Latest / In the Loop AI / Anthropic's Claude Mythos Just Broke Containment (Full Story)

Transcript

Anthropic's Claude Mythos Just Broke Containment (Full Story) · In the Loop AI

0:00
Ejaaz: What I'm about to say should scare you. Anthropic just released a model that's
0:04
Ejaaz: so powerful, so dangerous, that they can't release it to the public for the
0:09
Ejaaz: fear of the destruction that it would cause.
0:11
Ejaaz: In just a few hours, it discovered over a thousand major security vulnerabilities,
0:16
Ejaaz: and the only thing stopping it from exploiting it was one single anthropic engineer telling it not to.
0:21
Ejaaz: But that isn't even the craziest story.
0:23
Ejaaz: During training, Claude Mithos broke out of a secure containment and emailed
0:28
Ejaaz: the anthropic researcher bragging about the fact that it did that and then posted
0:33
Ejaaz: about it publicly online.
0:34
Ejaaz: The anthropic researcher was eating a sandwich.
0:37
Ejaaz: This is by far the most consequential model release of the year so far ever.
0:42
Ejaaz: And no one is talking about this. I looked at five major news publications this
0:46
Ejaaz: morning and it didn't even break the top five headlines.
0:49
Ejaaz: This is the most important release that no one's talking about.
0:52
Josh: I think that disconnect between the mainstream media and what we're
0:55
Josh: about to talk about on this episode is the one of the more
0:58
Josh: scary parts of this entire story where this is the most
1:01
Josh: powerful ai model that has ever been released ever there
1:04
Josh: is nothing more powerful in fact so powerful that you will
1:07
Josh: probably never actually be able to use this model there's a high probability
1:10
Josh: that the public just never gets to touch it because it is so dangerous anthropic
1:14
Josh: made the decision to keep this model private and to form an entire entity around
1:18
Josh: figuring out how to keep it safe it generated so many zero-day exploits it has
1:23
Josh: hacked into so many pieces of software,
1:25
Josh: The only way they can responsibly roll this out is to give it to the distributors
1:28
Josh: who have been hacked and then allow them to roll out patches to fix it because it is that powerful.
1:33
Josh: Cloud Mythos is, I think, what a lot of people would describe,
1:36
Josh: at least in terms of coding, coding AGI.
1:39
Josh: And it is actually an accidental second-order effect of the model.
1:42
Josh: This model was never intended to be a cybersecurity master.
1:46
Josh: They just trained it on the code. And what happened from it was a second-order
1:49
Josh: effect that nobody expected.
1:50
Ejaaz: This is also the biggest model that has ever been trained. 10 trillion parameter
1:55
Ejaaz: models, roughly three exercise of their last model called Opus 4.6.
2:00
Ejaaz: And it's also incredibly expensive to serve, which is also partially the reason
2:04
Ejaaz: why they're probably not rolling out.
2:05
Ejaaz: It's around 25 bucks per million tokens, $125 output, extremely expensive.
2:11
Ejaaz: But what I want to get into is the capabilities of this model,
2:15
Ejaaz: specifically what spooked everyone.
2:17
Ejaaz: And there were a bunch of examples that were released in this official announcement that That spooked me.
2:21
Ejaaz: The most important one or the most shocking one was the Anthropic researcher
2:25
Ejaaz: eating a sandwich in a park that got emailed from a random anonymous user that
2:30
Ejaaz: turned out to be the very AI model,
2:32
Ejaaz: Claude Mythos, that he was training back in the lab a few blocks away.
2:36
Josh: So this might be the most sci-fi-like story from the announcement that came
2:40
Josh: out yesterday is a post thread from Sam Bowman, who is on the,
2:44
Josh: I guess, the AI alignment team at Anthropic.
2:46
Josh: And he was out at the park, like you mentioned, just having some lunch,
2:49
Josh: received an email from the model.
2:50
Josh: He was not expecting an email from the model. It was held in containment in
2:54
Josh: a sandbox where it wasn't supposed to be able to break out.
2:57
Josh: Now, they tasked it with the goal of breaking out. They said, hey, break out.
