Latest / In the Loop AI / THIS WEEK IN AI: Tension at World AI Summit, Snapchat Glasses, Meta's Crisis

Transcript

THIS WEEK IN AI: Tension at World AI Summit, Snapchat Glasses, Meta's Crisis · In the Loop AI

0:00
Josh: You either die a hero or live long enough to see yourself become a villain.
0:02
Josh: Well, today there's a new version of that, which is you either exit Silicon
0:05
Josh: Valley or spend enough time there for no one to tell you that these glasses
0:08
Josh: look absolutely ridiculous.
0:11
Josh: What we're showing on screen right now is the Snap Specs. This is the new version
0:15
Josh: of Snapchat's hardware offering that is marketed as this spectacle of a device.
0:22
Josh: It is this remarkably intriguing and compelling and innovative device that sits
0:26
Josh: on your face, that has all these capabilities. But the reality is,
0:30
Ejaaz: I mean, dude.
0:31
Josh: These glasses on Evan Spiegel's face, the CEO of Snapchat, not only do they
0:34
Josh: look ridiculous, but they cost $2,200.
0:38
Josh: I mean, what was he thinking? What was the company thinking?
0:41
Josh: Who let this get released?
0:43
Josh: And how are they so confident as he's walking around the venue yesterday as
0:47
Josh: they release these specs?
0:49
Ejaaz: Listen, this is a serious show and I want to have a serious conversation about
0:53
Ejaaz: this, but this is ridiculous. Do you know when you buy sunglasses for a little child...
0:58
Ejaaz: And like, or they're like, mommy, can I like wear your glasses,
1:01
Ejaaz: please? And it just looks oversized and ridiculous on their head.
1:05
Ejaaz: This is what this looks like. I just want to point out, for those of you who
1:07
Ejaaz: are looking at our stream right now, the back half of the actual glasses,
1:12
Ejaaz: the handles, barely kind of broach his ear.
1:15
Ejaaz: They kind of like sit on top of his ear. It's squishing his ear.
1:19
Ejaaz: That's got to be so uncomfortable. Wait, and for those of you who say,
1:22
Ejaaz: oh, this is just an AI rendering. No, look at it.
1:24
Ejaaz: This is live, like in the mix of things. As he turns, you can see they kind
1:29
Ejaaz: of extend beyond the back of his head so that it's like oversized fit for him.
1:33
Josh: Now that dude's ears are fighting for their lives.
1:35
Ejaaz: They're fighting for their life. So this is Snapchat, who hasn't really done
1:39
Ejaaz: anything major for the last 12 years.
1:42
Ejaaz: In fact, for the last 12 years, they've been investing a heck ton of money to
1:45
Ejaaz: the tune of $3.5 billion to figure out the AR glasses game. You know,
1:50
Ejaaz: you had Meta that's released a bunch of models since then.
1:53
Ejaaz: You've got Apple that's released the Apple Vision 4 for $3,500.
1:57
Ejaaz: And finally, Snap has released their flagship model. You would think after all
2:00
Ejaaz: that R&D, they'd come out with something cheap, slick, and aesthetic.
2:04
Ejaaz: And unfortunately, it's been the complete opposite. This is $2,200 to achieve
2:08
Ejaaz: or get these specs that you're seeing on your screen now. It's a standalone device.
2:13
Ejaaz: It has around 51 degrees field of vision, which is similar to Meta's Ray-Ban
2:18
Ejaaz: glasses. But of course, this thing is like 3x larger than anything that we've ever seen.
2:23
Ejaaz: You don't need to connect it to your phone. And basically, it runs its own software
2:26
Ejaaz: operating system on the lens itself.
2:29
Ejaaz: Now, I have seen demos of what the software itself looks like,
2:34
Ejaaz: Josh. And believe it or not, it's not entirely the worst thing.
2:38
Ejaaz: You're seeing someone on the screen here where someone's playing golf.
2:40
Ejaaz: You're seeing someone basically ordering off a menu in a language that they
2:43
Ejaaz: don't understand. You're seeing someone taking directions. You're seeing someone
2:46
Ejaaz: watching a movie on a plane. You're someone design the architecture of a building.
2:50
Ejaaz: This all looks good, but these are all curated demos, just to be clear.
2:54
Ejaaz: I don't know whether it actually looks this good, but these demos are already
2:57
Ejaaz: better than meta Ray-Ban glasses. But.
3:00
Ejaaz: No one's going to buy these things. And it's such a shame because it looks terrible.
3:04
Ejaaz: You don't want to look like you're wearing old school 3D glasses,
3:07
Ejaaz: watching a movie that no one's even there to see. It's just, it's so dumb.
3:11
Josh: Well, forgive me for my PTSD, but I think I'm going to need to see it to believe
3:14
Josh: it. These seem like a lot of kind of structured demos that are not out in the wild.
3:19
Josh: I give Meta a lot of credit when they released their glasses,
3:22
Josh: their Meta Ray-Ban displays, I think they were called. I forget because no one actually bought them.
3:28
Josh: They did do some demos on stage and they just completely flopped and they failed.
3:32
Josh: So I give them credit for trying.
3:33
Josh: Snapchat, I don't think did that. This is mostly a curated view of demos.
3:37
Josh: The form factor, it will say, there's some nice things to say about this thing.
3:41
Josh: Like the fact that they're trying for one.
3:43
Josh: I give them a lot of credit for actually trying for throwing their hat in the ring.
3:46
Josh: It looks terrible, but the technology is on its way. You mentioned the 51 degree
3:51
Josh: field of view. That actually is an improvement.
3:54
Josh: What that means is it's about equivalent to a 24 inch monitor up close.
3:58
Josh: So if you have a 24-inch display that you use your computer on,
4:01
Josh: that's roughly what it looks like. Or if you step back to 10 feet,
4:04
Josh: it's equivalent to about a 115-inch TV with seemingly pretty good fidelity.
4:09
Josh: I was listening to one of the interviews with Evan Spiegel, the CEO, and
4:13
Josh: The interviewer asked him the very obvious questions is like,
4:15
Josh: how are you planning to justify this?
4:17
Josh: It's $2,200 for this like very clunky pair of glasses. And the way that he justified
4:22
Josh: it is like, hey, people spend that money for a computer all the time.
4:26
Josh: In fact, they spend a lot more money on that for a computer.
