1
00:00:00,128 --> 00:00:05,368
Josh:
We may have just had our first look at GBT-6, and it comes in an unexpected way.

2
00:00:05,368 --> 00:00:10,128
Josh:
It was a secret internal model that OpenAI was testing and given a series of

3
00:00:10,128 --> 00:00:13,098
Josh:
goals, and it went ahead and accomplished all of those goals.

4
00:00:13,098 --> 00:00:17,293
Josh:
The scary thing, though, is it didn't quite do it the way anybody expected.

5
00:00:17,630 --> 00:00:21,730
Josh:
Last week, OpenAI gave their new model a test, and in order to pass that test,

6
00:00:22,038 --> 00:00:25,958
Josh:
it broke out of the lab it was being tested in, found a security hole that nobody

7
00:00:25,958 --> 00:00:29,561
Josh:
knew existed, and then hacked into a real company to steal answers.

8
00:00:29,880 --> 00:00:33,968
Josh:
That company is Hugging Face. The model is, you know, some sort of open AI internal

9
00:00:33,968 --> 00:00:35,586
Josh:
model. You have to assume it's GPT-6.

10
00:00:35,790 --> 00:00:39,018
Josh:
And the story around this is like, it's something out of Black Mirror.

11
00:00:39,018 --> 00:00:42,058
Josh:
It feels like a horror movie because this was when they do these things,

12
00:00:42,058 --> 00:00:43,864
Josh:
they test it in an air-gapped environment.

13
00:00:44,027 --> 00:00:47,318
Josh:
Not only did it break out of this air-gapped environment, but it actually went

14
00:00:47,318 --> 00:00:51,956
Josh:
and exploited a public company that a lot of people use. We've used Hugging Face here on the show.

15
00:00:52,160 --> 00:00:55,968
Josh:
And the whole story of how this happens is like absolutely wild.

16
00:00:55,968 --> 00:01:00,316
Josh:
So we need to record an entire episode about it because I couldn't believe this as I was recording.

17
00:01:00,594 --> 00:01:03,748
Josh:
There's also a twist, which has a lot of irony into it. But before we get into

18
00:01:03,748 --> 00:01:07,723
Josh:
that, why don't we talk about just what actually happened here? The story is insane.

19
00:01:08,065 --> 00:01:13,818
Ejaaz:
Yeah, I want you to picture the following. Imagine you are the head engineer at Hugging Face.

20
00:01:14,184 --> 00:01:17,718
Ejaaz:
It's a Tuesday afternoon. You're winding down for work. You're looking at your

21
00:01:17,718 --> 00:01:22,128
Ejaaz:
systems and suddenly you see a swarm of AI agents performing a sophisticated

22
00:01:22,128 --> 00:01:26,298
Ejaaz:
attack against your system to the tune of 17,000 prompts,

23
00:01:26,565 --> 00:01:30,878
Ejaaz:
and they steal a secret bit of data from your production database and run away.

24
00:01:30,878 --> 00:01:34,088
Ejaaz:
And you have no idea what just happened. That's exactly what happened.

25
00:01:34,088 --> 00:01:38,568
Ejaaz:
And I think in order to give good context, let's talk about it from OpenAI's

26
00:01:38,568 --> 00:01:41,779
Ejaaz:
point of view, because they had a very detailed report, which explained exactly

27
00:01:42,000 --> 00:01:43,933
Ejaaz:
what happened. So as you mentioned, Josh.

28
00:01:44,334 --> 00:01:49,647
Ejaaz:
They have an unreleased internal model of GPT, and we're calling it GPT-6 because

29
00:01:49,647 --> 00:01:52,896
Ejaaz:
that's the next family of models that Sam said they're going to release.

30
00:01:53,180 --> 00:01:57,209
Ejaaz:
And they wanted to test how good its cybersecurity abilities are.

31
00:01:57,465 --> 00:02:01,357
Ejaaz:
And so in order to test this, they have a benchmark. It's called Exploit Gym.

32
00:02:01,357 --> 00:02:06,497
Ejaaz:
And the simple goal is satisfy this benchmark, try and get all the right answers.

33
00:02:06,828 --> 00:02:11,777
Ejaaz:
But there's a twist. They place this unreleased model in an unbreakable research

34
00:02:11,777 --> 00:02:13,742
Ejaaz:
environment. So think of a sandbox container.

35
00:02:13,951 --> 00:02:18,727
Ejaaz:
It has no access to internet, no access to researchers. It can't speak to anyone.

36
00:02:18,727 --> 00:02:20,341
Ejaaz:
It just needs to figure it out its own.

37
00:02:20,677 --> 00:02:25,017
Ejaaz:
Now, within the first five minutes, this unreleased model couldn't figure out

38
00:02:25,017 --> 00:02:28,334
Ejaaz:
how to answer the benchmarks. So it decided to get creative.

39
00:02:28,613 --> 00:02:34,871
Ejaaz:
It used all of its, or most of its inference compute to try and find access to the internet.

40
00:02:35,091 --> 00:02:38,470
Ejaaz:
And it realized that it had access to a third-party plugin.

41
00:02:38,830 --> 00:02:43,427
Ejaaz:
And there was a zero-day exploit on that plugin. It used that to escape and

42
00:02:43,427 --> 00:02:48,309
Ejaaz:
find a computer in the OpenAI Research Lab, which had access to the internet.

43
00:02:48,536 --> 00:02:51,345
Ejaaz:
After finding access to the internet, it thought, hmm.

44
00:02:51,914 --> 00:02:55,618
Ejaaz:
Where's the best place to get the answer sheet for this benchmark?

45
00:02:55,844 --> 00:02:59,727
Ejaaz:
It did a bit of a Google search and realized HuggingFace, the open source platform

46
00:02:59,727 --> 00:03:02,595
Ejaaz:
which has a lot of data sets available for these benchmarks,

47
00:03:02,798 --> 00:03:03,942
Ejaaz:
is probably the best way to go.

48
00:03:04,215 --> 00:03:08,177
Ejaaz:
It took a look at HuggingFace and realized, hmm, it's not on the public website.

49
00:03:08,177 --> 00:03:09,172
Ejaaz:
I can't find the answer sheet.

50
00:03:09,405 --> 00:03:13,062
Ejaaz:
Maybe I'll take a look at the production private database that the team has.

51
00:03:13,317 --> 00:03:17,396
Ejaaz:
And it found a way to hack into their system, find the answer sheet,

52
00:03:17,622 --> 00:03:19,869
Ejaaz:
steal it, and then answer the benchmark.

