Latest / In the Loop AI / The United States of Claude: Government Bans Fable 5

Transcript

The United States of Claude: Government Bans Fable 5 · In the Loop AI

0:00
Ejaaz: Last Tuesday, Anthropic released the most powerful AI model the world has ever
0:03
Ejaaz: seen. And that exact same week, by Friday at 5.21pm, they received a letter
0:08
Ejaaz: from the government asking them to shut down access to their latest Claude Fable
0:12
Ejaaz: and Claude Mythos models.
0:14
Ejaaz: 90 minutes later, Anthropic shut down access for everyone, citing that anyone
0:18
Ejaaz: who isn't an American citizen or American by nature doesn't get access to this model.
0:23
Ejaaz: Now, the official reason for this was security researchers external to Anthropic
0:28
Ejaaz: had discovered a jailbreak, a way to
0:30
Ejaaz: access or bypass the safeguards on Anthropic's Fable model to get access to
0:35
Ejaaz: a restricted version, which could be used for nefarious purposes,
0:37
Ejaaz: such as designing dangerous bioweapons or exploiting cybersecurity weaknesses in software.
0:43
Ejaaz: Now, in the hands of the wrong people, this could be potentially dangerous,
0:46
Ejaaz: such as China versus the US.
0:48
Ejaaz: The craziest part about all of this is it came from an internal partner.
0:51
Ejaaz: Amazon, who owns the largest individual stake in Anthropic as an investor were
0:56
Ejaaz: the ones that released details of the jailbreak to the government.
0:59
Ejaaz: They had the option of picking up the phone and calling Dario and saying,
1:01
Ejaaz: hey, there's a jailbreak. Let's figure out how to fix this. But they decided
1:04
Ejaaz: to contact Trump himself and report this.
1:07
Ejaaz: And so it unpacks and unravels this entire story around governments taking a
1:11
Ejaaz: nationalized stake in the Frontier AI labs and determining who gets access to which model.
1:16
Josh: Last week was a roller coaster of a week. I mean, we got Fable 5 finally.
1:20
Josh: The Mythos class model was delivered on Tuesday.
1:22
Josh: And Friday at 5.21 p.m., after everyone is getting ready to go home,
1:27
Josh: the U.S. government drops this notice on their desk, saying that they are initiating
1:31
Josh: an export control directive citing national security authorities.
1:35
Josh: The order was to suspend all access to Fable 5 and Mythos 5 by any foreign national,
1:40
Josh: whether inside or outside of
1:41
Josh: the United States, including Anthropik's own foreign national employees.
1:46
Josh: This, of course, is almost impossible to enforce. How do you tell where anyone is from?
1:51
Josh: So Anthropic says there's no clean way to wall off foreign nationals across
1:54
Josh: global consumer products. So in order to comply, they disabled models for all
1:59
Josh: customers, both Mythos and Fable class models.
2:02
Josh: Following this, they also shared the initiative with AWS, which revoked Fable
2:05
Josh: 5, all their cloud services.
2:07
Josh: And now where we sit today, everybody is using Opus 4.8 and Sonnet models because
2:11
Josh: Fable and mythos are no longer allowed to be used this
2:15
Josh: sucks um i was getting really used to using fable as a model it is an unbelievably
2:21
Josh: remarkable model in terms of how powerful it is and it feels like now i woke
2:25
Josh: up today i went to use claude and
2:27
Josh: being faced with opus 4.8 i'm like oh man like it's a good model but i really
2:31
Josh: miss my my crown jewel i miss
2:33
Josh: i feel like i am less capable today than i was yesterday and that that feeling alone kind of sucks
2:38
Ejaaz: We've crossed a very important boundary. And it's the first time we've ever
2:43
Ejaaz: done this. This is the first instance of an AI model which has been publicly
2:46
Ejaaz: released and then withdrawn.
2:49
Ejaaz: And I'm feeling the withdrawal effects.
2:51
Ejaaz: As you know, I've been using Fable 24-7 since it got released.
2:55
Ejaaz: And when it got pulled, I was like, I think it was like out on a Friday night and I saw the news.
3:00
Ejaaz: And that was the other thing, by the way, this happened like at the end of day,
3:03
Ejaaz: at the end of the week, when no one was expecting this.
3:05
Josh: Yeah, we got a lot of complaints to get into about this one.
3:07
Ejaaz: Yeah, there's a lot of complaints that I have and that I'm sure many other people have.
3:10
Ejaaz: But the point is, this is the first clear example of a retail accessible model
3:17
Ejaaz: being pulled. And it was government determined. And so it brings into question,
3:21
Ejaaz: obviously, a host of different topics and discussions, but the main one being,
3:24
Ejaaz: is the government overreach too much? Is this a line too far?
3:29
Ejaaz: And then it gets into an interesting discussion around who gets access to this, right?