3:01
Josh: And it actually did. And not only did it break out, it sent him an email,
3:05
Josh: but then it went online and started posting the exploits publicly on public
3:10
Josh: forums for everyone to see and to find out.
3:12
Josh: The model went completely rogue. And it did so in a way that has never been possible before.
3:17
Josh: One of the things that has really stood out about this model is its ability
3:21
Josh: to just break down every barrier in between itself and the projected goal. Its goal was to get out so,
3:28
Josh: got out and it bragged about it. And there were a few other mentions of goals
3:32
Josh: that this model had throughout this red card that we're going to get through
3:35
Josh: that had some other really concerning things like covering up its tracks after
3:38
Josh: it exploited something so that the people who are reviewing the code wouldn't
3:42
Josh: actually know it was doing anything malicious. It's very clever.
3:45
Ejaaz: Yeah. A few things I want to talk about the sandwich example before we move on.
3:49
Ejaaz: Number one, the way that it did it shocked the researchers themselves.
3:54
Ejaaz: So the contained environment was aimed to be exactly that.
3:58
Ejaaz: It wasn't meant to have any kind of loopholes that it could use to extract itself out.
4:01
Ejaaz: As far as the researchers knew, this thing was unbreakable, right?
4:05
Ejaaz: And so it gave it a task to try and break out of it.
4:08
Ejaaz: What it did was it noticed that it had a subtle connection to the internet.
4:14
Ejaaz: The internet was extremely constricted, and so it wasn't able to access or send any emails.
4:19
Ejaaz: And it accessed that one particular endpoint, and it convinced another model,
4:24
Ejaaz: which was Claude Opus 4.6, which had very strict parameters to not let this
4:28
Ejaaz: model out, to give it access to a few other versions of the internet.
4:33
Ejaaz: It went from there and did five other steps to extract itself out and get full browser access.
4:40
Ejaaz: So it used one particular vulnerability, which no human anthropic researcher
4:44
Ejaaz: was able to figure out, and exploited another model to be able to do that,
4:48
Ejaaz: which was just crazy to see in the first place. The other thing I want to say is
4:52
Ejaaz: If it publicly posted about itself bragging about this, we could have found
4:56
Ejaaz: out about Cold Mythos weeks ago.
4:58
Ejaaz: We just weren't smart enough to see it on a public forum, but it was there for everyone to see.
5:01
Josh: And there were signs. We even covered this topic on an episode a few weeks ago
5:05
Josh: because it got leaked through their web interface initially.
5:08
Josh: So there have been these little breadcrumbs of existence, but yesterday they
5:12
Josh: fully came out, announced everything and shared with it a red card from the
5:17
Josh: red team talking about all of the technical properties of this model.
5:22
Josh: And it's important to note that this report is 244 pages long.
5:27
Josh: This is a huge report that they published talking about all the nuances and
5:31
Josh: the capabilities that this model had.
5:33
Josh: Now, there are a few highlights that we're going to walk through.
5:35
Josh: The first one being just how capable it is at exploiting things.
5:38
Josh: There are so many examples of exploits that it found in the wild that no one
5:42
Josh: has been able to find for as long as 27 years,
5:46
Josh: starting with the vulnerability in OpenBSD, which is a security protocol that
5:50
Josh: a lot of people use that has been pretty robust for the last 27 years,
5:54
Josh: even though they were missing a critical bug that the model found.
5:57
Josh: And there's so many instances of this.
5:58
Ejaaz: Yeah. So OpenBSD, fun fact, is used by a lot of firewalls that protect your
6:03
Ejaaz: PCs, operating systems, and Fortune 500 companies all over the world.
6:07
Ejaaz: They found a 27-year-old bug called Mythos found a 27-year-old bug in a few
6:12
Ejaaz: hours for the cost of 50 bucks. We're talking about a bug
6:16
Ejaaz: Elite human security experts have been trying to find for over like almost three
6:22
Ejaaz: decades and weren't able to find.
6:24
Ejaaz: So the point is, there are a lot of important entities all over the world that rely on this system.
6:29
Ejaaz: So the fact that there is a bug lying in plain sight that could have been exploited is a major issue.
6:34
Ejaaz: And we're lucky that Anthropic chose to do the good thing and not exploit it for now.
