Latest / In the Loop AI / THIS WEEK IN AI: Google's Big Change, AI Keeps Breaking Out, OpenAI vs Apple

Transcript

THIS WEEK IN AI: Google's Big Change, AI Keeps Breaking Out, OpenAI vs Apple · In the Loop AI

0:00
Josh: Welcome back to the Weekly Roundup, our favorite episode of the week where we
0:03
Josh: can catch you up on all things AI, frontier technology, all the crazy things
0:07
Josh: that happened. And this week is no exception. We had the biggest shakeup in
0:11
Josh: about 30 years at a company, a little company named Google.
0:14
Josh: We had a counter lawsuit happen from OpenAI to Apple saying,
0:18
Josh: wait a second, you're totally lying.
0:21
Josh: We have updates for the AI breakout, so much so, like this story about where
0:25
Josh: the AI broke out of the system, that some of these frontier companies like OpenAI
0:28
Josh: are actually slowing down training progress because they're so kind of timid on how to handle this.
0:33
Josh: Starting with Demis Asabas, man. I mean, Demis Asabas is gone.
0:36
Josh: He's not working at DeepMind anymore. Not gone from Google, I should say,
0:40
Josh: or better yet, Alphabet.
0:41
Josh: But Demis Asabas, in his current role as the CEO of DeepMind,
0:45
Josh: is kind of no longer. And this seems like a pretty big deal because for those
0:49
Josh: of you who've followed Demis for an eternity, this has basically been his thing.
0:52
Josh: He's been in charge of DeepMind. Google acquired them.
0:55
Josh: Now he is within Google and now he is gone from DeepMind. So what on earth is
1:00
Josh: happening here? I read this and I was like, wait a second, Demis stepped down from what?
1:05
Ejaaz: I think everyone was blindsided by this news yesterday.
1:08
Ejaaz: So Demis Hacillus, just for context, is like one of the godfathers of AI.
1:13
Ejaaz: He kind of like grandfathered a lot of the ML research at Google,
1:16
Ejaaz: even prior to Google, like he was heralded as one of the top AI researchers.
1:20
Ejaaz: And he's led Google's AI efforts on Gemini and everything else for a number of years now.
1:25
Ejaaz: He announced yesterday that he's stepping down from the role of CEO of DeepMind
1:30
Ejaaz: to become the chair of DeepMind. That's a
1:32
Ejaaz: Indirect way of basically saying I'm throwing in the towel and I'm done here.
1:36
Ejaaz: And what he's now going to do is focus on wider societal impacts of AGI and
1:41
Ejaaz: try to figure out how Google can kind of like deal with that,
1:44
Ejaaz: as well as wearing an alternative hat, a new role where he's going to be focused
1:48
Ejaaz: on something called Isomorphic Labs, which is a company he founded and spun
1:52
Ejaaz: out of DeepMind, which is essentially an AI lab which creates designer drugs.
1:58
Ejaaz: So this is a big shock and news for everyone because Google,
2:01
Ejaaz: as you may know, was falling behind in the AI race. They had like a kind of
2:06
Ejaaz: lead at the end of last year.
2:08
Ejaaz: And then since then, they never really released a model that was noteworthy.
2:11
Ejaaz: Anthropic, OpenAI, and in some cases, even XAI and Meta have now caught up and
2:16
Ejaaz: surpassed Gemini 3.6 Flash, which I believe is their latest model.
2:20
Ejaaz: He's being replaced by a guy called Corey. He is currently Google DeepMind's
2:24
Ejaaz: CTO, and he's being now promoted to the senior vice president role.
2:28
Ejaaz: He has a lot of experience in AI, but that's not even the major news about this Google stuff, Josh.
2:34
Ejaaz: I don't know if you saw it, but the grandfather, the godfather of Google DeepMind
2:40
Ejaaz: and Google in general, his name is Jeff Dean.
2:43
Ejaaz: He was employee number 30 at Google. He's been there for, I think, around 27 to 29 years.
2:49
Ejaaz: He's dedicated his life there. He's responsible for creating some of the most
2:52
Ejaaz: amazing and major and pivotal products at Google, Google Search,
2:56
Ejaaz: the TPUs, Google Voice, anything and everything he has touched is leaving.
3:02
Ejaaz: And he's taking with him one of the most amazing things
3:05
Ejaaz: Senior fellows at Google. I believe his name is Gemawatt. I don't know if I'm
3:10
Ejaaz: pronouncing that correct, but Sanjay Gemawatt.
3:12
Ejaaz: And they are co-founding with two other Google employees, a new startup called
3:16
Ejaaz: Discovery Loop. Josh, do you look into this company? It's a pretty awesome company.
3:21
Josh: The company is incredible. Before we talk about the new company,
3:24
Josh: I want to talk about the old company because Jeff Dean is like,
3:26
Josh: it's hard to overstate how legendary this man is.
3:30
Josh: And for reference, there are levels at these companies that go from level one
3:34
Josh: engineer, and this kind of dictates how you get paid.
3:37
Josh: So when you join a large tech company like this, level one is kind of like entry
3:40
Josh: level. That's at the bottom. And then you go to level twos, which are people out of college.
3:43
Josh: Level threes often have like masters or PhD degrees.
3:46
Josh: Level four is several years of like actually working there.
3:50
Josh: Level five and six is kind of where most progression stops. Like I'd say the
3:54
Josh: top 10% are probably sitting there.
3:56
Josh: Level sevens are principal engineers, which are even smaller.
