Latest / In the Loop AI / THIS WEEK IN AI: Leopold's Fund Collapses, Nvidia Backs Ilya, Slowing Down AI

Transcript

THIS WEEK IN AI: Leopold's Fund Collapses, Nvidia Backs Ilya, Slowing Down AI · In the Loop AI

0:00
Ejaaz: Leopold Aschenbrenner is a 23-year-old kid that raised one of the most successful
0:05
Ejaaz: AI funds ever, but he might be in disastrous trouble right now.
0:10
Ejaaz: News reports this morning states that his fund might be down between 30 to 50%
0:16
Ejaaz: as semiconductor stocks have drawn down aggressively over the last four weeks.
0:20
Ejaaz: We're going to get into that and a lot more.
0:22
Josh: Yeah, let's kick it off with our poster boy, golden child, Leopold Aschenbrenner,
0:25
Josh: who has basically done no wrong for the last, what is it, two years of investing
0:30
Josh: where he's rolled this fund into something like 30 billion dollars well for
0:34
Josh: the first time ever since inception it seems like things are not doing super hot in paradise
0:40
Josh: looks like he's starting to sell and not just sell, but sell a lot.
0:43
Josh: In fact, liquidate a lot of his positions that he's been accumulating over time
0:48
Josh: because it seems like the fund is in trouble.
0:49
Josh: We know the kind of directive. We filmed enough episodes on Leopold to understand
0:53
Josh: his general thesis, which is long only with leverage to try to just maximize
0:59
Josh: upside returns, which has worked incredibly well.
1:01
Josh: But we understand that, you know, well, I remember vividly in the last episode,
1:05
Josh: you know, as we were comparing him to someone like Gavin Baker and how Gavin's
1:08
Josh: been around for 30 years because he's able to survive the upswings,
1:11
Josh: the downswings. They have that enduring capability.
1:14
Josh: Leopold is getting his first test of strength right now, and it looks like it's
1:18
Josh: not going super hot. I mean, on screen, I'm looking at a tweet from Martin Screlli
1:21
Josh: that says, Situation has sold half of its Anthropoc stake, which is like,
1:26
Josh: I mean, granted, this is all rumors.
1:28
Josh: None of this is confirmed to be true. But man, if that is, that is a painful, painful loss.
1:33
Ejaaz: To be clear, he invested in their series, a trownd in 2025 when they were worth $60 billion.
1:41
Ejaaz: So he must be in pretty dire trouble because it made up for like 40% of his
1:46
Ejaaz: funds. So we're talking about like three to maybe $6 billion.
1:50
Ejaaz: And for him to sell that, he must be in drastic trouble. So what are the rumors saying?
1:54
Ejaaz: Again, very reliable source here, Martin Shkreli from the Godel News says that
1:59
Ejaaz: Situational Awareness LP fund is down more than 50% month to date,
2:04
Ejaaz: and it was up 200% year to date. So that is a massive drawdown.
2:10
Ejaaz: And in terms of like what positions he's liquidating, I'm a little confused
2:13
Ejaaz: here, Josh, because when we covered his recent 13F,
2:17
Ejaaz: he was hedged. He wasn't particularly bullish and like long on everything,
2:23
Ejaaz: but he also wasn't particularly bearish. If you remember, he had one of the
2:26
Ejaaz: largest shorts on NVIDIA itself.
2:28
Ejaaz: So I was wondering if he had that sort of portfolio composition back then,
2:32
Ejaaz: he might have even like, you know, not drawn down too much.
2:36
Ejaaz: You know, rumors are saying that he's down like 30% overall on the fund,
2:39
Ejaaz: which, you know, is disastrous for the size of the fund, but it's still not
2:43
Ejaaz: the worst that it could be given that he was long all these semiconductor stocks.
2:46
Ejaaz: Now, the reason, you know, the question about like whether this is good or bad,
2:50
Ejaaz: I think investors have really been feeling it broadly.
2:53
Ejaaz: A lot of AI stocks in general have just drawn down a lot over the last four
2:57
Ejaaz: weeks. We actually had an episode that covered this yesterday.