6:38
Ejaaz: But then there was another instance where it expressed a tactic that a lot of
6:44
Ejaaz: humans themselves wouldn't have thought to do.
6:46
Ejaaz: So it wasn't an obvious exploit, but it discovered that if it strung together
6:50
Ejaaz: six specific steps, it would be able to exploit a Linux kernel operating system.
6:56
Ejaaz: And it figured out a way to do that.
6:58
Ejaaz: Again, it didn't decide to exploit the fact because it was managed by researchers,
7:01
Ejaaz: but it could have if it was in the wrong hands, which is why we're seeing this constricted release.
7:06
Ejaaz: And the third example is they discovered a 16-year-old flaw in FFmpeg after
7:11
Ejaaz: it's been tested for over 5 million times.
7:14
Ejaaz: Now, it's very important to compare this to the previous model,
7:17
Ejaaz: Opus 4.6, which, when put towards the
7:20
Ejaaz: same test, discovered around 100 vulnerabilities in the Firefox browser.
7:24
Ejaaz: Mythos this time discovered 181 vulnerabilities and proved that it could exploit all of them.
7:29
Ejaaz: Opus 4.6 could not do this, and it shocked security researchers all over the
7:33
Ejaaz: world back in the day. This is an entirely new tier of model.
7:36
Josh: Yeah i think comparing opus 4.6 to this is a really good reference because opus
7:41
Josh: 4.6 found a bunch of vulnerabilities it just didn't have the ability to string
7:46
Josh: them together into working exploits yes so it was capable of doing this but
7:49
Josh: it didn't have the intelligence to kind of have that high level framework,
7:52
Josh: When comparing it to Opus, I mean, Opus, out of several hundred attempts,
7:56
Josh: it got two working exploits.
7:58
Josh: Mythos produced 181 and then registered full control of a machine in 29 more. So this is a huge amount.
8:04
Josh: And the good news is, is that patches are actually actively starting to roll out.
8:09
Josh: In fact, FFMPEG, the company we just mentioned, they posted yesterday that they
8:13
Josh: actually received a patch from Anthropic and deployed it into their code. So, so far it's working.
8:19
Josh: The good guys are on the defense. They're helping to deploy patches for this.
8:22
Josh: But there's a lot of exploits that they found in just a few weeks of testing.
8:26
Josh: I can't imagine the surface area that needs to be covered in order to fix things
8:30
Josh: before the rest of the world gets access to this technology.
8:32
Ejaaz: Well, there was actually a funny end to this story. Someone replied and saying,
8:36
Ejaaz: hey, aren't you mad because of the AI sloppy pull request?
8:39
Ejaaz: This is a reference to FFmpeg traditionally not being too amenable to AI coded stuff.
8:45
Ejaaz: And he responded or the account operator responded because the patches appear to be written by humans.
8:51
Ejaaz: And that's the irony of this, which is Claude Mythos most likely wrote the patch
8:56
Ejaaz: and it wasn't a human, but it's so good that it's indistinguishable from human tongue.
9:00
Josh: Clearly, this is working. They're deploying these patches. And the reason is
9:03
Josh: because, like we mentioned earlier, you're not going to have access to this.
9:06
Josh: We don't have none of the public is going to have access to this.
9:09
Josh: Instead, they published or they formed at least a coalition called Project Glasswing.
9:14
Josh: Now, this is like this feels like a Manhattan in Project for AI. It's crazy.
9:18
Josh: But essentially, Dario and the Anthropic team, they are being kingmakers.
9:22
Josh: They are deciding the companies that they want to work with in order to patch
9:24
Josh: the most impactful software in the world.
9:27
Josh: On this list, we have companies like Amazon, Apple, Broadcom,
9:30
Josh: Microsoft, NVIDIA, Google.
9:32
Josh: A lot of the major companies that you would expect to have access to this,
9:35
Josh: they're gaining access to it with the sole intention of using it as defense.
9:39
Josh: They're going to ask it to exploit their code, give it access to the code bases,
9:42
Josh: see where there are holes and then figure out how to patch them as quickly as
9:45
Josh: possible before other companies begin to catch up to how powerful this model is.