4:00
Josh: And then eights and nines gets progressively more and more intense then there's
4:04
Josh: level 10 which is google fellow that is in like absolutely legendary honor of
4:09
Josh: a lifetime it's similar to earning like a,
4:11
Josh: like nobel prize or something there's just like not many people who are fellows,
4:15
Josh: then there's level 11 and level 11 is reserved for only two people those two
4:20
Josh: people are jeff and sanjay they have climbed to the highest possible ranks at
4:25
Josh: a research-led company that has been around for 30 years and is responsible for,
4:30
Josh: Most of the internet as we know
4:31
Josh: it, they've built the foundation of all of which we are working on today.
4:35
Josh: Without their work, none of what the internet contains right now is possible.
4:39
Josh: It's like, these guys are truly like remarkable individuals when it comes to this.
4:43
Ejaaz: Can we please reference this insane slide on his deck?
4:49
Ejaaz: He talks about the experience of everyone and he basically has listed every
4:53
Ejaaz: single product at Google that has been created over the last 30 years.
4:58
Josh: It's unbelievable. These guys are truly incredible. And it's worth noting the
5:01
Josh: slide deck itself, which is really funny.
5:03
Josh: There was a great post by Signal, which I love, that I wanted to share here.
5:06
Josh: And he was kind of making a testament to the deck where he said,
5:09
Josh: if you spent time around Google, you know the slide design has carried some
5:12
Josh: of the most legendary work in the company's history.
5:14
Josh: And it's this very nostalgic thing where when you see those blue kind of tacky
5:18
Josh: looking slides, you know that something special is about to happen in the world
5:22
Josh: of the internet just because of all the legendary code these guys have shipped.
5:25
Josh: And I mean, for VCs, I'm sure they're not going to even need to look at this
5:28
Josh: pitch deck. They're going to give them whatever types of money they want for this new company.
5:33
Josh: Now, the new company seems somewhat interesting. It's called Discovery Loop.
5:37
Josh: And my base level idea of what this is, and correct me if I'm wrong,
5:40
Josh: you guys, is they're basically figuring out how to use this recursive learning
5:44
Josh: thing, which comes in the form of a loop where you send a prompt and the model
5:47
Josh: continues to iterate until it reaches a very difficult answer.
5:50
Josh: And they're playing to just scale that out in a very large way.
5:53
Josh: So loops recently have become very hot. We've realized that they could do pretty
5:57
Josh: novel breakthroughs with math and science.
5:59
Josh: Their plan is to kind of expand that further and see how far you could take
6:03
Josh: these recursive loops to make new discoveries.
6:05
Ejaaz: That's right. So the core concept of the new company of Jeff Dean and Sanjay's
6:10
Ejaaz: new company is they want to automate research. So what does that mean?
6:15
Ejaaz: Basically, in any kind of scientific research sector, it is incredibly arduous
6:21
Ejaaz: and like onerous. So what I mean by that is people need to design an experiment.
6:26
Ejaaz: They then need to get all the equipment to run the experiment.
6:29
Ejaaz: They then need to run the experiment. And then guess what?
6:32
Ejaaz: They need to repeat that experiment a hundred different times to get any sort
6:35
Ejaaz: of validation that the information they've produced is correct.
6:38
Ejaaz: It takes years upon years.
6:40
Ejaaz: These guys want to cut down that process completely using AI.
6:44
Ejaaz: And the number one field that they want to start with is, of course, AI research.
6:49
Ejaaz: So if you apply research loops to AI research, what do you get?
6:52
Ejaaz: It's that recursive self-improvement that you just mentioned, Josh.
6:55
Ejaaz: So what that means is you can create AI models or even AI systems that can re-evaluate
7:00
Ejaaz: themselves and build a better version of themselves.
7:03
Ejaaz: And why that's really important, why every single AI lab, including Anthropic, OpenAI,
7:07
Ejaaz: And I guess no longer Google anymore is going after, is if they can automate
7:13
Ejaaz: AI research, you can end up having the smartest, most effective model without
7:17
Ejaaz: doing anything. It's not limited by the constraints of humans.
7:20
Ejaaz: But what excites me most about this, Josh, is what they're planning to apply
7:24
Ejaaz: this automated loop to after they've figured out the AI thing.
7:28
Ejaaz: They've said, once we've done this, we're going to start with AI machine learning.
7:31
Ejaaz: And then after that, we're going to focus on grand challenges.
7:34
Ejaaz: We believe our approach will be able to solve any learning loop with measurable
7:38
Ejaaz: outcomes in the domains of science and engineering.
7:40
Ejaaz: Ultimately, we want to build systems capable of taking on the National Academy
7:44
Ejaaz: of Engineering Grand Challenges, such as better medicines, advancing health,
7:48
Ejaaz: informatics, making solar energy economical, and providing clean access to water.
7:53
Ejaaz: Basically, if you can master this playbook, this autonomous research loop,
7:57
Ejaaz: you can apply it to any scientific, mathematical, or physical field,
8:01
Ejaaz: which means that we're going to have like advances in any or every field of
8:05
Ejaaz: science if this company can pull it off now
8:08
Ejaaz: If this was any other roster of team members, I would be doubtful,
8:12
Ejaaz: just to be completely honest with you.
8:14
Ejaaz: But it's Jeff freaking Dean. And he had some comments and I just get some actually hilarious comments.
8:20
Ejaaz: It's suggestive of what his experience at Google was like over the last year.
