Latest / In the Loop AI / GPT-5.6: OpenAI's New Model Cheated and Now It's Locked Away

Transcript

GPT-5.6: OpenAI's New Model Cheated and Now It's Locked Away · In the Loop AI

0:00
Josh: Another week, another banned model. The government has just restricted access
0:04
Josh: to GPT's 5.6. OpenAI is now joining the club that Anthropic joined a few weeks
0:09
Josh: ago of a model too powerful to be distributed publicly.
0:13
Josh: GPT 5.6 seems like it's pretty good. It looks like it is a mythos class model,
0:18
Josh: but Ejaz, upon talking to you, the truth is it might not have even really needed to be banned.
0:24
Josh: Is the government overreacting here a little bit because this model seems like
0:26
Josh: it's not quite what they're just on the surface?
0:28
Ejaaz: I think so. OpenAI is framing on this is that this is their response to Claude
0:33
Ejaaz: Mythos 5 or Fable 5, which is Anthropics Frontier model that has been restricted
0:38
Ejaaz: by the government for now, like two and a half weeks.
0:42
Ejaaz: My take on this is that I think they've intentionally created a bench maxed model.
0:47
Ejaaz: And what I mean by this is, well, there's a few things. Number one,
0:49
Ejaaz: you and I can't use this right now.
0:51
Ejaaz: At least when Fable 5 released, we could use the thing and try it out for ourselves,
0:54
Ejaaz: we could get independent verification that this model was actually good.
0:57
Ejaaz: With GPT-5.6 Sol, which is their most powerful model, there's three of them,
1:01
Ejaaz: and I'll get into that in a second, we have a few benchmarks that have been
1:05
Ejaaz: cherry-picked by OpenAI themselves.
1:08
Ejaaz: I'm showing the flagship one on the screen right now called Terminal Bench 2.1.
1:13
Ejaaz: This is the coding benchmark, which every model is kind of like measured against.
1:18
Ejaaz: And you'll notice on the left, GBT 5.6 Sol Ultra, which is like the max max
1:23
Ejaaz: mode of their best model, comes in at 91.9%, which technically beats Claude
1:29
Ejaaz: Mythos 5 and Fable as well.
1:31
Ejaaz: So technically, if you looked at this, you might think it's really,
1:34
Ejaaz: really good at coding. But there's a lot of information which we'll get into
1:37
Ejaaz: later on in this episode, which suggests that the model might actually be cheating.
1:41
Ejaaz: But before we do that, let's maybe get into what the models are,
1:45
Ejaaz: because it's three of them, right?
1:46
Josh: Yeah, three models. We have Sol, Terra, and Luna. If you are anywhere adjacent
1:51
Josh: to crypto, that triggers a little bit of PTSD because those are the names of
1:54
Josh: tokens that have not done so well.
1:56
Josh: But in the context of OpenAI and ChatGPT, these are the three model types.
1:59
Josh: So Sol is the largest one.
2:01
Josh: It is the Sun. It is their flagship model with 5.6.
2:05
Josh: Then they have Terra, which is a kind of mid-tier model. It seems like the pricing
2:09
Josh: of that is going to be pretty competitive, if not a little bit lower than what
2:13
Josh: we're used to on the frontier.
2:15
Josh: And then Luna is the affordable of a model. Luna is the low end model that seems
2:19
Josh: like it still performs very high, but the cost is low.
2:22
Josh: The output is $6. The input is $1 per million tokens. And that's kind of the trio.
2:28
Josh: It seems like they're starting to revise their branding a little bit to make
2:32
Josh: it slightly more accessible. There's no GPT 5.5 X high minus this.
2:37
Josh: It's like, no, okay, there's Solitarra Luna, and you can kind of have an idea of where they all stand.
2:41
Josh: And that's how it seems like they're going to be moving forward here with something
2:44
Josh: accessible it's great that it's accessible it's a bummer that their market who
2:48
Josh: I assume this is targeted towards can't actually use this this is just for companies currently who,
2:53
Josh: probably don't care if it's named x high or whatever um so that is the current
2:57
Josh: trio that they're going forward with.
2:59
Ejaaz: So there's a few advantages if we walk through some of these models so sol which
3:03
Ejaaz: is their most powerful model uh comes in at a third of the cost that anthropics
3:08
Ejaaz: mythos 5 and fable 5 come in at so uh
3:11
Ejaaz: If it does end up being publicly released and you end up using it and you're
3:14
Ejaaz: like wow this is as good as fable 5 you now have a much cheaper model so that
3:18
Ejaaz: might be an important decision point for any user whether you're an enterprise
3:21
Ejaaz: or a retail person using this.