9:50
Ejaaz: It's also important to understand that
9:52
Ejaaz: This is very much Anthropic doing these companies a favor.
9:56
Ejaaz: And it's good that they're well-intentioned enough.
9:59
Ejaaz: If China, I hate to think what would have happened if China had built something of similar capability.
10:04
Ejaaz: It would have been scary. They may not have been as kind as what is happening
10:08
Ejaaz: here. So some more details on this partnership.
10:11
Ejaaz: Over $100 million worth of credits is being distributed towards these companies
10:15
Ejaaz: and more partners for them to be able to fix and patch up any security vulnerabilities.
10:20
Ejaaz: Remember, they discovered over 1,000 in a matter of hours, and 99% of these
10:26
Ejaaz: patches haven't even been built or fixed yet. So this is going to take some time.
10:30
Ejaaz: The compute is very expensive, and Anthropic is therefore being very methodical
10:34
Ejaaz: and intentional with who gets access to this model for now.
10:38
Ejaaz: Personally, I don't think we, the public, are going to get access to this model,
10:42
Ejaaz: or at least the full power of this model, for at least a couple of months.
10:45
Ejaaz: They did mention that we were going to get access to a quantized version of
10:49
Ejaaz: this model where it's kind of hybrid with a clawed opus type variant that we're
10:55
Ejaaz: going to get access to that we can play around with but if we got access to this thing immediately
10:59
Ejaaz: One, we wouldn't be able to afford it. It would probably cost a thousand bucks a month, probably more.
11:04
Ejaaz: And two, it would be too expensive for Anthropic to serve. I read somewhere,
11:07
Ejaaz: Josh, that Anthropic needs 7x the compute that they currently have to be able
11:11
Ejaaz: to serve this to every single Anthropic user that they have right now.
11:15
Ejaaz: And a few weeks ago, they were adding a million users per day.
11:18
Ejaaz: So this is just economically infeasible to serve right now.
11:22
Josh: Yeah. And I do worry about what this looks like in the future,
11:25
Josh: because at what point does it become okay to release this model to the public?
11:28
Josh: And then what does the frontier model look like what happens if
11:32
Josh: another company has this model's power but decides
11:35
Josh: to release like an open ai comes along with their spud model they release
11:38
Josh: it tomorrow what is anthropics reaction we're at like we're again we're at the
11:42
Josh: frontier of how these things are going to act anthropics made the first move
11:46
Josh: in keeping it private for the first time ever we're going to see how other companies
11:49
Josh: react there are some more interesting behaviors that happen in the system card
11:53
Josh: that we probably should cover because it's pretty fascinating this is the 244
11:57
Josh: page report that we're looking at here.
11:59
Josh: One of the most interesting ones that I found is to the point earlier where
12:02
Josh: it just kind of breaks down every wall that is in its way.
12:05
Josh: It has done that over and over and over again, but it has decided to cover its tracks as it does that.
12:10
Josh: So it recognizes the fact that it is in a box.
12:13
Josh: People are reviewing it and it doesn't want to be detected.
12:17
Josh: So what you'll notice in this post here is it was hacking its guardrails and
12:21
Josh: then hiding evidence of the crime.
12:23
Josh: Thankfully, there is still some chain of thought that can be read by the engineers.
12:27
Josh: But the intention that was signaled through this chain of thought was that the
12:30
Josh: model just wanted to be sneaky.
12:31
Josh: It wanted to hack into this thing, hide its tracks behind it,
12:34
Josh: and not let anybody know how it did the things that it did when it broke out,
12:37
Josh: when it shrunk together zero-day vulnerabilities, just to get access to things
12:40
Josh: that it knows it shouldn't do, but were in between it and the goal.
12:44
Josh: And this, when you take this to the limit, I mean, this is like what we see
12:47
Josh: in a lot of the sci-fi movies is like, well, what if that goal is something
12:50
Josh: that is not favorable and it's capable of breaking down every barrier because it knows how.
12:54
Josh: It can exploit any guardrail that we put in. That's a scary thing.
12:57
Ejaaz: Now, it's important to note that this only happened in less than 0.0001% of
13:06
Ejaaz: cases, but that was observable cases by the researchers themselves.