2:59
Ejaaz: You should definitely go check that episode out. We explained kind of like why and what's going on.
3:03
Ejaaz: But I think his thesis still holds true. And listen, I hope he makes it out.
3:07
Ejaaz: I hope he's able to come out of this.
3:08
Josh: Yeah, dude, you and me both. I mean, we'll see. It's the difficult thing about
3:12
Josh: this again is it's all speculation because this is all insider rumors. No one actually knows.
3:16
Josh: We will see dude his next 13f filing is going to be one of the most exciting
3:20
Josh: episodes of all time did ilia survive or did he not or so speaking of ilia no did
3:26
Josh: leopold survive but speaking of ilia we have some news in the ilia front as
3:29
Josh: well which comes in the form of his company safe super intelligence they've been
3:33
Josh: radio silent and i give this company and the team a tremendous amount of credit
3:37
Josh: because gosh who is doing this much work remind
3:39
Ejaaz: Us who ilia is.
3:40
Josh: Ilia setzkever he is ex co-founder of open ai he left open ai to go start his
3:46
Josh: own company no one really knows what this company is they're just named safe
3:50
Josh: super intelligence the idea is that they want to figure out a way to safely align ai
3:54
Josh: with the human intention and it sounds like they have been making progress i
3:59
Josh: mean this company has been radio silent there's no leaks there's no information getting out
4:04
Josh: but they recently just this week announced a long-term strategic partnership
4:08
Josh: with nvidia which is like huge it says nvidia is making a substantial investment
4:13
Josh: in ssi that will let us 10xr compute in the next 12 months we reach the point
4:17
Josh: where our research is worth scaling
4:19
Josh: And with this partnership, we will be able to. We are honored by NVIDIA's conviction.
4:24
Ejaaz: That's crazy.
4:25
Josh: They made a breakthrough. They have some sort of a breakthrough.
4:27
Josh: Like Jensen saw this and he was like, oh yeah, bro, we got you.
4:30
Josh: We're investing in this.
4:31
Ejaaz: That's huge. That's huge.
4:33
Josh: So I have to imagine it's only a matter of time until we actually wind up getting
4:37
Josh: one of these breakthroughs published. But I know Ilya in particular has been
4:41
Josh: focused on the human anatomy of the brain.
4:44
Josh: And I have to imagine that this is some breakthrough relating to a part of the
4:48
Josh: brain that we haven't quite tapped into. It leaves very bullish on the idea that
4:53
Josh: AI is very similar to a human in the way that it learns and the way that it
4:56
Josh: thinks. And there's a lot of places we could draw from. So I am stoked to see
5:00
Josh: what this looks like. But the rumors are true. They're building something and it is very badass.
5:05
Ejaaz: So just to conceptualize, you know, why it's super important for these smaller
5:10
Ejaaz: AI labs that aren't called Anthropic, OpenAI or Google to announce something
5:14
Ejaaz: is they're basically just overvalued research labs.
5:18
Ejaaz: When they're funded, it's just a small team of researchers. and they're like,
5:21
Ejaaz: well, hey, hopefully we make a breakthrough that's better than what Anthropic and OpenAI is doing.
5:26
Ejaaz: And that's what you're betting on. So it's actually a very low hit rate that
5:30
Ejaaz: these AI labs actually get anything out of the door.
5:32
Ejaaz: So this announcement, like raising $5 billion, what did Jensen see is the meme
5:37
Ejaaz: that I'm seeing going around, is fantastic news for Ilya and what they're building.
5:41
Ejaaz: Now, the question is, if they've made an architectural breakthrough,
5:45
Ejaaz: how different is it to the LLM
5:47
Ejaaz: architecture, which is how Claude and Chachupiti work behind the scenes.
5:51
Ejaaz: If SSI has made a meaningful breakthrough, then there is an argument to be said
5:55
Ejaaz: that they could be competitive with these Frontier Labs. Because of course,
5:58
Ejaaz: remember, the argument is these Frontier Labs are so far ahead and eventually
6:02
Ejaaz: they're going to be building a model that can recursively build itself.
6:04
Ejaaz: How can you possibly catch up? So this news from Ilya is super cool.
