1
00:00:00,020 --> 00:00:04,420
Ejaaz:
China just got exposed for stealing our AI. In a new report from Anthropic,

2
00:00:05,000 --> 00:00:10,280
Ejaaz:
three top Chinese AI labs were exposed for having 16 million fraudulent conversations

3
00:00:10,280 --> 00:00:15,800
Ejaaz:
with Claude with one specific goal, to try and steal its capabilities to train their own models.

4
00:00:16,200 --> 00:00:20,580
Ejaaz:
Now, the week before, Google said the same thing about China attacking their Gemini models.

5
00:00:20,720 --> 00:00:24,440
Ejaaz:
The week before that, OpenEye said the same thing. The top three American AI

6
00:00:24,440 --> 00:00:28,100
Ejaaz:
labs are blaming China for trying to hack their own AI models.

7
00:00:28,660 --> 00:00:32,420
Ejaaz:
But here's the twist in the story. What China's actually doing may not actually

8
00:00:32,420 --> 00:00:34,360
Ejaaz:
be illegal in the first place.

9
00:00:34,480 --> 00:00:39,160
Ejaaz:
In fact, this is something that every AI company is doing to get ahead in the AI race.

10
00:00:39,260 --> 00:00:42,800
Ejaaz:
In this episode, we're going to explore what all these reports confirm and whether

11
00:00:42,800 --> 00:00:45,640
Ejaaz:
distillation, the hacking vector, is actually a bad thing.

12
00:00:46,100 --> 00:00:48,820
Josh:
Yeah, so it starts with this blog post that Anthropik published earlier this

13
00:00:48,820 --> 00:00:52,240
Josh:
week that says, it's a title, Detecting and Preventing Distillation Attacks.

14
00:00:52,520 --> 00:00:56,340
Josh:
And I guess maybe it's helpful to just kind of define distillation as a concept

15
00:00:56,340 --> 00:00:58,440
Josh:
before we get into what they're accusing China of.

16
00:00:58,740 --> 00:01:02,040
Josh:
And basically, the way it works is there is a teacher and a student model.

17
00:01:02,200 --> 00:01:07,220
Josh:
So the teacher is the large model. That would be Anthropik's Claude Opus model. It's this huge model.

18
00:01:07,440 --> 00:01:10,860
Josh:
They've spent hundreds of millions, billions of dollars training it and turning

19
00:01:10,860 --> 00:01:12,360
Josh:
it into the model that we use every day.

20
00:01:12,740 --> 00:01:16,060
Josh:
That model provides these high quality outputs to the student,

21
00:01:16,300 --> 00:01:18,980
Josh:
which is the smaller model that is getting distilled from.

22
00:01:19,120 --> 00:01:23,060
Josh:
So basically the smaller model, the distilled model learns to mimic the outputs

23
00:01:23,060 --> 00:01:26,880
Josh:
of the larger model, but does so at a fraction of the cost because it's able

24
00:01:26,880 --> 00:01:31,220
Josh:
to kind of cherry pick the types of outputs that it gets by prompting it very specifically.

25
00:01:31,460 --> 00:01:35,180
Josh:
So anybody with sufficient access to a model and enough prompts can actually

26
00:01:35,180 --> 00:01:39,780
Josh:
get enough information to emulate the large model with a much smaller data set.

27
00:01:39,860 --> 00:01:43,680
Josh:
Now the outputs are not always as good as the large model, but they're significantly

28
00:01:43,680 --> 00:01:45,360
Josh:
cheaper and oftentimes very close.

29
00:01:45,480 --> 00:01:48,280
Josh:
So in the case that you get an extra breakthrough or two on top of that,

30
00:01:48,460 --> 00:01:49,940
Josh:
you can build a pretty impressive model.

31
00:01:50,120 --> 00:01:52,880
Josh:
And allegedly, that's what's happening with these models from China,

32
00:01:53,160 --> 00:01:54,360
Josh:
at least according to Anthropic.

33
00:01:54,660 --> 00:01:59,260
Ejaaz:
At least that's what they say. But why is what you just described a good thing?

34
00:01:59,380 --> 00:02:03,500
Ejaaz:
It's because all the hundreds of billions of dollars that are invested in building

35
00:02:03,500 --> 00:02:08,180
Ejaaz:
out the best AI model isn't sustainable for the long-term future.

36
00:02:08,300 --> 00:02:11,300
Ejaaz:
In fact, if you want to have a model that's small enough to fit on your phone,

37
00:02:11,520 --> 00:02:15,580
Ejaaz:
but as intelligent enough as the top models, it needs to be distilled through

38
00:02:15,580 --> 00:02:16,920
Ejaaz:
that process that you just explained.

39
00:02:17,140 --> 00:02:20,940
Ejaaz:
So it's going from a big model to a smaller model that is just as intelligent

40
00:02:20,940 --> 00:02:22,780
Ejaaz:
in certain specific ways.

41
00:02:22,940 --> 00:02:27,060
Ejaaz:
The stats from this China hack on Anthropic, Josh, are kind of insane.

42
00:02:27,220 --> 00:02:32,920
Ejaaz:
So I mentioned 16 million exchanges, but they spun up 24,000 fake Anthropic

43
00:02:32,920 --> 00:02:35,240
Ejaaz:
accounts. Now, I have to specify it.