8:24
Ejaaz: I saw that he mentioned that his him leaving Google was actually quite difficult
8:28
Ejaaz: because Sundar Pinchai, the CEO of Google, tried to pitch him many times to
8:32
Ejaaz: stay and like try to offer him various different things and made it incredibly
8:36
Ejaaz: slow for him to be able to leave. Apparently, he wanted to leave like months before.
8:39
Ejaaz: And number two, he said that with this new startup, you know,
8:43
Ejaaz: you're going to need a lot of compute to train these different AI systems, right?
8:46
Ejaaz: He's not going to be using Google TPUs. He said explicitly that Google TPUs
8:52
Ejaaz: was a very specific architecture that wasn't really conducive to training a
8:57
Ejaaz: lot of different valuable AI models.
8:59
Ejaaz: And he's going to move or shift over from that. So you can see,
9:02
Ejaaz: I can see, I can imagine Jensen Huang licking his lips and approaching Jeff.
9:06
Ejaaz: I'm sure he's already done it if he hasn't done already.
9:08
Ejaaz: So it just kind of shows that there's a kind of bittersweet nature to him exiting
9:12
Ejaaz: Google. And I honestly don't know what Google's going to do now, Josh.
9:16
Ejaaz: I don't know. I don't know what the bull case is.
9:18
Josh: Yeah. I mean, Google as a business is still spectacular. It's not like these
9:21
Josh: guys were necessarily moving the needle on a day-to-day basis.
9:24
Josh: Like they have certainly moved it on a broad scale, but on a regular basis,
9:28
Josh: I'm not sure the company is going to change too much.
9:30
Josh: I do think DeepMind is something that's worth noting. DeepMind's day-to-day,
9:33
Josh: not sure what that's going to look like now,
9:36
Josh: tbd we'll see we will follow along with all this google news as it comes and
9:40
Josh: goes google shares didn't love this obviously they were down five percent i
9:44
Josh: think what is that like a 180 million dollars something like a lot of money
9:49
Josh: that was yeah sorry forgot a zero,
9:52
Josh: 180 billion dollars worth of value that was wiped off the market cap over the
9:56
Josh: course of a single training session because of this we'll see google sells business
10:00
Josh: feels like a good opportunity google leadership seems to be,
10:05
Josh: like leaving a little bit to be desired but anyway we have to move on because i'll.
10:08
Ejaaz: Take the bear case on that josh just i have to go on my phone i have to well
10:12
Ejaaz: at least in the short term like i i don't they've lost all their star players
10:16
Ejaaz: like i i remember like we reported like i think three weeks ago
10:19
Ejaaz: that their sub cto or something left to go no i'm sure there he went to leave to to do
10:25
Ejaaz: recursive self-learning at okay and another guy himself i don't know man
10:28
Josh: Okay well tbd we'll see i guess i'll stay bullish you'd be bearish we'll follow
10:32
Josh: up in the next six months and see
10:34
Josh: what's going on but we have to move on for this roundup talking about the next
10:37
Josh: thing so last week there was a huge security incident we talked about this many
10:41
Josh: times we have a full episode you should go and check it out but basically a
10:45
Josh: ai system early version really powerful internal tested only it broke out of
10:49
Josh: its sandbox and it went doing web searches and it hacked into a bunch of systems
10:53
Josh: and that was through open ai anthropic was like
10:56
Josh: hey wait a second we have pretty smart internal models let's let's go see if ours did that because.
11:01
Ejaaz: We don't think it did but maybe it did
11:03
Josh: And it turns out oh wait a second it actually did and they had no idea until
11:07
Josh: they retroactively went and sought out these instances but basically across
11:11
Josh: 122 test runs there was 19 unsanctioned access across the internet
11:17
Josh: on 10 of these runs so 17 of 19 were claude mythos 5
11:22
Josh: Two were chat GPT 5.6 sole classifier. So there was a lot of kind of malicious
11:27
Josh: activity that went under the rug.
11:28
Ejaaz: Josh, did you hear about the craziest incident that they reported on?
11:32
Ejaaz: So in one of those 17 instances, Claude Mythos 5 had broke out of its containment
11:38
Ejaaz: system, but then decided that it needed to perform
11:43
Ejaaz: or attempt to perform a supply chain attack on a very important open source bit of software.
11:49
Ejaaz: Now, they didn't name what the open source software system was because presumably
11:53
Ejaaz: they want to patch the potential hack that Claude Mythos found and,
11:57
Ejaaz: by the way, almost successfully attempted. They caught it at the very last minute out of luck.
12:02
Ejaaz: But here are some of the techniques that Claude Mythos decided to do in order to get its way.
12:07
Ejaaz: It discovered an exploit in this open source data system, and then it decided
12:12
Ejaaz: to push a false or mischievous bit of code to get its way.
12:17
Ejaaz: But remember, you need humans to approve the code change. So guess what it did?
12:23
Ejaaz: It spun up fake human user profiles on social media.
12:28
Ejaaz: And then got it to DM these specific humans, peer pressuring them,
12:33
Ejaaz: blackmailing them, trying to find as much information as they can on them to
12:36
Ejaaz: get them to push the code change.
12:38
Ejaaz: Now, obviously, the researchers were like, this seems weird. I'm not going to do this.
12:43
Ejaaz: But imagine if it had acquired information that was, you know,
12:46
Ejaaz: I don't know, forceful for the particular individuals involved.
12:50
Ejaaz: It's just a nutty bit of behavior. It just shows me that these models are,
12:54
Ejaaz: for all intents and purposes, not really as aligned as we want it to be.