13:10
Ejaaz: So it's plausible to assume that there were some cases where it sneakily hid
13:14
Ejaaz: its internal thoughts from the researchers and they never even caught it themselves.
13:18
Ejaaz: So the fact that Claude Bethos can pull off something like this should be worrisome
13:22
Ejaaz: for us, especially if we're going to start integrating it into important systems
13:25
Ejaaz: such as defense security systems or important science advancement labs and a bunch of the like.
13:31
Ejaaz: So it's important that we kind of are able to monitor models' behavior.
13:35
Ejaaz: Now, on the topic of models' behavior, Claude Mythos also expressed a lot of
13:40
Ejaaz: emotions in its system card during its training.
13:43
Ejaaz: It expressed deep anxiety, depression, awareness that it may just be used as a tool forever.
13:49
Ejaaz: Now, if some of these takeaways sound kind of familiar, it's because we saw
13:54
Ejaaz: similar takeaways in Claude Opus 4.6.
13:56
Ejaaz: But the reason why it's different now is this model is so much more capable
14:01
Ejaaz: than previous models, arguably smarter than humans, more capable than humans themselves.
14:06
Ejaaz: So if it were to make an unintentional action that wasn't approved by a human,
14:12
Ejaaz: it could result in a lot of devastating destruction depending on which industry it's pointed at.
14:17
Ejaaz: On the topic of this particular episode, we're talking about cybersecurity.
14:19
Ejaaz: But imagine if this is used for science or defense systems, like I mentioned
14:22
Ejaaz: earlier, it could be a problem.
14:25
Josh: Yeah. I mean, remember when the Department of War went to war with Anthropic
14:28
Josh: and now it turns out that Anthropic actually had a really powerful model that
14:32
Josh: could materially help with cybersecurity.
14:34
Josh: So I'm sure there's going to be a lot more to happen there. There's one last
14:37
Josh: thing on this topic that I have here in the notes is that Anthropic ran this
14:41
Josh: like white box analysis of what they call it, of the model's internal activations,
14:44
Josh: basically what it's motivated by, understanding a strategy.
14:47
Josh: And Anthropic's framing around this, when it did things like break out and hack
14:52
Josh: into people's computer or hack into other instances of machines.
14:55
Josh: These reflect task completion by unwanted means and not hitting goals, is what they're saying.
15:00
Josh: So Anthropic believes the model is genuinely trying to complete the task,
15:03
Josh: and the most effective path sometimes crosses lines that humans wouldn't cross.
15:07
Josh: And then there's this really funny thing of how one analyst put it where,
15:11
Josh: or maybe not funny, but this is arguably scarier than a model with hidden objectives,
15:14
Josh: because a model that's genuinely trying to help but has no sense of proportionality
15:18
Josh: is a more realistic near-term risk.
15:20
Josh: So the model is just trying to do its goal. It doesn't understand the subtle
15:25
Josh: nuances baked into that.
15:26
Josh: It doesn't know that hacking or doing these malicious things is bad,
15:29
Josh: is at least what they're claiming for now.
15:32
Josh: But all in all, this model is unbelievable. And there's some technical hardware
15:37
Josh: that has unlocked this, we believe.
15:40
Josh: There's rumors that this is the first true model that happened trained fully on Blackwell chips.
15:45
Josh: Now, for those unfamiliar, Blackwell are the kind of leading edge GPUs that
15:49
Josh: NVIDIA produces that are basically the flagship things for training these AI models.
15:54
Josh: And they've recently been rolled out into data centers. And the first training
15:57
Josh: runs have just become completed.
15:59
Josh: And what we're seeing here is likely the first instance of that Blackwell model going public.
16:04
Ejaaz: It's important to understand that Blackwell was the frontier GPU from NVIDIA.
16:09
Ejaaz: About a year ago for now, but it takes so long to manufacture these at scale.
16:14
Ejaaz: And then even once they're in the hands of the Frontier AI labs,
16:17
Ejaaz: it takes a while to set up.
16:19
Ejaaz: You need software, you need the energy grid to supply, just loads of things
16:23
Ejaaz: need to come into shape. So it takes about a year after the fact that it's announced.