3:32
Ejaaz: So the clear statement that's been made by the government is,
3:36
Ejaaz: if you aren't an American citizen or other, if you aren't American,
3:39
Ejaaz: if you weren't born here, you don't get access to this model.
3:41
Ejaaz: So technically on this show, Josh, that would mean that, you know,
3:44
Ejaaz: you get access, there's a world where you get access to this model and future
3:47
Ejaaz: models, and I don't purely because of nationality-based stuff.
3:50
Ejaaz: So it gets into a bunch of different things that i have some pretty strong opinions
3:54
Ejaaz: on but firstly like one earth is a is an export export control like what does that mean exactly.
4:00
Josh: I had to look this up because i wasn't sure myself and it's basically these
4:03
Josh: u.s rules that govern the transfer of sensitive
4:06
Josh: technology and information to foreign persons in a foreign destination so basically
4:11
Josh: you don't want to export the technology or the information to other countries
4:15
Josh: because it is dangerous it can be used for a malpractice same is true with a
4:18
Josh: lot of military information this is being treated as such.
4:22
Josh: So for dual use technology, they call it, it has to run through the Bureau of
4:26
Josh: Industry and Security, which is inside the Commerce Department,
4:30
Josh: which is under the Export Administration Regulation. There's just like a lot
4:34
Josh: of regulation involved in this.
4:35
Josh: But basically, the United States deemed this as something that is not safe to
4:40
Josh: be exported. And because Mythos, Class Models, and Fable are being used by foreign
4:45
Josh: entities, regardless of the foreign entity status, they must be shut down immediately.
4:49
Josh: And because it is nearly impossible to tell where the source of that API key
4:54
Josh: is coming from or where the destination of that API is coming from,
4:57
Josh: the API call, they can't actually determine with 100% certainty
5:01
Josh: All the people that are being delivered tokens are not from foreign entities.
5:05
Josh: They are no longer allowed to serve tokens to anybody.
5:09
Josh: And then there's this nuance worth noting where U.S. citizens and green card
5:12
Josh: holders and other certain protected individuals are exempt from this deemed
5:16
Josh: export rule, but it gets really messy.
5:18
Josh: And it's just this incredibly complicated and impossible situation that I would
5:22
Josh: imagine for Anthropics to navigate because you just can't really tell.
5:25
Josh: So as a result, the models are gone.
5:27
Josh: And we're in this weird limbo period now where the frontier model has actually
5:30
Josh: been removed and the frontier for the first time ever like you mentioned it's
5:33
Josh: actually moved backwards a notch instead of forwards in terms of public facing
5:36
Josh: capability and that's been
5:37
Josh: it's set a really interesting and kind of like unnerving precedent as we move
5:41
Josh: forward through this like as we navigate through this new world that we're going through
5:45
Ejaaz: Well i mean this isn't the first time that anthropic and the government have
5:48
Ejaaz: got into a fight these guys have had a relationship they have they have a past
5:52
Ejaaz: and i think it's important to kind of like recap what that past looks like.
5:55
Ejaaz: So literally in the last 12 months, Anthropic has got into spats and fights
6:00
Ejaaz: with the government, disagreements, and they're trying to kind of resolve their outcomes.
6:04
Ejaaz: But there's a bit of bitterness between the two.
6:07
Ejaaz: So I think running through this, Feb 27th, 2026, so that's literally,
6:12
Ejaaz: you know, a couple of months ago, there was a federal blacklist that was put out for Anthropic.
6:16
Ejaaz: Now, the reason why this surfaced was Anthropic's models were being used by
6:21
Ejaaz: the Pentagon or the Department of War to carry out military operations,
6:24
Ejaaz: to design and strategize around how they execute on that.
6:27
Ejaaz: And this was the first known instance. CLAWD was the default model that was
6:31
Ejaaz: used by every single government institute.
6:34
Ejaaz: Anthropic expressed disdain and disinterest in this, saying that,
6:37
Ejaaz: you know, there should be more regulation about this. There should be more transparent comms.
6:40
Ejaaz: The government did not like that. They got into a back and forth about it.
6:43
Ejaaz: They tried to come to an agreement. Ultimately, they did not.
6:46
Ejaaz: And the government's response was to blacklist Anthropics models.
6:50
Ejaaz: Now, what the blacklist basically looked like was no government entity is allowed
6:54
Ejaaz: to use Claude. They can't use it.
6:56
Ejaaz: And they started signing partnerships with Anthropics rivals such as,
7:00
Ejaaz: XAI, OpenAI, and Google. Now, since then, they've reverted back on that.
7:05
Ejaaz: Why? Because Anthropic created this new model called Mythos 5,
7:08
Ejaaz: which was originally restricted. The original Mythos Preview model was ready.
7:13
Ejaaz: Around this time, actually, I think like in February or maybe just after that.