44
00:02:35,550 --> 00:02:40,450
Ejaaz:
Anthropic does not allow Chinese users to access their models for the specific

45
00:02:40,450 --> 00:02:42,970
Ejaaz:
reason that adversaries to the U.S.

46
00:02:43,250 --> 00:02:45,910
Ejaaz:
Could get access to superintelligence that they're building.

47
00:02:46,430 --> 00:02:50,590
Ejaaz:
So DeepSeek, I'm going to name some names now. DeepSeek, one of the top AI labs

48
00:02:50,590 --> 00:02:54,010
Ejaaz:
which caused the stock market to crash at the end of 2024, I believe,

49
00:02:54,190 --> 00:02:59,390
Ejaaz:
were responsible for 150,000 of those exchanges. Moonshot AI, 3.4 million.

50
00:02:59,870 --> 00:03:02,770
Ejaaz:
Minimax, which is a favorite that you and I have spoken about on the show,

51
00:03:03,790 --> 00:03:06,950
Ejaaz:
13 million exchanges of those 15-minute conversations.

52
00:03:07,150 --> 00:03:11,390
Ejaaz:
There is an argument here that the open source gold rush that has been happening

53
00:03:11,390 --> 00:03:14,110
Ejaaz:
in China was mainly because they were stealing US secrets.

54
00:03:14,370 --> 00:03:17,190
Josh:
Wouldn't that be funny if that was the case? And then if that's also the case,

55
00:03:17,290 --> 00:03:20,450
Josh:
then what do you do about it? I mean, Anthropic kind of came out and they were

56
00:03:20,450 --> 00:03:22,030
Josh:
very upset about this, clearly.

57
00:03:22,330 --> 00:03:26,210
Josh:
But at the end of the day, it's like kind of on them. the onus is on them to

58
00:03:26,210 --> 00:03:29,550
Josh:
protect their systems and prevent this from happening there's a really great

59
00:03:30,230 --> 00:03:33,370
Josh:
great post that you have on screen and it's a joke it says my

60
00:03:33,370 --> 00:03:36,570
Josh:
son asking me a lot of questions it's a distillation attack obviously

61
00:03:36,570 --> 00:03:40,770
Josh:
and i think it's it's kind of funny where like the irony is and we can get into

62
00:03:40,770 --> 00:03:45,010
Josh:
the hypocrisy of the whole thing is that anthropic as a company very much has

63
00:03:45,010 --> 00:03:49,330
Josh:
done this in the past in order to get where they are and they are kind of the

64
00:03:49,330 --> 00:03:53,290
Josh:
person who's crying wolf now saying wait we're getting attacked this is not

65
00:03:53,290 --> 00:03:54,890
Josh:
allowed we should not be able to do this.

66
00:03:55,250 --> 00:03:57,990
Ejaaz:
Yeah, I mean, Anthropic is doing this with their own models,

67
00:03:58,150 --> 00:04:00,910
Ejaaz:
right? They've distilled Claude Opus into their Haiku model.

68
00:04:01,210 --> 00:04:05,470
Ejaaz:
Google's distilled Gemini Ultra into Gemini Nano. This is a common practice.

69
00:04:05,630 --> 00:04:09,170
Ejaaz:
So then the question becomes, which part of this is illegal?

70
00:04:09,270 --> 00:04:10,370
Ejaaz:
What has China done that is illegal?

71
00:04:10,510 --> 00:04:13,190
Ejaaz:
Well, it's two things that Anthropic has claimed. Number one,

72
00:04:13,530 --> 00:04:16,890
Ejaaz:
they've got this fancy terms of service, which they're lawyers that have been

73
00:04:16,890 --> 00:04:19,790
Ejaaz:
paid millions of dollars that have drafted up, which says, hey,

74
00:04:19,870 --> 00:04:22,110
Ejaaz:
hey, hey, if it's our model that you're doing this to,

75
00:04:22,710 --> 00:04:25,890
Ejaaz:
you can't do that. We've patented this thing. It's going to be illegal and we're

76
00:04:25,890 --> 00:04:26,850
Ejaaz:
going to sue you in a court of law.

77
00:04:27,130 --> 00:04:32,370
Ejaaz:
The issue is China is in China and they don't abide to the US legal system at all.

78
00:04:32,510 --> 00:04:35,890
Ejaaz:
Which brings me to the second thing that they violated, which is a geographical

79
00:04:35,890 --> 00:04:39,170
Ejaaz:
restriction. They don't let anyone in that region access clause.

80
00:04:39,310 --> 00:04:43,290
Ejaaz:
So the fact that China has been able to pull this off from the top AI labs means

81
00:04:43,290 --> 00:04:45,550
Ejaaz:
that they've illegally spun up accounts to do this.

82
00:04:45,710 --> 00:04:49,010
Josh:
Well, you know who doesn't care about laws is China.

83
00:04:49,270 --> 00:04:55,070
Josh:
Like they could not care less. In fact, this is the time for wartime CEOs.

84
00:04:55,310 --> 00:05:00,270
Josh:
Like in very many ways, this is the largest war that's being fought between

85
00:05:00,270 --> 00:05:02,170
Josh:
the US and China. And it's around AI.

86
00:05:02,510 --> 00:05:06,510
Josh:
And I think for them to say, that's against our terms of service, this is wrong.

