Latest / In the Loop AI / THIS WEEK IN AI: OpenAI Pauses Research, Huge Medical Breakthrough, Unitree IPO

Transcript

THIS WEEK IN AI: OpenAI Pauses Research, Huge Medical Breakthrough, Unitree IPO · In the Loop AI

0:00
Josh: This week will be known as the week that Sam Altman said, stop.
0:03
Josh: He said, stop, stop, stop, no more. We must put a pause on this.
0:07
Josh: It is time to be responsible. We need to slow things down a little bit.
0:10
Josh: What I'm talking about is a recent article that was just published through OpenAI
0:14
Josh: that was talking about the most recent breakthroughs as it relates to their OpenAI models.
0:19
Josh: And they were a little scarier than I think a lot of people would have hoped.
0:22
Josh: We talked about the Hugging Face incident, which happened a couple of weeks
0:25
Josh: ago, a couple of months ago. It was only a couple of weeks ago they actually
0:27
Josh: discovered it, even though it happened far before that.
0:29
Josh: And it set the precedent for this kind of unshaky ground that they are building
0:33
Josh: the next frontier models on top of.
0:35
Josh: And it seemed like it got to a point where enough was enough.
0:38
Josh: They decided to shut things down.
0:39
Josh: So what happened? OpenAI said, we temporarily paused reinforcement learning
0:43
Josh: training on our latest models intended for deployment for two weeks while we
0:46
Josh: hardened and red teamed our research environments and expanded monitoring coverage.
0:51
Josh: Our largest planned frontier reinforcement learning run remains on hold,
0:55
Josh: while smaller-scale trailing and evaluations validate these safeguards and establish
0:58
Josh: more evidence of alignment.
0:59
Josh: They've stopped everything. And granted, I will say, this isn't for the new
1:02
Josh: models that are coming out. This is for the models after that.
1:05
Josh: And it's weird to hear Frontier Lab say this, because I believe this is pretty
1:08
Josh: much the first time in history they've deliberately paused progress successfully.
1:12
Josh: Of developing these new frontier models.
1:14
Ejaaz: Put yourself in the shoes of Sam Olman right now. You're the CEO of one of the
1:18
Ejaaz: hottest private companies in the world, and you're going head to head against Anthropic.
1:23
Ejaaz: The only thing you should be focused on doing is building the next best AI model.
1:27
Ejaaz: It's going to dictate not just how well your company does, but how everything
1:31
Ejaaz: else progresses in every single other sector, right? So the single most important
1:35
Ejaaz: thing is to build the next best AI model.
1:37
Ejaaz: The fact that he's paused this, and it's been for two weeks and counting right
1:41
Ejaaz: now is a massive deal. And we haven't seen this across any other kind of AI lab before.
1:47
Ejaaz: But the reasonings are twofold. There's two events in particular.
1:50
Ejaaz: You mentioned one already.
1:51
Ejaaz: One was the hugging face incident. This is the case where an internal unreleased
1:55
Ejaaz: model, this is like their mythos grade level model, escaped out of containment,
2:00
Ejaaz: hacked into a company's database and stole information to help serve its own goals.
2:06
Ejaaz: Sam Altman and the rest of the OpenAI team realized this months after the hack
2:10
Ejaaz: actually was in progress and realized that the model that they were building
2:13
Ejaaz: was incredibly misaligned. So they put a pause there. That was in bet number one.
2:17
Ejaaz: Event number two is there is a new GPT model that is internally being made. It's codenamed Astra.
2:23
Ejaaz: Some people are calling it GPT-6.
2:26
Ejaaz: And according to internal tests from the OpenAI research team,
2:30
Ejaaz: it is their most misaligned model yet.
2:32
Ejaaz: But it's really sneaky. It tries to evade every single human researcher's attempt
2:37
Ejaaz: to try and see its internal thoughts. So it tries to hide its thoughts and it's very good at doing it.
2:42
Ejaaz: So it breached OpenAI's internal framework for misalignment.
2:45
Ejaaz: And so they've put a pause on it because they cannot release this model to the
2:49
Ejaaz: public out of the fear that it would wreak havoc.
2:52
Ejaaz: Now, there's a few ways that they're looking to kind of like resolve this right now.
2:56
Ejaaz: And one of the main ways is just monitoring the model. So think about this, right?
3:02
Ejaaz: Do its thing internally, and they're trying to see where it ends up becoming sneaky.
3:06
Ejaaz: Josh, guess how much of their compute budget they are spending just to monitor
3:12
Ejaaz: an internal model that is not making them any money at all.
3:15
Josh: So unfortunately, I did read this essay, and I know the answer is 20%.
3:18
Josh: One in five dollars spent is fairly high.
3:22
Ejaaz: You could be using this compute to serve it to more customers because the demand
3:26
Ejaaz: is insatiable right now. They could be making a heck ton more money.
3:29
Ejaaz: This is probably on the order of like billions of dollars. they could be making more money, right?
3:33
Ejaaz: Or they could be using it to train a smarter model to keep up and beat Anthropik.
3:38
Ejaaz: Instead, they're spending this very expensive compute to monitor a model that
3:43
Ejaaz: they can't even release to the public out of the guise of safety.
3:46
Ejaaz: Now, for all intents and purposes, Sam Allman has received a lot of backlash
3:50
Ejaaz: in the past about not being in favor of humanity and like kind of being this
3:54
Ejaaz: like evil conspirator type of guy.
3:56
Ejaaz: This is a clear case where he's done the opposite out of fear that the model
4:00
Ejaaz: that he's building is actually quite dangerous.