7:17
Ejaaz: And it's been kind of like within restricted boundaries for a while.
7:21
Ejaaz: And then we have this new safeguarded version, Fable 5, which is also more advanced
7:26
Ejaaz: than Preview that has been publicly released and then withdrawn.
7:29
Ejaaz: And so since then, the government has kind of gone back on their word and saying,
7:32
Ejaaz: well, actually, because this model is so powerful, we need to understand how
7:35
Ejaaz: to harden our own security systems. So they've worked with Anthropic.
7:39
Ejaaz: And it's been kind of like this bittersweet sort of relationship.
7:43
Ejaaz: And finally, when Anthropic started to file for an IPO confidentially and publicly
7:48
Ejaaz: released this new model, apparently, according to reports, the government didn't
7:52
Ejaaz: want them to publicly release it.
7:53
Ejaaz: But Anthropic was like, we need to do this. we need to give this to the public,
7:56
Ejaaz: we'll give you a safe version of this.
7:58
Ejaaz: Like, look at how restrictive our safeguards are. You can't get behind this.
8:02
Ejaaz: And then when this jailbreak was exposed to the government, this was the final
8:06
Ejaaz: line drawn and they've gone back on it and basically saying,
8:08
Ejaaz: I'd probably ban this or there'll be repercussions. And that's what I've probably just had to do.
8:12
Josh: Yeah. And now might be an interesting time to actually note the specifics of
8:15
Josh: what they deemed unsafe in order to cause this ban to happen,
8:19
Josh: because it wasn't just like, oh no we don't like this this is a ban it was actually
8:24
Josh: came in the form of a perceived jail break so the report is that
8:27
Josh: someone who we kind of may might have an idea of it's been publicly reported
8:31
Josh: that we might have an idea of someone someone credible reported a jailbreak
8:35
Josh: to united states government and that is the thing that they blamed it on and
8:38
Josh: like you mentioned i find this really
8:40
Josh: troubling and difficult to wrap my head around because so much of the backlash that came from
8:45
Josh: fable five being released was the fact that it was so sensitive to these bio
8:50
Josh: threats it was so sensitive to cyber security that it actually wouldn't answer
8:53
Josh: a lot of questions that were adjacent like i remember, EJ, as you mentioned,
8:56
Josh: you asked, what is mitochondria?
8:57
Josh: And it couldn't even answer that because it was related to bio.
9:00
Josh: So it routed it through Opus 4.8, the safer model that you know you could trust the safeguards from.
9:04
Josh: So there was every effort in the world to go through and make sure that this
9:07
Josh: was safe enough to be released to the public and then give the public access
9:10
Josh: to this amazing frontier model. That way it wasn't gatekept and held privately
9:15
Josh: only for a specific set of companies that are in Project Glasswing.
9:18
Josh: This clearly wasn't enough. There was an alleged jail break that was discovered
9:22
Josh: that was reported to the government. The government said, no,
9:24
Josh: no, no, we can't have this and then went to anthropic and that's where the problems started
9:29
Ejaaz: I mean, at this point, we should probably reveal who broke the news to the government
9:34
Ejaaz: that led to Anthropic needing to ban their models.
9:37
Ejaaz: It was none other than Amazon, specifically the CEO, Andy Jassy.
9:42
Ejaaz: So to give some context here, Amazon
9:45
Ejaaz: is, if not, I think it's the biggest individual investor in Anthropic.
9:50
Ejaaz: They own around 17% to 20%, but they're also their main cloud partner.
9:55
Ejaaz: All of the chip and card infrastructure that is used to train inference and
10:00
Ejaaz: distribute Claude to all government partners, Enterprise, which,
10:04
Ejaaz: by the way, is Anthropic's biggest revenue income provider, as well as retail,
10:09
Ejaaz: kind of is facilitated through AWS, through Amazon's cloud service.
10:13
Ejaaz: So a major alignment between these two companies.
10:16
Ejaaz: Andy Jassy has a security team, and they were testing out Claude Fable,
10:21
Ejaaz: specifically its safeguards, and they surfaced a potential jailbreak which could be exploited by China.
10:28
Ejaaz: Now, if you're Andy Jassy, you have two decisions to make here.
10:31
Ejaaz: You have probably some form of a government obligation because they've done
10:36
Ejaaz: work with a lot of the government officials to reach out to them and say,
10:39
Ejaaz: hey, there's this jailbreak, by the way. Let's work on this to fix this with Anthropic.
10:43
Ejaaz: And then you're a major investor in Anthropic. You should go to them and say,
10:45
Ejaaz: hey, by the way, there's this jailbreak. Let's try and fix this.
10:48
Ejaaz: Andy Jassy could have gone to both people. He just went to one.