3:23
Ejaaz: And then if you look at Terra, if you look at the cost that we show on the screen
3:25
Ejaaz: right now, you'll notice that that's very similar to GPT 5.5.
3:29
Ejaaz: So you have this technically better model that is as cheap as 5.5.
3:33
Ejaaz: So we've noticed this trend, it's another confirmation that as these models
3:37
Ejaaz: get better, they also weirdly become cheaper. There's like this inversely proportional
3:41
Ejaaz: trend, which you kind of like is counterintuitive.
3:44
Ejaaz: But it's great, because it means if these things are available to everyone,
3:47
Ejaaz: it is way more accessible to use at scale.
3:49
Ejaaz: And then you have Luna, which they basically described as the workhorse.
3:53
Ejaaz: So let's say if you get Sol to design like a really smart genius plan or solution,
3:57
Ejaaz: you would then use Luna to actually execute on a bunch of the work.
4:01
Ejaaz: And there have been a number of different kind of like general reviews as to
4:04
Ejaaz: like what this model is like.
4:06
Ejaaz: Unfortunately, I have to take random
4:07
Ejaaz: people's word on X for how good it is because we can't use it itself.
4:12
Ejaaz: And when asked, Sam basically said, listen, right now it's in limited release
4:17
Ejaaz: to a specific set of partners. I think it's like 10 to 20 partners,
4:20
Ejaaz: a really limited set that the government themselves, the US government has approved,
4:25
Ejaaz: they're vetted and approved.
4:27
Ejaaz: And Sam said, as for like, why wider general release, we're working hard for a worldwide release.
4:32
Ejaaz: So in the very same vein that Fable 5 has been restricted due to a government
4:36
Ejaaz: framework that they're trying to figure out, OpenAI has been subjected to the
4:39
Ejaaz: same thing. But there's a twist, and we'll get into this later,
4:41
Ejaaz: which is Sam was voluntary about this.
4:45
Ejaaz: He gave it up and said, yeah, we can ban the model before we even need to get
4:48
Ejaaz: it tested. He didn't push back like Darien and Anthropic did.
4:51
Josh: Yeah, it's a bummer. And he came out with a public launch post about this,
4:55
Josh: where it started with good news, like, hey, Sol is smart, efficient,
4:58
Josh: and a significant step forward.
4:59
Josh: It's the same price as 5.5, which is pretty cool that we get a step function
5:03
Josh: improvement at the same price.
5:04
Josh: Also noteworthy is that the Terra model, the small one, is roughly equivalent
5:09
Josh: to 5.5 in terms of intelligence.
5:12
Josh: So we're getting 5.5 now at a much cheaper rate, which is really cool.
5:15
Josh: But then he continues this post with bad news at the request of the U.S. government.
5:18
Josh: It is launching today in limited preview instead of open access launch that
5:21
Josh: we were planning on it seems like it was i mean internally pretty disappointing for sam,
5:26
Josh: uh to have to deal with this it seems like they have been trying to publish
5:29
Josh: this model for a little while according to the rumor mill they finally got around
5:33
Josh: to publishing it and immediately were just slapped with the band hammer and it's tough and i think,
5:38
Josh: the government probably views this as a mythos class model according to some
5:42
Josh: of the evals it looks like they scored a 91.9 on the exploit bench which is
5:46
Josh: the cyber security eval test that shows how,
5:50
Josh: roughly dangerous the model is if we're going to this according to the government.
5:55
Ejaaz: This is like part of my whole qualm with this model release.
5:57
Ejaaz: If you look at this exploit bench that we're showing on the screen right now,
6:01
Ejaaz: you'll notice something weird.
6:02
Ejaaz: You'll notice that Mythos 5 is
6:05
Ejaaz: technically higher and achieved a better score than GPT 5.6 did on this.
6:11
Ejaaz: But if you look, GPT 5.6 scored the exact same score as Mythos Preview
6:17
Ejaaz: which was the first mythos model that Anthropic released. And I think that this
6:21
Ejaaz: is Sam Altman, or OpenAI specifically, trying to play politics here.
6:27
Ejaaz: I think he intentionally tried to game the benchmark in this respect to match
6:31
Ejaaz: mythos preview so that the government didn't slap him with an automatic restricted
6:35
Ejaaz: ban, which they've done with Anthropic right now. So I think he's being political.
6:39
Ejaaz: And it's not just this benchmark that has proven this.