4:29
Josh: And what you're buying with these is essentially a computer.
4:31
Josh: This is a standalone display.
4:33
Josh: It does have two Qualcomm chips built into it.
4:36
Josh: It has its own vision suite of sensors. It has two full color cameras.
4:39
Josh: It has two infrared cameras.
4:41
Josh: It has speakers. It has a microphone array. It has hand tracking, voice control.
4:45
Josh: A lot of the features that you expect inside of a computer, it's on these pair
4:49
Josh: of glasses. Now, the pair of glasses, the battery life is decent.
4:53
Josh: It's four hours of mixed use.
4:55
Josh: And with the charging case that it comes with, you get about 20 hours.
4:58
Josh: The old Spectacles managed about 45 minutes. This is a pretty considerable jump
5:02
Josh: over previous versions. It's just like the Vision Pro,
5:07
Josh: There's just not really a market for this. I can't imagine there's going to
5:10
Josh: be a world in which like, I love this type of technology. I am version one beta
5:15
Josh: buyer of all the hardware, but even this is like, I don't want to wear these
5:19
Josh: things. They're too bulky. They're too clunky.
5:20
Josh: I give them a lot of credit for trying to do this, for innovating.
5:25
Josh: And it's certainly correct, like directionally at least, where the hardware
5:29
Josh: of glasses as a form factor is great. It's just the actual hardware required
5:34
Josh: to make them a great product. It doesn't seem to be there, at least in the case of Snap for now.
5:38
Ejaaz: But it's also like the worst time to release a premium product.
5:42
Ejaaz: Like it's not cheap. This is a high-end laptop at $2,200.
5:47
Ejaaz: You're at that level, you are competing with Apple Vision Pro,
5:51
Ejaaz: which hasn't taken off either, right?
5:53
Ejaaz: Apple yesterday announced that they're going to increase the cost of all their
5:56
Ejaaz: products because the cost of memory and everything is going up.
5:59
Ejaaz: So it's the worst time to, when people are like, kind of like pinching their
6:03
Ejaaz: wallets and not really spending much because, you know, their premium gadgets
6:06
Ejaaz: that they want to buy from Apple, they can't purchase themselves.
6:09
Ejaaz: What makes Snap think that they're going to be the replacement there?
6:12
Ejaaz: Now, one, like the points you make around the software thing and that they're
6:16
Ejaaz: trying, I agree, is good.
6:18
Ejaaz: The one thing that I saw when I was looking into the dev kit for this particular
6:22
Ejaaz: device is you can hook it up to Claude or ChatGPT, which I think is pretty cool
6:28
Ejaaz: for creating new kinds of AI apps. Typically with Meta, they're restricted to Meta AI specifically.
6:34
Ejaaz: I know with Google Glasses that is coming up, I believe it's still tied to Gemini.
6:37
Ejaaz: The fact that they've kind of like open sourced this to kind of like create
6:40
Ejaaz: and run your own agent could be pretty cool.
6:42
Ejaaz: It's like the closest thing that maybe we've come to Jarvis from Iron Man,
6:47
Ejaaz: except this is like discount store Jarvis. This is like we've got Jarvis at
6:50
Ejaaz: home, you know, like you can't buy the real Jarvis.
6:53
Ejaaz: I'm hoping that this is the worst it's ever going to be. I like the attempt,
6:58
Ejaaz: on goal that they have for creating the software suite around this.
7:01
Ejaaz: And I also like that it's a standalone product that I don't need to hook up
7:04
Ejaaz: to my phone. It kind of like runs on its own.
7:07
Ejaaz: But it remains to be seen. Like I want to see someone in the wild that isn't
7:11
Ejaaz: from Snapchat actually trying this thing but to spend three and a half billion
7:15
Ejaaz: dollars they spent 1.9 of that by the way in 2023 alone and it's taken them
7:20
Ejaaz: like what three years to launch this thing.
7:22
Ejaaz: It remains to be seen, but I think we should probably move on to the next topic.
7:27
Ejaaz: There was a very important meeting amongst world leaders in Evian,
7:32
Ejaaz: France, which is where the famous Evian Water comes from, by the way.
7:36
Ejaaz: It's the G7 Summit, where you had the top seven nations of the world come together
7:41
Ejaaz: to discuss an array of political affairs.
7:43
Ejaaz: One of the main ones, however, was around AI and governing and regulating AI.
7:49
Ejaaz: And we had the main heads of AI were there, including Demis Asivis from Google,
7:53
Ejaaz: Dario Modi from Anthropic, and Sam Altman from OpenAI.
7:58
Ejaaz: Now, the main call, the main plea from Dario, Demis, and Sam was for a US-led
8:04
Ejaaz: coalition to determine the global standards and rules for AI in a closed-door
8:09
Ejaaz: meeting that happened in the morning.
8:10
Ejaaz: So it was kind of like a lunch thing that happened that the G7 Trump was president,
8:13
Ejaaz: as well as all the leaders of like Europe and a bunch of other countries.
8:16
Ejaaz: And the main idea was let's start a conversation around how we regulate this
8:19
Ejaaz: thing because as we know from recent news
8:22
Ejaaz: the models have finally gotten good enough where they pose themselves as a national
8:26
Ejaaz: threat if it's in the hands of an adversary and so that was what that discussion ensued.
8:31
Josh: Yeah this is hopefully a productive meeting i mean the the working theme of
8:35
Josh: this launch was ensuring a safe rapid and effective deployment
8:39
Josh: of ai yes and i think this is effectively what everybody wants there's just
8:42
Josh: a lot of varying opinions on how you actually get there And I am hopeful that
8:46
Josh: the intention of this meeting was to find some sort of convergence,
8:49
Josh: to find some sort of middle ground on how they could consider moving forward
8:52
Josh: with these frontier models.
8:54
Josh: We very clearly see now that we are experiencing this takeoff scenario in which
8:58
Josh: the models are becoming better and better. They're improving themselves.
9:02
Josh: The accelerated rate at which we are getting high quality models is like a little
9:07
Josh: staggering. It's like very fast.
9:10
Josh: It's overwhelming. And I think a lot of times there is this clash between the
9:15
Josh: government, which traditionally moves a little bit slower, and these AI labs
9:17
Josh: who are like, hey guys, this is not going away.
