cool release um are we going to be using it is this
24:40
Josh: something in the daily driver toolkit maybe thumbnail.
24:44
Ejaaz: Creation no no why
24:45
Josh: Would you do that when you got Chachi Petit image gen baby that's true or nano banana,
24:50
Josh: like these are both just significantly better so there's no real use case for
24:54
Josh: anyone who wants to generate images
24:56
Josh: if you just have a $20 a month subscription anywhere else but it is cool to
25:00
Josh: see more progress coming out of meta I mean we have brain wave reading and now
25:03
Josh: we have image generating this week from them so they are pushing the needle forward.
25:07
Ejaaz: We have two lost items on the docket and then we're going to round this episode
25:11
Ejaaz: out since we're on the topic of image gen I hate to break it to you,
25:15
Ejaaz: but there might be a replacement for GPT Image 2 coming out soon.
25:19
Ejaaz: CDream from ByteDance, those are the owners and creators of TikTok,
25:23
Ejaaz: if you didn't know, has come out with their first image model that competes
25:27
Ejaaz: directly with GPT Image 2.
25:29
Ejaaz: And honestly, if this demo is anything to go by, it is incredibly impressive,
25:32
Ejaaz: but not because of the quality necessarily, but one for the cost.
25:37
Ejaaz: It's super cheap to create and also really easy and intuitive to edit.
25:41
Ejaaz: I mean, look at some of these things. Like I could create websites de novo from
25:44
Ejaaz: scratch using some of these things. I could use different languages,
25:47
Ejaaz: whatever I want. I could swap in animals for other types of animals.
25:50
Ejaaz: I can go for a certain palette color.
25:51
Ejaaz: It just seems like a creator or
25:54
Ejaaz: designer's dream. Definitely one worth experimenting with and one that I'm
25:59
Ejaaz: honestly excited to get my hands on
26:00
Josh: Yeah this this one is really good i think anything that by dance releases has
26:05
Josh: been like pretty impressive and this is no shortage the most interesting thing
26:08
Josh: like you said is the ability to edit things quickly and to kind of workshop things to your
26:12
Josh: own ability you can separate elements almost as if it's photoshop so if you're
26:15
Josh: a designer who's been incredibly frustrated because you can't edit specific
26:18
Josh: features this is the model for you you can now basically select any part of
26:23
Josh: the image you want update it change it to however you'd like
26:25
Josh: really powerful for customization so
26:28
Josh: i'd say this probably is less for the average person more for the pro or the
26:32
Josh: designer who wants to get a little bit more out of the image who wants a very specific output,
26:36
Josh: very impressive overall.
26:37
Ejaaz: And the final item, Cloudflare, which is not really a company that you hear
26:41
Ejaaz: often because it's in the back end of the internet, has opened up a new monetization
26:45
Ejaaz: gateway, a new payments API.
26:47
Ejaaz: Now, if you operate any kind of website, it's likely that Cloudflare is the
26:51
Ejaaz: guys behind it that is making sure there's no bots attacking you,
26:53
Ejaaz: that there's no spam accounts, et cetera, et cetera, so they can govern visits of a website.
26:58
Ejaaz: The reason why this is important is, let's say you create a cool bit of content.
27:01
Ejaaz: Let's say you create an awesome video on limitless that, you know,
27:04
Ejaaz: might be for paid subscribers in the future, you can now have an AI agent troll
27:09
Ejaaz: this site and pay a little micropayment.
27:12
Ejaaz: Typically, websites in the internet aren't built for this.
27:14
Ejaaz: But with protocols like x402 from our friends over on the Web3 side of the world,
27:19
Ejaaz: it enables micropayments.
27:21
Ejaaz: So you can pay a couple of cents to get access to an article instead of having
27:23
Ejaaz: to pay like a $100 per year subscription and stuff like that.
27:27
Ejaaz: So the reason why I wanted to point this out is I thought it was cool.
27:30
Ejaaz: I think it probably marks the open of a new type of economy which is agents
27:34
Ejaaz: paying for stuff without you even knowing about it and having the means to do
27:37
Ejaaz: so just thought it was worthy of an update
27:40
Josh: Yeah i found it interesting how cloudflare uh they say 52 percent of crawler
27:43
Josh: requests are now for ai training which is funny so half of the crawlers in the
27:46
Josh: world are used just for training better models which is up from 22 percent in spring of 2025,
27:52
Josh: um and now ai crawlers request content 100 to 10 000 times more than they refer
27:57
Josh: traffic back so this is kind of
27:58
Josh: built around that thesis that idea it's interesting it's cool i think um
28:03
Josh: it's a testament to where things are to come we just had the cloud fair publication
28:06
Josh: a few weeks ago that said the internet's now dead uh over half of the traffic
28:10
Josh: is bought so this is a continuation of that okay well if half the
28:13
Josh: internet traffic is bought let's build a payments gateway for them and that's
28:17
Josh: exactly what this release is so just a continuing of the trend on this topic and with that.
28:22
Josh: That is all for this week. You are fully caught up in the know with all things
28:27
Josh: AI, frontier technology. This is our fourth episode of the week.
28:29
Josh: If you missed any previous ones, go back and check them out.
28:32
Josh: We had one about this company that we love, Etched.
28:34
Josh: We had the Fable 5 Return. We had,
28:37
Josh: ai finding all the trades that we missed over the last 200 episodes of recording
28:41
Josh: which was a very fun episode so there's some good stuff for you to dig into
28:44
Josh: this week if you haven't caught up if you have caught up
28:47
Josh: go and enjoy the weekend touch grass it is an amazing weekend at least in new
28:50
Josh: york it's looking sunny and beautiful i hope wherever you are the same is true
28:54
Josh: and yeah if you enjoyed this episode don't forget to share it with a friend
28:57
Josh: don't forget to rate us on your favorite podcast platform,
28:59
Josh: he does any final parting thoughts before we head out.
29:02
Ejaaz: I think the final thing to say is if you love our episodes, if you enjoy our
29:05
Ejaaz: content, and if you would like to support us in any way, we are in the market for sponsors.
29:10
Ejaaz: We're looking for companies and features and products that we're excited to
29:13
Ejaaz: use and talk about to our audience of hundreds and hundreds and hundreds of
29:17
Ejaaz: thousands of people per month.
29:19
Ejaaz: So if that is you or if that is someone that you know that think might be them,
29:23
Ejaaz: please let them know. Please send them a message. Send us a message.
29:26
Ejaaz: We're available on X, on email, whatever. We'll leave it in the description below.
29:30
Ejaaz: And yeah, hopefully speak to you
29:31
Ejaaz: soon. Thank you so much for listening and we'll see you on the next one.
29:34
Josh: See ya.