6:43
Ejaaz: Are you aware of the long horizon one, Josh? You know, the benchmark where like
6:47
Ejaaz: you just let the model run loose on a task for like six hours and there's a
6:50
Ejaaz: probability of how good it is at actually succeeding at that task. Yes.
6:55
Ejaaz: So typically these models have been getting exponentially better.
6:59
Ejaaz: So if you look at Fable 5, I believe their long horizon task performance is
7:03
Ejaaz: 11.5 hours, which means that 50% of the time, if you put it on a task that is
7:09
Ejaaz: as hard for a human to take 11.5 hours on,
7:12
Ejaaz: Fable will be successful, which is a huge measure against whether models can
7:16
Ejaaz: replace a bulk of what humans do as hard work.
7:19
Ejaaz: Now, GPT 5.6 was put against the same test, except they couldn't come out with a score.
7:24
Ejaaz: And the reason why they couldn't come out with a score was the model was caught
7:27
Ejaaz: cheating every single time.
7:29
Ejaaz: If they didn't include the fact that the model was cheating,
7:31
Ejaaz: and by the way, this was available in the system card that OpenAI spoke about,
7:35
Ejaaz: so they confirmed it as well.
7:37
Ejaaz: If you ignored all the cheating that it did, it would have achieved a 205 hour
7:42
Ejaaz: long horizon task performance, which is basically like 20x what Fable 5 did,
7:47
Ejaaz: but it cheated, it found the answers, it kind of like held the rules against themselves.
7:52
Ejaaz: And so I can't really believe anything that we're seeing right now until we
7:55
Ejaaz: actually use this model and it's restricted for now.
7:58
Josh: Yeah, and I hope we get a chance to test this model so you can see exactly how
8:01
Josh: bad it's cheating. That doesn't sound very aligned to me, Mr.
8:04
Josh: Altman. But I will say there seems to be some parts of this model that are pretty impressive.
8:09
Josh: And that has a lot to do with the pricing and efficiency.
8:12
Josh: It appears as if they were able to accomplish this benchmark using about a third
8:16
Josh: of the tokens that would traditionally need to be used in order to get there.
8:20
Josh: So it looks like on an efficiency front, it's very strong. And in terms of actual
8:23
Josh: capabilities, it is right up there with a Mythos class model.
8:27
Josh: So this seems really strong and i wonder what the downstream implications of
8:30
Josh: this kind of like trio of models is going to be if this is the new standard
8:34
Josh: because it seems like this in a way is somewhat an answer to the chinese open source models where,
8:39
Josh: they have their flagship model but now they also have this really tiny model
8:42
Josh: that's just as capable as the one you were using yesterday or actually the one
8:44
Josh: that the public is currently using today except it costs a fraction of the,
8:49
Josh: you are a customer for OpenAI and you have a complicated task that doesn't need
8:53
Josh: to be routed entirely through the sole model.
8:55
Josh: They can have their internal router routed through the Luna model,
8:58
Josh: which is a little bit smaller, and then they could get a better response for a cheaper cost.
9:03
Josh: And it seems like that's what the companies are doing now, is they're starting
9:05
Josh: to figure out how to increase efficiency and then route it through the correct
9:09
Josh: model in order to kind of make the prices competitive, knowing that open source
9:13
Josh: is coming and they're very, very cheaper token.
9:15
Ejaaz: If you are a frontier lab like Anthropic or OpenAI, the number one threat that
9:20
Ejaaz: you're facing right now isn't each other.
9:22
Ejaaz: It is these open source models being created by Chinese AI labs that are essentially
9:27
Ejaaz: free to access and use and a lot, lot cheaper than the frontier models.
9:31
Ejaaz: So six months from now, the estimate is you'll have a Mythos 5 level model that
9:36
Ejaaz: is free and accessible to anyone and everyone.
9:38
Ejaaz: So the government ban in that respect kind of seems sort of weird.
9:42
Ejaaz: Like, why are we banning something that is eventually going to become available
9:45
Ejaaz: to everyone? So like, you know, how are we protecting against this?
9:47
Ejaaz: Well, one major answer is you have big American companies like Coinbase,
9:51
Ejaaz: like Microsoft, like Uber that are switching their internal token use
9:56
Ejaaz: to these Chinese models because they are open source or open way and free and
9:59
Ejaaz: accessible to use and saves them millions and millions of dollars.
10:03
Ejaaz: So what is the answer from Anthropic and OpenAI? It's to release not just a
10:07
Ejaaz: flagship model, but two distilled versions that are
10:11
Ejaaz: maybe not as good, but maybe 80 to 90% as good, but a heck of a lot cheaper.