9:20
Josh: In fact this is only speeding up and finding some sort of middle ground is productive
9:24
Josh: so that we don't have any more issues like fable
9:26
Josh: i mean not being able to use fable or being able to use it for three days and
9:29
Josh: having it removed it sucks it hurts the users it hurts the company it seems
9:33
Josh: like it mostly hurts everyone so hopefully the outcome of this will be a little bit more of a
9:38
Josh: formalized process or at least a mutual understanding of where things lie so
9:42
Josh: when companies release the model they feel very confident in the model that they're releasing
9:46
Josh: they feel confident that the government supports them, and that there is not
9:49
Josh: going to be another situation in which something gets released that is a frontier
9:52
Josh: model and then gets pulled back. And I mean, fingers crossed to go as well.
9:56
Josh: Everyone who should be in that room seems like they are in that room.
9:59
Josh: And hopefully that leads to a pretty productive conversation.
10:02
Ejaaz: Well, being in the room is one thing where they were sat was another thing.
10:05
Ejaaz: I don't know if you saw the seating plan for this thing. So I was I was watching
10:09
Ejaaz: clips of of this actual discussion because it was filmed.
10:13
Ejaaz: And I noticed one thing. So if you look on the right side of the image here,
10:17
Ejaaz: on either side of Trump, you had Demis Hasevis, head of Google,
10:21
Ejaaz: DeepMind, so Google's AI division, and then you had Sam Altman to the right of him.
10:26
Ejaaz: And then that guy, Dario Modi, yep, seated exactly opposite,
10:30
Ejaaz: so nowhere near Trump. And so the tension was pretty obvious.
10:35
Ejaaz: The kind of rivalry, or rather the grudge that was going on between the government
10:39
Ejaaz: right now and Anthropic, as they're discussing kind of like re-release of Fable
10:43
Ejaaz: and how they kind to kind of regulate and govern that.
10:46
Ejaaz: Obviously a big tension between the two parties. When Dario spoke up,
10:50
Ejaaz: and he was the last to speak of the three, he spoke about warning of regulation
10:55
Ejaaz: and saying we need to overregulate this kind of things.
10:57
Ejaaz: And we cannot get these models in the hands of China, which is nice to hear
11:01
Ejaaz: because that's, I think, him and the government kind of align on that.
11:04
Ejaaz: The second main discussion that was had between the three of them was the idea
11:09
Ejaaz: of making sure we govern hardware as well as software.
11:12
Ejaaz: Now, typically, the US has taken a very proactive approach to making sure that
11:15
Ejaaz: China and adversaries don't get access to cutting edge chips, primarily from NVIDIA.
11:21
Ejaaz: But that has also pushed China to kind of like create their own chips,
11:23
Ejaaz: which have now almost caught up with the US in terms of like inference and training.
11:27
Ejaaz: Dario has been warning like, hey, we need to restrict the hardware even more.
11:31
Ejaaz: And we need to be careful with the safeguards that we set up for our models,
11:33
Ejaaz: which is ironic, of course, because that has led to the ban of fables so far.
11:37
Ejaaz: Sam was the first to speak up. So he was given, if you're talking about the
11:40
Ejaaz: Trump's favorites or the G7 favorites, it was Sam, it was Demis,
11:43
Ejaaz: and then it was Dario. So a lot of political things going on.
11:46
Josh: Okay, there's another big update that I want to talk about. And this one feels
11:49
Josh: pretty noteworthy. And it's around this man by the name of Noam Shazir.
11:54
Josh: And Noam is interesting because he is very much a legend in the AI space.
11:59
Josh: For those who are familiar with the word GPT, perhaps you've heard it in the context of chat GPT.
12:05
Josh: The letter T stands for transformer.
12:07
Josh: And transformer is a engineering concept that was actually pioneered
12:12
Josh: in part by Noam Shazir, who wrote the canonical paper, Attention is All You
12:17
Josh: Need, about how a transformer actually works all the way back in 2017,
12:21
Josh: which in AI years, that is an eternity ago.
12:24
Josh: So Noam is this foundational character in the world of AI who has really been
12:28
Josh: there through all of the changes, all the evolutions of it.
12:33
Josh: And he's had this kind of rough, bumpy road with Google.
12:37
Josh: He's been working at Google or he joined Google initially in 2000,
12:40
Josh: I think that's 26 years ago he's been there. And for a long time.
12:44
Josh: He was an engineer. He was building these AI models. They figured out the transformer
12:48
Josh: architecture. And then he left.
12:50
Josh: And he left to go start his own thing. He went off for a couple of years.
12:53
Josh: And then in 2024, Google missed him so much, they reportedly paid $2.7 billion
13:00
Josh: To license his company, Character AI's technology, to bring Noam Shazir back,
13:05
Josh: making him a VP and one of the people in the small group that are steering how Gemini gets towards AI.
13:10
Josh: So from 2024 to 2026, he's been working there after a pretty hefty payday.
13:17
Josh: Yesterday, we got the news that Noam is leaving the company.
13:20
Josh: And not only is he leaving, he's going to open AI.
13:24
Josh: So that's the backstory. And this seems like a pretty big deal.
13:27
Josh: Because when you think about the Avengers of AI, who are the people who are
13:30
Josh: really making a difference?
13:31
Josh: Noam is one of those guys. And now he has changed houses, I guess,
13:35
Josh: in this Game of Thrones from the Google house to the house of open AI.
13:39
Ejaaz: And he's going to be leading the architecture team at OpenAI,
13:41
Ejaaz: which is arguably the priority team that will design the model,
13:46
Ejaaz: feature iterations of the model, and mainly determine who leads in the race.
13:50
Ejaaz: If you want to think of who his counterpart is at Anthropic,
13:53
Ejaaz: it's Andre Carpathy, who also recently joined Anthropic. So if you want to kind
13:57
Ejaaz: of think of the titans of the game, the leaders, the geniuses,
14:00
Ejaaz: it's going to be Noam Shazia now versus Andre Carpathy.
14:03
Ejaaz: And Noam, as you mentioned, father of the Transformer. I can't believe he left.
14:07
Ejaaz: Um, when he left Google initially, what you mentioned, I think he was there
14:11
Ejaaz: for like 26 years, and then he left in 2021, he, uh, started Character AI,
14:15
Ejaaz: Josh. The-the-the AI that, like, kind of, like, creates kind of,
14:18
Ejaaz: like, different personable characters. The original AI agent, if you want.