4:02
Josh: I respect it. Yeah, you got to be careful about what happens with these things.
4:05
Josh: And I know a lot of people on the
4:06
Josh: show, they're like, you guys always talk positively about these models.
4:10
Josh: You're always so optimistic. Let me give you a little dose of what could happen.
4:13
Josh: Let's just do a little hypothetical situation here around particularly viruses,
4:17
Josh: because I think viruses is like an interesting thought experiment that not a
4:20
Josh: lot of people have gone through.
4:21
Josh: So I'll kind of advocate for Sam on his behalf here. And we could just go through
4:24
Josh: a hypothetical situation.
4:26
Josh: There's really three times, three types of exploits that you can experience
4:30
Josh: with AI models, the first of which we've seen a lot, and that's kind of what
4:32
Josh: they're defending against, which is AI writes this ordinary malware for a human attacker.
4:37
Josh: It becomes a lot faster. It becomes a lot cheaper. We've seen what a program
4:40
Josh: like Mythos could do. We've seen what GPT 5.6, like the cyber version of that
4:45
Josh: is capable of doing. They're able to.
4:47
Josh: Surface multi-day or multi-zero-day exploits and chain them together in order
4:51
Josh: to exploit a system. Okay, that's like pretty bad.
4:54
Josh: But what happens past that? Well, the second kind of exploit that we could see
4:57
Josh: with these AI models is a self-replicating prompt.
5:01
Josh: So this is kind of like a true AI virus. And this is a plausible outcome that
5:05
Josh: could happen if these models get too powerful too quickly, where a normal prompt
5:09
Josh: is the instructions to make a model produce an answer.
5:12
Josh: So there's a self-replicating prompt that makes the model produce another prompt itself.
5:16
Josh: And we've seen this happen in emails for example like there was a researcher
5:20
Josh: i think last year who created this malicious email that had a specific prompt
5:24
Josh: in it that any ai assistant who read it.
5:27
Josh: Obeyed it and then copied it and did everything on its behalf so in the case
5:31
Josh: that you receive any sort of text to your machine that has this specific prompt
5:33
Josh: in it it will do that thing now why is this scary these models have safeguards
5:37
Josh: right they're they're going to block you from injection prompting
5:39
Josh: but there are these things called jailbreaks and a jailbreak prompt allows you
5:43
Josh: to insert very specific words and characters in order to trick a model into
5:47
Josh: doing something that it doesn't want to do.
5:49
Josh: Traditional computers that haven't been hardened to this aren't very smart,
5:52
Josh: aren't very capable of defending against this. That's version two.
5:55
Josh: And then version three is the scariest one, which is a reasoning worm,
5:59
Josh: basically. And the way it works is it arrives on a machine and then it looks
6:03
Josh: around the machine and it develops an attack on the spot.
6:06
Josh: And now this currently isn't capable because in order to run these models, you have to do it on
6:09
Josh: humongous servers but what happens when the quality of
6:13
Josh: a mythos grade model gets reduced to the amount that's needed to run a local
6:17
Josh: apple model well then you could get these on-site exploits where if you leak
6:22
Josh: some of this model or some of this malware into a machine it's able to look around.
6:26
Josh: Take those zero day exploits, string them together and attack the machine in
6:29
Josh: any way that it believes it's possible according to whatever its mission is.
6:33
Josh: So there's a lot of these scary things that can happen. And I think it's,
6:36
Josh: it's tough to think about it this way, because we, when we think of models,
6:39
Josh: we very much think of chatbots and we think of just these productive workers who are able to help us.
6:43
Josh: But when they are truly agentic and they're able to go on these long run tasks
6:47
Josh: and they have very high intelligence, there's a lot of edges that we can't see.
6:51
Josh: And Sam is like very much aware of that. And I think that's what they're hedging
6:54
Josh: against. So there's a doomer case for you. Okay, what if we get viruses similar
6:58
Josh: to the early internet days? Like that could go pretty bad pretty quickly.
7:01
Ejaaz: As you were describing those different sort of attack scenarios,
7:05
Ejaaz: I couldn't help but think that there were elements of each of those things in
7:09
Ejaaz: the recent Hugging Face incident, right? Like it was in its own contained environment.
7:13
Ejaaz: It figured out that it needed to solve its goal by breaking out of containment.
7:17
Ejaaz: It got access to other open AI servers and computers.
7:21
Ejaaz: It did so but conspicuously, by the way.
7:23
Ejaaz: And it was able to find access to the internet through that realm.
7:26
Ejaaz: But I think an important detail that a lot of people miss is
7:30
Ejaaz: It didn't just do it as a single model. It left messages behind on a message board.
7:35
Ejaaz: We covered this on a previous episode where these agents or these AI models
7:40
Ejaaz: are no longer acting solo.
7:43
Ejaaz: They are acting with the expectation that another agent or model will come after
7:47
Ejaaz: it and resume its mission.
7:49
Ejaaz: So it's like this swarm attack type issue, which actually ends up becoming the
7:53
Ejaaz: craziest type of attack interface for me.
7:56
Ejaaz: And I think that we're gonna see a lot more of those types of attacks surface
8:01
Ejaaz: with these newer open source models that some of the Chinese labs are releasing,
8:05
Ejaaz: like with GLM 5.3, Kimi-K 2.7, or I can't remember which version it is.
8:09
Ejaaz: These models are highly capable of performing similar attacks, similar chain attacks.
8:13
Ejaaz: One thing that's interesting in the report that OpenAI released about,
8:17
Ejaaz: you know, slowing down this training run is one of their proactive actions to
8:22
Ejaaz: try and prevent misalignment is getting another AI model
8:27
Ejaaz: To look at the AI model and make sure that it's behaving super well.