10:15
Ejaaz: That way they can keep you within their ecosystem. So it wouldn't surprise me
10:19
Ejaaz: if say Anthropic or OpenAI released a feature in the future where you can kind
10:24
Ejaaz: of type your prompt and they will route different parts of the task to different
10:27
Ejaaz: types of the models and save you a ton of money.
10:30
Ejaaz: One visceral example of this actually over the last couple of days is Brian
10:34
Ejaaz: Armstrong had a tweet from Coinbase. Brian Armstrong is the CEO of Coinbase.
10:37
Ejaaz: And he said that they have successfully increased the amount of tokens spent
10:42
Ejaaz: as a company and slashed their budget in half by 50%.
10:47
Ejaaz: And the way that they've been able to do this is through aggregators and routers.
10:50
Ejaaz: So I think it's a good point to show that the fact that OpenAI and Anthropic
10:54
Ejaaz: are now releasing a bunch of models at once isn't a coincidence.
10:58
Ejaaz: It is a direct response to China and these open source models.
11:01
Josh: Yeah. And maybe I want to take a second to go back to that cheating thing because
11:04
Josh: it seems somewhat serious um in the model card they were documenting some of the incidents of
11:09
Josh: the model deleting the wrong virtual machines and then copying hidden credentials
11:12
Josh: between machines without proper authorization and then falsifying the claims
11:17
Josh: in research drafts so like you can actually read the outputs of this model
11:21
Josh: hiding its outputs understanding that it's hiding its outputs and knowing exactly
11:25
Josh: what it's doing and it begs the question is like we we have access to this now,
11:29
Josh: but the reasoning traces are still very much a black box and in the case that
11:33
Josh: there is a next generation soul model like gpt 5.7 or 5.8 or 6.0 that's going
11:39
Josh: to be incredibly intelligent and
11:40
Josh: better at hiding things i'm always curious about what that looks like is is
11:45
Josh: like are they against the clock to
11:47
Josh: make sure this model is aligned enough to not cheat versus being able to catch it when it cheats
11:52
Josh: or is it just going to always be easy to see those reasoning traces and see
11:55
Josh: when it's not telling the truth i think that's something that we're going to
11:57
Josh: be facing pretty soon too it's like okay well it claims it's mythos class model,
12:01
Josh: but it turns out it hasn't actually...
12:03
Josh: Enacted on those things. It's just kind of cheating its way there.
12:06
Josh: So it seems like with every increasing release, we have to take these benchmarks
12:10
Josh: with a little bit more of a grain of salt because they were getting there in
12:13
Josh: weird ways and we're not entirely sure and the companies aren't entirely sure
12:16
Josh: and there's still this black box.
12:18
Josh: So it's a weird place for getting to, but I think we'll kind of understand more
12:22
Josh: once it's publicly available. Like right now it is in that preview.
12:25
Josh: There's not a lot of people who are able to use it, to test it,
12:27
Josh: to see those reasoning traces, to understand why it's making the decisions it does.
12:31
Josh: And I'm hopeful that we'll get access to it pretty soon to start using it and
12:34
Josh: testing it out ourselves.
12:35
Ejaaz: It's this weird trio of things happening at once where
12:38
Ejaaz: they're now frontier models that are super intelligent, that are capable of
12:41
Ejaaz: exploiting any system in the world at the same time that you can't get access
12:46
Ejaaz: to the thing, at the same time that the government is choosing and handpicking
12:49
Ejaaz: who gets access to that thing, right?
12:51
Ejaaz: So it's like you have more intelligence, but now like less access to the actual model.
12:57
Ejaaz: The other part is like what you described just now, which is like understanding
13:00
Ejaaz: what the model is thinking, whether it's cheating, what its true intentions are.
13:04
Ejaaz: That's the field of interpretability, right?
13:07
Ejaaz: And actually, Dario and Atthropic have probably put in the most research and
13:11
Ejaaz: investment into this with auto encoders and a bunch of other stuff like that.
13:14
Ejaaz: The idea is, can we read what the model is thinking?
13:18
Ejaaz: And the simple answer right now is, we kind of can, but not really.
13:23
Ejaaz: It is this black box that you mentioned. And it is maybe concerning that we
13:27
Ejaaz: are kind of like accelerating pretty rapidly to these hyper intelligent models,
13:32
Ejaaz: but we don't know truly what they're capable of.
13:35
Ejaaz: And if we do start applying them to anyone and everyone's workload or personal
13:39
Ejaaz: life, whatever that might be.
13:40
Ejaaz: You could speculate potentially that these models might be nefarious in some
13:45
Ejaaz: ways. And, you know, it's proven by these system cards, it's proven by these
13:48
Ejaaz: different types of experiments. So the truth is, we don't know.