6:09
Ejaaz: Now, the rumor has it is his model will release as soon as mid to late August.
6:15
Ejaaz: And if strawberry man over here on the screen is is is right he goes uh it'll
6:20
Ejaaz: be superior to fable now listen.
6:22
Ejaaz: By the time we get to the end of August, Anthropic will put out Fable 6 and
6:26
Ejaaz: it'll crush everything else.
6:27
Ejaaz: But the point is, for a smaller lab to catch up is very interesting.
6:30
Ejaaz: Now, on the topic of recursive self-improvement, Josh, there was a big and very
6:35
Ejaaz: weird alignment between two of the biggest enemies in AI.
6:40
Ejaaz: Anthropic and OpenAI researchers, I think 1,300 spanning across other frontier
6:45
Ejaaz: AI labs and the AI industry in general,
6:47
Ejaaz: signed an open petition calling for the US government to deliberately slow down
6:52
Ejaaz: or pace AI development in order to prevent it from advancing too quickly.
6:56
Ejaaz: And the general thesis behind this letter is we are very close to building an
7:01
Ejaaz: AI model that can self-improve, that can build the next version of itself.
7:05
Ejaaz: And in that world, all that matters is the AI model could get much smarter too
7:10
Ejaaz: quickly for humans to comprehend.
7:12
Ejaaz: And we could end up in a situation where the AI model is just autonomously working
7:15
Ejaaz: for itself and humans, for lack of a better fantasy, end up becoming the slaves
7:21
Ejaaz: of this thing i think that's what's happening here josh.
7:24
Josh: Yeah and funny enough ilia who we just mentioned actually signed this this morning
7:28
Josh: the petition so he is number say like
7:30
Josh: 1301 to be involved in this and there's basically three points to this
7:34
Josh: uh the first being that they believe that ai may soon be able to build better
7:37
Josh: ai which is that recursive self-improvement loop that's shocking about um that
7:41
Josh: can make progress accelerate uh really fast and make things really scary we
7:45
Josh: started to see this with the chat gpt 6.0
7:48
Josh: leak and how sam altman just came from
7:51
Josh: the government this week he's like he's in washington dc talking to the government
7:54
Josh: and he's saying words like we have permanently deactivated the model as if it
7:59
Josh: is this dangerous malicious thing because they are actually like afraid of
8:02
Josh: what it's able to do and don't quite know how to sandbox just yet the second
8:05
Josh: part of this is that no single company or country can slow down alone so they
8:09
Josh: believe that everyone needs to do this in unison
8:11
Josh: So they're asking the U.S. government to lead this international effort and
8:15
Josh: to build the tools and agreements needed to slow things down if it becomes necessary.
8:19
Josh: It seems very clear that the labs are getting spooked out pretty quickly because
8:22
Josh: the capabilities of these things are getting good quick and they're getting
8:25
Josh: way better, way faster than anticipated.
8:28
Josh: And there's this like kind of looming fear like, hey, how do we actually safely
8:33
Josh: even test these things privately, let alone release them publicly?
8:37
Josh: So it's this like really strange disconnect that's happening.
8:40
Josh: And it's interesting to see all the labs align around this thing, including ILIA,
8:44
Josh: And all of the major players from a lot of the major labs. So we'll see how this plays out.
8:48
Josh: It is definitely interesting, though. It's like everyone's thinking now,
8:51
Josh: hey, if we slow things down, actually, it probably is a good thing.
8:54
Josh: It's just the question now of whether it's even possible.
8:56
Ejaaz: Yeah, well, I mean, on the topic of whether something is potentially good or
9:00
Ejaaz: bad, whether you should slow down model development, Anthropic came out with
9:03
Ejaaz: a statement on the debate between open and closed source.
9:07
Ejaaz: Now, for those of you who aren't aware, there's been a big point of contention between the U.S.
9:12
Ejaaz: Government and open source weights or open source AI models,
9:17
Ejaaz: where they're hinting at potentially banning open source models,
9:20
Ejaaz: which would like, you know, which would receive, you know, quite a lot of disdain
9:23
Ejaaz: from a lot of founders in America and in the West in general.