87
00:05:06,670 --> 00:05:11,610
Josh:
Like that is not a grounds for defending yourself because clearly they have

88
00:05:11,610 --> 00:05:14,390
Josh:
no disregard, they have no regard for any sort of law.

89
00:05:14,610 --> 00:05:18,750
Josh:
I mean, you'd look at Seed Dance 2.0 and how it violates every copyright law under the sun.

90
00:05:18,890 --> 00:05:23,030
Josh:
And yet people don't care. It's the best video generation model in the world that exists.

91
00:05:23,250 --> 00:05:26,990
Josh:
So it is a challenge to claim that because they're violating terms of service,

92
00:05:27,090 --> 00:05:29,190
Josh:
this is an illegal thing that you shouldn't be able to do.

93
00:05:29,350 --> 00:05:33,150
Josh:
And they've not just cut off China, but I think it's important to note they've

94
00:05:33,150 --> 00:05:34,910
Josh:
also cut off other frontier AI labs.

95
00:05:35,010 --> 00:05:39,570
Josh:
They famously had this beef with XAI recently where they cut off all of the

96
00:05:39,570 --> 00:05:40,950
Josh:
Claude Code access to other labs.

97
00:05:41,070 --> 00:05:45,290
Josh:
So Anthropica has been very controlled and closed down in who's actually able

98
00:05:45,290 --> 00:05:48,970
Josh:
to access their models. And it sounds like someone was able to bypass that and

99
00:05:48,970 --> 00:05:50,570
Josh:
they just got pretty upset about it.

100
00:05:51,000 --> 00:05:55,840
Ejaaz:
Well, I mean, the Pentagon is relying on the likes of Anthropic,

101
00:05:56,100 --> 00:06:00,380
Ejaaz:
XAI, and OpenAI to fund the warfare effort against China, right?

102
00:06:00,680 --> 00:06:03,800
Ejaaz:
To your point, we're in like a wartime position. Like these AI models are being

103
00:06:03,800 --> 00:06:05,620
Ejaaz:
used as a geopolitical weapon.

104
00:06:05,820 --> 00:06:10,340
Ejaaz:
And so whoever owns the best model per se can advance the quickest.

105
00:06:10,520 --> 00:06:14,640
Ejaaz:
So it's like an economically dependent thing. And this whole drama with the

106
00:06:14,640 --> 00:06:19,180
Ejaaz:
Pentagon has been, the Pentagon has been using Claude for pretty much quite

107
00:06:19,180 --> 00:06:23,160
Ejaaz:
a lot of covert activity, including the recent capture of Nicolas Maduro,

108
00:06:23,660 --> 00:06:26,160
Ejaaz:
the former, I guess, president of Venezuela.

109
00:06:26,620 --> 00:06:31,260
Ejaaz:
And the issue now is that Anthropic is restricting Pentagon's access,

110
00:06:31,660 --> 00:06:34,920
Ejaaz:
like American owned self-defense against these kinds of things.

111
00:06:35,080 --> 00:06:37,900
Ejaaz:
And so the Pentagon is getting fed up and issuing them an ultimatum and saying,

112
00:06:37,980 --> 00:06:42,200
Ejaaz:
listen, if you don't figure this out, we're going to classify you as a threat to the country.

113
00:06:42,360 --> 00:06:48,140
Ejaaz:
Now, I have to give credit to Anthropic for maintaining their identity evenly

114
00:06:48,140 --> 00:06:49,840
Ejaaz:
across every single facet, but

115
00:06:49,840 --> 00:06:53,040
Ejaaz:
I don't think it's the smart way to do it because at the end of the day,

116
00:06:53,200 --> 00:06:55,900
Ejaaz:
there are going to be things that require more uncensored versions and you just

117
00:06:55,900 --> 00:06:57,420
Ejaaz:
need to be compliant with that fact.

118
00:06:57,660 --> 00:07:03,120
Ejaaz:
Because to your point earlier, Josh, Claude, OpenAI, ChatGPT has become a national

119
00:07:03,120 --> 00:07:05,480
Ejaaz:
asset. And so it needs to be treated as such.

120
00:07:06,000 --> 00:07:08,900
Josh:
Yeah, it's a matter of national security. And the thing about Anthropic that's

121
00:07:08,900 --> 00:07:13,000
Josh:
unique to Anthropic, and I'm not sure many other companies in the AI space is

122
00:07:13,000 --> 00:07:16,440
Josh:
their mission statement, where if you talk to any employee who works at Anthropic,

123
00:07:16,720 --> 00:07:20,200
Josh:
they'll tell you the purpose of the company is safety and alignment.

124
00:07:20,440 --> 00:07:24,920
Josh:
And I think while it's a valiant effort and incredibly important,

125
00:07:25,220 --> 00:07:28,520
Josh:
it doesn't really bode well for the current state of affairs in which

126
00:07:28,900 --> 00:07:31,960
Josh:
velocity momentum and just raw speed to get

127
00:07:31,960 --> 00:07:35,580
Josh:
to the best model possible is actually beneficial so

128
00:07:35,580 --> 00:07:38,680
Josh:
i think what we're seeing here is there's just these increasing conflicts

129
00:07:38,680 --> 00:07:41,880
Josh:
with i mean the secretary of defense and the pentagon wanting

130
00:07:41,880 --> 00:07:44,640
Josh:
access to do things that they deem to

131
00:07:44,640 --> 00:07:49,220
Josh:
be a matter of national security and like xai wanting to go and build code using

132
00:07:49,220 --> 00:07:52,260
Josh:
their tools they're like no no no no that's not how we want this used we're

133
00:07:52,260 --> 00:07:55,700
Josh:
not going to allow that and then the rumor is is that apparently the pentagon

134
00:07:55,700 --> 00:07:59,420
Josh:
actually just kicked out anthropic and now grok and the xai team is responsible

135
00:07:59,420 --> 00:08:02,220
Josh:
for being the AI provider for the Pentagon.