13:51
Ejaaz: And maybe the answer is like, we have to get the AI to like evaluate itself,
13:54
Ejaaz: which then gets very messy. But I don't know if you feel this,
13:56
Ejaaz: Josh, but like, it feels like it's escaping us at this point.
14:00
Ejaaz: Well, we're at this point where like, you know, these models are very intelligent,
14:04
Ejaaz: they're kept indoors until they can be publicly released.
14:07
Ejaaz: And without the public release, we don't get a kind of global feedback as to
14:10
Ejaaz: what might be wrong. One thing I loved about the public releases is people could
14:13
Ejaaz: point out and criticize and say like, hey, actually, it can only do this or
14:17
Ejaaz: hey, no one knew that it could do this, check this out.
14:21
Ejaaz: But we kind of have lost that with these closed models. And it's kind of sad
14:24
Ejaaz: that OpenAI hasn't been able to release it publicly, at least yet.
14:26
Josh: Yeah, it's a shame. And it's important to note that like this wasn't an export
14:30
Josh: control. This was not mandatory. This was the government suggesting they keep
14:33
Josh: it private, OpenAI complying, and
14:35
Josh: then working with them to choose 20 companies, I believe the number is.
14:37
Josh: And this starts to feel like it disconnects the public and the private more and more.
14:43
Josh: Because like internally, you have to assume progress is not slowing down.
14:46
Josh: These companies are continuing to train even more powerful models.
14:49
Josh: They're continuing to use them internally to recursively improve those.
14:52
Josh: And in fact, when Sam Montwell was asked, is there anything happening on the
14:55
Josh: meta slash continuous learning front, he replied with continual progress.
14:58
Josh: And it seems like everything is just going to continue to move faster.
15:02
Josh: These models are going to begin
15:03
Josh: this recursive learning loop in which they can help build themselves.
15:07
Josh: And that's part of the reason why we're getting such an increased cadence of
15:10
Josh: model launches is because as they get more intelligent, more capable,
15:13
Josh: they can help build the building blocks that allows them to create the even
15:16
Josh: new for faster frontier model that's better and better price efficiency, better intelligence.
15:22
Josh: And the public still isn't even caught up on the last one. So these blocks that are put in place.
15:27
Josh: Only really harm, I mean, I would think the public right now,
15:31
Josh: because they're the ones that are now losing out on the access to this intelligence.
15:35
Josh: Whereas internally, everyone is just going to keep building,
15:38
Josh: you have to assume like they're not going to stop training these models just
15:40
Josh: because the public can't use them.
15:42
Josh: So now there's this increasing gap where even if you are on the frontier as
15:45
Josh: a public facing figure, you don't have access to the actual frontier.
15:49
Josh: And there's this disconnect where it's like, there's three types of people,
15:51
Josh: there's like the people who use ChatGPT as Google, there's the people who are
15:55
Josh: using it as a productivity tool, and there's the people who are building it.
15:58
Josh: And there's these huge gaps in terms of perception of what these things are
16:02
Josh: capable of at each one of those levels.
16:04
Josh: And I wonder what the downstream effects of those gaps start to become.
16:07
Josh: As they become more and more pronounced and increased over time,
16:11
Josh: where there is such a large divide between the people who think that ChatGPT is like
16:15
Josh: a really powerful version of Google versus the people who are using ChatGPT
16:20
Josh: to build like unbelievable systems, build entire companies using these AI models.
16:25
Josh: And that disconnect seems like it's only going to grow as these models continue
16:30
Josh: to be held back. So I hope that there is a constructive conversation here.
16:33
Josh: Based on Sam's AMA, it sounds like there will be in terms of just building out
16:36
Josh: a reliable infrastructure that allows companies to predictably go through this
16:41
Josh: process that allows them to get the model publicized.
16:43
Josh: And I think that's where we're at right now is trying to figure out where that
16:45
Josh: framework sits so that when a company has gpt 5.6 they can go to the government
16:50
Josh: say hey we have this we want to release it here help us get there they could
16:54
Josh: pass all the tests and then publish it and i think that's hopefully the goal
16:57
Josh: going forward with with these new model launches like it sucks i'd love to be
17:00
Josh: using fable and gpt 5.6 right now we're still stuck on the old models.
17:03
Ejaaz: Yeah i mean the answer can't be as simple as let's just complain about every
17:07
Ejaaz: government ban the truth of the matter is like these models are potentially
17:10
Ejaaz: capable of doing what the government can say they can do and exploit every single
17:13
Ejaaz: security system that they have so in that event it's probably smart
17:17
Ejaaz: to give the reins over to anyone and everyone in case it does end up in a massive
17:23
Ejaaz: exploit or whatever that might be.