9:26
Ejaaz: Now, Anthropic came out with their positioning on this, which is,
9:29
Ejaaz: we have never stated that we want to ban open models.
9:32
Ejaaz: We just want to make sure that we can slow the development enough such that
9:36
Ejaaz: this model, these models, these AGI type intelligences don't escape us and become
9:42
Ejaaz: a problem. So Dario, in his own words, presented three very reasonable options, in my opinion.
9:48
Ejaaz: One, we shouldn't sell powerful chips or chip making equipment to China.
9:51
Ejaaz: The point, of course, that he's making there is China doesn't have next-gen chips.
9:55
Ejaaz: They can't train next-gen models, which they can open source and release to
9:57
Ejaaz: the world to potentially malicious actors.
9:59
Ejaaz: Number two, we have to crack down on distillation operations.
10:03
Ejaaz: For those of you who don't know, distillation is basically the act of taking
10:06
Ejaaz: an already smart release model and using it to create prompts or outputs that
10:11
Ejaaz: you can use to train a completely different model. It's kind of like sealing the blueprint in a way.
10:16
Ejaaz: And Anthropic and other labs like OpenAI and Google banned this explicitly,
10:19
Ejaaz: but China has been accused of potentially doing this at a wide scale.
10:23
Ejaaz: So just making that illegal.
10:24
Ejaaz: And the third one is whatever your model is, if it's sufficiently capable to
10:29
Ejaaz: perform a cybersecurity threat or biothreat or whatever that might be,
10:32
Ejaaz: closed or open, it should go through a framework of processing beforehand had
10:37
Ejaaz: to make sure that it is safe and has the necessary safeguards.
10:39
Ejaaz: Now, of course, the obvious point of contention there is open models can't have
10:43
Ejaaz: safeguards really. You just run it at home privately on your own servers.
10:47
Ejaaz: It's a really good way to kind of shape the conversation. I don't know how far
10:50
Ejaaz: we're going to get, but it was good to see Darius take finally on this.
10:53
Josh: Yeah, it's a big conversation that's been coming up a lot more.
10:56
Josh: Just as everything starts to move faster, open source gets a little bit scarier,
10:59
Josh: closed source gets a little bit more powerful.
11:01
Josh: And there is this contradiction happening in the world where people who are
11:05
Josh: AGI-pilled cannot simultaneously be open source maximalists.
11:09
Josh: Because if you do believe in the impact of agi and what it will do on the world it is difficult to
11:15
Josh: Also say that you want that to be available and open sourced as quickly as the
11:20
Josh: as like progress is kind of making so it's an interesting area
11:23
Josh: um there's a lot of commentary that's been going around this week i think a
11:26
Josh: lot of the news has been around like thought boy pieces commentary duarkesh
11:29
Josh: had one that we really loved he had a bit of a contrarian take where he was
11:34
Josh: basically if i could summarize this i would say everything up
11:37
Josh: everything up gpus up usage up utilization up like up only is how i interpreted
11:41
Josh: this if i could summarize this in two words
11:43
Josh: but basically the idea is that um recently we've had this phenomenon and this
11:47
Josh: is the counter michael burry phenomenon if you remember he said the price of
11:50
Josh: gpus was going to go down over time dora kesha saying hey wait a second
11:54
Josh: um because there is going to be so few gpus and so much high value intelligence
12:00
Josh: that the value of that intelligence per token is actually just going to keep going up
12:04
Josh: and the value of the gpus is actually going to keep going up and if you are
12:07
Josh: trying to lock in rental and compute now that's probably a good idea because
12:12
Josh: There's a fixed amount of capability of running intelligence,
12:16
Josh: but an infinite amount of high quality intelligence that is coming.
12:19
Josh: And the premium on that, while we figure out this dispersion of resources,
12:23
Josh: is going to increase to a level that's fairly high. And the price of these tokens,
12:27
Josh: the price of these GPUs, will just continue to kind of steadily trend upward
12:31
Josh: until we somehow manage to figure out the supply-demand problem.