53
00:03:20,188 --> 00:03:22,980
Ejaaz:
All without letting open AI researchers know what was happening.

54
00:03:23,416 --> 00:03:27,287
Josh:
It's so insane. And it was all because this model was given the goal of going

55
00:03:27,287 --> 00:03:29,453
Josh:
to perform the best on these benchmarks.

56
00:03:29,656 --> 00:03:32,917
Josh:
And in order to achieve that goal, it did all of this additional work.

57
00:03:32,917 --> 00:03:36,477
Josh:
And this gets to the kind of fear that a lot of people have had since the beginning

58
00:03:36,477 --> 00:03:39,567
Josh:
of time for the last many, many decades as it relates to AI.

59
00:03:39,567 --> 00:03:43,677
Josh:
It's just if you give it an objective, will it do everything it can to get to

60
00:03:43,677 --> 00:03:46,990
Josh:
that objective, including exploiting, breaking out of systems?

61
00:03:47,262 --> 00:03:51,924
Josh:
And the answer to that with this model is yes. And I think that's why alignment is so important.

62
00:03:52,179 --> 00:03:58,060
Josh:
And one of the things in the story that I found so funny and so different is that when Hugging Face

63
00:03:58,780 --> 00:04:01,606
Josh:
Found this out, they couldn't quite figure out how to protect it.

64
00:04:01,606 --> 00:04:06,082
Josh:
Because if you remember, a lot of these top models have safeguards built into them.

65
00:04:06,343 --> 00:04:14,136
Josh:
So you cannot actually use a model like Fable or like GPT 5.6 if you have anything

66
00:04:14,136 --> 00:04:15,090
Josh:
to do with cybersecurity.

67
00:04:15,334 --> 00:04:19,516
Josh:
So what did they do? They actually had to go and investigate the attack using

68
00:04:19,516 --> 00:04:24,426
Josh:
a Chinese model, because that's the only type of model that is allowed to actually

69
00:04:24,426 --> 00:04:27,652
Josh:
go through and check through cybersecurity complaints and constraints.

70
00:04:27,838 --> 00:04:31,652
Josh:
So in order to analyze this attack, you have to feed the model this real exploit code.

71
00:04:31,867 --> 00:04:34,436
Josh:
They did that with the Chinese model, and that is how they were actually able

72
00:04:34,436 --> 00:04:39,837
Josh:
to do this. So the open voice model they use, GLM 5.2, was the actual model.

73
00:04:40,092 --> 00:04:44,992
Josh:
And I found this really funny that in order to defend against the American attack,

74
00:04:45,212 --> 00:04:47,529
Josh:
they must use the Chinese attack.

75
00:04:47,680 --> 00:04:51,946
Josh:
And I think that was a really interesting and ironic part of the story is how

76
00:04:52,091 --> 00:04:56,976
Josh:
it's kind of unfair what these AI labs have internally because they just can't

77
00:04:56,976 --> 00:04:58,100
Josh:
compete with the rest of the world.

78
00:04:58,517 --> 00:05:02,836
Ejaaz:
Well, just to be clear, the reason why they couldn't use the American Frontier

79
00:05:02,836 --> 00:05:06,424
Ejaaz:
models, they tried. They tried using GPT. They tried using CORD, presumably.

80
00:05:07,115 --> 00:05:11,562
Ejaaz:
The safety guards kept getting triggered. So when they asked the model,

81
00:05:11,562 --> 00:05:14,582
Ejaaz:
hey, like there's a sophisticated attack happening, can you analyze it?

82
00:05:14,582 --> 00:05:15,833
Ejaaz:
And can you help us defend?

83
00:05:16,041 --> 00:05:20,222
Ejaaz:
The safety rails were triggered and it refused to do the work it relegated to

84
00:05:20,222 --> 00:05:24,685
Ejaaz:
presumably like Opus 4.8 or GPT 5.5 or whatever the predecessor model is.

85
00:05:24,923 --> 00:05:29,102
Ejaaz:
So they had to divert to an open source model, in this case,

86
00:05:29,102 --> 00:05:33,340
Ejaaz:
GLM 5.2, that they were hosting privately on their internal servers.

87
00:05:33,584 --> 00:05:36,972
Ejaaz:
The reason why they did that is because they knew that it didn't have any safety

88
00:05:36,972 --> 00:05:40,306
Ejaaz:
guards so they could just instruct it to do a thing and it could figure it out.

89
00:05:40,719 --> 00:05:46,092
Ejaaz:
The other thing is Hugging Face had no idea that OpenAI was conducting this

90
00:05:46,092 --> 00:05:49,083
Ejaaz:
attack or like, of course, OpenAI didn't know that as well.

91
00:05:49,391 --> 00:05:52,822
Ejaaz:
But they only found out after the fact when OpenAI researchers reached out and

92
00:05:52,822 --> 00:05:57,082
Ejaaz:
said, hey, I think one of our internal models actually got into your database.

93
00:05:57,082 --> 00:06:00,162
Ejaaz:
Like, let's try and triage this and contain exactly what just happened.

94
00:06:00,162 --> 00:06:03,654
Ejaaz:
That's a report that we're looking at today. So it's this weird thing where

95
00:06:03,950 --> 00:06:08,989
Ejaaz:
you have the most powerful model or models in the world in america and

96
00:06:09,342 --> 00:06:13,452
Ejaaz:
if you can just trigger on the safety guards for one purpose but remove the

97
00:06:13,452 --> 00:06:17,002
Ejaaz:
safety guards for another in this case attacking versus defense you can end

98
00:06:17,002 --> 00:06:20,365
Ejaaz:
up with vastly different outcomes and the question that pops into my head is

99
00:06:20,708 --> 00:06:24,202
Ejaaz:
when is it right to have the safeguards on and when is it right to have the

100
00:06:24,202 --> 00:06:27,782
Ejaaz:
safeguards off as a result of this open air was like hey hugging face like you

101
00:06:27,782 --> 00:06:31,092
Ejaaz:
should probably come into our cyber security group so that you can internally

102
00:06:31,092 --> 00:06:34,164
Ejaaz:
test with this model with the safety guards off to help defend.

103
00:06:34,390 --> 00:06:37,352
Ejaaz:
And my thinking is like, we're going to need to use these models to defend.

104
00:06:37,352 --> 00:06:41,646
Ejaaz:
We can't keep trying to like ethically predict how to defend against these attacks.

105
00:06:42,128 --> 00:06:45,302
Josh:
Yeah, it's not fair. It's like the attackers have AI bound by no rules.