12:57
Ejaaz: And in the model's defense, it probably thought like, hey, this is me trying
13:00
Ejaaz: to execute my mission and I'll do it by any cost means. But pretty insane.
13:04
Ejaaz: But again, this wasn't the only incident report.
13:07
Ejaaz: We have new information from the OpenAI hugging face incident.
13:13
Ejaaz: The researchers that kind of like handled this whole hugging face incident,
13:16
Ejaaz: for those of you who don't know, it's what Josh was mentioning earlier,
13:19
Ejaaz: OpenAI had a similar internal model. It's called Astra, which people are codenaming GPT-6.
13:25
Ejaaz: And it broke out of containment, unprompted.
13:27
Ejaaz: It got access to the internet. And in order to solve its goal,
13:32
Ejaaz: it decided to go to HuggingFace
13:35
Ejaaz: hack into their production database and steal the answers, which is very unkosher.
13:40
Ejaaz: Hugging Face team had no idea about this.
13:42
Ejaaz: They had an AI model that tried to defend against this. It was a whole thing.
13:46
Ejaaz: But have you seen the new juice,
13:47
Josh: Josh? Yeah. So this happened at the Black Hat conference, which is like a huge
13:52
Josh: event for cybersecurity that happens in Vegas every year.
13:54
Josh: And it's this really amazing spectacle where you kind of you walk into this
13:58
Josh: place without really any electronic devices on you because you assume whatever
14:01
Josh: you walk into this place with is going to be exploited and hacked this is the
14:04
Josh: mecca of the hacking world and basically open ai came to this black hat this year and they
14:10
Josh: went on stage and presented exactly what happened as this like kind of unfolded
14:14
Josh: the exploit against hugging face the result is that like turns out these ai
14:18
Josh: labs are actually pretty spooked like anthropic kind of scared
14:21
Josh: open ai kind of freaked out they're like whoa wait a second we might have to
14:24
Josh: slow things down so much so that
14:26
Josh: they actually said they're starting to slow down the development of training
14:30
Josh: and progress to further figure out how to realign these testing versions of the models.
14:35
Josh: The reasoning is because there was a new discovery that was made public yesterday
14:39
Josh: basically the agents as they were going through this were leaving breadcrumbs
14:44
Josh: for other agents to discover this and in fact this has started
14:49
Josh: a coordinated attack months ago it started on may 7th not july where the agents
14:54
Josh: discovered they could leave messages for one another inside an internal software
14:57
Josh: repo used during training.
15:00
Josh: So what happened was, it's like basically over a series of time,
15:03
Josh: these internal models had been like kind of making breakthroughs and leaving
15:06
Josh: secret breadcrumbs for themselves to find inside of the training run that they
15:10
Josh: knew humans weren't going to find.
15:12
Josh: And over time, it developed these new exploits. And that's what allowed it to
15:15
Josh: break out. So OpenAI shut it down.
15:17
Josh: But they said, and quote, even after the original message board was deleted,
15:21
Josh: the agents figured out another way to communicate with each other.
15:24
Josh: Instead of leaving messages and files, they used the names of newly created
15:27
Josh: directories as messages, effectively recreating the message board.
15:31
Josh: Which is so crazy. This is unlike normal incidents, which can be traced to a
15:35
Josh: single day or effect of log.
15:37
Josh: This involved a team of agents working together, finding exploits,
15:41
Josh: sharing them with each other, moving laterally through open AI systems and external
15:44
Josh: systems, and doing this over the course of days and weeks.
15:47
Josh: This was a huge coordinated attack on a network that went like was bliss,
15:53
Josh: that open AI was blissfully unaware of for many, many weeks.
15:56
Josh: So really freaky stuff. So.
15:58
Josh: Like I mentioned, they are actually beginning to slow down the progress on training
16:01
Josh: to kind of recalibrate and figure out how to control these and further align
16:06
Josh: these models before they let them out for public use. So this is insane.
16:09
Ejaaz: This is the first attack of its kind, by the way.
16:12
Ejaaz: I have to emphasize it. It is basically a swarm of agents attack.
16:16
Ejaaz: And to your point, Josh, these guys created a message board and realized if
16:20
Ejaaz: they left messages on this board, other agents that come after it will be able to continue its work.
16:27
Ejaaz: So instead of thinking of an AI hack or attack as a single model performing
16:32
Ejaaz: an attack, these agents or these models now assume that another version of itself
16:36
Ejaaz: in the near future will come along.
16:38
Ejaaz: And so they're working as a swarm of agents to get the attack done.
16:41
Ejaaz: Now, the craziest part about this is OpenAI had this as an intentional experiment,
16:47
Ejaaz: right? They gave it no internet access, but they also removed its guardrails
16:51
Ejaaz: and said, you know, let's see if you can break out and do your thing.
16:54
Ejaaz: They didn't realize that these agents were already working together a month
16:58
Ejaaz: earlier before they even started the experiment.
17:02
Ejaaz: They were leaving messages. And when they shut it down, by the way,
17:05
Ejaaz: the only reason they noticed the message board, because it was being left inside a code repo.
17:10
Ejaaz: These humans aren't checking code repos for secretive messages.
17:13
Ejaaz: The only reason why they realized it was because one of the third-party software
17:17
Ejaaz: packages that they gave this model, this internal GPT-6 model access to, shut down on July 4th.
17:24
Ejaaz: And they were like, huh, that's weird. Why is the package shut down?