136
00:08:02,440 --> 00:08:05,380
Josh:
So I found that interesting too. It's just like a little side development.

137
00:08:05,680 --> 00:08:09,540
Ejaaz:
Well, I mean, like what you're getting at there is that some of these AI models

138
00:08:09,540 --> 00:08:13,320
Ejaaz:
or AI companies in America are kind of being super hypocritical.

139
00:08:13,580 --> 00:08:16,540
Ejaaz:
Like this tweet actually explains it really well.

140
00:08:16,940 --> 00:08:22,100
Ejaaz:
Hey, did you hear about the little like $1.5 billion lawsuit that Anthropic

141
00:08:22,100 --> 00:08:28,160
Ejaaz:
had to pay out over pirating or illegally downloading 7 million books to train their own models?

142
00:08:28,160 --> 00:08:32,600
Ejaaz:
Open AI is facing similar lawsuits against newspapers or across newspapers,

143
00:08:32,900 --> 00:08:34,140
Ejaaz:
code repositories and authors.

144
00:08:34,480 --> 00:08:37,580
Ejaaz:
I'm pretty sure Anthropic got sued for using Reddit data to train their models.

145
00:08:37,740 --> 00:08:41,060
Ejaaz:
Google trained their entire model over the index data that they took.

146
00:08:41,200 --> 00:08:43,380
Ejaaz:
Now, the question then becomes, is that fair?

147
00:08:43,640 --> 00:08:48,520
Ejaaz:
Who are paying the authors and creators of the content where these AI labs that

148
00:08:48,520 --> 00:08:51,780
Ejaaz:
have like amassed hundreds of billions of dollars worth of valuation,

149
00:08:51,780 --> 00:08:53,600
Ejaaz:
who's paying those creators?

150
00:08:53,800 --> 00:08:56,480
Ejaaz:
No one is, right? So you could argue that that is a form of distillation.

151
00:08:56,680 --> 00:08:59,680
Ejaaz:
Now, obviously, that's looking at it in a very black and white face,

152
00:08:59,760 --> 00:09:00,900
Ejaaz:
but I do think it's hypocritical.

153
00:09:01,000 --> 00:09:04,860
Ejaaz:
And most importantly, the memes are just so, so good here.

154
00:09:05,020 --> 00:09:08,860
Ejaaz:
You've got people that are asking Claude in Chinese, what model are you?

155
00:09:08,980 --> 00:09:11,260
Ejaaz:
And them replying, hey, I'm DeepSeek.

156
00:09:11,480 --> 00:09:13,900
Ejaaz:
And then you've got this one here where it says, I can't believe someone would

157
00:09:13,900 --> 00:09:15,240
Ejaaz:
just steal from Anthropic like this.

158
00:09:15,660 --> 00:09:19,240
Ejaaz:
Anthropic spent millions of man hours handwriting code, text,

159
00:09:19,520 --> 00:09:20,380
Ejaaz:
art, and books. Obviously,

160
00:09:20,790 --> 00:09:23,590
Ejaaz:
you know tongue-in-cheek this isn't actually real the point that's being made

161
00:09:23,590 --> 00:09:28,130
Ejaaz:
is that all information is kind of taken or stolen or interpreted in some way

162
00:09:28,130 --> 00:09:31,610
Ejaaz:
shape or form so what makes it any different for china in this regard.

163
00:09:31,610 --> 00:09:35,050
Josh:
The crux the argument is that the same foundation

164
00:09:35,050 --> 00:09:39,690
Josh:
that anthropic built its models on is the foundation that chinese models are

165
00:09:39,690 --> 00:09:44,790
Josh:
building their foundation on it's just one level kind of up where they clearly

166
00:09:44,790 --> 00:09:49,650
Josh:
stole maybe not stole content but they clearly used the content that we've produced

167
00:09:49,650 --> 00:09:51,910
Josh:
as humans over time to train their model.

168
00:09:52,170 --> 00:09:54,890
Josh:
What Deep Seek is doing is the next layer up. It's taking the,

169
00:09:55,030 --> 00:09:58,090
Josh:
I guess, the quantized version of all of the human intelligence that we've developed

170
00:09:58,090 --> 00:09:59,790
Josh:
and then distilling that one layer up.

171
00:09:59,850 --> 00:10:04,370
Josh:
It's easy to see why they would be upset, but it's also easy to see why everyone

172
00:10:04,370 --> 00:10:06,150
Josh:
is kind of deeming them as hypocritical.

173
00:10:06,410 --> 00:10:09,810
Josh:
It's like, again, you know that you are a nation state actor,

174
00:10:10,190 --> 00:10:12,810
Josh:
like relatives to the rest of the world in one of the most important wars that's

175
00:10:12,810 --> 00:10:14,890
Josh:
being fought. You know that you are going to be getting attacked.