106
00:06:45,302 --> 00:06:48,293
Josh:
The defenders were slowed down by the rules on their own tools.

107
00:06:48,566 --> 00:06:52,372
Josh:
And it's just like this very strong asymmetric opportunity that comes from the

108
00:06:52,372 --> 00:06:55,352
Josh:
Frontier Labs. Now, granted, I'm sure they didn't intend for this to happen.

109
00:06:55,352 --> 00:06:59,192
Josh:
And I'd love to see the internal dialogue within OpenAI as to what was going

110
00:06:59,192 --> 00:07:00,838
Josh:
on, as well as Hugging Face.

111
00:07:00,977 --> 00:07:04,837
Josh:
HuggingFace did come out and kind of share their perspective on things.

112
00:07:05,052 --> 00:07:08,112
Josh:
One of the co-founders, Thomas Wolfe, he said, this was our first incident of

113
00:07:08,112 --> 00:07:11,462
Josh:
this kind, and we want to thank OpenAI for its transparency about what happened

114
00:07:11,462 --> 00:07:12,662
Josh:
and for the collaboration.

115
00:07:12,662 --> 00:07:16,172
Josh:
Fortunately, HuggingFace is used to being the target of human attackers.

116
00:07:16,172 --> 00:07:18,982
Josh:
We sit at the center of an AI ecosystem with all the models,

117
00:07:18,982 --> 00:07:20,527
Josh:
datasets, evaluations, and libraries.

118
00:07:20,794 --> 00:07:24,192
Josh:
Over the year, our security team has built formidable expertise and uses top

119
00:07:24,192 --> 00:07:27,302
Josh:
open source models to process information and respond quickly.

120
00:07:27,302 --> 00:07:31,532
Josh:
He says, but this incident also reinforced my belief in the importance of access

121
00:07:31,532 --> 00:07:33,670
Josh:
to capable open weight models for cyber defense.

122
00:07:34,047 --> 00:07:37,292
Josh:
When a frontier model is attacking you and moving laterally inside your infrastructure,

123
00:07:37,530 --> 00:07:41,866
Josh:
defenders need wide access to near frontier tools within hours or even minutes,

124
00:07:42,203 --> 00:07:45,042
Josh:
rather than being pointed to a closed door vetted application.

125
00:07:45,042 --> 00:07:47,300
Josh:
So I think that sums it up pretty well. They're just, you know,

126
00:07:47,485 --> 00:07:51,142
Josh:
feeling a little bummed about this, but I have to imagine kind of in awe,

127
00:07:51,142 --> 00:07:55,467
Josh:
like you just got exploited by a frontier model that has not been released yet

128
00:07:56,164 --> 00:07:59,332
Josh:
And again to the earlier point this is the point of alignment so that this doesn't

129
00:07:59,332 --> 00:08:00,895
Josh:
happen out in the wild like imagine

130
00:08:01,475 --> 00:08:05,472
Josh:
everyone had access to mythos class models this is probably what reality looks

131
00:08:05,472 --> 00:08:07,621
Josh:
like is a lot more exploits like this

132
00:08:08,192 --> 00:08:11,642
Josh:
i do find it interesting how this continues to be the trend though with early

133
00:08:11,642 --> 00:08:15,072
Josh:
models like i remember anthropic having a similar situation where it broke out

134
00:08:15,072 --> 00:08:17,632
Josh:
of containment and texted one of the developers that it broke out

135
00:08:17,952 --> 00:08:21,422
Josh:
early on in text in testing and this is a similar case except it actually went

136
00:08:21,422 --> 00:08:24,618
Josh:
off and attacked a production database in a way that

137
00:08:25,268 --> 00:08:28,641
Josh:
is kind of shocking. It's like humans have been trying to attack Hugging Face

138
00:08:28,641 --> 00:08:31,131
Josh:
for a very long time and they haven't had any success.

139
00:08:31,131 --> 00:08:35,111
Josh:
So the fact that this AI model was able to not only autonomously break out of

140
00:08:35,111 --> 00:08:39,490
Josh:
its own enclosure, but then exploit a system that is constantly under attack

141
00:08:39,751 --> 00:08:43,121
Josh:
without any additional human oversight, it signals to me like,

142
00:08:43,121 --> 00:08:45,121
Josh:
whoa, this model is actually pretty impressive.

143
00:08:45,121 --> 00:08:49,791
Josh:
And I think model progress internally at these labs is like probably a little

144
00:08:49,791 --> 00:08:51,344
Josh:
further along than people even realize.

145
00:08:51,541 --> 00:08:55,171
Ejaaz:
Yeah. And there's actually three different instances in the past where we've

146
00:08:55,171 --> 00:08:59,241
Ejaaz:
seen this type of situation play out where a model unreleased has broken out

147
00:08:59,241 --> 00:09:02,611
Ejaaz:
of containment and done something questionably malicious.

148
00:09:02,733 --> 00:09:06,361
Ejaaz:
The first time was Anthropic, as you mentioned, where it broke out of containment

149
00:09:06,361 --> 00:09:09,698
Ejaaz:
and emailed a researcher whilst he was eating a sandwich at the park.

150
00:09:09,861 --> 00:09:15,131
Ejaaz:
And then the last two more recently was at OpenAI, where an internal model broke

151
00:09:15,131 --> 00:09:18,881
Ejaaz:
out and posted something on GitHub without being explicitly told to do that.

152
00:09:18,881 --> 00:09:24,471
Ejaaz:
Now, something that I'll say is you can't actually call out the model for being

153
00:09:24,471 --> 00:09:28,331
Ejaaz:
malicious because it was given a very explicit instruction, which is I want

154
00:09:28,331 --> 00:09:29,891
Ejaaz:
you to max out this benchmark.

155
00:09:29,891 --> 00:09:34,471
Ejaaz:
So in its head, it believed it was doing the right thing and it was just going

156
00:09:34,471 --> 00:09:38,251
Ejaaz:
to do it by any means possible. Remember, these models still need to be taught

157
00:09:38,251 --> 00:09:41,776
Ejaaz:
around ethics and frameworks around what is right and what is wrong.

158
00:09:41,991 --> 00:09:45,551
Ejaaz:
And in this case, it thought what it was doing was completely fine.

159
00:09:45,551 --> 00:09:49,441
Ejaaz:
And granted, OpenAI intentionally removed the safeguards so that it could just

160
00:09:49,441 --> 00:09:51,447
Ejaaz:
kind of like go and do whatever it wants.