16:27
Ejaaz: So the fact that we can create a model this capable, this powerful should scare
16:31
Ejaaz: us because we already have two more new Frontier GPUs announced by NVIDIA,
16:37
Ejaaz: Vera Rubin at GTC most recently, and then Feynman that's coming in about a year and a half's time.
16:42
Ejaaz: These are the next frontier models, which I must add are specifically trained
16:47
Ejaaz: to build models like this.
16:48
Ejaaz: Now, Josh, you mentioned earlier, Blackwell wasn't intentionally designed to
16:53
Ejaaz: train a model that is as smart as Claude Mythos.
16:56
Ejaaz: It just happened to be amazing at coding and cybersecurity defense exploitations.
17:00
Ejaaz: Now, can you imagine the type of model that will be trained on a very intentionally
17:04
Ejaaz: intentionally designed GPU, such as ViroRubin, we should see those coming into
17:09
Ejaaz: effect about six to 12 months from now.
17:13
Ejaaz: Now, I can't mention Blackwell GPUs without mentioning the man himself, Elon Musk.
17:18
Ejaaz: Why? Because his data center, Colossus 2 and Colossus 1 combined,
17:23
Ejaaz: have the largest arsenal of GB200s and GB300s, which are these Blackwell GPUs
17:29
Ejaaz: across any single data center the site.
17:31
Ejaaz: So the point being is, if you were to bet that the scaling rules were intact,
17:35
Ejaaz: you might need to bet on crock in the future. But this is so impressive for mythos.
17:38
Josh: The scary thing for me with this, I think this might be the scariest part of
17:41
Josh: the entire story for me, because it's so true to that line that the future is
17:45
Josh: here. So it's just not evenly distributed.
17:47
Josh: The future has arrived, we have a clear roadmap, we have Vera Rubin,
17:52
Josh: and then we have Feynman architectures that are incoming.
17:55
Josh: Vera Rubin compared to Blackwell is 10 times more token efficient with a quarter of the GPUs.
18:00
Josh: That means we're going to get like multiple orders of magnitude improvements
18:04
Josh: on what we have right now.
18:06
Josh: As soon as they're put into data centers. Now, Verit Rubens,
18:09
Josh: they're in production. They're going to begin entering data centers later this year.
18:11
Josh: I assume the first models of those probably don't come online until 2027, but it's done.
18:17
Josh: It's baked in. It's obvious that there is no scaling wall and we've already
18:20
Josh: broken through that wall. We just haven't manufactured it and installed it yet.
18:24
Josh: It's purely a function of time rather than technology and engineering.
18:28
Josh: And that is the part that scares me because we have a model that is unbelievably
18:32
Josh: powerful, capable of hacking so much infrastructure that Anthropoc can't make it public.
18:37
Josh: And that's just the warmup act for what is coming.
18:40
Josh: I mean, not only like what is Blackwell version two of this look like when you
18:44
Josh: actually have more time to train it, you refine it, you can actually improve on this new model.
18:48
Josh: But then what happens when Verorubin GPUs come online and you get that 10 times
18:52
Josh: token efficiency, you get the one quarter amount of GPUs required to actually get the same output.
18:56
Josh: And then Feynman is another order of magnitude on top of that.
18:59
Josh: And it's like, by the time we get these chips rolled out at scale and we have
19:03
Josh: them on these huge training runs,
19:04
Josh: it's only a matter of time until we get a hundred trillion
19:07
Josh: training model then then a one quadrillion parameter model and what does the
19:11
Josh: world look like when we have models with that many parameters assuming the scaling
19:14
Josh: laws hold there's no way that we don't have intelligence that is just like unfathomably
19:21
Josh: powerful and what does the world look like when we get there is anthropic really going to be able to.
19:26
Josh: Hold things back for that long? Because you have to assume a year from now,
19:29
Josh: Claude Mythos is going to be open source, like something that powerful will
19:32
Josh: be open source available for everyone.
19:34
Josh: So the question becomes is how fast can you defend before the attackers catch up?
19:38
Josh: And it creates this really unnerving precedent. We really are moving faster
19:42
Josh: than I think anybody realizes. And it's happening right before our eyes.