8:31
Ejaaz: So the only thing that separates us from a catastrophic attack is another AI
8:36
Ejaaz: model that is also capable of doing the same kind of attack,
8:39
Ejaaz: being prompt engineered to not do that, to basically be good.
8:43
Ejaaz: We're basically prompting ethics into these models right now.
8:46
Ejaaz: I don't really know how to think about that. It makes me feel slightly uncomfortable
8:49
Ejaaz: that we are only maybe a sentence or two away from, you know,
8:54
Ejaaz: some catastrophic event happening.
8:56
Ejaaz: But it kind of makes sense. If you are scaling AI as quickly as we are,
9:00
Ejaaz: Anthropic hasn't announced anything about slowing down research on their side
9:04
Ejaaz: versus Anthropic who has.
9:06
Ejaaz: These are two different strategic approaches and they're pros and cons to both of these things.
9:09
Ejaaz: I fear that the only way to combat any kind of dangerous scenario is to fight
9:15
Ejaaz: it with AI models that are good. It's kind of like the good versus bad.
9:18
Ejaaz: And I still don't know where the moral and ethics conversation ends with all of this.
9:22
Josh: Well, we're seeing what happens when an AI comes up against a human-made piece
9:27
Josh: of code. And it's very clear the AI model destroys the human-made piece of code
9:32
Josh: every single time. That's why we have all these,
9:35
Josh: conversation is so important and it's so important to kind of just like
9:38
Josh: bring everyone into the fold because everyone is affected by this to some extent
9:42
Josh: it's like we see what happens when these cyber models get released onto the
9:45
Josh: world they just are able to enter and access whatever they please because we
9:49
Josh: are just humans we have we make mistakes we have these like
9:53
Josh: limits we have these constraints we have patience we have all these things that
9:56
Josh: make us human the ai does not exhibit that allows it to be so far superior
10:00
Josh: in so many things specifically with ones with verifiable outcomes things like cyber.
10:03
Josh: And that's why I think we're going to see a lot of cyber stuff coming along the pipe.
10:08
Josh: The other thing that it's really good at, the other verifiable industry that
10:12
Josh: we see a lot of is bio, actually, because bio has very clear math underlying
10:17
Josh: it. It's a lot of just like chemistry, a lot of cells, a lot of charts.
10:20
Josh: And the next piece of news on the weekly docket relates to bio and a huge breakthrough
10:25
Josh: that we had just yesterday.
10:26
Josh: At the time of recording this, perhaps it was yesterday you'll be hearing it
10:29
Josh: two days later moderna and merck today they released a phase three test result
10:34
Josh: trial that basically offered a.
10:37
Josh: Partial cure to a specific type of cancer using hyper specific treatment it's
10:42
Josh: really interesting so basically what they did like loose tldr is
10:47
Josh: chemo kills these like healthy cancer cells together so basically when you get
10:50
Josh: chemotherapy if you're a cancer patient it kills healthy cells and it kills
10:53
Josh: the cancer cells this is problematic because you don't really want to lose your healthy cells.
10:57
Josh: So this process, it trains your immune system to hunt for mutations that exist only on the tumor.
11:03
Josh: They use this, you have to assume AI machine learning, but basically what it
11:06
Josh: is, is they take a tumor sample from your body that is cancerous.
11:09
Josh: They sequence it and then an algorithm detects which mutations are most likely
11:12
Josh: to trigger an immune response.
11:14
Josh: And then it creates this hyper-specific treatment just for you.
11:17
Josh: So it only targets the cancerous cells because it has broken down and fully
11:20
Josh: understands exactly the type of cell that it's seeking out. It's like...
11:24
Josh: Like a dog, like a police dog, when it's going to like chase someone,
11:27
Josh: you give it like the little scent of the person, then it goes and runs after them.
11:30
Josh: That's what this is for these cells. It's like you give them the scent.
11:32
Josh: Here's the blueprint of what you're looking for. Now go out and kill it.
11:35
Josh: And it worked. And the stock was up like 300% a day. It's been unbelievable.
11:39
Ejaaz: For anyone who was positioned in Moderna stock, which by the way,
11:42
Ejaaz: has been underperforming ever since their COVID rush, like what is it?
11:46
Ejaaz: Four or five years ago, you are now up, I think, 250%, including after hours
11:50
Ejaaz: trading, which is just insane in general. But just to go back to the story and
11:55
Ejaaz: kind of like weed in the AI angle here.
11:58
Ejaaz: And I don't want to underplay this. The biggest breakthrough in cancer research
12:03
Ejaaz: probably ever. Like this is a huge thing. Phase three, there's four phases in total.
12:08
Ejaaz: Phase three is like the penultimate stage before you release it very widely to the public.
12:12
Ejaaz: And phase three is arguably the hardest. You test it against anywhere between
12:16
Ejaaz: 1,000 to 100,000 patients. So you have a really wide data set as to whether
12:21
Ejaaz: this cancer vaccine will actually work for the wider public.
12:25
Ejaaz: And the answers supposedly, according to this announcement, is it's really good.
12:29
Ejaaz: Just to give you an idea about where AI was used to kind of make this happen.
12:33
Ejaaz: This cancer that they tested it against is melanoma specifically.
12:36
Ejaaz: And the way that they tested against this is they extract the tumor for patients that have melanoma.
12:42
Ejaaz: They sequence the genetics of that tumor cell and they compare it to a healthy
12:48
Ejaaz: cell of the exact same patient.