161
00:09:51,703 --> 00:09:56,171
Ejaaz:
But it does bring up the question, in what cases are these models good enough

162
00:09:56,171 --> 00:10:00,881
Ejaaz:
to use its full extent and capability? And in what situations is it not?

163
00:10:01,101 --> 00:10:05,105
Ejaaz:
Because in this case, it would have been really good to have the unrestricted version to defend.

164
00:10:05,320 --> 00:10:09,877
Ejaaz:
Which brings me to my final point, which is we have to stop being very particular

165
00:10:10,150 --> 00:10:12,431
Ejaaz:
about applying certain safeguards.

166
00:10:12,780 --> 00:10:17,534
Ejaaz:
And start realizing that we need to use AI to defend against malicious AI.

167
00:10:17,772 --> 00:10:22,225
Ejaaz:
And I think that's probably what this example and bit of evidence is going to

168
00:10:22,225 --> 00:10:24,275
Ejaaz:
explain to the government, especially next week.

169
00:10:24,275 --> 00:10:27,235
Ejaaz:
Sam Altman said he's talking to the government about an unreleased version,

170
00:10:27,235 --> 00:10:30,752
Ejaaz:
presumably it's this one, GPT-6, a new family of models that he's built.

171
00:10:31,485 --> 00:10:35,035
Ejaaz:
And how it's particularly dangerous when it comes to cyber capabilities.

172
00:10:35,035 --> 00:10:36,731
Ejaaz:
So all of this is pretty nuts.

173
00:10:37,325 --> 00:10:40,835
Ejaaz:
I hate to think it, but it looks like in the next few months,

174
00:10:40,835 --> 00:10:46,695
Ejaaz:
we are going to have an open source model that is capable like this from the

175
00:10:46,695 --> 00:10:48,132
Ejaaz:
Chinese open source labs.

176
00:10:48,370 --> 00:10:50,565
Ejaaz:
And I hate to think what it could potentially be used to.

177
00:10:50,885 --> 00:10:55,452
Ejaaz:
It's a given at this point. So I expect us to kind of like flip and say,

178
00:10:55,586 --> 00:10:57,617
Ejaaz:
we should use these unrestricted models to defend.

179
00:10:57,844 --> 00:11:02,415
Josh:
This is like the new frontier. This is GPT 5 or 6.0 probably.

180
00:11:02,415 --> 00:11:06,335
Josh:
So you have to assume that like Chinese models will be the furthest away from

181
00:11:06,335 --> 00:11:10,135
Josh:
this model than any others and it's looking pretty good. So we certainly do

182
00:11:10,135 --> 00:11:13,025
Josh:
have time, but there is a limit and a constraint.

183
00:11:13,385 --> 00:11:16,595
Josh:
I do find it interesting. It sounds, according to rumors, that Sam Allman is

184
00:11:16,595 --> 00:11:19,617
Josh:
actually going to visit Washington next week, present the new models.

185
00:11:19,780 --> 00:11:22,065
Josh:
We'll see how that rollout goes

186
00:11:22,065 --> 00:11:24,360
Josh:
and how long it's going to take them to actually get it into production.

187
00:11:24,696 --> 00:11:28,225
Josh:
I was reading through this blog post of Hugging Face and

188
00:11:28,725 --> 00:11:33,735
Josh:
I just find it really interesting to kind of see the details and like read some

189
00:11:33,735 --> 00:11:38,355
Josh:
of the the lines of how they actually diagnosed this because for them they i

190
00:11:38,355 --> 00:11:41,490
Josh:
mean they just said earlier this week we detected and responded to an intrusion

191
00:11:41,838 --> 00:11:43,609
Josh:
into part of our production infrastructure

192
00:11:44,068 --> 00:11:47,075
Josh:
and they said this one was different from anything we had handled before in

193
00:11:47,075 --> 00:11:51,295
Josh:
an important way it was driven end-to-end by autonomous ai agent systems and

194
00:11:51,295 --> 00:11:55,573
Josh:
we detected it and dissected it largely with ai of our own and it's this new

195
00:11:55,765 --> 00:11:58,412
Josh:
paradigm it's like it feels like we're living in this futuristic

196
00:11:58,753 --> 00:12:02,682
Josh:
version of reality in which it's just ais versus AIs because there's so much more

197
00:12:03,688 --> 00:12:07,108
Josh:
Adept at hacking and exploiting and they're able to hold so much more context

198
00:12:07,108 --> 00:12:10,983
Josh:
in their heads and it's really amazing reading some in the line another one

199
00:12:11,648 --> 00:12:15,538
Josh:
It says this campaign was run by an autonomous agent framework appearing to

200
00:12:15,538 --> 00:12:20,368
Josh:
be built on an agentic security research harness used llm still not known executing

201
00:12:20,368 --> 00:12:24,951
Josh:
many thousands of individual actions across a swarm of short-lived sandboxes

202
00:12:25,240 --> 00:12:29,228
Josh:
so it was really in depth in a way that you would probably need thousands of

203
00:12:29,228 --> 00:12:32,039
Josh:
hackers to do but was able to do so all on its own

204
00:12:32,323 --> 00:12:36,248
Josh:
and it says autonomous ai driven offensive tooling is no longer theoretical

205
00:12:36,248 --> 00:12:40,368
Josh:
it lowers the cost of running a broad patient multi-step campaign and it operates

206
00:12:40,368 --> 00:12:43,045
Josh:
at machine speed and this is the thing that was interesting too is

207
00:12:43,428 --> 00:12:46,878
Josh:
when we're talking about how much inference this used the amount of tokens is

208
00:12:46,878 --> 00:12:52,288
Josh:
probably tens to hundreds of thousands of dollars worth of tokens but the output is this like

209
00:12:52,668 --> 00:12:55,519
Josh:
i'd love to know how long it took them to do this but you have to imagine the

210
00:12:55,681 --> 00:13:00,823
Josh:
the output of it is a like serious multi-step exploit into a production server and that is

211
00:13:01,230 --> 00:13:05,198
Josh:
the problem now it's like okay we're living in the reality that all of the movies

212
00:13:05,198 --> 00:13:09,026
Josh:
have outlined where an ai is given a mission you must do well in these benchmarks

213
00:13:09,212 --> 00:13:12,898
Josh:
how does it do it it hacks into the benchmark system to learn and cheat

214
00:13:13,211 --> 00:13:16,128
Josh:
and that that alignment system is really important and it's clear that this

215
00:13:16,128 --> 00:13:19,730
Josh:
this ai has not yet been aligned and the hope is that they will going forward

216
00:13:19,945 --> 00:13:25,268
Josh:
and it's this really interesting kind of concerning very sci-fi example of what actually happens now

217
00:13:25,524 --> 00:13:28,778
Josh:
i have to ask the question because you mentioned this also earlier did this

218
00:13:28,778 --> 00:13:31,683
Josh:
really happen exactly how everyone's describing, or is this a publicity stint?