176
00:10:15,050 --> 00:10:17,950
Josh:
You know that these people are going to be coming for you to build their own

177
00:10:17,950 --> 00:10:20,230
Josh:
models in the race for this AGI and beyond.

178
00:10:20,610 --> 00:10:27,230
Josh:
And to think that it's not going to happen and to be upset when it does just seems wrong.

179
00:10:27,510 --> 00:10:30,870
Josh:
And I think that's probably where a lot of the backlash is coming from is because,

180
00:10:31,090 --> 00:10:35,850
Josh:
I mean, again, it's on them to solve for these issues before they happen or

181
00:10:35,850 --> 00:10:37,510
Josh:
accept the consequences if they don't.

182
00:10:37,710 --> 00:10:41,870
Josh:
And that's just what happens here. I mean, there's like, this is a bar fight.

183
00:10:42,010 --> 00:10:44,910
Josh:
There are no rules in this fight. It is the only thing that you're trying to

184
00:10:44,910 --> 00:10:48,690
Josh:
do is get to AGI as fast as possible. And clearly, China doesn't care.

185
00:10:49,030 --> 00:10:52,850
Ejaaz:
Can I say something in China's defense? And maybe this is a hot take.

186
00:10:53,410 --> 00:10:56,130
Ejaaz:
Their models be banging recently, okay?

187
00:10:56,390 --> 00:11:01,830
Ejaaz:
Like they have been churning out new model updates from the likes of Alibaba with Quen 3.5.

188
00:11:02,170 --> 00:11:05,030
Ejaaz:
By the way, if you haven't tried this model out, apparently it's really amazing

189
00:11:05,030 --> 00:11:08,770
Ejaaz:
with agents. It's absolutely crushed benchmarks. Once again, open source.

190
00:11:09,070 --> 00:11:12,910
Ejaaz:
We've got Minimax AI that we mentioned earlier, which was the biggest perpetrator

191
00:11:12,910 --> 00:11:15,150
Ejaaz:
of this distillation attack against Anthropic.

192
00:11:15,390 --> 00:11:17,310
Ejaaz:
It's the most used model on Open Router.

193
00:11:17,610 --> 00:11:22,530
Josh:
Also, what's interesting is like Minimax 2.5 is the most popular Chinese model for Open Claw too.

194
00:11:22,530 --> 00:11:25,430
Josh:
And i personally used it like when i was running into

195
00:11:25,430 --> 00:11:28,050
Josh:
the oauth issues with claude because they were they were kind

196
00:11:28,050 --> 00:11:30,910
Josh:
of threatening again they were threatening to ban users for using oauth for

197
00:11:30,910 --> 00:11:34,610
Josh:
going around things they're just the hardos like they have no fun uh but when

198
00:11:34,610 --> 00:11:38,670
Josh:
they were threatening to like break people's accounts and ban them i switched

199
00:11:38,670 --> 00:11:42,270
Josh:
over to minimax 2.5 and it actually worked very well and it's a fraction of

200
00:11:42,270 --> 00:11:46,130
Josh:
the cost and i was like hey if you're gonna push me away i'm gonna go here to

201
00:11:46,130 --> 00:11:49,550
Josh:
these models that get the job done for me and minimax was that one I

202
00:11:49,550 --> 00:11:53,430
Ejaaz:
Have a question for you like where are you geographically located right now are you in China.

203
00:11:54,040 --> 00:11:55,080
Josh:
No, I'm certainly not.

204
00:11:55,580 --> 00:11:58,840
Ejaaz:
Okay, so it looks like they're just giving you free access to do these things.

205
00:11:58,920 --> 00:12:02,360
Ejaaz:
There's no geographical jurisdictions that they're kind of like placing on your

206
00:12:02,360 --> 00:12:03,900
Ejaaz:
restrictions. They're just letting you do the thing.

207
00:12:04,060 --> 00:12:06,520
Ejaaz:
It's awesome. Like all of these models are open source.

208
00:12:07,020 --> 00:12:11,420
Ejaaz:
These are kind of embellishments that America should be propagating,

209
00:12:11,540 --> 00:12:13,400
Ejaaz:
but they're not. They're playing the opposite.

210
00:12:13,680 --> 00:12:16,580
Ejaaz:
They're playing kind of secretive and it's not working out in their favor.

211
00:12:16,720 --> 00:12:20,500
Ejaaz:
I mean, you've got Minimax, all these latest Chinese labs, by the way, GLM-5,

212
00:12:20,640 --> 00:12:23,980
Ejaaz:
Kimi-K 2.5, minimax are crazy good

213
00:12:23,980 --> 00:12:27,180
Ejaaz:
at computer use and agentic tooling kimike 2.5

214
00:12:27,180 --> 00:12:30,020
Ejaaz:
actually for the open claw fans out there released a

215
00:12:30,020 --> 00:12:32,680
Ejaaz:
browser extension and it's actually really good because the

216
00:12:32,680 --> 00:12:35,660
Ejaaz:
major issue with using open claw was that there was security issues

217
00:12:35,660 --> 00:12:39,200
Ejaaz:
well they created a sandbox environment that you can now use it so they're innovating

218
00:12:39,200 --> 00:12:43,580
Ejaaz:
at scale and to the case that they might be stealing certain secrets i don't

219
00:12:43,580 --> 00:12:47,160
Ejaaz:
think this is regarded as a hack or a stealing thing i actually just think they're

220
00:12:47,160 --> 00:12:50,980
Ejaaz:
trying to get better models out to more people and hey if america can use it

221
00:12:50,980 --> 00:12:54,620
Ejaaz:
it's hardly a geopolitical thing so i don't know i'm kind of in the defense

222
00:12:54,620 --> 00:12:58,740
Ejaaz:
of china here and maybe that's a hot take and then kind of finally i just want to one.