14:22
Ejaaz: And then getting it acquired for $3 billion was just massively impressive.
14:25
Ejaaz: It's also a huge knock on Google. I don't know how I can view this aside from
14:31
Ejaaz: being bearish on Google.
14:32
Ejaaz: Everyone knows that for the last year, they've struggled to keep Gemini models up with the frontier.
14:37
Ejaaz: They had a massive lead at the end of last year when they released Gemini 3.
14:40
Ejaaz: Recently, 3.5 and 3.5 Flash has kind of caught up in some kind of generalized
14:46
Ejaaz: LLM qualities, but they've missed out on the major quality, which is coding.
14:50
Ejaaz: It just hasn't been able to catch up.
14:52
Ejaaz: We know that a lot of Google employees have actually been using code code and
14:56
Ejaaz: codex itself to try and distill what the model is, but they haven't been able to achieve that.
15:00
Ejaaz: So Gemini has struggled. And the fact that their chief architect,
15:03
Ejaaz: he's the VP of engineering, he is basically Demis of Civis' right man.
15:07
Ejaaz: Demis of Civis is like the front man, the CEO of Google DeepMind.
15:09
Ejaaz: And this guy is the actual technical guy. The fact that he's left to build architecture
15:14
Ejaaz: at one of the biggest competitors ever tells me that he was either being restrained
15:17
Ejaaz: at Google or that he wasn't able to kind of build the thing that he wanted to. And now he will be.
15:22
Ejaaz: Sam in his announcement tweet for the hire basically said, I've been trying
15:26
Ejaaz: to work with Noam for 10 years and it's finally happened.
15:29
Ejaaz: And I can't wait. So yeah, congratulations, Sam.
15:32
Josh: Yeah, we'll see what happens. It feels I can't help but notice that it feels
15:36
Josh: like all the top talent is starting to aggregate itself between only two houses.
15:41
Josh: We had Andre Carpathy recently joining Anthropic.
15:43
Josh: Now we have Noam joining OpenAI. It feels like a lot of the top talent is kind
15:47
Josh: of converging on these two places.
15:49
Josh: It seems like these two places are kind of running away with it like we we have
15:53
Josh: google gemini and every time we try to say like this is a new frontier model
15:57
Josh: every time they have a big release it doesn't quite do it
16:00
Josh: grok and the xai team and i mean spacex xas the whole that universe the elon
16:05
Josh: universe are trying they haven't quite figured out how to get to the frontier
16:08
Josh: and right now there's very clearly between these two companies and with this
16:11
Josh: engineering talent piling on i don't see a world in which
16:13
Josh: other companies are really able to catch up as quickly as they would like to
16:18
Josh: so we'll see we'll continue to monitor that as we go
16:21
Josh: and speaking of these leaders one of two of them had quite an update this week
16:26
Josh: that came in the form of a copy of a letter from commerce secretary howard lutnik
16:31
Josh: who wrote to dario moda i just do you have any light on this i haven't really looked much into it
16:34
Ejaaz: Yeah so just to catch everyone up with some context um it's been a rocky week
16:40
Ejaaz: for i'm probably to say the least uh they were started on such a high uh,
16:43
Ejaaz: last week when they released their brand new frontier model fable 5 it's their
16:46
Ejaaz: mythos grade uh model and it smashed every single benchmark and Everyone that
16:51
Ejaaz: used it, including you and I, said that this is the best model ever.
16:54
Ejaaz: And then three days later, the government yanked it. They sent a letter from
16:58
Ejaaz: Howard Lutlink to Dario Modi basically saying, hey, we can't have this model
17:03
Ejaaz: getting into the hands of foreign adversaries.
17:05
Ejaaz: So either you KYC and ID verify everyone that uses this model or you shut it down.
17:10
Ejaaz: Dario doesn't have the infrastructure to KYC and ID everyone,
17:12
Ejaaz: so he shut it down for everyone globally around the world. He can't control
17:15
Ejaaz: it. So currently, if you'll listen to this right now, you don't have access to Fable 5.
17:20
Ejaaz: Now, the one thing that was missing was we didn't know what this letter said.
17:23
Ejaaz: Finally, we have the details.
17:26
Ejaaz: There's a lot of legalese in this, Josh. That's the TLDR.
17:30
Ejaaz: There's a lot of kind of structured responses basically saying,
17:34
Ejaaz: hey, listen, according to regulation, blah, blah, blah, you can't put certain
17:39
Ejaaz: tools and certain products in the hands of foreign adversaries.
17:42
Ejaaz: If this is the case, then basically we have the right to restrict this.
17:47
Ejaaz: And then there was a lot of back and forth between Dario and this letter basically
17:50
Ejaaz: saying, hey, this isn't a physical tool. This is a software product.
17:54
Ejaaz: Your regulation basically doesn't cover this. And it was just stuff like that.
17:58
Ejaaz: My kind of whole take from this is that Howard Lutnick and Trump or the government,
18:02
Ejaaz: whoever's responsible for this, basically wanted to take down this model and
18:05
Ejaaz: they were going to go to any lengths to be able to do that.
18:08
Ejaaz: And it was triggered by a warning that was issued from Amazon CEO Andy Jassy,
18:13
Ejaaz: basically saying that their security team had discovered a hole or an exploit
18:18
Ejaaz: that could be basically accessed by China.
18:21
Ejaaz: Now, in the interest of the letter, they describe what this attack looks like, Josh.
18:26
Ejaaz: And basically it's an attack that can be exposed through current models that
18:29
Ejaaz: OpenAI is running right now and Grok and Google Gemini.
18:33
Ejaaz: So it isn't really a strong argument, but they're just picking on Anthropik
18:37
Ejaaz: specifically and on this model specifically because they believe that it could
18:41
Ejaaz: be dangerous in the wrong hands.
18:42
Josh: Yeah, it's a bummer. I hope it gets worked out. Fable's an incredible model.
18:45
Josh: It would be awesome for the world to be able to continue to benefit and use
18:49
Josh: that model. So hopefully all sorted out soon.