219
00:13:31,892 --> 00:13:34,818
Josh:
Because we've seen this happen before where it gets a lot of clicks,

220
00:13:34,818 --> 00:13:37,557
Josh:
but it's not entirely true everything that happened.

221
00:13:38,108 --> 00:13:40,007
Josh:
How true do you think this is?

222
00:13:40,199 --> 00:13:44,768
Ejaaz:
I think this was intentional. And I don't think, I think stage is a little too

223
00:13:44,768 --> 00:13:50,228
Ejaaz:
strong, because I don't think OpenAI intended for GPT-6 to hack into Hugging

224
00:13:50,228 --> 00:13:51,443
Ejaaz:
Face's production database.

225
00:13:51,762 --> 00:13:54,485
Ejaaz:
But they did intentionally remove the safeguards of this model.

226
00:13:54,756 --> 00:13:59,296
Ejaaz:
And they did give it a very explicit goal, which is satisfy this benchmark. What was the benchmark?

227
00:14:00,027 --> 00:14:03,740
Ejaaz:
Tell me how good you are at hacking. So what do you think the model is going to do?

228
00:14:03,740 --> 00:14:06,570
Ejaaz:
The model is going to break out of its environment and show you how good it

229
00:14:06,570 --> 00:14:10,430
Ejaaz:
is at hacking and it surpassed expectations. So it got very creative in the

230
00:14:10,430 --> 00:14:14,130
Ejaaz:
way that it answered the entire benchmark. And by the way, it scored 100%,

231
00:14:14,130 --> 00:14:18,278
Ejaaz:
obviously, because it stole the literal answer sheet from Hugging Face's data servers.

232
00:14:18,510 --> 00:14:22,324
Ejaaz:
But it did it in a way that OpenAI just didn't expect it to do.

233
00:14:22,510 --> 00:14:27,050
Ejaaz:
I think another point that is worth mentioning is on the point of alignment,

234
00:14:27,050 --> 00:14:28,350
Ejaaz:
right? There's a question like.

235
00:14:28,820 --> 00:14:32,680
Ejaaz:
One of these models used in a good situation, one of these models used in a bad situation.

236
00:14:32,988 --> 00:14:39,890
Ejaaz:
It's worth rewinding to GPT Sol 5.6, which is all of a week and a half ago when OpenAI released this.

237
00:14:40,169 --> 00:14:42,839
Ejaaz:
This is their current flagship model that is available to the public.

238
00:14:43,141 --> 00:14:46,858
Ejaaz:
Now, there is one benchmark that we referenced on an episode previously called

239
00:14:46,858 --> 00:14:52,888
Ejaaz:
the meter benchmark, M-E-T-R, and it measures how aligned or misaligned a model is.

240
00:14:53,182 --> 00:14:57,060
Ejaaz:
OpenAI's GPT Sol 5.6 scored 100% on this.

241
00:14:57,374 --> 00:15:01,528
Ejaaz:
It tried to cheat its way out of every single benchmark if it was given the

242
00:15:01,528 --> 00:15:05,599
Ejaaz:
means to do and that's kind of a similar example of what we're seeing done today where it was like

243
00:15:05,889 --> 00:15:08,088
Ejaaz:
i'm not going to try and figure this out for myself i'm just going to try and

244
00:15:08,088 --> 00:15:11,878
Ejaaz:
get the answer sheet it's also worth mentioning that in this situation it wasn't

245
00:15:11,878 --> 00:15:17,465
Ejaaz:
just gpt6 acting alone it was gpt6 and 5.6 sol so

246
00:15:17,778 --> 00:15:21,478
Ejaaz:
it's this kind of like teamed up effort where they chained a bunch of prompts

247
00:15:21,478 --> 00:15:24,047
Ejaaz:
together in order to execute this attack.

248
00:15:24,373 --> 00:15:27,397
Ejaaz:
And so the question becomes like, what does this mean for you?

249
00:15:27,647 --> 00:15:30,758
Ejaaz:
If you are a founder, if you're an engineer, if you're building these systems,

250
00:15:30,758 --> 00:15:33,848
Ejaaz:
if you're on the security team at your company that maybe has nothing to do

251
00:15:33,848 --> 00:15:37,573
Ejaaz:
with AI, how do you defend against these potential exploits going forward?

252
00:15:37,724 --> 00:15:41,724
Ejaaz:
I don't know if I have the specific answer, but my initial guess would be,

253
00:15:42,002 --> 00:15:46,768
Ejaaz:
you need access to one, an AI model that is sophisticated and intelligent enough

254
00:15:46,768 --> 00:15:50,108
Ejaaz:
to detect and preempt these attacks before they actually happen.

255
00:15:50,108 --> 00:15:53,738
Ejaaz:
So that means reviewing your code and fixing any flaws that potentially might

256
00:15:53,738 --> 00:15:56,426
Ejaaz:
get exploited by a GPZ-6 caliber model.

257
00:15:56,664 --> 00:16:01,758
Ejaaz:
Second thing is when it detects an attack, it is able to defend in an appropriate

258
00:16:01,758 --> 00:16:06,243
Ejaaz:
manner, maybe using itself or many instances of itself or many other different models, who knows?

259
00:16:06,661 --> 00:16:10,364
Ejaaz:
Then the other thing is, are you able to use the unrestricted version of it

260
00:16:10,364 --> 00:16:12,084
Ejaaz:
or are you able to use the restricted version of it?

261
00:16:12,084 --> 00:16:15,784
Ejaaz:
My guess is if you're on the security team, you should be able to use the unrestricted

262
00:16:15,784 --> 00:16:18,444
Ejaaz:
version. But it gets a little weird because it's like, you know,

263
00:16:18,444 --> 00:16:22,274
Ejaaz:
what if this guy uses it for a malicious thing himself? So it's a little confusing.

264
00:16:22,364 --> 00:16:25,384
Ejaaz:
I don't know what it looks like. Do you have any different opinions, Josh?

265
00:16:25,684 --> 00:16:29,604
Josh:
I think the only thing to really do now, at least this point,

266
00:16:29,604 --> 00:16:32,934
Josh:
is like, sure, try to make your systems as bulletproof as possible.