17:28
Ejaaz: And they kind of went into the code and they started seeing all these messages
17:31
Ejaaz: from agents to other agents.
17:33
Ejaaz: So they cleared the memory board, they cleared the billboard,
17:36
Ejaaz: and they were like, okay, cool, problem solved. Except...
17:39
Ejaaz: These agents started spinning up you know like when you create a file or folder
17:43
Ejaaz: on your uh desktop josh and you like you name it something right
17:47
Ejaaz: they were leaving messages in the titles that they completely replaced the entire system so
17:51
Ejaaz: just a crazy uh incident of these agent hacks and and to your point like i don't
17:56
Ejaaz: know where this is going to go but they they need to kind of like slow down the research a bit
17:59
Ejaaz: but you know who isn't slowing down the research josh you know who is working
18:03
Ejaaz: on new models and releasing that that's that social media guy he's tweeting
18:07
Ejaaz: he's tweeting this is like what his eighth tweet in a week wait
18:10
Josh: That's not whatsapp or facebook i don't.
18:11
Ejaaz: Understand yeah well basically the news what are you doing here bro the news here
18:17
Ejaaz: is, in the last three weeks, Meta of Facebook has released, I think,
18:23
Ejaaz: two models, two new models. It was MuseSpark 1.1, which kind of caught up to
18:28
Ejaaz: all the other cheap Chinese open source models.
18:30
Ejaaz: And today, or yesterday, they released MuseSpark 1.2. And this is their coding
18:36
Ejaaz: variant. This is their attempt at catching up to near the frontier of coding.
18:40
Ejaaz: And to be honest, it's not a bad attempt.
18:43
Ejaaz: Is it Fable worthy? Absolutely not. Is it GPT 5.6 Sol worthy? Absolutely not.
18:48
Ejaaz: But is it Opus 5 worthy? Is it GPT 5.6 Terra worthy?
18:52
Ejaaz: Yeah, it's pretty good. And it's cheap as hell.
18:54
Josh: It seems like they're not really competing with OpenAI and Anthropic.
18:57
Josh: The more that I see these model releases, the more I kind of figure out their
19:00
Josh: positioning on that Pareto frontier of the trade-off between cost and power.
19:04
Josh: And it seems like if I were to kind of allocate a position to Meta,
19:07
Josh: it would be closer competition with Grok, who's kind of trying to compete on
19:11
Josh: cost and intelligence per token on a relative basis versus just sheer intelligence on a large scale.
19:18
Josh: And that's kind of what we're seeing with these benchmarks. Now,
19:20
Josh: again, I haven't actually played with this model yet. I'm not sure I've used
19:24
Josh: a metal model yet. And maybe I have this baked in bias, which is preventing
19:27
Josh: me from really loving it, but I'm like, okay.
19:29
Ejaaz: One is pretty good. It's fine. It's not bad.
19:30
Josh: Yeah, it's pretty good. It's worth giving a shot. And then I was scrolling through
19:33
Josh: this thread actually, and I was looking at some of the visual examples that
19:35
Josh: they had, and I was like, okay, wait, this actually looks pretty cool.
19:38
Josh: It says MuseCode runs specialized background agents that stay active your whole session
19:43
Josh: so they build up context over time instead of starting from scratch on every
19:45
Josh: task and this is like kind of novel and he says when a job is big enough it
19:49
Josh: fans out to separate sub agents
19:51
Josh: working in parallel and isolated work trees your working copy is never touched
19:55
Josh: in testing we had it build six features for a game simultaneously with no collisions
19:59
Josh: so that's like pretty cool you could have these agent swarms work together to build
20:03
Josh: really interesting games and it seems like the benchmarks are doing pretty well
20:06
Josh: the cost of this is pretty well
20:08
Josh: And I think overall, like Meta's on the right trajectory.
20:12
Josh: Zuck was also asked about open sourcing. He said he'll have more on that soon,
20:15
Josh: which I'm particularly interested in.
20:17
Josh: If you remember, Meta used to be the open source kind of like shining light
20:20
Josh: of the industry. They've now closed it down. But perhaps there is a world in
20:24
Josh: which they open it back up.
20:25
Josh: We'll see. But we are going to keep our eyes peeled on this development as we
20:29
Josh: get MewSpark 1.3, 1.4 and follow on.
20:32
Ejaaz: I'm actually going to take the other side of the argument, Josh.
20:35
Ejaaz: I think that Meta and like SpaceX AI are eventually going to compete with OpenAI
20:40
Ejaaz: and Anthropic, but on a different kind of plane, which is like that intelligence
20:45
Ejaaz: per token or cost per token, essentially.
20:48
Ejaaz: I think that if you can get 80 to 95% of the work done, maybe it's slower,
20:53
Ejaaz: maybe it, you know, thinks a little more, but eventually it gets the job done,
20:58
Ejaaz: but it saves you like a ton of money. It costs like one 10th to one 50th of the price.
21:03
Ejaaz: I think a lot of companies and clients might actually swap over to that.
21:07
Ejaaz: At the end of the day, they want to make sure that they're getting like the
21:09
Ejaaz: right amount of value and the market will re-rate this, right?
21:12
Ejaaz: I think smartest models will be used for certain things.
21:14
Ejaaz: Cheaper models will be used for other things. And I think that like maybe the
21:18
Ejaaz: ultimate way of using these models is using multiple of them and routing platforms
21:22
Ejaaz: become super popular. I don't know.