223
00:12:58,740 --> 00:13:03,240
Josh:
Thing on that china note i'm not sure that it's out of their own goodwill of

224
00:13:03,240 --> 00:13:07,820
Josh:
their heart i think it's like the reason they're open source is probably because

225
00:13:07,820 --> 00:13:12,980
Josh:
they're behind i have i would imagine that if they did have an nvidia equivalent

226
00:13:12,980 --> 00:13:16,020
Josh:
in China that was creating top tier GPUs and

227
00:13:16,020 --> 00:13:17,300
Ejaaz:
They did have the.

228
00:13:17,300 --> 00:13:20,460
Josh:
Yeah they would did have these leading models like Opus 4.6 and

229
00:13:20,460 --> 00:13:23,600
Josh:
GPT 5.3 they would close it

230
00:13:23,600 --> 00:13:27,320
Josh:
down because there is so much value in owning that but because China's behind

231
00:13:27,320 --> 00:13:31,940
Josh:
there's value in being open and sharing it and gaining as much adoption as possible

232
00:13:31,940 --> 00:13:36,040
Josh:
as quickly as possible and it seems like it's more strategic and tactical than

233
00:13:36,040 --> 00:13:41,000
Josh:
out of the the goodness of their heart but yeah I mean again open source really benefits everyone.

234
00:13:41,360 --> 00:13:45,240
Josh:
And as a US citizen, I've used plenty of the Chinese models and they work awesome

235
00:13:45,240 --> 00:13:46,780
Josh:
because they're just so cheap and effective.

236
00:13:47,860 --> 00:13:52,260
Ejaaz:
To be clear, it doesn't benefit everyone. It benefits the users of those models, right?

237
00:13:52,400 --> 00:13:56,140
Ejaaz:
Because the American AI labs, their valuations are going to tank if you have

238
00:13:56,140 --> 00:14:01,000
Ejaaz:
a Chinese open source, much cheaper version that can run on much less expensive hardware.

239
00:14:01,220 --> 00:14:05,020
Ejaaz:
So it makes sense that the Chinese models are basically going the open source

240
00:14:05,020 --> 00:14:07,980
Ejaaz:
route so that they can kind of like chip away at American valuations.

241
00:14:08,200 --> 00:14:11,180
Ejaaz:
And then, as you said, Josh, entrench users in it. But it's

242
00:14:11,180 --> 00:14:14,100
Ejaaz:
not even just american llms there's

243
00:14:14,100 --> 00:14:17,200
Ejaaz:
chinese models that are specifically just good in

244
00:14:17,200 --> 00:14:20,040
Ejaaz:
china and beat a lot of the american models like what you're looking on the

245
00:14:20,040 --> 00:14:26,680
Ejaaz:
screen now is not transformers 6 it is a 30 second video from c dance 2.0 which

246
00:14:26,680 --> 00:14:31,600
Ejaaz:
is a chinese video model which is just at the front of its like own race it's

247
00:14:31,600 --> 00:14:35,540
Ejaaz:
basically at the top of its kind and it's super cheap to produce like hollywood

248
00:14:35,540 --> 00:14:37,060
Ejaaz:
cinematic effects right now.

249
00:14:37,460 --> 00:14:40,060
Ejaaz:
C-Dance 3, the stats were leaked the other day.

250
00:14:40,660 --> 00:14:43,880
Ejaaz:
10 to 18 minutes of continuous cinematic video.

251
00:14:44,100 --> 00:14:47,720
Ejaaz:
So we're going from 30 seconds to almost like a casual episode on,

252
00:14:47,980 --> 00:14:51,360
Ejaaz:
I don't know, like on your network TV's worth in a matter of seconds.

253
00:14:51,360 --> 00:14:52,660
Ejaaz:
It's just kind of insane to see.

254
00:14:52,800 --> 00:14:56,260
Ejaaz:
And I don't think that, you know, this is a nudge against China.

255
00:14:56,420 --> 00:15:00,720
Ejaaz:
I just think like, you know, this thing is accessible to anyone and everyone. Should be at scale soon.

256
00:15:00,860 --> 00:15:03,780
Josh:
When you're in a bar fight, the dude who like

257
00:15:03,780 --> 00:15:06,740
Josh:
smashes the bottle over the counter and starts waving

258
00:15:06,740 --> 00:15:10,000
Josh:
it around as a weapon like that's the guy that wins the

259
00:15:10,000 --> 00:15:12,700
Josh:
person who are armed and willing to break the rules and to do

260
00:15:12,700 --> 00:15:16,440
Josh:
whatever it takes to win that's the person that wins and china time and time

261
00:15:16,440 --> 00:15:20,620
Josh:
again has proven that that's what they're willing to do and it creates really

262
00:15:20,620 --> 00:15:25,300
Josh:
difficult moral dilemma between companies like anthropic that i genuinely do

263
00:15:25,300 --> 00:15:30,420
Josh:
believe have people's best interests at heart but none of the incentives align with that mission.