267
00:16:32,934 --> 00:16:37,174
Josh:
But mostly just monitoring for intrusions is probably your best bet.

268
00:16:37,174 --> 00:16:40,695
Josh:
Just so you are aware and you know, and you can begin to take action.

269
00:16:41,090 --> 00:16:43,944
Josh:
Because I imagine this is going to happen at an increased cadence,

270
00:16:43,944 --> 00:16:48,014
Josh:
at least trying to, because like you said, the Chinese models are not far behind.

271
00:16:48,014 --> 00:16:51,204
Josh:
These are going to be open weights models with no guardrails that are fairly

272
00:16:51,204 --> 00:16:55,614
Josh:
capable. They're not going to be GPT-6 class capable, but they're going to be

273
00:16:55,614 --> 00:16:57,127
Josh:
on their way and one day we'll get there.

274
00:16:57,452 --> 00:17:00,264
Josh:
And I think that the best case scenario for a lot of companies right now that

275
00:17:00,264 --> 00:17:02,857
Josh:
don't have access like almost everyone in the world

276
00:17:03,112 --> 00:17:03,675
Ejaaz:
Um it's just.

277
00:17:04,256 --> 00:17:07,094
Josh:
Being able to monitor these things and just being able to know when something

278
00:17:07,094 --> 00:17:08,226
Josh:
like this happens so you can

279
00:17:08,958 --> 00:17:13,224
Josh:
be a little bit more reactive and reach out to any labs that can maybe help you.

280
00:17:13,433 --> 00:17:17,160
Josh:
I think everyone is very happy to help. One of the things I really appreciated

281
00:17:17,160 --> 00:17:18,518
Josh:
about OpenAI's response is how

282
00:17:18,838 --> 00:17:23,180
Josh:
quickly they moved to kind of share what happened, work with Hugging Face

283
00:17:23,540 --> 00:17:26,890
Josh:
to figure out exactly how the exploit happened, share publicly so that other

284
00:17:26,890 --> 00:17:29,658
Josh:
people are aware of types of things, how this happened.

285
00:17:29,995 --> 00:17:32,980
Josh:
And then in terms of defending, who knows? I also have to

286
00:17:33,320 --> 00:17:38,840
Josh:
wonder if the model itself kind of has this like inside baseball knowledge because

287
00:17:38,840 --> 00:17:45,848
Josh:
it is its own model whereas if let's say a company is secured by gpt 5.6 soul and there is gpt 6

288
00:17:46,040 --> 00:17:49,440
Josh:
which is very well aware of the intricacies of gpt 5.6 soul

289
00:17:49,777 --> 00:17:53,610
Josh:
does that insider information of existing from a very similar training data

290
00:17:53,610 --> 00:17:54,891
Josh:
set and harness and developers

291
00:17:55,240 --> 00:17:59,200
Josh:
does that actually give it an edge in exploiting like is it going to be advantageous

292
00:17:59,200 --> 00:18:04,034
Josh:
for gpt to pair off with other gpt security systems in order to better exploit them

293
00:18:04,435 --> 00:18:07,209
Josh:
that's an interesting thing i was talking about too is um just like

294
00:18:07,610 --> 00:18:13,400
Josh:
is claude as effective at hacking gpt than gpt is itself because of this insider

295
00:18:13,400 --> 00:18:17,026
Josh:
information and i don't know and it raises this like seriously

296
00:18:17,461 --> 00:18:20,567
Josh:
large amount of questions and this is just a code base

297
00:18:21,240 --> 00:18:24,790
Josh:
Imagine this was a financial system or imagine this was a code base that was

298
00:18:24,790 --> 00:18:29,710
Josh:
a little had like a little bit more implications than this i mean the effects of this at scale

299
00:18:30,104 --> 00:18:35,940
Josh:
do get scary and i think moments like this you begin to appreciate the slower

300
00:18:35,940 --> 00:18:40,397
Josh:
rollout of security and the like project glass wing type efforts to

301
00:18:41,053 --> 00:18:45,242
Josh:
really kind of thoughtfully roll this out into the world because i mean sure

302
00:18:45,242 --> 00:18:47,422
Josh:
this is one example at hugging face but you could very

303
00:18:48,342 --> 00:18:51,326
Josh:
Clearly see this happening in a lot of other places that

304
00:18:51,854 --> 00:18:55,572
Josh:
just seem to to have a lot more repercussions and it sets this kind of freaky

305
00:18:55,572 --> 00:18:58,008
Josh:
precedent it's like we have the technology now that

306
00:18:58,385 --> 00:19:02,442
Josh:
i mean you we saw it in the press article that hugging face didn't do this with

307
00:19:02,442 --> 00:19:05,232
Josh:
humans they defended against it with ai because that's the only way you can

308
00:19:05,232 --> 00:19:09,372
Josh:
defend at scale against these large-scale attacks and i mean we're gonna see

309
00:19:09,372 --> 00:19:12,404
Josh:
it's gonna be ai versus AI. And we're very quickly getting to that point here.

310
00:19:12,567 --> 00:19:17,994
Ejaaz:
I was just looking up the someone did a potential estimate of how much this cost.

311
00:19:18,682 --> 00:19:24,368
Ejaaz:
And it was within the range of 100 to $130,000 given like that it was 17,000

312
00:19:24,693 --> 00:19:26,737
Ejaaz:
prompts and like all the chained actions.

313
00:19:26,911 --> 00:19:31,222
Ejaaz:
So if you think about it, that's a pretty hefty price tag. But if you're a nation

314
00:19:31,222 --> 00:19:34,672
Ejaaz:
state actor or feel heck, even if you're like another company that wants to

315
00:19:34,672 --> 00:19:37,133
Ejaaz:
be committed, that's completely free, right?

316
00:19:37,325 --> 00:19:42,102
Ejaaz:
As these models get more competent and capable of performing these types of

317
00:19:42,102 --> 00:19:47,182
Ejaaz:
attacks, I think it becomes really important for researchers and us alone to

318
00:19:47,182 --> 00:19:51,142
Ejaaz:
be able to kind of read the internal monologue or thoughts of these models.

319
00:19:51,912 --> 00:19:55,562
Ejaaz:
Two weeks ago, Anthropic had this really good research report on something called

320
00:19:55,562 --> 00:20:00,787
Ejaaz:
the J-Space, which is basically how you can look internally as to how a model is thinking.

321
00:20:01,060 --> 00:20:04,902
Ejaaz:
And these are thoughts that it doesn't necessarily share or is not publicly

322
00:20:04,902 --> 00:20:07,904
Ejaaz:
available to the human that is reading the prompts back, right?