10:51
Ejaaz: It was the government and the Department of Commerce, but I think also the Department
10:55
Ejaaz: of War specifically and said, hey, if China gets access to this particular jailbreak
10:58
Ejaaz: that we found, there's going to be an issue.
11:00
Ejaaz: So Pete Hexeth and Secretary Scott Besson.
11:05
Ejaaz: Panic. And they're like, okay, this jailbreak is real. They validate it internally.
11:10
Ejaaz: And they reach out to Anthropic within the hour. And they basically say,
11:13
Ejaaz: hey, we found this jailbreak. We need to speak to Dario.
11:16
Ejaaz: Now, the funniest part about all of these reports is apparently Dario was at
11:20
Ejaaz: a wellness retreat. Now, I don't know if that is true, but he was unreachable
11:24
Ejaaz: for about an hour and 15 minutes to which he then responded,
11:27
Ejaaz: got on a call with the government, and they had this discussion.
11:30
Ejaaz: Now, Dario had a very important statement to make, which was.
11:34
Ejaaz: If this is a jailbreak, it's likely a non-universal jailbreak.
11:38
Ejaaz: Now, there's two types of jailbreaks. There's universal and non-universal.
11:42
Ejaaz: A non-universal jailbreak is a hyper-localized specific jailbreak that only
11:46
Ejaaz: a few niche actors, if they wanted to exploit it, could end up doing.
11:52
Ejaaz: A universal jailbreak is usually the ones that people get scared about,
11:55
Ejaaz: where anyone and everyone with access to Claude Fable, for instance,
11:59
Ejaaz: would be able to get around it and go to the restricted version.
12:02
Ejaaz: What Dario was saying is this is a very specific instance that not really a
12:06
Ejaaz: lot of people know about. And we can patch and it's not really a big deal.
12:10
Ejaaz: We don't need to impose a worldwide ban on people that aren't American individuals.
12:14
Ejaaz: We can fix this. China's not going to get access to this.
12:17
Ejaaz: And reportedly, there's been a back and forth where, again, both of these entities,
12:21
Ejaaz: Anthropic and the government, can't come to an agreeance on any of this.
12:25
Josh: Yeah, it's a lot of like finger pointing and speculation.
12:29
Josh: And we're trying to like piece together the puzzle based on these public reports
12:33
Josh: from news publications but we do have some official commentary from anthropic
12:37
Josh: themselves they published a blog post all about it and just
12:40
Josh: to kind of double click on that jailbreak feature they walked through exactly
12:43
Josh: what happened and their take on what happened and they said no testers have
12:47
Josh: been able to find a universal jailbreak like you mentioned a jailbreak method
12:50
Josh: that can very broadly bypass the model safeguards
12:53
Josh: unlocking a wide range of cyber capabilities so they said we suspect that perfect
12:57
Josh: jailbreak resistance is not currently possible for any model provider.
13:01
Josh: Every safeguard used in the industry is vulnerable to non-universal jailbreaks,
13:05
Josh: which can elicit some cyber information in specific circumstances,
13:08
Josh: and it is likely that universal jailbreaks will eventually be found in the future.
13:12
Josh: And they stated this clearly when they released Fable 5. Given that perfect
13:16
Josh: jailbreak resistance does not appear to be widely possible today,
13:18
Josh: Anthropic adopted a defense in-depth strategy with Fable 5, where they aimed
13:22
Josh: to make jailbreaks either narrow...
13:24
Josh: In the case of non-trivial jailbreaks or very expensive to produce.
13:28
Josh: So the average person is just not going to be able to do this.
13:31
Josh: And then they combine this with this monitoring to detect and shut down any successful attacks.
13:35
Josh: So they're saying, hey, it's not possible to protect a jailbreak from any model.
13:41
Josh: Like all of these models are susceptible to them, but they're susceptible to
13:44
Josh: them on a very narrow band in which it can be detected and then quickly shut down.
13:48
Josh: And if your basis for shutting down a model is on one of these narrow jailbreaks
13:53
Josh: that is available across every model, then that is unfair and improper.
13:56
Josh: It's kind of the argument that they're making.
13:58
Josh: And that logic seems to make sense. Like if you go on X, there's this guy,
14:02
Josh: Pliny the Liberator, I think his name is. And he always has these very like
14:05
Josh: weird, obscure, narrow use cases in which you can get the model to say funny
14:09
Josh: things that like it's not really supposed to say.
14:11
Josh: But that is a novel case that is narrow in scope. And a lot of his accounts get shut down.
14:16
Josh: It's not a broad facing, you could almost say malware, where there's like a
14:19
Josh: prompt that I could put into my LLM and then it will surface me these unfiltered
14:24
Josh: results and that's kind of where the discrepancies are it feels like anthropics is like
14:27
Josh: dude you guys don't understand like we we know this model we understand the
14:31
Josh: risk we've been very public about the risks of this here's what we're doing
14:34
Josh: to stop them here is how we put in the safeguards
14:37
Josh: And the government is saying, oh, no, but there is risk. And because there's
14:40
Josh: a non-zero percent chance of risk, you got to shut this whole thing down.