264
00:15:30,620 --> 00:15:33,440
Josh:
There is no incentive for being safe when...

265
00:15:34,240 --> 00:15:39,120
Josh:
The opposers on the other side of the planet have no regard for it because should

266
00:15:39,120 --> 00:15:44,420
Josh:
being safe slow down our progress that only allows them to catch up or accelerate ahead.

267
00:15:44,600 --> 00:15:48,740
Josh:
And then we are living under a world in which it is run by Chinese rules from

268
00:15:48,740 --> 00:15:52,460
Josh:
Chinese models. And it's this impossibly difficult dilemma that they're trying to navigate.

269
00:15:52,580 --> 00:15:55,820
Josh:
And I really have a lot of empathy for that because it's a difficult place.

270
00:15:55,940 --> 00:15:59,540
Josh:
You want to create this safe super intelligence, this safe AGI that doesn't harm the world.

271
00:15:59,660 --> 00:16:05,580
Josh:
But at the same time, you do need to be a wartime presence. You need to lock down your endpoints.

272
00:16:05,720 --> 00:16:10,700
Josh:
You need to have detection for 24,000 fake accounts that are extracting tons of data for you.

273
00:16:10,820 --> 00:16:16,340
Josh:
Like this is a serious issue. And I really hope that this is kind of like a

274
00:16:16,340 --> 00:16:22,280
Josh:
warning cry or just like a refocusing for a lot of these AI labs in how important it is to...

275
00:16:22,990 --> 00:16:28,150
Josh:
Keep your stuff locked down or just do whatever needs to be done to win this race?

276
00:16:28,310 --> 00:16:32,790
Ejaaz:
To round this up, I see a few things happening going forwards.

277
00:16:33,190 --> 00:16:37,790
Ejaaz:
Number one, I think companies like Anthropic and maybe even OpenAIR and Gemini

278
00:16:37,790 --> 00:16:42,570
Ejaaz:
or Google to an extent are going to start locking down their APIs in a few ways.

279
00:16:42,750 --> 00:16:47,190
Ejaaz:
Google started locking down their thing to OpenClaw, their API to OpenClaw this week.

280
00:16:47,490 --> 00:16:50,570
Ejaaz:
Anthropic started doing the same after announcing this distillation attack.

281
00:16:50,570 --> 00:16:54,050
Ejaaz:
Now, this is not going to be good for net net for users because,

282
00:16:54,050 --> 00:16:58,130
Ejaaz:
you know, they say that they're preventing Chinese hacks, but really like it's

283
00:16:58,130 --> 00:17:00,310
Ejaaz:
the software engineer in America that suffers from this.

284
00:17:00,430 --> 00:17:02,590
Ejaaz:
And I would say it would have the opposite effect that they want,

285
00:17:02,690 --> 00:17:05,230
Ejaaz:
which is these software engineers who can't afford, you know,

286
00:17:05,550 --> 00:17:08,870
Ejaaz:
to spend tens of thousands of dollars every month to access top tier models

287
00:17:08,870 --> 00:17:10,210
Ejaaz:
are just going to go to these Chinese models.

288
00:17:10,310 --> 00:17:12,370
Ejaaz:
So it's going to have the opposite effect of what you actually want.

289
00:17:12,370 --> 00:17:15,430
Ejaaz:
I think the other thing that we have to recognize, which is just the uncomfortable

290
00:17:15,430 --> 00:17:19,870
Ejaaz:
truth, is this isn't a conversation about AI models and the AI race.

291
00:17:20,010 --> 00:17:22,570
Ejaaz:
I think this is a geopolitical discussion.

292
00:17:22,910 --> 00:17:25,630
Ejaaz:
This is America versus China, as it always has been.

293
00:17:26,070 --> 00:17:29,810
Ejaaz:
And to Dario's point in, what was it, the name of it, Davos,

294
00:17:30,090 --> 00:17:35,330
Ejaaz:
he stated that, you know, giving or selling GPUs or selling model access to

295
00:17:35,330 --> 00:17:38,750
Ejaaz:
China is the equivalent of giving them the keys to nukes, right?

296
00:17:38,750 --> 00:17:42,130
Ejaaz:
Because if you assume that these AI models are going to become intelligent enough,

297
00:17:42,350 --> 00:17:45,170
Ejaaz:
they're going to be used against each other's adversaries.

298
00:17:45,310 --> 00:17:47,070
Ejaaz:
So you can't necessarily or

299
00:17:47,070 --> 00:17:49,750
Ejaaz:
you don't necessarily want to give China access to these side of things.

300
00:17:49,910 --> 00:17:56,210
Josh:
The progress of AI and the safety of AI will fall to that lowest common denominator where like...

301
00:17:56,420 --> 00:17:59,660
Josh:
We want a good video model. Well, China doesn't care for copyright.

302
00:17:59,820 --> 00:18:00,820
Josh:
They go and create seed dance.

303
00:18:01,320 --> 00:18:04,940
Josh:
Anthropic doesn't want to cooperate with the Pentagon and it wants to make sure

304
00:18:04,940 --> 00:18:07,520
Josh:
that the Pentagon does things a little safer than the Pentagon would like.

305
00:18:07,680 --> 00:18:09,980
Josh:
Well, Grok is there to step in and to fill that void.