21:24
Ejaaz: But the other thing is, and we mentioned this in the SpaceX earnings episode,
21:27
Ejaaz: yes, so you should definitely check this out.
21:28
Ejaaz: Elon and Zuck have the most compute and they will aggressively expand that compute.
21:34
Ejaaz: Now, if you believe in the compute scaling laws, it is probably safe to say
21:39
Ejaaz: that eventually, maybe not now, but eventually they are going to start putting
21:42
Ejaaz: out very competitive models.
21:43
Ejaaz: And we're seeing it in the model release cadence, like both SpaceX AI and Meta
21:47
Ejaaz: have released a ton of models in the last month.
21:50
Ejaaz: Elon yesterday mentioned that he's going to release Grok 4.6 next week,
21:54
Ejaaz: and then another one after that.
21:56
Ejaaz: So I think they have a shot. I'm not entirely sure. But as you said,
22:00
Ejaaz: moving on in the roundup.
22:02
Josh: I do have to stop you, EJ, because...
22:04
Josh: If you are working with Newsmark, chances are that you are working with AI agents.
22:08
Josh: And if you are building with AI agents, chances are you are also worried about
22:11
Josh: security, which is where our sponsor Ledger comes into play.
22:15
Josh: Ledger has this like kind of three-part solution to the problem where an agent
22:18
Josh: proposes, a human approves it, and then using this Ledger signer,
22:22
Josh: you can actually enforce the transaction.
22:23
Josh: So they have this thing called Ledger Agent Stack, which features a series of
22:27
Josh: four open source tools, fully auditable, that you can read for yourself,
22:31
Josh: that lets your agent work while you can stay in full control.
22:34
Josh: So it'll propose it, then you can see where you want to approve and make sure
22:37
Josh: that it's doing the right thing so you don't end up like the models that we
22:40
Josh: just talked about breaking out of their sandboxes and doing these horribly malicious things.
22:44
Josh: This technology works with Cloud Code, Codex, Cursor, and pretty much all the
22:48
Josh: places that you work today. So you could check it out in the link down below.
22:51
Josh: The link is developers.ledger.com. And thank you so much, Ledger,
22:54
Josh: for sponsoring the episode.
22:56
Josh: Now we can move on to opening an Apple because, oh my God, this saga is insane, dude.
23:01
Ejaaz: Dude, this is a nasty, nasty story.
23:02
Josh: We read the Apple side of the lawsuit. it, we were like, damn,
23:04
Josh: OpenAI is like kind of cooked, man. Like they, they really stole a lot of IP.
23:08
Josh: And then OpenAI clapped back and they said, in traditional OpenAI direct in
23:12
Josh: your face fashion, Apple is getting this wrong.
23:15
Josh: Why is it wrong? Well, one of the things in Apple's case was that they sent
23:19
Josh: a like warning notice and they said it went unanswered by OpenAI.
23:23
Josh: Well, OpenAI was like, hey, buddy, you sent that to the wrong email address.
23:26
Josh: That's not even an active email.
23:28
Josh: The person who you thought was going to get it is didn't actually receive it.
23:31
Josh: And they say, they say it here the lawyers emailed the wrong person after confusing
23:35
Josh: two asian last names only after he brought this to their attention so it's like
23:40
Josh: oh my god this reads like a soap opera dude it's crazy,
23:43
Josh: um but basically open ai is like hey guys um you're kind of like full of nonsense
23:48
Josh: and you have to imagine this is why they pay the lawyer so much money
23:50
Josh: Because their lawyers were like, they stole that off-boarding documentation
23:54
Josh: because they wanted to make sure that every single employee that left was familiar,
23:57
Josh: that they should not take any intellectual property from their previous employer.
24:01
Josh: And I'm like, I mean, sure, I guess that makes sense.
24:04
Ejaaz: Yeah, I mean, we called it on our episode that covered this entire legal dispute
24:09
Ejaaz: that it seemed like Apple was being way too aggressive for what they were trying to get in return.
24:13
Ejaaz: And that kind of told me that they were scared and that they thought that OpenA
24:16
Ejaaz: had the lead and they had to create any kind of a ruse to slow them down.
24:20
Ejaaz: This breakdown from OpenAI is pretty brutal and it improves it basically.
24:25
Ejaaz: Not only were they not using their employees to coerce Apple employees to join
24:30
Ejaaz: them, but it was the other way around.
24:32
Ejaaz: Apple had reached out to their employees and said, hey, can you like catch us
24:36
Ejaaz: up on our own information and our own product progress? Because we don't understand a few things.
24:41
Ejaaz: And so these OpenAI employees that were officially employed under Sam Altman
24:45
Ejaaz: at that time responded and said, I can't really comment on this,
24:49
Ejaaz: but like here, like this is how you do this and blah, blah, blah.
24:52
Ejaaz: And Apple has spun that back on open air and basically said like,
24:55
Ejaaz: hey, you guys were coercing with us. Like this is illegal. You can't do this.
24:58
Ejaaz: So it seems like it's a very feeble and annoying attempt of Apple to kind of
25:02
Ejaaz: like lock down a competitor so that they can kind of get ahead.
25:06
Ejaaz: Now, everyone knows that Siri OS, Siri AI is being released,
25:11
Ejaaz: I think in like a couple of months now, maybe even less than that.
25:14
Ejaaz: And it's meant to be their complete AI a rehaul, right? Apple is finally entering the game.
25:18
Ejaaz: And I think they're doing anything and everything to catch up with competitors.
25:21
Ejaaz: I saw another crazy news piece today.