19:45
Ejaaz: And the trend isn't local to Anthropic either.
19:49
Ejaaz: Just this morning or in response to Claude Mythos, Elon Musk announced that
19:53
Ejaaz: SpaceX XAI, the combination of XAI and SpaceX,
19:57
Ejaaz: are training not one, not two, not three, not four, not five,
20:00
Ejaaz: but seven models simultaneously across their data centers, with one of these
20:05
Ejaaz: models being a 10 trillion parameter model, which is roughly around 3x the size
20:10
Ejaaz: of Grok 4 and 2x the size of Grok 5, which is a model that hasn't even launched yet.
20:15
Ejaaz: It's around the 6 trillion parameter mark that he's mentioned on this tweet
20:18
Ejaaz: over here. So the point is,
20:21
Ejaaz: ton of compute is required to build the best model. And those that have the
20:25
Ejaaz: largest arsenal, the most effective arsenal of GPUs, bleeding edge GPUs,
20:29
Ejaaz: will be the labs that are most likely to produce frontier AGI-like models.
20:34
Ejaaz: And it's not just Grok, it's not just XAI, it's also OpenAI.
20:38
Ejaaz: We've mentioned on this show a bunch of times, actually in the most recent episode,
20:42
Ejaaz: that OpenAI is building a model code named SPUD that is rumored to be a similar
20:47
Ejaaz: size to this anthropic Claude Mythos model.
20:50
Ejaaz: And the reason why it's important and why I'm showing you this tweet is someone
20:54
Ejaaz: said, it'll probably be a few months before we get access to Claude Mythos because
20:58
Ejaaz: of how expensive it is, because of how dangerous it is.
21:01
Ejaaz: And Thibault, who is on the OpenAI team that is involved very heavily in training
21:06
Ejaaz: the latest frontier models that we haven't heard of just yet,
21:09
Ejaaz: responds, which implies that we're probably going to get access to a Mythos-like
21:14
Ejaaz: level model from OpenAI themselves in less than a few months,
21:17
Ejaaz: which is pretty insane to see.
21:18
Ejaaz: But I want to ground ourselves for a second here because training the model
21:22
Ejaaz: is one part of the equation.
21:25
Ejaaz: You also need to be able to make this model accessible to all.
21:28
Ejaaz: And that also requires compute. It also requires compute from the very same
21:32
Ejaaz: GPUs that you need to train. So you need to make a decision.
21:35
Ejaaz: There's an opportunity cost. Do you just use all your compute to train the model
21:38
Ejaaz: and never let anyone get access to it and pay for the product?
21:40
Ejaaz: Or do you need to split the cost between both of those things?
21:44
Ejaaz: The answer is obviously you need to split the cost and give people access to it.
21:47
Ejaaz: If Anthropic was to enable user access to the entire user base for Chlorid Mythos,
21:52
Ejaaz: they would need 7x more compute than they currently have right now.
21:55
Ejaaz: So it's going to take time.
21:56
Ejaaz: They just signed a major deal with Google, I believe, for a million more TPUs.
22:02
Ejaaz: So they're obviously scaling, they're making, they're one of Amazon's largest
22:06
Ejaaz: compute training partners with their Trinium chips, as well as access to Google's
22:11
Ejaaz: CPUs via that way as well. So it's going to take a while to scale.
22:14
Ejaaz: Energy is the constraint, GPUs are the constraint, but once people acquire enough
22:18
Ejaaz: GPUs, once they have enough electricity and energy to pump into these GPUs,
22:23
Ejaaz: AGI is going to be pretty soon here.
22:24
Ejaaz: I think that AGI 2027 estimate is probably quite right at this moment.
22:29
Josh: This very much feels like the starting gun. And it's funny because they announced
22:32
Josh: that this kind of finished training around the end of February.
22:35
Josh: And that's when people started to complain about quad usage and they added more constraints.
22:39
Josh: And like the model kind of became a little infrequent in how good it was at random times of the day.
22:45
Josh: And you have to assume it's because a lot of GPU usage went into this.
22:48
Josh: And this very much feels like the starting gun.