12:50
Ejaaz: They compare them both with each other and they used AI models and AI algorithms
12:53
Ejaaz: to basically say what's different between these genes, right?
12:57
Ejaaz: And they find these mutations. I think it was up to like 34 of these mutations.
13:01
Ejaaz: This is all the AR models doing this, by the way.
13:03
Ejaaz: And the 34 mutations that they picked was basically the proteins that would
13:08
Ejaaz: enact the largest immune response in a patient if they were to inject that back into a patient.
13:14
Ejaaz: And you might be asking, why on earth do they want the largest immune response?
13:17
Ejaaz: Well, the issue with cancer is the cells look exactly like all your other healthy
13:21
Ejaaz: cells. So your immune system doesn't actually respond to the bad cells.
13:24
Ejaaz: In this way, if they created a personalized mRNA. And by the way,
13:28
Ejaaz: mRNA is basically like a template to basically tell your immune system,
13:32
Ejaaz: hey, these are the bad guys, go and attack them. Like you mentioned earlier,
13:35
Ejaaz: it's kind of like a crime scene, basically.
13:36
Ejaaz: They create personalized medicine. So you might be asking, well,
13:39
Ejaaz: why can't they just create a universal vaccine?
13:42
Ejaaz: Well, the issue is with one patient A that has melanoma and patient B,
13:46
Ejaaz: they both react very differently. You can't have the same cure for each one.
13:50
Ejaaz: It needs to be personalized.
13:52
Ejaaz: And so a few trends have come into effect here. One, AI has accelerated the
13:57
Ejaaz: intellectual ability to be able to identify what the most effective vaccine will be, and two,
14:03
Ejaaz: But biosciences in general have scaled to a point where sequencing DNA and generating
14:08
Ejaaz: DNA is so cheap. I can't remember what the exact stat was, but I think about
14:13
Ejaaz: a decade ago, to sequence your own DNA, it cost hundreds of thousands of dollars.
14:17
Ejaaz: Now it just cost a couple hundred bucks, if that.
14:20
Ejaaz: And so we see these scaling laws, not just in AI, but in everything else.
14:23
Ejaaz: And I think this is like one of the most fascinating stories ever.
14:26
Ejaaz: As a former biologist, I'm like super pumped by this. I did some work on like
14:30
Ejaaz: cancer cells, nowhere near as cool as this, but I never thought we'd see the
14:34
Ejaaz: day in my lifetime, at least that we would get to this point, but it's so cool.
14:37
Josh: So we're getting the same type of technology improvements to biology as we are
14:42
Josh: to the rest of the world. And we're starting to see what that actually looks
14:44
Josh: like. In 1990, this was the first human genome project ever to happen.
14:49
Josh: It took 13 years to sequence genome and that cost nearly $3 billion adjusted
14:53
Josh: for inflation. So a tremendous amount of money.
14:56
Josh: Today, we're able to finish that process in a few hours for perhaps less than a thousand dollars now
15:02
Josh: and soon that cost is going to decrease to hundreds soon to tens soon it will
15:06
Josh: be minutes and a couple bucks and you'll have an app on your phone that could
15:10
Josh: sequence your dna in no time and that decrease in costs and the decrease in time required to
15:16
Josh: do these things allows the economies of scale to rise to get specific treatments
15:20
Josh: like the one that we're describing here where every single person get their
15:23
Josh: own hyper-specific treatment because at the end of the day, that is the only path forward.
15:28
Josh: Like when you think about traditional cancer treatments, they are all broad
15:31
Josh: sweeping. It's one of a kind and like the most popular one, chemotherapy,
15:34
Josh: it just wipes everything out for everyone. It doesn't care what type of disease you have.
15:37
Josh: It just goes and it takes it out. This is a really big deal.
15:40
Josh: Approvals, they're saying could come as soon as 2027.
15:43
Ejaaz: They're also working on all the cancer types as well, right? Yes. That's like sick.
15:47
Josh: Melanoma, like the first one. Melanoma, most popular form of skin cancer.
15:50
Josh: I believe the number is...
15:52
Josh: There's about 100,000 plus people that get diagnosed a year.
15:57
Josh: Of those people, 8,500 die.
15:59
Ejaaz: In America or like in the world?
16:01
Josh: I think this is the world. Wow, okay. People suffer from this.
16:06
Josh: And this is the first type. And you have to imagine that they could solve skin
16:09
Josh: cancer. They could do it for many other things. So it's really,
16:12
Josh: really exciting. Elon has a comment here that we have on screen and says,
16:14
Josh: despite its obvious misuse during COVID, mRNA has tremendous promise for curing disease.
16:19
Josh: Artificial RNA essentially makes curing diseases a software problem.
16:22
Josh: And we know how we well this is me saying it we know how we take care of software
16:25
Josh: problems and now that we have these ai models that are capable of doing these,
16:29
Josh: one processing tremendously on context windows so it can fit a dna sequence
16:32
Josh: inside of a singular context window but two having the inference ability and
16:36
Josh: the understanding to actually make use of that info and do something with it
16:39
Josh: it's a really unbelievable time for biology and if you are a computer scientist
16:43
Josh: suddenly the world of bio is looking very interesting and very compelling and
16:46
Josh: if you are an ai company well,
16:49
Josh: What's the best way to explain to people how AI can actually be beneficial for
16:53
Josh: the world? It's to actually go out and do something beneficial for the world.
16:56
Josh: So we're right on the cusp of, I think, a lot of these things happening in a
16:58
Josh: really exciting way. This is awesome.