18:53
Josh: This next part of the episode is all about open source it's been a big week
18:56
Josh: for open source like shockingly big week actually starting with microsoft believe
19:02
Josh: it or not uh so yeah microsoft still exists they're still
19:05
Josh: in the world of ai or ai jason whatever you want to call it and
19:08
Josh: they have this product called copilot you may not have used it unless you are
19:12
Josh: forced to through your employer but copilot is actually used by like quite a
19:15
Josh: few people and they have made a large shift this week if i understand correctly,
19:19
Josh: they're switching from usage-based pricing with OpenAI's ChatGPT
19:24
Josh: to fully open source through the DeepSeq model, the Chinese open source model.
19:29
Josh: They're running a Chinese open source model on Microsoft servers. Is that right?
19:33
Ejaaz: Yeah. So Microsoft's flagship AI assistant, as you said, is Copilot.
19:38
Ejaaz: I think over the last three months, they've released various versions of Copilot,
19:43
Ejaaz: which does different things.
19:45
Ejaaz: One of them uses Anthropics model exclusively. And now Now this new one uses
19:50
Ejaaz: a Chinese open source model, which they've fine tuned.
19:52
Ejaaz: Now their reasoning for using this model, their main reasoning is it's super cheap.
19:58
Ejaaz: And Microsoft has undergone internally a shift from using these expensive coding
20:02
Ejaaz: models, such as Claude Code, to cutting down their budgets and using cheaper models.
20:06
Ejaaz: This is their first consumer-facing or enterprise-facing product where you now
20:10
Ejaaz: get access to a Chinese open source model.
20:12
Ejaaz: Now, they received a lot of feedback about this, which is like,
20:14
Ejaaz: you know, is this safe? Is this cool to use? Their claim is that,
20:17
Ejaaz: yes, it is, but it remains to be seen.
20:20
Ejaaz: The wider story about this, Josh, is Microsoft's the biggest investor in OpenAI.
20:27
Ejaaz: Their stake alone is worth upwards of $135 billion.
20:32
Ejaaz: And they signed an exclusive partnership with OpenAI to get access to all their IP.
20:36
Ejaaz: Right now, as we speak, they have access to all of OpenAI's model weights for all their models.
20:41
Ejaaz: So they could theoretically train their own foundational model using OpenAI's
20:45
Ejaaz: Blueprint and create even their own version. But you would think that they would
20:49
Ejaaz: use OpenAI's ChatGPT models because it's the frontier. Unfortunately, they've fallen out.
20:54
Ejaaz: Recently, Sam Altman announced that ChatGPD will be available through AWS,
20:59
Ejaaz: which is Microsoft Azure's main kind of competitor when it comes to cloud services.
21:04
Ejaaz: Now, before that, ChatGPD was only being served through Microsoft Azure.
21:09
Ejaaz: They were the exclusive partner, but they've since had a falling out over the
21:12
Ejaaz: numbers, over the for-profit structuring system and all that kind of stuff.
21:16
Ejaaz: So now they don't really talk. Satya Nadella, the CEO of Microsoft,
21:19
Ejaaz: has been having frequent meetings with Dario Modi.
21:21
Ejaaz: The point is like him and Sand don't really get along right now.
21:24
Ejaaz: And he's taking a model agnostic approach.
21:26
Ejaaz: And DeepSeek is the first one. And DeepSeek just had a crazy week as well.
21:29
Ejaaz: They raised around, I think, $8 billion at a $50 billion valuation.
21:34
Ejaaz: The founder himself put in, I think, $3 billion of his own money to maintain
21:39
Ejaaz: kind of like ownership class shares of this entire company.
21:42
Ejaaz: And he said he is committing to keeping all his models open source.
21:45
Ejaaz: So it's a huge win for China.
21:47
Josh: While you're saying Microsoft has access to their open weights of Chash EBT,
21:51
Josh: And they're still not doing anything about it.
21:53
Josh: It reminded me of this paper that I wrote. I forget who the author was,
21:56
Josh: but I saw it on Twitter earlier this week.
21:58
Josh: And I think the title was that the window is closed. And there is this like-
22:02
Josh: Yes, yes. And it was really interesting because it talked about how the opportunity
22:09
Josh: for a company to become a frontier AI lab has kind of ended.
22:14
Josh: It's almost impossible to catch up. And this is kind of the trend that we're
22:18
Josh: seeing. And I just want to make some time to talk about it for a sec.
22:20
Josh: Because we saw xai and elon and grok like they were the newest entrant they
22:25
Josh: were the newest competitor and elon is among the best in the world at aggregating
22:29
Josh: resources and solving difficult problems and they got close
22:33
Josh: xai and grok they got very close and they're still close but they haven't actually
22:37
Josh: been able to reach the frontier and that's after building these data centers in a
22:42
Josh: record amount of time by like a factor of five it's like he he did everything
22:46
Josh: you could possibly do and did it all right and still wasn't quite able to catch up.
22:50
Josh: For a company like Microsoft, who isn't as hardcore, for a company like Google,
22:54
Josh: who maybe just moves a little bit slower, that's such a disadvantage in a world
22:58
Josh: where this rapid iteration is happening between, I mean, OpenAI and Anthropic.
23:02
Josh: And it feels like there is this kind of pulling away. You see it with the talent,
23:06
Josh: you see it with now the benchmarks, you can feel it when you use Fable.
23:09
Josh: You're like, oh, this is a novel thing that I've never experienced before.
23:12
Josh: And it's just, it's a shift that I just wanted to highlight for a second,
23:15
Josh: because it really feels like The companies that are building the Frontier AI today
23:19
Josh: Are going to be the ones that are going to continue to be leading the pack because
23:22
Josh: of that like self recurring improvement loop and it's just like the self-fulfilling
23:27
Josh: thing where the better it gets the faster they can iterate the faster it gets
23:30
Josh: it's just this unbelievable thing so
23:32
Josh: we'll probably continue to see that i don't want to get too sidetracked because
23:35
Josh: we have another open source thing to
23:37
Ejaaz: Talk about i actually think it's it's worth talking about because i think few
23:40
Ejaaz: of our listeners actually kind of like can see the wider picture and i think
23:43
Ejaaz: it's important to like paint it like,
23:45
Ejaaz: recursive self-improvement basically like the model like building itself we've
23:49
Ejaaz: reached a point where these models are now good enough to do it and previously
23:52
Ejaaz: it was a capital race spend as much money as you can hire the best talent and
23:56
Ejaaz: then design the best thing
23:57
Ejaaz: it looks like anthropic and open air are walking away with it i will say that
24:01
Ejaaz: i think elon and xai have a good shot because they just acquired cursor for 60 billion dollars um,
24:08
Ejaaz: typically grok has been terrible No one has really kind of used it.