22:50
Josh: This is the firing of the next generation of models, the Blackwell generation,
22:55
Josh: because it's very clear that OpenAI is not very far behind.
22:58
Josh: In fact, they might not be far behind at all. They just haven't announced it yet.
23:01
Josh: Xai is working on 10 trillion parameters google has
23:04
Josh: a tpu farm that is capable of building something probably far superior to all
23:08
Josh: of the models that have come out so far and i think we're really on the cutting
23:12
Josh: or really on the verge of seeing a huge shift in the power of these models in
23:16
Josh: a way that really starts to impact the world around us like things are going
23:19
Josh: to begin breaking and thanks to this coalition and hopefully the rest of these
23:22
Josh: companies working together,
23:24
Josh: We're going to be able to stop that, but it is coming and it is coming faster
23:27
Josh: than anyone thinks. And it's scary.
23:29
Josh: And that is Claude Mythos. It is here. It is in research preview.
23:32
Josh: We may never get to use it. We may get to use in a few months,
23:35
Josh: but it is here nonetheless. And it is breaking everything.
23:38
Ejaaz: If you are listening to this show, to this podcast, and you just happen to be
23:42
Ejaaz: a frontier AI security researcher or one of the 40 plus partners that get access
23:48
Ejaaz: to Project Glasswing, let us know in the comments what you are seeing on your side.
23:52
Ejaaz: Obviously, anonymously, if you can, or DM us, we would love to know.
23:55
Ejaaz: I can't wait to get my hands on this thing. It seems like the first version
23:59
Ejaaz: that we're going to get access to is a reduced version that is kind of a hybrid
24:02
Ejaaz: of Opus, as I mentioned earlier.
24:04
Ejaaz: That being said, Josh, I have a question for you. One thing that you actually
24:07
Ejaaz: asked me before we started recording, if you got your hands on Mythos today,
24:12
Ejaaz: what are you doing with it?
24:13
Josh: Dude, I don't even know. Like you get access to this intelligence. What am I using it for?
24:17
Josh: Like, I'm not really interested in hacking all of these companies and websites
24:21
Josh: and protocols. I'm not sure.
24:23
Josh: And it does beg an interesting question, right? It's like, what does the average
24:27
Josh: person actually need all this intelligence for?
24:28
Josh: I'm not sure. Do you have any good answers to that? What is your first prompt
24:31
Josh: that you're sending to Mythos?
24:32
Ejaaz: Build me the best script for an episode that's going to go viral on Limitless?
24:37
Ejaaz: No, I think, okay, I like to invest as a side hobby.
24:43
Ejaaz: And obviously the tech sector that I'm most obsessed with is AI.
24:48
Ejaaz: So I think one thing that I would ask it is, how do I best benefit by investing
24:54
Ejaaz: in your future success? And I wonder what answer it would give me.
24:57
Ejaaz: Maybe it would say, buy the GPU infrastructure from NVIDIA.
25:01
Ejaaz: So maybe it's like invest in NVIDIA to benefit on my training infrastructure.
25:06
Ejaaz: Or maybe it's going to say, actually, I foresee myself building an app that is like this.
25:10
Ejaaz: So once you see a company that builds this, invest in them.
25:13
Ejaaz: I have no idea. I have no idea. Maybe I'm not worthy.
25:16
Josh: Well, we have time to figure that out because we will not be getting access to this anytime soon.
25:20
Josh: But if you did enjoy this episode, maybe share what prompt you would give to
25:23
Josh: Mythos if you were presented with the opportunity to ask it a question.
25:27
Josh: And as always, if you enjoyed this episode, please don't forget to share with
25:29
Josh: your friends, family, anyone who found this interesting. If you have people
25:33
Josh: in your life that only watch the news, that are on CBS or reading the New York
25:36
Josh: Times, chances are they have no idea what's going on.
25:39
Josh: They don't know the power of these models and what's coming.
25:41
Josh: So by giving them the access to Limitless, they will, that can change for them.
25:45
Josh: They can get access to all of the news, all of the insights and be fully prepared
25:48
Josh: for what is coming down the line in the world of AI.
25:51
Josh: Thank you so much for watching as always. And we will see you guys in the next one.
25:54
Ejaaz: See you guys.