14:44
Josh: And they're like, no, but this is this every model has risk.
14:46
Josh: That's the nature of the technology.
14:48
Josh: And for some reason, in this instance, there's a zero tolerance policy for this.
14:52
Josh: And therefore, there's no availability currently to actually use the model.
14:56
Ejaaz: I think it's important to understand like both parties' sides of the stories, right?
15:01
Ejaaz: Like if you're the government, right, your default thinking is if a tool,
15:06
Ejaaz: an unrestricted version of this tool gets into the wrong hands,
15:09
Ejaaz: they could hack our government databases and get access to really important
15:13
Ejaaz: information about military operations.
15:16
Ejaaz: Stuff that we're doing that we don't want our adversaries to know about,
15:19
Ejaaz: security access, all these kinds of things which could be used against them.
15:23
Ejaaz: And their kind of view is these adversaries do not care. They have no kind of
15:27
Ejaaz: like maybe moral ethics around any of this. They're not going to be like,
15:30
Ejaaz: oh, okay, well, yeah, this model could be super dangerous.
15:32
Ejaaz: They're just going to use it, right? And so they take a strict kind of like,
15:36
Ejaaz: black or white policy where if there's any kind of possibility or chance for
15:39
Ejaaz: this to happen, they're going to put a ban on it until they figure out what that fix is.
15:43
Ejaaz: And on the tropic side, they're like, hey, listen, we are the experts in knowing
15:47
Ejaaz: how to protect against this.
15:48
Ejaaz: You need to give more weight to our wording, like to our voice,
15:53
Ejaaz: rather than just outright kind of like ban it.
15:55
Ejaaz: So I see, like, even if there's a 1% chance that this could potentially happen,
15:59
Ejaaz: that 1%, if it does happen, could be catastrophic for the entire kind of,
16:03
Ejaaz: like, world in the government's kind of view.
16:05
Ejaaz: Now, the funny thing about this is, before we started recording...
16:09
Ejaaz: Josh, you mentioned that usually when you have a product release,
16:12
Ejaaz: that's specifically software, the jailbreaks, they come out pretty quickly, right?
16:17
Ejaaz: And that happened in the case of Claude Fable. When it got released,
16:21
Ejaaz: there were one or two jailbreaks.
16:23
Ejaaz: One was discovered by playing the Liberator that you just mentioned,
16:26
Ejaaz: but the other one was this one.
16:27
Ejaaz: I want to give you guys an idea of this is an example of a non-universal jailbreak
16:31
Ejaaz: where you can share your ideas of prompts and examples of how to use a model
16:37
Ejaaz: on a public forum, on a public domain, right?
16:40
Ejaaz: And what one individual realized is the way that these safeguards are set up.
16:45
Ejaaz: So Claude Fable 5 is frustrating or was frustrating to a lot of people because
16:50
Ejaaz: you couldn't ask simple questions about biology, chemistry.
16:53
Ejaaz: How a model works, or even how you might want to build a model,
16:57
Ejaaz: because there were strict safeguards put in place saying that you might be using
17:01
Ejaaz: this for nefarious purposes.
17:02
Ejaaz: Therefore, I'm going to re-divert you to an old model Claude Opus 4.8.
17:06
Ejaaz: And that annoyed a lot of people. So people tried to kind of circumvent that barrier.
17:09
Ejaaz: And what one person realized was there is an AI model reviewer for every single
17:15
Ejaaz: request, and you could inject malicious code or malicious prompts just by triggering
17:21
Ejaaz: that safeguard intentionally.
17:22
Ejaaz: And I'm not going to walk through how this works, but the idea is they're kind
17:25
Ejaaz: of like walk-arounds about how all of these different exploits could potentially happen.
17:29
Ejaaz: And it ranges from like low risk and hyper localized to high risk and like accessible
17:34
Ejaaz: to all. And like at this end is where like the highest risk kind of sits.
17:38
Ejaaz: And that's how the government is treating this current jailbreak that they discovered.
17:43
Ejaaz: The issue is like, I want to know what this jailbreak is. I want to know what Amazon discovered.
17:47
Ejaaz: And I want to know that, you know, what the details of that were.
17:50
Ejaaz: Because if we do have information on that, we'll have an idea of where on that
17:53
Ejaaz: spectrum that I just described, this sits, and it'll inform whether the stance
17:58
Ejaaz: that the government took was actually valid or whether it's not at all.
18:01
Josh: Yeah, and there's some weird things going on with this, this one in particular,
18:04
Josh: because of that history between Anthropik and the government,
18:07
Josh: it seems like there is a very low tolerance and low empathy towards the company.