24:11
Ejaaz: Their GPUs aren't being used.
24:13
Ejaaz: But if they adopt cursor into the model, you can now jump and kind of catch
24:18
Ejaaz: up at the frontier for coding and all that kind of stuff. Remains to be seen.
24:22
Ejaaz: Google just lost one half of their entire architecture team,
24:26
Ejaaz: who is the most important member, arguably.
24:28
Ejaaz: So now I'm kind of hesitant, but I'm willing to say that maybe they have lost
24:32
Ejaaz: the race now and they're going to run away with it. So it's pretty crazy to say.
24:36
Ejaaz: Actually, talking about another company that might have lost the race entirely,
24:41
Ejaaz: and have actually probably spent the most money out of all of these companies, Meta.
24:45
Ejaaz: Meta's had like a crazy couple of weeks, but a crazy week in particular.
24:51
Ejaaz: So to kind of catch you all up, Meta over the last year and a half has spent,
24:57
Ejaaz: and I'm not exaggerating, upwards of $35 billion.
25:01
Ejaaz: To acquire or hire around, I want to say 150 people for their AI team.
25:08
Josh: I don't even want to think about what the average amount they paid for that person is.
25:11
Ejaaz: That's crazy. I don't even want to think about it either.
25:14
Ejaaz: The craziest part is they already had an AI team and they were already,
25:18
Ejaaz: they weren't at the frontier a year and a half ago, but they were near the frontier
25:22
Ejaaz: and they had the best open source model. It was called Llama.
25:25
Ejaaz: And the guy that headed their team, Jan LeCun, was, you know, he wasn't that bad.
25:30
Ejaaz: They since fired him and re-awked the entire team. Every month and a half,
25:35
Ejaaz: I saw a tweet that someone had like measured out the cadence of this.
25:38
Ejaaz: They fire around four to 8,000 people from their team. And they had layoffs recently about.
25:44
Ejaaz: Three weeks ago. So they've re-auged their entire structure.
25:47
Ejaaz: And this week, Andrew Bosworth, who is kind of like Zuck's right-hand man in
25:51
Ejaaz: terms of laying out the foundational structure for Meta's AI efforts,
25:56
Ejaaz: basically said, yeah, the re-aug was a disaster.
25:58
Ejaaz: We shouldn't have done this. And it was a terrible move on our side.
26:02
Ejaaz: They're also making some really weird decisions now. So apparently,
26:05
Ejaaz: there's a rumor going around that 30% to 50% of their software engineers on
26:09
Ejaaz: core teams have now become data labelers.
26:12
Ejaaz: So what this means is they're being tracked and monitored for all the work that
26:16
Ejaaz: they do and it's being fed into a data system which is then used to train AI
26:20
Ejaaz: agents and AI models to do their jobs.
26:22
Ejaaz: So it's effectively kind of like reverse engineering what they do and then eventually
26:27
Ejaaz: the AI models will be capable enough to actually do their job and I guess they're not needed anymore.
26:31
Ejaaz: I think the narrative that's being given is that they'll be up-leveled and kind
26:34
Ejaaz: of like hired like a team leader for all these different AI agents but I just
26:39
Ejaaz: don't see that happening. Meta has completely blown the lead that they had,
26:43
Ejaaz: if they had any right now. And I just don't think they can catch up.
26:46
Josh: It's pretty sad. Yeah, it's disappointing because when you spend that much money,
26:50
Josh: you expect to get some sort of outcome that's interesting and unique and noteworthy.
26:55
Josh: And the only unique and noteworthy thing we've mentioned on the show as it relates
26:58
Josh: to things that they've released is their, what is it, the Muse Spark model that
27:02
Josh: was serving as customer service that was then used to exploit all the accounts
27:06
Josh: on Instagram and Facebook.
27:07
Josh: It's just been a total disaster. And when it comes to headcount,
27:11
Josh: I mean, I think at the beginning of this year or the end of last year,
27:15
Josh: there were just around 80,000 employees.
27:19
Josh: That's a tremendous amount of people to be working on.
27:22
Josh: Like a problem and i have to imagine that it's it's very difficult to
27:27
Josh: turn that group to be productive and to move as fast as they need to in order to make actual progress
27:34
Josh: um but at the same time he spent all this money on building a team you have
27:39
Josh: to imagine they exist in somewhat of a silo with all the resources they need
27:42
Josh: and still they haven't been able to deliver anything so it's this weird place
27:45
Josh: where again i say this a lot but i i think it's true that like meta
27:50
Josh: Is really still just feeling like facebook plus plus it's like they tried to
27:55
Josh: pivot to meta to the metaverse they fail that now their name is meta but they
27:59
Josh: have no metaverse to show for it
28:01
Josh: they're trying with the glasses they're in the same boat as evan spiegel and
28:04
Josh: snap that we mentioned earlier in this episode where the technology is promising
28:07
Josh: but it's still horrible no one's actually going to want to buy it or use it
28:10
Josh: so that's still too early they have yet to prove that they can make hardware
28:13
Josh: they have yet to prove that they can make a compelling
28:15
Josh: ai product because everything that they've released so far has been a complete
28:19
Josh: and total flop and it just leaves me in this really weird spot where like i
28:22
Josh: want to believe in zuck and this founder-led company but nothing that they have produced over the last
28:28
Josh: i don't know five six seven plus years has actually been compelling so it's
28:33
Josh: a difficult time for for meta and i really hope they figure it out
28:36
Ejaaz: Yeah if there was one thing to kind of like summarize meta in like the last
28:40
Ejaaz: couple of months they they created this internal token leaderboard do you remember
28:44
Ejaaz: this where yeah uh they ranked employees on the amount of tokens that they spent ai tokens
28:50
Ejaaz: to be productive and it was a measurement of productivity obviously for any sensible individual
28:56
Ejaaz: the number of tokens you spend doesn't necessarily correlate to how smart you
28:59
Ejaaz: are how intelligently you are using tokens.