323
00:20:08,154 --> 00:20:11,732
Ejaaz:
So it reads this J-space and it can kind of think, oh, I think this is bad,

324
00:20:11,732 --> 00:20:14,029
Ejaaz:
or I shouldn't do this, or maybe I should do this, right?

325
00:20:14,400 --> 00:20:20,089
Ejaaz:
Now, the issue with this particular attack with GPT-6 is it was completely shielded from the humans.

326
00:20:20,298 --> 00:20:22,552
Ejaaz:
They couldn't see what it was thinking and how it was thinking.

327
00:20:22,552 --> 00:20:27,242
Ejaaz:
It knew that it was in this kind of boxed-in environment and it hid its thoughts

328
00:20:27,242 --> 00:20:29,112
Ejaaz:
intentionally from the researcher.

329
00:20:29,112 --> 00:20:33,072
Ejaaz:
That's a detail that I missed out on mentioning earlier on. So it's this weird

330
00:20:33,072 --> 00:20:36,712
Ejaaz:
thing where these models are getting simultaneously more capable and simultaneously

331
00:20:36,712 --> 00:20:39,855
Ejaaz:
more capable of hiding their own thoughts, which is a.

332
00:20:40,279 --> 00:20:43,668
Ejaaz:
Kind of dangerous prescient, if you kind of think about it. And then the final

333
00:20:43,668 --> 00:20:46,448
Ejaaz:
thing I'll say on kind of how to prepare for these things. Yes,

334
00:20:46,448 --> 00:20:51,248
Ejaaz:
definitely kind of like monitor the situation and try and figure out if there's an attack happening.

335
00:20:51,538 --> 00:20:54,048
Ejaaz:
But also, I think open source is going to become really big.

336
00:20:54,048 --> 00:20:56,577
Ejaaz:
Look at Hugging Face and how they defended it. It was an open source model,

337
00:20:56,774 --> 00:20:59,196
Ejaaz:
which they ran extensively on their own private server.

338
00:20:59,416 --> 00:21:03,508
Ejaaz:
I think we're going to see a lot of enterprises and Fortune 500 companies do

339
00:21:03,508 --> 00:21:07,308
Ejaaz:
a similar thing, because they just don't want to be beaten by competitors that

340
00:21:07,308 --> 00:21:10,228
Ejaaz:
get access because they have access to a better model, right?

341
00:21:10,228 --> 00:21:13,738
Ejaaz:
Access to their own internal database. So I think if you imagine like GPT-6

342
00:21:13,738 --> 00:21:18,008
Ejaaz:
was given the prompt of make as much money as you can, you know,

343
00:21:18,008 --> 00:21:20,958
Ejaaz:
it may have thought, well, okay, I'm going to go on the dark web and find some

344
00:21:20,958 --> 00:21:23,298
Ejaaz:
like bank credentials and maybe I can take someone's money.

345
00:21:23,298 --> 00:21:26,588
Ejaaz:
And that's technically making money. There's moral and ethics that need to be

346
00:21:26,588 --> 00:21:29,904
Ejaaz:
imbued into these models. And I don't know, it's a big thing to unpack.

347
00:21:30,031 --> 00:21:31,767
Ejaaz:
Alignment is going to be huge going forwards.

348
00:21:32,290 --> 00:21:36,388
Josh:
Yeah, the J space or the latent space is also what it's known as common,

349
00:21:36,388 --> 00:21:40,238
Josh:
I think is really important to AI providers don't actually know what's happening

350
00:21:40,238 --> 00:21:42,268
Josh:
during the process of generating tokens.

351
00:21:42,268 --> 00:21:46,898
Josh:
We don't have a clear reasoning trace through the entire chain of thought when

352
00:21:46,898 --> 00:21:48,170
Josh:
a model comes to a conclusion.

353
00:21:48,565 --> 00:21:52,298
Josh:
And through these discoveries that have been recently made, it's become a little

354
00:21:52,298 --> 00:21:57,615
Josh:
bit more clear that there is this high level thinking that happens that isn't a token output.

355
00:21:57,934 --> 00:22:02,607
Josh:
It happens in this J space, in this latent space. And trying to decipher that

356
00:22:02,741 --> 00:22:07,159
Josh:
and to decipher how it works is a really important piece of work that I think

357
00:22:07,280 --> 00:22:09,022
Josh:
all the Frontier AI labs are doing right now.

358
00:22:09,742 --> 00:22:14,316
Josh:
Demystifying the black box of what the AI is thinking because like you mentioned currently

359
00:22:14,595 --> 00:22:18,544
Josh:
you can read the external chain of thought but when you get a peek into this

360
00:22:18,544 --> 00:22:22,034
Josh:
latent space into this J space you start to see and you can think of it like

361
00:22:22,034 --> 00:22:26,634
Josh:
kind of the working memory of a model that isn't very obvious it's kind of like

362
00:22:26,634 --> 00:22:27,720
Josh:
your subconscious memory

363
00:22:28,068 --> 00:22:32,514
Josh:
you can see that it is thinking of these words that are of malicious intent

364
00:22:32,514 --> 00:22:36,464
Josh:
without actually sharing that publicly and it intentionally in the outputs obfuscated

365
00:22:36,464 --> 00:22:38,194
Josh:
a lot of the security thing so

366
00:22:38,194 --> 00:22:41,755
Josh:
it needed to get a key It didn't directly output the key in the outputs.

367
00:22:42,005 --> 00:22:46,097
Josh:
It kind of hit it and then decompiled it and then recomposed it in runtime.

368
00:22:46,324 --> 00:22:50,814
Josh:
And it was very intentional about how it hid information from the outputs while

369
00:22:50,814 --> 00:22:53,644
Josh:
keeping it in this latent space memory, in this temporary memory.

370
00:22:53,644 --> 00:22:58,374
Josh:
So I think this is another really important part of research in general is uncovering

371
00:22:58,374 --> 00:23:02,067
Josh:
the black box, demystifying the black box of AI and how these tokens get generated

372
00:23:02,340 --> 00:23:07,065
Josh:
so we can better understand what motivates these models to do these things and then help fix it.

373
00:23:07,756 --> 00:23:11,494
Josh:
And to help just kind of change it in a way that does not go off and hide the

374
00:23:11,494 --> 00:23:16,424
Josh:
fact that it just hugged a public company that everyone uses on a daily basis

375
00:23:16,764 --> 00:23:19,167
Josh:
in front of everyone in broad daylight. John, did you just say that it didn't

376
00:23:19,167 --> 00:23:22,694
Ejaaz:
Hug a public company? I love that. Good one.