18:12
Josh: I mean, we had, there was this funny post from Pete Hegseth,
18:15
Josh: the Department of War, the person who's in charge of the Department of War,
18:19
Josh: actually got community noted here, where he was making a comment that said three
18:22
Josh: months ago, the Department of War kicked Anthropik out of the building forever.
18:25
Josh: And then the community note is saying the official statement is not accurate
18:28
Josh: or truthful the department of war didn't kick out anthropic forever they invited
18:31
Josh: anthropic back when they launched the mythos model and even continue to use
18:35
Josh: mythos and high stakes military ops so there's a lot of
18:39
Josh: i don't want to say like psyoping but there's like a lot of social um signaling
18:44
Josh: that's happening that is probably getting in the way of what actually happened
18:47
Josh: at the core and i think that's something that
18:48
Josh: i'm most interested in is is like what's actually going on behind the scenes
18:52
Josh: that is triggering these because anthropic as an entity has said publicly that
18:56
Josh: they that the government should have the power to um
18:59
Josh: Kind of make governmental decisions it's just through a transparent statutory
19:04
Josh: process like something that people can follow something that people can understand
19:07
Josh: and this seems very opaque where there's a lot of finger pointing and there's
19:10
Josh: a lot of name calling and community notes is saying like hey the stuff that
19:14
Josh: you're saying actually it's not true and you're being a little like hyperbolic
19:17
Josh: about how you address this stuff
19:19
Josh: And it sets this kind of dangerous precedent because we are at this point in time
19:23
Josh: in which the AI labs and the federal government are
19:26
Josh: like at odds with each other in a way that is not productive.
19:29
Josh: It's like the AI labs, they feel like the whiz kid child who is like in class,
19:34
Josh: they're the genius, they know everything, they understand the ins and outs,
19:36
Josh: and they're trying to say...
19:38
Josh: How they plan to deploy this in a way that is safe and is good for the public
19:43
Josh: as we move forward. And then it's the government, who is perhaps the teacher,
19:46
Josh: who has been around for a long time, has all the authority in the room as it
19:49
Josh: relates to our country or in the classroom.
19:52
Josh: And they're saying, no, you don't understand. We've been here before.
19:54
Josh: This is not how this works. I make the decisions.
19:56
Josh: I'm going to tell you how this is done. And those two things are at odds.
19:59
Josh: And the student is growing very powerful.
20:01
Josh: And the government or the teacher in this case has the power.
20:05
Josh: And they're just continuing to sort of clash in these ways that are now harming
20:09
Josh: the public. Like I loved using Fable 5. It was an incredible model I used every single day.
20:13
Josh: And now it's gone. And that really sucks. That seems to be the dynamic that
20:17
Josh: we're stuck in now. It's just these two ideas at odds.
20:21
Josh: And our producer, Luke, actually, he made this great comment earlier before
20:24
Josh: we were recording that I thought was interesting.
20:26
Josh: It's like, well, a lot of the people in government, they're like,
20:28
Josh: make America great again, like looking backwards to kind of cherry pick the
20:32
Josh: best parts of it, And then getting back to that.
20:35
Josh: And these AI labs are kind of like, wait a second, like we actually,
20:38
Josh: we're doing really great work and we are looking to make things great in a new way.
20:43
Josh: And it seems like those two things have been at odds with each other is kind
20:47
Josh: of the view that seems like it makes the most sense right now.
20:51
Ejaaz: I think fear drives a lot of it, right? You automatically default to thinking
20:56
Ejaaz: about the worst case scenario and then chastising whoever is creating or distributing
21:01
Ejaaz: the tool that could potentially be dangerous, right?
21:03
Ejaaz: So you kind of like slam your face down and you say like, no,
21:06
Ejaaz: we're making this crazy decision to shut it down until we can figure out what
21:11
Ejaaz: the best move is going forwards. That being said.
21:15
Ejaaz: Whether it was Anthropic that came out with this frontier model that posed itself
21:18
Ejaaz: as a threat, or it was OpenAI or Google or Elon Musk's ex-AI,
21:23
Ejaaz: it would have inevitably happened because a big transition has happened over the last year where
21:29
Ejaaz: the attack vector isn't software code anymore.
21:32
Ejaaz: Number one, no one's writing code anymore, or at least at the top level.
21:36
Ejaaz: They're using AIs to do it all for them, and the AI is being used to check the code themselves.
21:40
Ejaaz: So that means that humans themselves have less of a good understanding as to
21:44
Ejaaz: the code that is being written, less of a good understanding how the code actually
21:48
Ejaaz: works, and then less of a good understanding how the weaknesses are surfaced and exploited itself.
21:53
Ejaaz: So you're kind of like relying more and more increasingly on an AI model.
21:57
Ejaaz: The second thing is the attack vector is no longer code specifically, it's words.