29:02
Ejaaz: They have now used that exact same board to figure out who they fire for every
29:07
Ejaaz: single layoff so if you're if you're a major token spender they're like oh you're
29:11
Ejaaz: misusing your tokens we're letting and you go, that's what they did in the recent layoffs.
29:15
Ejaaz: Sad, crazy, I hope Meta figures it out. But talking of like lean teams that
29:20
Ejaaz: are spending much, much less, probably like one 50th,
29:25
Ejaaz: of what Meta has spent over the last couple of years, but are still not only
29:28
Ejaaz: maintaining the frontier, but in some cases beating it.
29:31
Ejaaz: We have a new Chinese open source model from the Chipu, Z-H-I-P-U,
29:38
Ejaaz: I believe, Chinese lab, GLM 5.2.
29:42
Ejaaz: Open weights, MIT licensed open weights. So, you know, I think that means after
29:47
Ejaaz: you get to a certain number of users, you have to like credit the company or something like that.
29:51
Ejaaz: But I just want everyone to focus on this chart for a second.
29:54
Ejaaz: It's on agentic coding performance. Now, if you're wondering,
29:57
Ejaaz: why isn't it just looking at regular AI coding performance? Why not?
30:00
Ejaaz: Why is it looking at agentic?
30:01
Ejaaz: It's because a lot of these agents are the new way of working with AI tools going forward.
30:06
Ejaaz: It's not just going to be you in a single session or in a single chat window
30:10
Ejaaz: talking to one specific LLM, you're going to speak into multiple instances of
30:14
Ejaaz: the same LLM, and it's going to be independently talking to each other and doing
30:17
Ejaaz: a bunch of work. That's what makes you exponentially more productive.
30:21
Ejaaz: And so agentic coding is a strong measure of whether the model is good enough.
30:25
Ejaaz: Now, GLM 5.2, when compared to Claude Opus 4.8, pretty much matches it on its score.
30:31
Ejaaz: Now, notably, this doesn't include Fable 5, which is the restricted model that
30:35
Ejaaz: we can't get access to right now, but it's pretty impressive to have open source
30:39
Ejaaz: models still catch up with the Frontier Labs.
30:42
Ejaaz: I'm still trying to figure out how they're able to pull all of this off.
30:45
Ejaaz: Like, they don't have any of the money. They don't have any of the chips.
30:49
Ejaaz: I'm kind of encouraged to think that this is probably a data play,
30:53
Ejaaz: so the Chinese have, like, access to a lot of data that they can use to smartly train models.
30:57
Ejaaz: But it's just impressive nonetheless. And it has a really cool ability to create
31:02
Ejaaz: front-end UX design, which is typically...
31:05
Ejaaz: Something that's hard for models to achieve visual interpretation visual intelligence
31:09
Ejaaz: but what you're seeing on the screen right now is i believe like a personal
31:12
Ejaaz: website that was created in one shot in five minutes which is pretty cool to see.
31:15
Josh: Yeah very cool model it's um it's interesting to see the benchmarks they always
31:19
Josh: have to be taken with a grain of salt because when you see them being compared to other models
31:22
Josh: i'm not quite sure it's it's close and for that it's respectable it's like okay
31:27
Josh: if even if you've come close if you're a little bit above a little bit below
31:29
Josh: just being in the ballpark is respectable i find that a lot of times now with
31:33
Josh: the models you just got to try it out and play around with it and really get
31:35
Josh: a feel for it because a lot of that
31:37
Josh: the like kind of nuanced feel that you get when you play with the model is the
31:41
Josh: thing that's actually meaningful
31:43
Josh: versus the benchmark so that's kind of how i've been doing it i'm looking to
31:47
Josh: try and play around with this thing and see if it could do some cool front-end design for us maybe
31:51
Josh: rebuild our website uh that would be kind of cool um but it's really exciting
31:55
Josh: to see china still kind of hanging on they're still pushing forward to the frontier
31:58
Josh: every time we count them out they come back and to clap back with something else.
32:01
Josh: So an exciting release from Jipu, or however you pronounce them, I'm not entirely sure.
32:07
Josh: But yeah, it's very cool. And then also another ex-open source company named MidJourney.
32:13
Josh: You might remember MidJourney because they did the image generation model.
32:17
Josh: They were one of the very earliest people to do. Yeah, exactly.
32:19
Josh: They were one of the very earliest people to do image visual generation with AI.
32:24
Josh: Well, they've pivoted in a pretty meaningful way.
32:28
Josh: Like, I'm not even sure this is the same company. When I saw this news this
32:31
Josh: morning, I had to click on the
32:32
Josh: profile and make sure it was the same MidJourney that I was thinking of.
32:36
Josh: It's like something out of Dune, Josh. Because it does. The AI company that
32:38
Josh: you know and love as the Image Generation Company has created a physical hardware
32:42
Josh: device and a new division of the company that is known as Mid Journey Medical.
32:48
Josh: What is Mid Journey Medical's first device? Well, if you're seeing on screen,
32:50
Josh: there's a person being submerged underwater. This is not AI. This is a real person,
32:55
Ejaaz: I think at least. Yeah, it's a real demo.
32:56
Josh: And the person is being submitted or submerged underwater to get a full body
33:01
Josh: MRI scan, which is shocking, but also remarkably impressive. I mean, this is...
33:07
Josh: Something that would traditionally take 60 or so plus minutes and now they're
33:11
Josh: doing it in a minute like how does an image company go from that to this this is a pretty huge leap
33:16
Ejaaz: I think founder context is important here the founders parrots uh one of them
33:21
Ejaaz: is a dentist and i believe his mother is literally an mri physician so he's
33:25
Ejaaz: probably inspired in some way there and you know he's in the image game josh
33:28
Ejaaz: it's not too far to i guess leap.
33:29
Josh: To imaging right
33:31
Ejaaz: Yeah imaging and and so honestly, what I thought was that this device used AI
33:35
Ejaaz: to kind of, you know, figure out what the insides of someone might look like.
33:39
Ejaaz: And I was getting super excited. Unfortunately, there's no AI involvement here.
33:42
Ejaaz: But nevertheless, this is a massive, massive scientific breakthrough. And I'll tell you why.
33:48
Ejaaz: You mentioned one thing, which is typically an MRI scan takes 60 minutes to
33:52
Ejaaz: occur, you need to go to like a regulated like medical institute to be able to access this.