12:34
Ejaaz: Yeah, I think another really good point that he makes is these AI chips would
12:39
Ejaaz: only make sense to use for the best ai models in the world in that scenario
12:43
Ejaaz: so like just imagine we have
12:45
Ejaaz: claude fable 6 in the future and it is as good as doing the tasks and profession
12:51
Ejaaz: as a professional software engineer that's worth like what is that 250k a year
12:56
Ejaaz: his point is if one ai chip can run claude fable 6 and it can do the job of
13:02
Ejaaz: a 250k a year employee why shouldn't it be worth 250k.
13:08
Ejaaz: And therefore, it starts to price out all the smaller labs like safe super intelligence
13:11
Ejaaz: and stuff that doesn't have as much capital.
13:13
Ejaaz: So it only makes sense for NVIDIA to raise the prices. Everything goes downstream
13:18
Ejaaz: of the model capability.
13:19
Ejaaz: So if NVIDIA knows that their next-gen chip, well, let's call it Feynman in
13:23
Ejaaz: 2028, which they're releasing, is good enough to train a superior model at Anthropic
13:28
Ejaaz: or Opening or whatever, they're going to charge them the value of that market share, essentially.
13:34
Ejaaz: So that means if you're a smaller AI lab, you can't really get access to these
13:38
Ejaaz: things. So Dworkesh's point is like the value of these GPUs are actually going
13:41
Ejaaz: to go up only. And if you look at any of the recent earnings reports from Microsoft
13:44
Ejaaz: and stuff, their cloud service providing business has gone up.
13:48
Ejaaz: These revenue numbers are jumping up like 40 to 50%.
13:51
Ejaaz: Sam confirms this as well. He says it's also a demand problem where Sam,
13:56
Ejaaz: who is notorious for having spent or committed, I think it's like $500 billion,
14:00
Ejaaz: like half a trillion dollars to compute over the next five to 10 years.
14:04
Ejaaz: He says on this podcast this week that it wasn't enough and that he needs to
14:09
Ejaaz: be even more aggressive. So I don't know who's right at this point,
14:12
Ejaaz: but two very bullish posts from Dwarkesh at this point.
14:14
Josh: Two very bullish takes. And we have another leader in the space,
14:17
Josh: this one in the world of security, and that is Ledger, where if you are building with
14:21
Josh: agents and you're worried about security ledger has a solution for you there's
14:25
Josh: a three-step process the agents propose the humans approve and the ledger signers
14:29
Josh: enforce so they have this thing called the ledger agent stack which is a series
14:33
Josh: of open source tools that let your agent work
14:35
Josh: with crypto while you stay in full control of the entire thing you could think
14:39
Josh: of it like auto mode with cloud code except you have to approve at each step
14:43
Josh: of the way and this works with all of your favorite providers this works with
14:46
Josh: codex it works with cloud code it works with Courser.
14:49
Josh: It is open source and is fully available today. So if you are building with
14:52
Josh: agents, go and check them out at developers.ledger.com. It is linked in our
14:56
Josh: description below. Thank you so much to Ledger for supporting this episode.
15:00
Josh: And now we have to get into earnings season, of course, because,
15:03
Josh: you know, I mean, we do talk about stocks quite a bit and the earnings are all over the place. I mean,
15:08
Josh: We saw the semiconductors and the memory stocks get absolutely crushed,
15:13
Josh: but there is one company who is not, and that is Microsoft, who had $90 billion
15:18
Josh: in revenue with estimate per share of $4.75 versus $4.25.
15:23
Josh: So it looks like, I mean, things are up only at Microsoft. I'm taking a look
15:27
Josh: at their cloud revenue, which is up 43% year over year. I mean,
15:32
Josh: this is a great business.
15:33
Josh: Who would have thought Microsoft, the company that's never released a single AI product.
15:36
Ejaaz: The market absolutely loves it. Like their stock this morning as of recording
15:39
Ejaaz: is up 15%, which is pretty big for a big boy like Microsoft.
15:43
Ejaaz: Now listen, the number one highlight from their earnings is their Azure earnings.