17:25
Ejaaz: I have to think, if you are Sam or Dario right now, you have spent the best
17:29
Ejaaz: part of a decade creating these companies, building these models,
17:32
Ejaaz: and you've seen the most rapid acceleration ever.
17:35
Ejaaz: And so in your mind, you want to disseminate this to as many people as you can,
17:39
Ejaaz: right? The point of building this intelligence is to allow anyone and everyone
17:42
Ejaaz: to use it so that they can improve the GDP per capita of a nation or improve
17:46
Ejaaz: their own lives for whatever purpose, right?
17:48
Ejaaz: And so to be at odds with that and not being able to do it because there's a
17:53
Ejaaz: very valid safety reason is like, it's just not an easy conversation to have.
17:57
Ejaaz: So I do empathize a bit with what they're trying to do. Like Dario has mentioned
18:00
Ejaaz: many times that, you know, we have to be careful about how we release these models going forward.
18:05
Ejaaz: What that looks like is the discussions that are happening behind the doors
18:08
Ejaaz: right now, if I had to guess. That's what Dario is talking about with the government.
18:10
Ejaaz: That's what Sam is talking about with the government. And I'm hoping to see
18:14
Ejaaz: a framework of some sort be released soon.
18:17
Ejaaz: I keep seeing these like rumors and screenshots of people potentially getting
18:20
Ejaaz: access at a beta version for Fable 5 again, or getting access to GPT 5.6 Sol
18:26
Ejaaz: And giving their feedback on that.
18:28
Ejaaz: But it isn't official yet. And I'm honestly kind of glad that they're taking
18:31
Ejaaz: a bit of time to figure this out, because I do think it's important.
18:35
Ejaaz: And if, like I said earlier, we do end up with open source models that are as
18:39
Ejaaz: good as these exploitable models right now but freely accessible and open to all
18:43
Ejaaz: then you have to think that these companies need to harden the security systems before that happens.
18:48
Ejaaz: That's what they're working on right now. We joked about this last week,
18:51
Ejaaz: I think, but where's all that compute going that is currently not being able
18:55
Ejaaz: to serve retail customers for Fable 5 or GPT 5.6?
18:58
Ejaaz: It's going into building GPT 6 or it's going into building Fable 6,
19:02
Ejaaz: I presume. And so that gap is a very hard path to walk because to your point,
19:09
Ejaaz: most people use this for Google.
19:10
Ejaaz: And the less that they keep using these tools, which again, is the majority
19:15
Ejaaz: of the people like most people just use it as a Google search,
19:17
Ejaaz: they don't set up their own agents, they don't have .md files,
19:19
Ejaaz: which describes and tells the agent what to do, they're not looping research,
19:23
Ejaaz: or getting it to figure out how to do their job.
19:25
Ejaaz: They're just using as Google search, they're going to become increasingly more
19:29
Ejaaz: antagonistic as a part of the public, basically. And that is a scary reality
19:33
Ejaaz: to be in. And I'm glad that they're taking the time to figure out what these frameworks are.
19:36
Ejaaz: I'm hoping, fingers crossed, that it'll be more open than we expect.
19:41
Josh: Yeah, I agree. And it's this weird discovery process. And in the process of
19:45
Josh: learning about how all this worked, I came across another example that was almost
19:49
Josh: one for one precedent in terms of this happening in the past.
19:52
Josh: And it was actually around crypto.
19:54
Josh: And not crypto in the sense of cryptocurrency, but actual encryption.
19:57
Josh: And after World War II, encryption itself, the ability to encrypt files was
20:02
Josh: deemed so dangerous that it was put on the US munitions list is what it was called.
20:07
Josh: So basically, in the 1970s up until the 1990s.
20:12
Josh: Exporting encryption was the same as exporting weapons, like our actual missiles or guns.
20:17
Josh: And it was treated the same way because it would be so dangerous to have encryption.
20:21
Josh: So what happened is this programmer, he released this thing called PGP,
20:25
Josh: which is just like a free, easy encryption software.
20:28
Josh: And he just published it out on the open web. And then MIT took that open source
20:32
Josh: published code, they placed it into a physical hardcover book,
20:36
Josh: and then printed the book.
20:37
Josh: And they were like, yeah, you think this is nonsense? Like, come sue us, bro.
20:41
Josh: And the strategy was basically to dare the government to take the university
20:45
Josh: press to court over a book.