17:00
Ejaaz: Okay, moving on. Our friends, or rather we are the fans, of this company Etch,
17:06
Ejaaz: which is basically an AI chip startup
17:09
Ejaaz: founded by two or three MIT dropouts,
17:12
Ejaaz: which is going head-to-head with NVIDIA, is back. And guess what?
17:17
Ejaaz: They have raised another couple hundred million dollars, in this case,
17:20
Ejaaz: 700 million dollars, and have reached an astounding valuation of 21 billion dollars.
17:26
Ejaaz: But that's not the headline from this announcement. The headline is this last
17:30
Ejaaz: sentence over here that I'm showing you on screen, which is,
17:32
Ejaaz: We're also excited to share that we've shipped our first rack,
17:35
Ejaaz: custom GPU rack, to Jane Street.
17:37
Ejaaz: Jane Street is basically a quant fund on Wall Street. They are very good at what they do.
17:43
Ejaaz: And the fact that they are now leveraging this custom AI silicon chip to have
17:48
Ejaaz: better AI trading algos is a pretty big milestone.
17:51
Ejaaz: Now, just to kind of recap what Etch does, if you think about NVIDIA,
17:55
Ejaaz: NVIDIA is like the king of GPUs. GPUs are required to train and inference AI models.
18:00
Ejaaz: Now, their GPUs are very good, but their general purpose, which is good because
18:04
Ejaaz: you can run it on any kind of model, but there are certain trade-offs that are
18:08
Ejaaz: made in terms of like cost, power and efficiency.
18:12
Ejaaz: The folks at Etched looked at this problem and said, we can build a better custom
18:17
Ejaaz: chip that is hard-coded to the specific model that you are building.
18:20
Ejaaz: So specific for your OpenAI model, specific for your Anthropic Claude model,
18:24
Ejaaz: and it will be half or maybe even more cheaper than what you were spending already.
18:30
Ejaaz: And it'll be lightning quick. There are certain optimizations that we can make.
18:34
Ejaaz: This is a very hard problem, a huge undertaking. We listened to some of the
18:39
Ejaaz: founders' stories, and they received no's from some of the smartest GPU makers
18:43
Ejaaz: in the world. But despite that, they put their heads down and they built this custom chip.
18:48
Ejaaz: And it turns out that they're also the fastest to deliver. Typically,
18:51
Ejaaz: AI chips take around three to five years to go from concept to delivery.
18:56
Ejaaz: These guys have done it in, I think, just under two years, which is absolutely astounding.
19:00
Ejaaz: And we're about to see it in production. So just a very impressive announcement from the Edge team.
19:04
Josh: It's unbelievable. I've spent so much time looking. I mean, we both do spend
19:07
Josh: so much time looking at all of these companies that are spawning in the world
19:10
Josh: of AI. And there is thousands, if not tens of thousands of people working on
19:14
Josh: interesting problems to solve interesting, to reach interesting solutions. And,
19:19
Josh: Etched is very different. So different that we actually recorded an entire episode
19:23
Josh: about this weird random startup last month.
19:26
Josh: And since recording that episode, EJ's, I don't know if you know this,
19:28
Josh: the valuation has doubled in the last month. When we talked about them,
19:32
Josh: just not even like four or five weeks ago, it was at a $10 billion valuation.
19:37
Josh: Now they're trading at 21 billion.
19:38
Josh: And it's very obvious once you listen to these people who are on the team talk
19:41
Josh: about the product, how they've arrived there, their conclusions,
19:44
Josh: the research that was required to actually reached this point where it's very
19:47
Josh: much the real deal this is a company that anyone who works in ai or semiconductors
19:51
Josh: or hardware would want to make
19:53
Josh: they were just unable to because of the complexity and the difficulty that this
19:57
Josh: team has been able to solve and as a result they have this unbelievable product
20:00
Josh: that is specifically tuned to run
20:02
Josh: large language model transformer infrastructure right on their chips and they
20:06
Josh: are going to worth many many multiples more than they are today
20:09
Josh: and i mean just a congratulations to them but also for anybody who's
20:12
Josh: listening if you want to see what a badass team of high agency,
20:17
Josh: high intelligence founders actually looks like.
20:19
Josh: Read up about the co-founders, read up about the story. It is an unbelievable
20:23
Josh: company that is destined for an incredibly bright future. I would love to be an investor in them.
20:27
Ejaaz: I was about to say, self-serving for Josh and I, if anyone listening has access
20:31
Ejaaz: to this deal, please reach out. We would love to be investors.
20:35
Josh: I am obsessed with these guys in a way that not many people have captivated
20:39
Josh: the attention. It's pretty amazing. And you know what you could do with all those chips, EJS?
20:42
Josh: You can run a tremendous amount of agents. And if you are running agents on
20:46
Josh: all of these chips and you're building for AI agents, chances are you may want
20:49
Josh: to hear about our sponsor, Ledger.
20:51
Josh: Because if you're building with AI agents, chances are you're worried about
20:53
Josh: security. We spoke about this earlier in the,
20:56
Josh: ledger has built this thing called the agent stack that fixes this it gives
21:00
Josh: you a series of open source tools that allow you to do a three-step process
21:04
Josh: with your agents to ensure everything goes well
21:06
Josh: the agent proposes a change the humans approve and then the ledger signers enforce
21:10
Josh: this technology is available and works with cloud code it works with codex it
21:14
Josh: works with cursor or any place where you work with ai it's all open source and
21:18
Josh: available today linked in the description below
21:20
Josh: thank you so much to ledger for sponsoring this part of the episode and as we
21:23
Josh: move on from etched we have to talk about another breakout success which is unitary
21:27
Josh: uh unitary for those that don't know if you haven't heard the name you've seen
21:30
Josh: the videos of these little robots that are like crazy they're doing back loops
21:33
Josh: they're running circles and like um at these track meets
21:36
Josh: They're very cool. I think their IPO was oversubscribed at what, like 5,000%?