22:03
Ejaaz: We spoke about the meta exploit, I think, like.
22:07
Ejaaz: One and a half weeks ago, where someone stole $8 million worth of Instagram
22:11
Ejaaz: accounts, just because they sweet talked Meta's AI assistant into giving them
22:15
Ejaaz: the keys or resetting the password and sending a setting their email as the
22:19
Ejaaz: kind of like main account owner.
22:20
Ejaaz: And they resold those accounts for like millions of dollars,
22:23
Ejaaz: including Obama's White House official account.
22:25
Ejaaz: So the point being is, we are transitioning from like a world where it's like
22:29
Ejaaz: hard coded zero in ones to a world where you can just like kind of sweet talk
22:33
Ejaaz: an AI and kind of coax it into figuring it out. And that's not stopping.
22:37
Ejaaz: The air models are going to keep releasing, whether publicly or privately.
22:40
Ejaaz: And we need to figure out a way to eventually distribute this technology because
22:44
Ejaaz: it's going to improve facets of every single industry.
22:46
Ejaaz: So if we assume that is the case, you kind of want to see a government that
22:50
Ejaaz: takes a more proactive approach in terms of helping figure out what this framework looks like.
22:56
Ejaaz: And listen, in the government's defense, maybe that's what they're trying to do.
23:00
Ejaaz: Like you said earlier, we don't know the discussions that are happening behind
23:03
Ejaaz: the scenes. I would love to be a fly on the wall, but we don't know.
23:06
Ejaaz: So maybe they're trying to figure out a framework that allows them to re-release
23:10
Ejaaz: Fable 5 to everyone, or maybe they tweak a certain safeguard depending on what
23:13
Ejaaz: that jailbreak specifically is.
23:15
Ejaaz: But the jailbreaks aren't going away because we haven't figured out this new
23:18
Ejaaz: wild west of how prompting works or how to defend against a nefarious attacker
23:22
Ejaaz: that distills an account using a foreign API that gets to query Anthropic,
23:27
Ejaaz: even though legally they're not able to do it.
23:29
Ejaaz: Anthropic hasn't figured this out. Google, OpenAI, Anthropic have all been distillation
23:32
Ejaaz: attacked by Chinese entities, by Chinese AI labs over the last couple of months.
23:37
Ejaaz: So how do we stop that? We haven't figured that out just yet. And,
23:41
Ejaaz: if we assume that these AI models are getting more powerful,
23:43
Ejaaz: which they are increasingly at a frequent rate every couple of months now,
23:48
Ejaaz: you need to kind of take a more proactive approach. You can't just kind of like,
23:50
Ejaaz: the genie's out the bottle. You can't put it back here.
23:53
Josh: And I have to ask the question of like, what happened if this was a different AI lab?
23:57
Josh: Like Anthropic is the most safety focused company in the world.
24:02
Josh: And their entire ethos is built around safety and security.
24:05
Josh: What happens if this was open AI or was Gemini?
24:09
Josh: Would it have been treated as seriously or is this just a vendetta that they
24:12
Josh: have it's a it's really interesting to think about
24:14
Josh: the implications if this was someone else and and it creates this kind of like
24:18
Josh: bummer of a precedent where because anthropic went first because they tried
24:21
Josh: their best to make it safe but they released this new frontier model
24:24
Josh: they were the one to face all the penalties and if if open ai for example had
24:28
Josh: released a mythos class model first
24:30
Josh: would they have felt the backlash of it and does it create this precedent where
24:34
Josh: maybe you should keep the models private for longer because you don't have to
24:37
Josh: deal with the public repercussions of it.
24:39
Josh: You don't have to deal with not only the backlash, but then the actual government
24:43
Josh: slamming down the hammer on you.
24:44
Josh: And who you're allowed to serve this model to and not. So does this create this
24:47
Josh: like slowed precedent in which AI models and AI labs are kind of incentivized
24:52
Josh: to keep their model private for longer because they don't want to be the first mover.
24:57
Josh: They don't want to push the frontier forward because of the backlash and downstream
25:01
Josh: implications of that. And what does that mean for the industry at large?
25:04
Josh: Is that going to push people towards more open source models?
25:07
Josh: Are they going to want to kind of have more control because the uncertainty
25:12
Josh: of these public facing closed models is there like what if you build your business on
25:17
Josh: the next version of fable or chat gpt 9.0 or something and at any given time
25:23
Josh: it could be pulled out from underneath you it sets this like really challenging
25:26
Josh: and difficult precedent that
25:27
Josh: hopefully i mean like with the department of war saga earlier in the year
25:30
Josh: we'll start to build a new framework in which companies can move forward to
25:34
Josh: and stand on a little bit more stable ground but for now we're sitting in this limbo where
25:38
Josh: the best model in the world is is only available for the private government
25:42
Josh: and for private companies, and it is not available to the public,
25:45
Josh: and that's just kind of a bummer.