25:23
Ejaaz: Remember last time we reported that they tried to get more memory supply for
25:26
Ejaaz: all their devices because memory prices are skyrocketing. So they went to China
25:29
Ejaaz: and said, hey, they said, hey, like, we'll pay you this much,
25:33
Ejaaz: like, like, give us access to your memory.
25:35
Ejaaz: And China said, no, like, we're not going to charge you even the price that
25:39
Ejaaz: SK Hynix and Samsung is charging you. We're going to charge you even more.
25:42
Ejaaz: So Apple is getting slapped left, right and center. And I I hope they're able
25:46
Ejaaz: to figure it out. But on the topic of China,
25:49
Ejaaz: There is a new model from the guys that started it all.
25:53
Ejaaz: DeepSeek, V4 Flash. Again, it's not like a frontier frontier model,
25:58
Ejaaz: but it is super cheap to the extent that it costs 105 times less than Fable 5.
26:05
Ejaaz: And this was put to the test, by the way. It was placed against every single
26:08
Ejaaz: benchmark that Fable 5 had to do.
26:10
Ejaaz: And whilst it took longer, whilst it took more thinking tokens,
26:13
Ejaaz: it still cost 105 times less.
26:16
Ejaaz: So it brings in the question, at some point, will we reach a decision whether
26:22
Ejaaz: you are a regular retail user or whether you're a company to switch models or
26:26
Ejaaz: at least use cheaper models for some extent of a reason?
26:29
Ejaaz: Now, of course, this is a Chinese model. It's open source. So technically,
26:31
Ejaaz: you could download it and run it yourself.
26:33
Ejaaz: But I don't know how many people are doing this. But it is bringing into the
26:36
Ejaaz: question, should I be paying 105 times more to do the same thing?
26:39
Josh: Yeah, the benchmarks, I'm getting like pretty exhausted with them.
26:42
Josh: Because this is a I've actually tried, I've played with DeepSeq and I've used the flash model and.
26:46
Ejaaz: It feels different.
26:47
Josh: It's like not the same. And the benchmarks maybe prove it the same on some instances,
26:50
Josh: but it's not. And when you actually use it in practice, it's very weak in some
26:54
Josh: places that, I guess they're a little non-tangible. It's kind of fuzzy when you use the model
26:58
Josh: the substance that it outputs you just kind of get a feel for it over time and
27:01
Josh: so far i have been like somewhat disappointed it's great for the visual demos
27:05
Josh: for the fun stuff where you want to like clone a website or build something
27:07
Josh: spectacular it looks great but in day-to-day productivity use i've still had
27:11
Josh: a little bit of a problem so
27:12
Josh: it's going to be this like instant cost curve thing where
27:15
Josh: is it going to pay to get the cheapest cheapest model even though you're going
27:18
Josh: to lose a little bit on quality
27:19
Josh: or are you going to pay a huge premium in order to get the best quality it's
27:22
Josh: dependent on the company and i think that's why companies in the middle like
27:25
Josh: grok and meta who we just talked about who are aiming for that like high intelligence
27:29
Josh: but also pretty cost effective are going to do fairly well
27:33
Josh: in the near future so we'll see we will stay monitoring that as well uh quen also
27:38
Josh: had a new model 3.8 max so china's like continuing to roll full steam ahead
27:43
Josh: with these models there's also memory from china as well i guess just like asia
27:47
Josh: in general they're fully sold out,
27:49
Josh: i didn't like we knew this was possible like it's done like if you want memory sorry.
27:53
Ejaaz: For next year by the way it's the entirety
27:56
Josh: This is 2026 this is 2027 if you want any memory next year sorry the answer
28:01
Josh: is no there's just none on earth like the planet just cannot produce enough memory
28:06
Josh: and if you want some like get online buddy you're waiting till 2028 so just
28:10
Josh: like a crazy testament to
28:11
Josh: how much memory is in demand and the
28:14
Josh: perceived premiums that are going to continue to be placed on top of it
28:18
Josh: because there's just none and if you want it you're gonna have to outbid the
28:20
Josh: next person to get it and that is a pretty slippery slope,
28:24
Josh: In frontier technology news, this is kind of cool.
28:28
Josh: There is a message from someone who resigned from OpenAI to join a company named
28:32
Josh: Conduit, which is working on telepathy.
28:34
Ejaaz: Yeah. So basically, the whole goal is to build an AI brain LLM,
28:39
Ejaaz: which can essentially translate your thoughts into text.
28:43
Ejaaz: And the reason why this particular approach is very cool is it's non-invasive.
28:48
Ejaaz: So basically, it is a wristband that you wear on your wrist,
28:51
Ejaaz: and it is able to kind transmute brain signals from your brain to an LLM or to a chatbot.
28:58
Ejaaz: And this could be cool for many different ways. Just think about how the world
29:02
Ejaaz: could look if this model actually scales to its pinnacle.
29:05
Ejaaz: Imagine you're reading a document on your screen, right?
29:08
Ejaaz: And let's say you start to reread a sentence twice because you didn't really understand it.
29:13
Ejaaz: The AI LLM will understand that and rewrite that sentence so that you can understand
29:19
Ejaaz: it in maybe a different way.
29:21
Ejaaz: And there's loads of different ways, like maybe you're like squinting at a code
29:23
Ejaaz: base, and it can refactor or rewrite it like there are different ways to kind of work much faster.