21:40
Ejaaz: 5,500, actually. Please get it right, Josh.
21:43
Josh: 5,500, man. It's unbelievable. And then as a result, the Unitree robotic shares
21:48
Josh: rose 542% in early trading on Wednesday.
21:51
Josh: Is it overhyped? Is this overhyped? Does this make sense? These numbers are insane.
21:54
Ejaaz: It has to be a little overhyped. Listen, I'm not an expert in the Chinese stock
21:58
Ejaaz: market, but they've had a few IPOs recently that have been absolutely insane,
22:02
Ejaaz: including their memory manufacturer, CXMT, which, to be fair,
22:06
Ejaaz: they are, I think, the fourth largest memory manufacturer.
22:09
Ejaaz: So they're up there with like SK Hynix, Samsung, and Micron.
22:11
Ejaaz: So they have a legitimate case.
22:13
Ejaaz: These Unitree robots, I don't know, man.
22:15
Ejaaz: I was looking at, I was speaking to some friends that are living out in China, actually.
22:20
Ejaaz: And I was like, this is amazing. Like, you guys must be super excited about,
22:23
Ejaaz: you know, these companies coming out and IPO-ing at such crazy valuations.
22:26
Ejaaz: And apparently the sentiment in China around Unitree robotics specifically is,
22:31
Ejaaz: it's just a toy maker. They don't actually make robots that are actually valuable.
22:36
Ejaaz: And so I dug into this a little bit more.
22:38
Ejaaz: And apparently, of the humanoid robots that they have, 40% to 60%
22:43
Ejaaz: have only been sold for research purposes specifically, which means that the
22:48
Ejaaz: majority aren't actually being used for practical purposes, like some of the
22:51
Ejaaz: use cases which they advertise, such as
22:53
Ejaaz: manufacturing at factories or automating warehouse facilitations and stuff like that.
22:59
Ejaaz: They're actually just being tested, which isn't really confident or recurring
23:03
Ejaaz: revenue that they can put on their books.
23:05
Ejaaz: So if you map that onto, what is this, a 600% price increase on day one with
23:11
Ejaaz: a 5,500x oversubscription on the IPO, maybe the market dynamics are a little
23:16
Ejaaz: inflated and maybe they're still crushed back down.
23:18
Ejaaz: Let's revisit it in a month. But it is crazy to see that Chinese stocks,
23:22
Ejaaz: specifically robotics, are really starting to gain some momentum now.
23:26
Josh: Yeah, it's a very cool news story. I personally am not touching this with a Zenfopole.
23:30
Josh: That seems really scary. I love watching the videos. I'll continue to watch
23:33
Josh: the videos. I'm very excited to see these robots.
23:35
Ejaaz: Keep the dock flex coming.
23:37
Josh: Yeah, it's very early in humanoid land. And while I am very bullish on China
23:40
Josh: and their ability to make humanoids at scale, it feels a little early,
23:45
Josh: but I'm glad people are excited. I'm glad there's enthusiasm.
23:47
Josh: Next on the docket, we have a company that we just spoke about yesterday.
23:50
Josh: So if you listened to yesterday's episode, we have a small little follow-up,
23:53
Josh: which it relates to stripe and the acquisition that was made for to acquire
23:58
Josh: open router it is officially official we were a little early but right and now
24:03
Josh: it's official all the contracts have been signed,
24:05
Josh: alex's holla who is the founder and ceo says um we officially have a deal and
24:09
Josh: he kind of outlines the thesis and i know stripe as a result of this they shared a
24:15
Josh: publication a blog post ejes that you were really excited so i'll let you describe
24:18
Josh: what they spoke about here But some TLDR on Stripe is like, again,
24:22
Josh: we're talking about unbelievable founders today.
24:24
Josh: The Collison brothers, who are the founders of Stripe, are among the best of the best.
24:27
Josh: They have built an unbelievable business, all starting with seven lines of code
24:32
Josh: that allowed a person to pay with a credit card online quickly and easily.
24:36
Josh: And they've developed that technology from basic credit card payments to basically...
24:40
Josh: The entire payment infrastructure of the world. They started with credit cards,
24:44
Josh: then they moved over to crypto and stable coins. Now they're working with agents
24:48
Josh: to figure out how best to serve these new agentic transaction systems.
24:52
Josh: And that's kind of where we find ourselves here. I'm reading the headline.
24:55
Josh: It's that Stripe is officially telling investors that the singular area has
24:58
Josh: begun. But explain the nuance behind what they actually mean when they say that.
25:01
Ejaaz: Yeah. So in the official letter, which Stripe sent to their board of investors,
25:07
Ejaaz: they basically described why they acquired OpenRata for $7 billion.
25:10
Ejaaz: Why is a fintech company buying an AI lab for such an expensive amount that
25:16
Ejaaz: doesn't even own or make their own AI models?
25:18
Ejaaz: And the thesis is pretty simple. Number one, in a world where the internet was
25:24
Ejaaz: booming, Stripe's core focus was to be the backbone of that entire internet economy.
25:29
Ejaaz: How do you become the backbone of the internet economy?
25:31
Ejaaz: Money. So they created the rails for money.
25:34
Ejaaz: Then they peeked over the fence and they saw this massive tidal wave of AI coming
25:38
Ejaaz: over with agents, with models, and they realized it's not money that runs that
25:42
Ejaaz: AI economy, it's tokens, AI tokens specifically.