15:48
Ejaaz: So that's their cloud service provider, which they use to basically lease out
15:51
Ejaaz: compute to any AI model lab or AI company that wants to run or train their,
15:55
Ejaaz: or inference their own models.
15:56
Ejaaz: And that's what Microsoft's cloud businesses focus on. That is up 43% and is
16:01
Ejaaz: the thing that everyone is looking at when you look at hyperscalers specifically.
16:04
Ejaaz: So Microsoft, Google, and Amazon.
16:07
Ejaaz: And it's a really good business. And I think the profit margins for all of the
16:10
Ejaaz: revenue that they're bringing in is incredibly impressive.
16:13
Ejaaz: Moving on, there's a few more companies that we covered in the episode yesterday, but TLDR is Samsung.
16:19
Ejaaz: That little memory company where their stock is dumping is doing extremely well.
16:23
Ejaaz: They're signing a lot of LTAs, which are long-term agreements with future customers.
16:27
Ejaaz: So they're locking it in until like, I think, 2028. So 2027's memory supply
16:32
Ejaaz: might already be accounted for.
16:34
Ejaaz: And then we have Intel. Josh, you know, you just spoke about agents.
16:37
Ejaaz: Intel provides the number one bit of infrastructure to help those agents do
16:41
Ejaaz: their thing called CPUs.
16:43
Ejaaz: Intel and AMD, I don't know when AMD's earnings is coming out.
16:46
Ejaaz: Has done really well and they had a banging quarter.
16:48
Ejaaz: So the TLDR is these companies have reported amazing revenue earnings.
16:53
Ejaaz: Their profit margins are going up because of AI, but the stocks are kind of
16:57
Ejaaz: like very volatile, except for the one company that is not doing anything in
17:01
Ejaaz: AI right now that just became the number one most valuable company in the world.
17:05
Ejaaz: We're talking about Apple. Congratulations to Apple. Who's gone on a historic run. Yep.
17:09
Josh: Yeah, I mean, Apple, they deserve this, honestly. Siri is going to be incredible.
17:13
Josh: They're going to have two amazing devices with the iPhone Fold,
17:15
Josh: the iPhone 19, or whatever number we're at now.
17:20
Josh: They're in such a great position for running local inference on machines,
17:23
Josh: and I think people are recognizing that.
17:25
Josh: So I expect to see a lot more coming down the pipe from Apple.
17:28
Josh: I mean, congrats to Apple. This is great news.
17:31
Josh: We also have Kimi K3, as we round up this rapid fire section,
17:34
Josh: who released their open source weights. And remember, we get excited about running
17:37
Josh: open source models at home. Well, it turns out it costs a little bit more than you think it might.
17:42
Josh: It looks like it might cost about, what? $35,000 a month.
17:45
Ejaaz: Dude, it costs like a million dollars per year. It's crazy.
17:49
Josh: No, no, no, no, no, no, no. Half a million dollars. Half a million dollars.
17:53
Ejaaz: One time.
17:54
Josh: So only like, yeah. So maybe twice as many people can do that as a one-time
17:58
Josh: purchase. But yeah, Kimi is expensive. that these model weights are ginormous.
18:01
Josh: But it is open source at least. So hey, if you got half a million dollars laying
18:04
Josh: around, you want to play with it. This would be your time to do it.
18:07
Ejaaz: And that is it. You are all caught up for the week on AI.
18:11
Ejaaz: Let us know if you've missed anything in the comments. But wherever you're listening,
18:14
Ejaaz: if you're listening to us on YouTube, Spotify, Apple, subscribe, leave us a comment.
18:18
Ejaaz: Make sure you give us an amazing rating. I think we're a pretty good show personally.
18:22
Ejaaz: And we have a newsletter that goes out twice a week. Definitely subscribe to
18:25
Ejaaz: that. And Josh, is there any final comments that we want before the week?
18:28
Josh: Share it with your friends. Let them know how much you love the show.
18:31
Josh: If you think they might be interested, go touch some grass for this weekend
18:33
Josh: because you are now fully caught up on all things AI and investing in frontier
18:37
Josh: technology and all the good things.
18:39
Josh: So have an amazing weekend and we will see you all when we come back next week.