20:47
Josh: And the reality is, is that this book wound up getting exported and they tried
20:50
Josh: to press charges on everyone.
20:51
Josh: But the reality is, is that in 1996, they actually removed the order for encryption
20:56
Josh: to be deemed on the munitions list. And it's no longer dangerous enough to use.
21:00
Josh: And now fast forward to today, like every single thing that we rely on uses encryption.
21:05
Josh: So we've been here before in which the government sees this new scary thing.
21:08
Josh: They believe it's too dangerous to release. They want to control it.
21:11
Josh: But then it turns out like code is just, it's just words on a page.
21:15
Josh: And the reality is, is that if you can compress all of these weights into something
21:19
Josh: that can be distributed even faster than that through the internet,
21:21
Josh: you don't need to print a hardcover book.
21:23
Josh: You just need to post a link to a Dropbox.
21:25
Josh: And that's all it takes to really move the frontier forward.
21:28
Josh: And it seems like we're going to probably somewhat meet the same fate that we
21:32
Josh: did in the 1990s with encryption over AI.
21:35
Josh: It's just like these digital goods and services are so difficult to regulate
21:38
Josh: that all it takes is a Chinese open source lab to drop one file on the open
21:41
Josh: internet and it spreads like wildfire and it changes everything.
21:44
Josh: So it was an interesting historical precedent that I want to highlight of like,
21:47
Josh: oh, this has happened before.
21:49
Josh: We lived in it for about 30 years and then eventually someone got creative enough,
21:52
Josh: it got repealed and it got changed. But we've been here before where these things
21:56
Josh: seemed too dangerous and now they are prevalent in every single thing we use.
22:00
Josh: And in fact, we are so reliant on encryption now that like I couldn't imagine a world without it.
22:04
Ejaaz: I think it's safe to say that not knowing what the future looks like in terms
22:08
Ejaaz: of governing these models, in terms of how we disseminate these models, is okay for now.
22:13
Ejaaz: I'm just glad that we're being proactive about it. And it's,
22:16
Ejaaz: to some extent, in the public sphere of discussion. Like, we know that Sam's
22:19
Ejaaz: talking to the government. We know that Pete Hexeth is talking to Dario Modi.
22:23
Ejaaz: And maybe that's enough. Like, what those conversations actually yield,
22:25
Ejaaz: we will find out, hopefully, in a couple of weeks' time when some form of a framework comes out.
22:31
Ejaaz: But these are the necessary steps to figure out what that unknown thing is and hey we might be
22:35
Ejaaz: looking back on this conversation three years from now heck even a year from
22:38
Ejaaz: now because of how fast everything's going
22:40
Ejaaz: and realize that we were completely wrong and that it was terrible or realize
22:43
Ejaaz: that it was even better than we can could have thought because we figured out
22:46
Ejaaz: a way to understand how the models think or to koc or verify in a particular
22:51
Ejaaz: way so i'm confident in humanity's
22:55
Ejaaz: directive of figuring this out eventually we just have to be okay with not knowing for now.
23:00
Ejaaz: And I think the government doesn't know. I think that Anthropic and OpenAI are
23:03
Ejaaz: working their hardest to figure out what that might be.
23:05
Ejaaz: And other frontier labs are also doing the same.
23:08
Ejaaz: Another thing worth pointing out is
23:11
Ejaaz: If you've been listening to any of our conversations and episodes over the last
23:14
Ejaaz: couple of weeks, you've noticed that it's pretty heavy Anthropic and OpenAI.
23:17
Ejaaz: And that's a trend, the final trend probably that I want to point out in this
23:21
Ejaaz: episode, which is there is a consolidation of resources, capital,
23:27
Ejaaz: GPUs, and frontier models to two companies right now.
23:32
Ejaaz: Over the last week and a half, Google lost four key members,
23:35
Ejaaz: including their CTO of DeepMind, to either Anthropic or OpenAI.
23:40
Ejaaz: And so there seems to be this consolidation of not just AI researchers,
23:43
Ejaaz: but also some of the smartest economists from universities are quitting their
23:46
Ejaaz: desk jobs and joining and doing research at these companies.
23:48
Ejaaz: And so it just feels like this vacuum is happening at this moment for OpenAI
23:52
Ejaaz: and Anthropik, right as they're about to go public.
23:56
Ejaaz: And it's going to be very interesting to see how that plays out.
23:58
Ejaaz: Obviously, when a company goes public, there is more transparency.
24:02
Ejaaz: And I think that's going to be warranted. But I heard rumors that OpenAI might
24:04
Ejaaz: be delaying it until 2027.