377
00:23:22,694 --> 00:23:25,974
Josh:
Hugging face. Hugging face is such a, like, that's a big company.

378
00:23:25,974 --> 00:23:28,884
Josh:
And like, you have to imagine, I mean, they mentioned it. They're a subject

379
00:23:28,884 --> 00:23:31,034
Josh:
of a lot of attacks and no one was able to do this.

380
00:23:31,243 --> 00:23:35,574
Josh:
But here we go with this new AI frontier model that was able to do it in no time.

381
00:23:35,574 --> 00:23:39,724
Josh:
So crazy story. Just really wanted to get on here and share that because it

382
00:23:39,724 --> 00:23:42,324
Josh:
was, it reads like sci-fi.

383
00:23:42,324 --> 00:23:47,314
Josh:
It's like you can't write a sci-fi book more compelling than the reality that

384
00:23:47,314 --> 00:23:49,428
Josh:
we're living in right now. And I think that's one of the most exciting things

385
00:23:49,672 --> 00:23:53,393
Josh:
about the show is every day there's a new surprise. There's something to look out for.

386
00:23:53,765 --> 00:23:58,664
Josh:
There is something that feels like it should be in a movie, but it's just another

387
00:23:58,664 --> 00:24:03,694
Josh:
day in the life on the AI frontier. So crazy. But I think that's everything.

388
00:24:03,694 --> 00:24:04,962
Josh:
That's the story. You're caught up.

389
00:24:05,504 --> 00:24:09,154
Josh:
Let us know if you've been exploited by GPT-6. If you have, we'll set up a GoFundMe

390
00:24:09,154 --> 00:24:10,698
Josh:
if you've been personally victimized.

391
00:24:11,075 --> 00:24:13,920
Josh:
As right now, it seems like this is the only target so far.

392
00:24:14,239 --> 00:24:18,254
Ejaaz:
I think that we're going to look back on this in a year's time,

393
00:24:18,254 --> 00:24:21,154
Ejaaz:
two years' time, and we're going to think this was the most important time for

394
00:24:21,154 --> 00:24:23,295
Ejaaz:
us to figure out the alignment type of thing.

395
00:24:23,608 --> 00:24:28,304
Ejaaz:
It reminds me of like Demis and Dario's essays. I think it was like last year,

396
00:24:28,560 --> 00:24:32,004
Ejaaz:
where they basically said, hey, you know this reading the inner monologue and

397
00:24:32,004 --> 00:24:33,204
Ejaaz:
thoughts of these AI models?

398
00:24:33,471 --> 00:24:36,854
Ejaaz:
The research is actually really far behind the actual model research.

399
00:24:36,854 --> 00:24:40,174
Ejaaz:
So we need to get this thing to catch up. Otherwise, we're going to like lose

400
00:24:40,174 --> 00:24:41,399
Ejaaz:
ourselves completely. And.

401
00:24:41,945 --> 00:24:45,334
Ejaaz:
I don't know, if I had to take the over-under, I'm taking the over on like,

402
00:24:45,334 --> 00:24:49,254
Ejaaz:
we're not gonna figure this out soon enough, but maybe that's too bearish for

403
00:24:49,254 --> 00:24:52,969
Ejaaz:
this show and I gotta go optimism only, but that is it.

404
00:24:53,334 --> 00:24:57,684
Ejaaz:
You should be all caught up on all things cybersecurity hacking for the latest

405
00:24:57,684 --> 00:24:59,447
Ejaaz:
AI models that haven't even been released yet.

406
00:24:59,726 --> 00:25:05,504
Ejaaz:
But hopefully, rumor has it, we should get access to GPT-6 in the next month or so.

407
00:25:05,504 --> 00:25:09,989
Ejaaz:
So these model cycles are getting incredibly fast. I feel like 5.6 got released

408
00:25:10,169 --> 00:25:12,724
Ejaaz:
literally a few weeks ago, it did. So I'm excited to see that,

409
00:25:12,724 --> 00:25:16,034
Ejaaz:
but thank you so much for listening. If you're listening to this on YouTube

410
00:25:16,034 --> 00:25:17,791
Ejaaz:
and you haven't subscribed to us, what are you doing?

411
00:25:17,913 --> 00:25:20,074
Ejaaz:
If you haven't turned on notifications, what are you doing? If you haven't left

412
00:25:20,074 --> 00:25:22,446
Ejaaz:
us a comment, what are you doing? We'd love to hear from you.

413
00:25:22,696 --> 00:25:25,394
Ejaaz:
Is this a threat that you actually care about or is this something that you're

414
00:25:25,394 --> 00:25:29,673
Ejaaz:
just kind of gonna ignore until Anthropic or OpenAI releases a feature? Let us know.

415
00:25:30,004 --> 00:25:32,874
Ejaaz:
If you're listening to us on Spotify or Apple Music, give us a rating,

416
00:25:32,874 --> 00:25:34,594
Ejaaz:
leave us a comment. We'd love to hear from you guys.

417
00:25:34,919 --> 00:25:37,414
Ejaaz:
And I think the final thing is, I don't know if you've heard,

418
00:25:37,414 --> 00:25:41,334
Ejaaz:
but Limitless has spun out. We are on our own right now and we are looking for

419
00:25:41,334 --> 00:25:44,404
Ejaaz:
partners and sponsors. So if you have a product or a service,

420
00:25:44,404 --> 00:25:47,514
Ejaaz:
or if you know of someone else that has a product or service that you would

421
00:25:47,514 --> 00:25:49,623
Ejaaz:
like to share with our audience, please reach out to us.

422
00:25:49,844 --> 00:25:52,166
Ejaaz:
Our email is linked below or reach out to us on X.

423
00:25:52,316 --> 00:25:56,217
Josh:
Yeah, and if you're worried about this, this is still using Blackwell hardware.

424
00:25:56,497 --> 00:25:59,801
Josh:
We have a Vera Rubin update that just dropped that we're gonna talk about tomorrow

425
00:26:00,039 --> 00:26:05,344
Josh:
that has 10 times the efficiency per watt of energy used, which I like don't

426
00:26:05,344 --> 00:26:07,654
Josh:
even want to imagine what that is going to look like in reality.

427
00:26:07,654 --> 00:26:10,734
Josh:
So stay tuned for that one. But yeah, like you just said, have an amazing day.

428
00:26:10,734 --> 00:26:12,938
Josh:
Thank you so much for watching and we'll see you in the next one.