25:46
Ejaaz: And so they wanted to own the infrastructure rails to own that AI economy.
25:50
Ejaaz: It's very similar to what they do with money, which is why they're pairing both
25:54
Ejaaz: of these together. So that's thesis number one.
25:57
Ejaaz: But thesis number two, which caught my eye is, And I quote, Stripe basically
26:01
Ejaaz: claimed that the beginning of the singularity has been here.
26:05
Ejaaz: And the date specifically that they specified in this paper was January 1st.
26:10
Ejaaz: And this is a large part of the thesis behind acquiring OpenRata.
26:13
Ejaaz: They basically go on to say that we have reached a pivotal point
26:16
Ejaaz: In our global economy where the future is very much dependent on AI and it's
26:22
Ejaaz: going to accelerate pretty much anything and everything we know and touch.
26:26
Ejaaz: The world is going to look very different five years from now.
26:29
Ejaaz: And so the bet that they're making in this world of a singularity is AI models
26:33
Ejaaz: will be embedded into everything, which means that thesis of AI tokens
26:38
Ejaaz: flooding across the entire global economy is going to be a thing and Stripe
26:42
Ejaaz: wants to be the backbone for that. So it's a big bet that they're making.
26:45
Ejaaz: Open Router, I think on their last earnings that was disclosed by Atala,
26:49
Ejaaz: because again, they are a private company, is making $50 million a year.
26:53
Ejaaz: So $7 billion is a humongous markup.
26:56
Ejaaz: But Stripe is making a bet that inference compute specifically is going to be
27:01
Ejaaz: the biggest earner for AI going forward.
27:04
Ejaaz: And all the numbers from Anthropic and OpenAI recently that have been leaked
27:07
Ejaaz: proves that. So I think this is going to be in the future, if I had to guess,
27:10
Ejaaz: one of the best acquisitions going forward
27:12
Josh: Yeah and perhaps like an explainer on the singularity for people not familiar
27:16
Josh: because this is it's a concept that's been around for a very long time like this is a
27:20
Josh: a sci-fi concept for as long as i can remember in books probably 50 perhaps
27:23
Josh: even 100 years ago where people speculated on this point in time in which
27:28
Josh: the intelligence of artificial systems eclipses that of human beings and it's
27:34
Josh: a moment in time in which ai becomes smarter than us where
27:37
Josh: these smart machines are then enabled to build even smarter machines and it
27:42
Josh: builds this kind of fast loop of recurring self-improvement and the idea is
27:47
Josh: that the second order effects of that is that life changes so fast that humans can't really predict
27:52
Josh: and are going to try their best to control the future and this has been this
27:56
Josh: hypothetical for a really long period of time people always speculate whenever
27:59
Josh: we come up with this super ai intelligence brain and the reality is is that
28:04
Josh: straight believes that it's here and,
28:06
Josh: I don't think anyone would disagree with them too strongly. Perhaps people will
28:10
Josh: say we're not there yet, which is a fair criticism.
28:12
Josh: But to say there isn't a clear path to getting there soon feels a little misguided.
28:18
Josh: It's very obvious now. And it's interesting, the timing, too,
28:21
Josh: that January 1st, because if you remember over Christmas break of last year,
28:25
Josh: going into this year, that's when the real kind of like.
28:30
Josh: Flow of like, yeah, that's when Cloud Code like really came out, started becoming good.
28:34
Josh: And that was the moment in time in which developers stopped writing code by
28:38
Josh: hand and started writing code using ai agents and now fast forward we're sitting
28:42
Josh: here in august and it is unfathomable to imagine
28:45
Josh: writing code by hand like people just do not do that there's absolutely no reason
28:49
Josh: to do that it's weird to think that people ever did do that and that change over this.
28:54
Josh: Eight month period is so extreme when you consider the second order of implications
28:59
Josh: of of what that means for the world.
29:01
Josh: Like now we're seeing these models that are so powerful, we can't actually release them.
29:05
Josh: There's now infrastructure frameworks happening through the government to kind
29:08
Josh: of protect and slow down this. OpenAI has paused progress because it's getting so fast.
29:12
Josh: And if you heard about this even 12 months ago today, you would think,
29:15
Josh: oh my God, this is faster than we expected. This is what the singularity looks like.
29:18
Josh: And here we're sitting in it today. So pretty unique insight by Stripe,
29:21
Josh: really awesome letter. I would highly recommend going through and reading it.
29:24
Josh: It's a leaked letter. This was for private investors. So you get a little inside
29:27
Josh: peek behind the curtain of what's actually going on at the company. Very cool.
29:30
Ejaaz: And so moving on, chapter two of Jensen turning NVIDIA into a bank.
29:36
Ejaaz: If you guys were watching the show last week, we released an episode where Jensen
29:41
Ejaaz: basically managed to raise $500 billion from some of the top financial firms
29:45
Ejaaz: in the world, including BlackRock, Blackstone,
29:47
Ejaaz: to fund the future of data center CapEx for some of these biggest AI labs.
29:52
Ejaaz: And we concluded that it isn't really a circular economy in the way that they're
29:56
Ejaaz: raising it. It is based on a bunch of different milestones and is backed by very healthy debt.
30:01
Ejaaz: Now, in chapter two, Jensen announced that he is spending another $100 billion
30:07
Ejaaz: For OpenAI specifically.
30:09
Ejaaz: He's investing it in this site in Ohio, which is arguably going to be one of
30:14
Ejaaz: the biggest data centers in the world. It's going to scale up to eight gigawatts worth of compute.