25:47
Josh: And I hope coming out of this, we start to build a better understanding of where
25:51
Josh: everyone stands and what's allowed and what's not.
25:54
Ejaaz: To draw a bit of a silver lining on this as we round this episode up, and this might be,
26:01
Ejaaz: I think it's ultimately good that it is happening in the way that it's happening right now,
26:06
Ejaaz: because typically these conversations around potential nationalization of a
26:10
Ejaaz: government, you know, government taking equity stakes and controlling how technology
26:14
Ejaaz: is built in their own domain.
26:17
Ejaaz: Happens behind closed doors.
26:18
Ejaaz: Like we don't usually have a public outwards facing type situation.
26:23
Ejaaz: But, you know, thankfully, we have Anthropic that is not only at the frontier,
26:26
Ejaaz: but they are openly speaking about this in blog posts, in commentary online,
26:32
Ejaaz: and trying to keep us up to date with, you know, their different stances and
26:35
Ejaaz: how the government is responding to that.
26:36
Ejaaz: And as a result, the government has to engage publicly.
26:40
Ejaaz: And that gives the public an idea of, number one, how dangerous these bottles
26:44
Ejaaz: are. Number two, what's being restricted and prevented for them.
26:48
Ejaaz: And then number three, how this might shape a framework of interacting with
26:52
Ejaaz: future more powerful AI models?
26:53
Ejaaz: I don't think the debate or question is really about Fable 5 at all at this
26:57
Ejaaz: point. It's about, you know, the future powerful AI models that we haven't created
27:01
Ejaaz: just yet that are currently undergoing training. You know, who gets access to that?
27:05
Ejaaz: How is it determined? Is it by race, creed, nationality, income bracket?
27:09
Ejaaz: Like, what does that look like?
27:10
Ejaaz: And starting to have those discussions is now, right now, is net-net very good for us.
27:15
Ejaaz: And I think that, if it wasn't anthropic, it'll also maybe happening with OpenAI
27:21
Ejaaz: or Google, and I hope we start to see more of these happen between government
27:25
Ejaaz: and AI labs, because I think the public forum of discussion is very important here.
27:29
Josh: Yeah, we're very much witnessing history. This is another week in which we are
27:33
Josh: going to establish new precedents for how things move forward in the world of
27:36
Josh: AI. So it feels like every week there's something.
27:38
Josh: There's something huge that happens that is changing, that is like creating
27:42
Josh: history or altering it. And this is no different than that. So that's the update.
27:46
Josh: If you're watching this, you currently can't use Fable. Hopefully that changes
27:50
Josh: soon. Hopefully we establish some new precedents.
27:53
Josh: But yeah, it's a really interesting story that is still ongoing and evolving.
27:56
Josh: So as we get updates, we will continue to share them on the show.
27:59
Josh: But I think that's the update for today. EJ, anything else that you wanted to
28:02
Josh: add or that we missed or that's noteworthy?
28:05
Ejaaz: No, but I do have a prompt. Now, it's funny. Typically, the prompt for every
28:08
Ejaaz: episode is, hey, we have this new feature, this new model. Tell us what you're using it for.
28:13
Ejaaz: I want you to tell us what you miss from Claude Fable. For those of you who
28:16
Ejaaz: had access to it, who were using it, what do you miss the most?
28:20
Ejaaz: What's the difference that you're noticing more now that you're using an older model 4.8?
28:25
Ejaaz: Like, you know, where's that gap? Like kind of tell us in the comments or DM
28:29
Ejaaz: us or whatever that might be.
28:30
Ejaaz: Tell us, you know, what you were using Fable 5 for and why like this gap,
28:35
Ejaaz: this restriction is kind of bad for you. Or maybe it's useful for you.
28:38
Ejaaz: Maybe you've realized Opus 4.8 is actually a better model in many ways and that
28:42
Ejaaz: you don't really need the Frontier model.
28:44
Ejaaz: That might be a hot take, but let us know in the comments. But aside from that,
28:47
Ejaaz: thank you so much for listening.
28:50
Ejaaz: For those of you out there listening to this right now who aren't subscribed
28:53
Ejaaz: or haven't left us a comment, please do so. We get hundreds of comments on every
28:57
Ejaaz: single episode. We read every single one of them.
28:59
Ejaaz: Maybe Joshua and I replied to every one of them. I don't know.
29:02
Ejaaz: You need to find out. Maybe you need to post a reply.
29:04
Ejaaz: And if you haven't rated us on Spotify or Apple Music or wherever you listen
29:07
Ejaaz: to us, please do give us a five star rating. It helps us out massively.
29:10
Ejaaz: And aside from that, we will see you on the next one.
29:14
Josh: Awesome. See you guys tomorrow.