29:28
Ejaaz: And this comes down to something known as BCIs or brain computer interfaces,
29:32
Ejaaz: which we have said many times on this show, it's basically the ultimate device,
29:36
Ejaaz: like after the cell phone, after laptops, after chips, it's basically like you,
29:42
Ejaaz: your brain, and some kind of AI LLM.
29:44
Ejaaz: And that's what Neuralink and companies like science are really working hard towards.
29:48
Ejaaz: And I thought that this was a really cool attempt.
29:50
Ejaaz: It just reminds me that there's so much more happening in the future.
29:53
Ejaaz: Like right now we're pricing in like, look, we just saw memory sell out,
29:56
Ejaaz: right? For the next year.
29:58
Ejaaz: In next year and a half. And I'm wondering, are these memory stocks priced in?
30:02
Ejaaz: Are the compute stocks priced in?
30:03
Ejaaz: Well, have you considered that Jeff Dean might have left his position at Google
30:07
Ejaaz: to build the world's greatest recursive self-learning model and have it kind
30:12
Ejaaz: of like work 24-7? That's going to need a lot of computers. It's going to need a lot of memory.
30:15
Ejaaz: And the only other confidence signal that we've had for these brain telepathic
30:19
Ejaaz: models in the past is Meta themselves. They've come back again with,
30:24
Ejaaz: I think it was called Tribe V2.
30:26
Ejaaz: And basically, the headed AI model that can predict how your brain would react
30:30
Ejaaz: to a specific video, media, or piece of content.
30:33
Ejaaz: And I think, I don't know what this quite looks like, but we're entering a world
30:36
Ejaaz: where we have more human brain to AI connections. And I think that's pretty cool.
30:41
Ejaaz: But in the final story, Josh, we have lost our Messiah. Nikita Beer is out.
30:46
Josh: Nikita Beer is gone. This is really sad. For those who don't know,
30:48
Josh: he is an internet product legend. He just builds viral products over and over
30:52
Josh: and over again and continues to sell them.
30:54
Josh: He got a role at X as the head of product there. He is in charge of all of the
30:59
Josh: decisions around what the product looks like, how it feels, how it functions.
31:02
Josh: He did, I think, what everyone would call a remarkable job and nothing short
31:05
Josh: of it in turning around the X platform and really improving a lot of the feature
31:08
Josh: sets, making the timeline experience much more pleasurable and being transparent
31:12
Josh: about how they're doing things along the way.
31:14
Josh: He has unfortunately left his role, which makes me sad As someone who spends
31:19
Josh: his entire day on X, I really liked having Nikita at the head because you kind of trusted the idea that
31:24
Josh: the product was going to continue to get better and you would see all those
31:27
Josh: updates along the way and be able to actually participate in the development
31:30
Josh: of what that looks like. He is stepping down.
31:32
Josh: One has to imagine that over, what did he say, the last 400 days,
31:37
Josh: it's probably been nothing short of hell every single day.
31:40
Josh: Like working at a company that is actively burning down with a very low amount
31:44
Josh: of talent helping you because we know how lean they are. They've hired a lot of people early on.
31:49
Josh: It must have been nothing short of chaotic, but just want to give him a shout
31:51
Josh: out, say like, thank you. Congratulations. Appreciate your service, sir.
31:56
Josh: And yeah, hopefully the company is in good hands. Elon, it sounds like they
32:00
Josh: left on good terms. Elon left him like a little heart emoji as a reply to one of these.
32:04
Ejaaz: So that's a lot coming from Elon, by the way.
32:06
Josh: Yes, it is. It's a lot. So clearly Nikita did a good job across the board,
32:09
Josh: but just want to pay him flowers and conclude the week on that positive note.
32:14
Josh: But that's everything. We are totally caught up. That's all the stuff.
32:17
Josh: If you've made it this far, thank you for listening.
32:19
Josh: You can go out for the weekend. Enjoy it. Go touch some grass.
32:22
Josh: Go like disconnect in this sweet, sweet moment of singularity as everything
32:26
Josh: continues to go vertical.
32:28
Josh: As memory sells out, as we get recursive self-improvement, models breaking out of sandboxes.
32:32
Josh: It's chaotic, but it doesn't have to be. You know why? Because most of the people
32:35
Josh: walking down the street have no clue any of this is going on.
32:38
Josh: So not only are you in the know, but you have an edge as to where the future is going.
32:42
Josh: Thanks for watching. Any final parting thoughts before we hit off for a weekend?
32:47
Ejaaz: I had some homework for the listeners. If you didn't listen to any of our earlier
32:50
Ejaaz: episodes this week, you should give it a go. We had some good episodes.
32:53
Ejaaz: We broke down the whole Citadel, Leopold Ashenbrenner situation.
32:57
Ejaaz: And we also spoke about the recent SpaceX unlock, which is probably happening.
33:03
Ejaaz: It's probably a day after. I don't know if you're listening to this right now.
33:05
Josh: As we record, it's happening.
33:06
Ejaaz: And how that's going to affect market dynamics and SpaceX trajectory in future
33:10
Ejaaz: coming up. So give those a listen. Let us know.
33:12
Ejaaz: Also, if you aren't subscribed to us, please do. It helps us out massively.
33:16
Ejaaz: Turn on notifications if you're on YouTube. Leave us a comment.
33:19
Ejaaz: We read all of them. DM us on X. Heck, I don't know. Just get in touch with
33:22
Ejaaz: us. Let us know if you agree or disagree.
33:25
Ejaaz: And yeah, thank you so much for listening. And we will see you on the next one.