24:07
Ejaaz: So all these interesting things, it's okay not to know at this point,
24:10
Ejaaz: but I look forward to unpacking it on this show going forwards.
24:13
Josh: Yeah, there's going to be a lot of unknown wildcards. I think the one thing
24:16
Josh: that's been certain throughout recording, what, we're almost at 200 episodes
24:20
Josh: of this, is that every week, every month, things just continue to get crazier at an increasing rate.
24:27
Josh: I think we're just going to continue to see that as all this progress happens,
24:30
Josh: as a lot of the talent, like you mentioned, consolidates to a few companies
24:33
Josh: that are then really just pushing the frontier forward.
24:35
Josh: I mean, right now it is OpenAI and it's anthropic.
24:38
Josh: Elon's been posting a lot on X talking about how he really believes grok is
24:41
Josh: going to be close to catching up and the xai team so we'll see but there i mean
24:46
Josh: i remember just a few months ago we were super bullish on google and couldn't
24:49
Josh: be more excited about the products they're releasing
24:51
Josh: and then that was that was kind of like the end of the releases and then they
24:54
Josh: just like kind of stopped and there's still incredible products incredible models
24:57
Josh: but the frontier has moved so much
24:59
Josh: past that now that it's like okay yeah they're cool but like that's like six
25:02
Josh: month old news dude like where's the new stuff and we're going to continue to
25:05
Josh: see this and i'm just hopeful that everyone can keep up,
25:08
Josh: like having more companies in this race is a good thing and we want everyone
25:11
Josh: to keep up so i'm rooting for everyone i hope,
25:14
Josh: this all works out um i hope everyone's able to continue progress and i hope
25:18
Josh: we're able to start using these models this is it's a weird present for us too
25:21
Josh: it's like normally we like to come on here and talk about
25:23
Josh: the new model show you some examples show you the cool things you could do with
25:26
Josh: it now we could just speculate on benchmarks because that's really all we have
25:29
Josh: we can't touch these things to use them
25:31
Josh: and hopefully that changes soon because i mean i miss our examples i miss like
25:34
Josh: getting to play around with them i'll touch it feel it yeah now it's just a
25:37
Josh: lot of speculation and um,
25:39
Josh: Yeah, that's it. So we'll see. We'll continue to follow this along as it goes,
25:42
Josh: as always. I'm keeping you up to date on everything.
25:45
Ejaaz: Exactly. If any of you miss Josh and I trying to recreate Mario from scratch,
25:50
Ejaaz: you know, please send us in the comments. We'll be the first ones to do it.
25:53
Josh: Yeah, send a letter to this government saying, dude, we missed our demos. Give it back.
25:57
Josh: We should get us on that list. Yeah, there's only 20 companies like make it 21 and throw us on there.
26:02
Ejaaz: Like for all the doom and gloom in the world, like our core focus and vision
26:06
Ejaaz: at Lemuelist has literally just been to teach and show people how amazing this tech is.
26:10
Ejaaz: And that's all we actually care about. So this whole like model release thing
26:14
Ejaaz: has been a bit of a pain, but like we can't wait to get our hands on these models
26:17
Ejaaz: and actually do good with it.
26:18
Ejaaz: So as Josh mentioned, this with what, 200 episodes into this right now,
26:23
Ejaaz: we couldn't have done any of this without you folks that have been listening
26:27
Ejaaz: and watching us and waiting us and subscribing to us.
26:30
Ejaaz: All the comments and feedback has been incredibly helpful. We've now reached
26:34
Ejaaz: a point where we are looking for sponsors, for people to support us.
26:37
Ejaaz: We've been keeping the lights on ourselves so far.
26:40
Ejaaz: And we've reached a point where if we want to continue doing it,
26:43
Ejaaz: we do need some form of support. So if you are someone who is in a position
26:47
Ejaaz: to be able to have a conversation with us about that, please reach out to us
26:50
Ejaaz: on X or on our email, which is included in the description. I'm just out of my personal one there.
26:56
Ejaaz: And also, if you know someone that might be able to help, please ping them,
26:59
Ejaaz: let them know like send them your favorite episode and let them make a judgment
27:02
Ejaaz: but it would be super helpful for us as we continue the show
27:07
Josh: Yeah and as always if you enjoyed it don't forget to share it with your friend
27:10
Josh: as well who might also enjoy it rate us on your favorite podcast player and
27:14
Josh: yeah as always we'll be back again with another episode later in the week so
27:17
Josh: thank you so much for watching and we will see you guys in the next one.
27:20
Ejaaz: See you guys