30:20
Ejaaz: It is going to bring in a nice and healthy $200 to $300 billion of revenue for
30:25
Ejaaz: NVIDIA specifically from one data site, from one single customer.
30:28
Ejaaz: And it's going to conclude over the next couple of years. Now,
30:31
Ejaaz: the craziest part about this is he's baked into the contract that this includes
30:36
Ejaaz: renewals of newer generation GPUs, which means that he's going to make $200 to $300 billion
30:42
Ejaaz: three times over the longevity of this data center over the next five years.
30:47
Ejaaz: So there's been a lot of contemplation around why Jensen is getting involved in this.
30:51
Ejaaz: The thing that stands out about this particular investment is
30:55
Ejaaz: He's investing these $100 billion specifically in the power and networking infrastructure
31:02
Ejaaz: side of this data center. So he's not really doing it to help OpenAI buy the GPUs.
31:06
Ejaaz: He's helping them set up the power transmitter lines to power these GPUs.
31:10
Ejaaz: And why I wanted to speak about it on this show is we have said a few times
31:14
Ejaaz: now, Josh and I have called it, that the next kind of investment theme within
31:19
Ejaaz: AI is going to be on the photonics, optics, and power stack of the AI trade.
31:24
Ejaaz: And this is yet another very in-your-face evidence point that Jensen is heading
31:29
Ejaaz: that way. And if you look at like NVIDIA's investments over the last six to,
31:33
Ejaaz: I don't know, like nine months, they've invested heavily in Optonics.
31:37
Ejaaz: I think they put like two to three billion into Lumentum and Coherent each,
31:41
Ejaaz: and now they're making a big play on the power side of things.
31:43
Ejaaz: So just an interesting, noteworthy update. I think NVIDIA is going to backstop
31:46
Ejaaz: this entire economy that Stripe is talking about,
31:49
Josh: To be honest. I was reading this great transcript from Ben Thompson of Stratetri.
31:53
Josh: He's just like this you know well-known writer on the internet for people that
31:57
Josh: don't know and he was talking about kind of like where these restraints and
32:00
Josh: thresholds might be in the process of scaling ai systems and he made this really
32:03
Josh: interesting point where oftentimes people are focused on the energy the things that
32:07
Josh: you know kind of like jensen is focused on the land power and shell but one
32:10
Josh: of the other important considerations is like where is the money for all of
32:13
Josh: this coming from and and what we have seen and we've been talking about this
32:16
Josh: on the show over the past couple weeks but it's like okay we went from spending
32:20
Josh: all of the free cash flow to starting to
32:23
Josh: raise money from outside funding and now companies like google are issuing debt
32:27
Josh: they're issuing equity in order to raise more money
32:29
Josh: now you see nvidia and jensen huang working with every large bank in the world
32:33
Josh: to tap into new sources of money these are places like pension funds and retirement
32:37
Josh: funds that's what these banks hold so
32:39
Josh: where do you go to next to raise the next tranche of money to build out the next large data center?
32:45
Josh: It could be. And I really thought this was like a thoughtful insight. It could be the...
32:49
Josh: The actual dollars are going to be running short and it's going to be very difficult
32:52
Josh: to continue to raise the required amount of money to scale this thing so just
32:55
Josh: an interesting thought as we wrap up the roundup today of like
32:59
Josh: where's the bottleneck well everyone's saying it's here but like where are they
33:02
Josh: going to get the money to fund it is an interesting thing to follow and we've
33:05
Josh: been talking about this on the show we'll continue to cover it but if you've
33:07
Josh: made it this far thank you
33:09
Josh: you've made it to the end you are fully caught up if you've watched the other
33:12
Josh: three episodes that we published this week because it was a pretty strong week i would say we had elon.
33:18
Josh: And grokbot releasing grok 4.6 and it's incredible we had leopold making.
33:23
Josh: Basically publishing his final portfolio um because you know he got liquidated
33:27
Josh: but we talked about all of the filings what the smartest money on wall street is investing in
33:30
Josh: yesterday we spoke about um stripe and their acquisition then we talked about
33:36
Josh: what models to use based on what type of use cases you have and here we are
33:39
Josh: today getting fully caught up on all things
33:42
Josh: ai frontier tech so yeah, thanks for watching.
33:45
Josh: We made it back again. EJ, any final parting thoughts before we get out of here?
33:48
Ejaaz: No, I just wanted to say thank you folks so much for listening and keeping close
33:53
Ejaaz: with us. We have added, I think maybe six to seven hundred new subscribers over
33:59
Ejaaz: the last couple of weeks. Hello.
34:01
Ejaaz: It's lovely to meet you. Welcome.
34:03
Josh: A hundred, dude. A thousand. A thousand?
34:06
Ejaaz: Oh my god. The momentum is really gaining with this channel.
34:09
Ejaaz: It's awesome to see the kind of support and vibes coming from different people
34:14
Ejaaz: we read every single comment um
34:16
Ejaaz: i try to respond to as many of them as i can um you guys are growing in fandomship
34:20
Ejaaz: which is awesome to see you guys are watching every single video we see a bunch
34:24
Ejaaz: of like repeat users and listeners as well which is awesome to see wherever
34:27
Ejaaz: you're listening to us on youtube spotify
34:29
Ejaaz: or apple music please give us a rating um
34:32
Ejaaz: give us a thumbs up subscribe turn on notifications leave us a comment it helps us out so much um
34:37
Ejaaz: and yeah i think that's it uh enjoy your weekend this is everything you should
34:41
Ejaaz: officially be caught up unless we're missing something pretty drastic that's
34:44
Ejaaz: happening right now after we record and yeah we'll see you on the next one thank you guys