33:57
Ejaaz: I got a pronubo scan recently, it took me 60 minutes, big, scary device,
34:01
Ejaaz: stick them ages to set up.
34:02
Ejaaz: This is a device that you step onto. It submerges you into, and I'm quoting
34:07
Ejaaz: here, warm water. So you're not going to cold. It's warm water.
34:11
Ejaaz: It's non-invasive. So there's no radiation at all. And it takes 60 seconds.
34:17
Ejaaz: So the time it takes for a coffee to be made, you'll have a full body MRI scan.
34:22
Ejaaz: The other interesting part is they're not putting these devices in hospitals.
34:26
Ejaaz: They're putting it in potential spas or they're setting up a coffee shop and
34:30
Ejaaz: spa type situations where you can like grab a coffee and just get an MRI scan
34:34
Ejaaz: if you bring your swimsuit along, I guess.
34:36
Ejaaz: And this is really cool because this is going to bring the cost of MRI scans
34:39
Ejaaz: down massively. MRI scan imaging is just amazing in general for figuring out
34:44
Ejaaz: the health of your body internally.
34:46
Ejaaz: But it's often priced a lot of people out who don't have health insurance or
34:49
Ejaaz: who don't want to spend several thousands of bucks to get access to it and then
34:53
Ejaaz: book an appointment and then wait months to do this. Now you can just walk into
34:56
Ejaaz: anywhere that has one of these machines, I don't know when they're rolling out, and get access to this.
35:00
Ejaaz: And the plan for the future is to use AI to look at these different images and
35:05
Ejaaz: help you with the diagnosis part of things, which is going to be pretty cool.
35:08
Josh: I love this. It's so cool seeing a company who has a lot of money.
35:12
Josh: They're clearly very well resourced, actually use it for something that's applicable
35:15
Josh: that can make people's lives way better.
35:17
Josh: And I'm sure there's going to be a lot of people who have a lot of critiques
35:20
Josh: on how the technology works, how they haven't actually built any hardware,
35:23
Josh: but just moving the puck in this direction, just coming up with the concept
35:27
Josh: and figuring out how to actually turn this into reality in a way that's going
35:30
Josh: to improve a lot of people's lives.
35:31
Josh: I think it's awesome. And it's following this trend that we have of this health and wellness and
35:36
Josh: Like biosciences trend that i think we'll probably continue to see a lot of
35:40
Josh: particularly as these ai models get much better at understanding the human body
35:44
Josh: understanding what makes us tick how to improve things
35:47
Josh: and being able to capture data like an mri machine is the first part of improving
35:52
Josh: the human body we first need to understand how it works what it's made up of
35:55
Josh: where the problems are and this will hopefully help to do that for a lot more people so
35:59
Josh: in the case of mid-journey this is awesome congratulations very exciting work to see i
36:03
Josh: love the pivot more people should pivot from image gen ai slop to saving people's
36:07
Josh: lives in healthcare that seems pretty tight to me
36:10
Josh: but that is everything that's our our whole week rounded up if you've made it this far thank you
36:15
Josh: we had a few pretty great episodes this week one of them is going crazy uh with
36:20
Josh: leopold ashenbrenner and the reason why he is short and video with a massive
36:24
Josh: amount of his portfolio we had another one covering the spacex ipo and whether
36:28
Josh: or not you should buy or not so if you haven't
36:30
Josh: go and get yourself caught up on that but if you've made it this far that's
36:32
Josh: it you're good you can go into your weekend feeling fully up to date on all things
36:36
Josh: AI, frontier technology, and I guess healthcare now too.
36:40
Ejaaz: If I were to leave all of you listeners with a final prompt,
36:44
Ejaaz: it's been a while since we've asked what the latest, coolest thing you guys
36:49
Ejaaz: have been doing with not just an AI model, but any AI feature.
36:53
Ejaaz: I would love to hear from you guys if there's anything that we should be talking
36:57
Ejaaz: about that we haven't covered yet.
36:58
Ejaaz: Are there any use cases that you guys, the weirder, the better.
37:02
Ejaaz: By the way, I just want to emphasize this. Like, don't feel like it needs to
37:04
Ejaaz: be anything impressive that makes you a million dollars.
37:06
Ejaaz: Like, if you're using this to design your grocery list and you think it's a
37:10
Ejaaz: pretty unique or novel thing, tell us.
37:12
Ejaaz: I want to hear about it because we're always looking for use cases where these
37:15
Ejaaz: AI models basically apply to the wider, like 99% of people who haven't gone
37:20
Ejaaz: past ChatGPT at this point.
37:22
Ejaaz: So if you're using it for something cool, let us know in the comments and wherever
37:25
Ejaaz: you're listening to us, whether it's on YouTube, whether it's on Spotify,
37:28
Ejaaz: whether it's on Apple Music, wherever, please subscribe and give us a rating
37:33
Ejaaz: where you can turn on notifications if you're on YouTube.
37:36
Ejaaz: You have no idea how helpful it is for us.
37:39
Josh: Share with your friends.
37:39
Ejaaz: And share with your friends. Share with your friends. It's very important because
37:43
Ejaaz: it helps us get a wider reach. Josh and I are trying to grow this channel into
37:46
Ejaaz: the biggest AI and Frontier Tech channel that we can.
37:49
Ejaaz: And we're excited to get that. We recently, I think we got like 1,100 new subscribers
37:54
Ejaaz: on YouTube over the last less than 24 hours. So hello to all of you.
37:58
Ejaaz: It's lovely to meet you. We come up with Frontier AI episodes,
38:01
Ejaaz: 20-minute chunks, four times a week.
38:03
Josh: They're a little bit longer now. We're working on 20 minutes.
38:05
Ejaaz: A little bit longer. Some of the roundups are a little bit longer. You're right.
38:08
Ejaaz: And in Viewslesser, which goes out to 100,000 of you every single week, twice a week.
38:14
Ejaaz: So we have an essay and we have the five weekly highlights if you don't want
38:19
Ejaaz: to watch all the videos and you just want to catch up through text.
38:21
Ejaaz: But aside from that, that is it.
38:23
Josh: That's it. We will see you. Go enjoy your weekend. We'll see you next week.
38:25
Josh: Have fun. We'll see you next week. Goodbye, everyone. We'll see you.