306
00:18:10,300 --> 00:18:16,060
Josh:
And the reality is, is that while these morals are so important to stand on,

307
00:18:16,200 --> 00:18:20,780
Josh:
they're so incredibly difficult to enforce because the stakes are as high as they are.

308
00:18:20,940 --> 00:18:24,880
Josh:
And I think when we look at the Game of Thrones, how do we evaluate all the

309
00:18:24,880 --> 00:18:29,020
Josh:
positions of all these companies, it's becoming increasingly clear that the

310
00:18:29,020 --> 00:18:33,120
Josh:
moral compass is going to become increasingly complex as the stakes get higher.

311
00:18:33,440 --> 00:18:37,540
Josh:
And a company like Anthropic, who wants to be Anthropic, is going to have a

312
00:18:37,540 --> 00:18:41,600
Josh:
very difficult time maintaining that, even though it's probably critical for

313
00:18:41,600 --> 00:18:42,680
Josh:
the safety and well-being.

314
00:18:42,880 --> 00:18:46,960
Ejaaz:
The other thing I was thinking about is when these types of hacks,

315
00:18:47,120 --> 00:18:54,380
Ejaaz:
hacks using distillation, removes all the safety caps that American AI labs put in.

316
00:18:54,480 --> 00:18:57,800
Ejaaz:
So for example, if you had an uncensored version of Claude, you could use it

317
00:18:57,800 --> 00:19:00,760
Ejaaz:
to create or help you create biochemical weapons.

318
00:19:01,220 --> 00:19:05,360
Ejaaz:
But Anthropic puts in safeguards so that you aren't able to do such things, right?

319
00:19:05,780 --> 00:19:09,160
Ejaaz:
Chinese model labs that are distilling models into there to train their own

320
00:19:09,160 --> 00:19:11,040
Ejaaz:
models don't have that safety limit.

321
00:19:11,160 --> 00:19:14,940
Ejaaz:
You would need to rely on China being able to do that and not adding any nefarious backdoors.

322
00:19:15,100 --> 00:19:20,740
Ejaaz:
So I see the point around American model labs being responsible for their own

323
00:19:20,740 --> 00:19:23,880
Ejaaz:
thing and understanding that they are now a national level asset and they need

324
00:19:23,880 --> 00:19:25,200
Ejaaz:
to kind of respond effectively.

325
00:19:25,500 --> 00:19:29,700
Ejaaz:
But equally, we can't necessarily just be relaxed and let China do similar things

326
00:19:29,700 --> 00:19:31,980
Ejaaz:
like this. So it is a tricky one.

327
00:19:32,140 --> 00:19:37,420
Ejaaz:
I think, without doubt, the frontier of modern warfare against these two nations

328
00:19:37,420 --> 00:19:41,820
Ejaaz:
looks like an AI model attacking each other. I don't think it's got anything to do with weapons.

329
00:19:41,940 --> 00:19:44,600
Ejaaz:
It's got quite the opposite. That's why the Pentagon cares so much.

330
00:19:44,800 --> 00:19:48,980
Ejaaz:
That's why they're signing deals with Open Air and Grok to create drone warfare

331
00:19:48,980 --> 00:19:50,180
Ejaaz:
technology and so much more.

332
00:19:50,280 --> 00:19:53,260
Ejaaz:
So I think this is in the end, we're going to see way more attacks from this,

333
00:19:53,260 --> 00:19:55,660
Ejaaz:
Maybe even in switched roles, I don't know.

334
00:19:56,690 --> 00:19:57,210
Josh:
Interesting

335
00:19:57,210 --> 00:19:57,790
Ejaaz:
To see nevertheless.

336
00:19:57,790 --> 00:20:00,450
Josh:
Yeah it's um i mean again this game of

337
00:20:00,450 --> 00:20:03,630
Josh:
thrones is just going to keep getting more interesting higher stakes people

338
00:20:03,630 --> 00:20:06,630
Josh:
are going to start sacrificing more and more and this is just the most recent example

339
00:20:06,630 --> 00:20:09,510
Josh:
of anthropic being the one in the crosshairs but

340
00:20:09,510 --> 00:20:12,250
Josh:
i'm sure it's just a matter of time until others are as well but i

341
00:20:12,250 --> 00:20:15,050
Josh:
think that concludes the episode today that is the update the

342
00:20:15,050 --> 00:20:17,810
Josh:
anthropic drama that's everything you need to know about it

343
00:20:17,810 --> 00:20:20,490
Josh:
and i guess the the prompt for today which i'm

344
00:20:20,490 --> 00:20:24,270
Josh:
curious about is like kind of where do you stand on the issue it's complicated

345
00:20:24,270 --> 00:20:28,990
Josh:
because in a way everyone is right and everyone is wrong like everyone is breaking

346
00:20:28,990 --> 00:20:33,310
Josh:
the rules but does like what are the rules actually are they actually able to

347
00:20:33,310 --> 00:20:38,990
Josh:
be enforced i don't know but yeah i'm curious to hear just general takes on on the issue here it's

348
00:20:38,990 --> 00:20:42,030
Ejaaz:
A good one like how do you feel like for those of you who are playing around

349
00:20:42,030 --> 00:20:47,190
Ejaaz:
with like kimmy k 2.5 or minimax like myself do you feel like more likely to

350
00:20:47,190 --> 00:20:49,910
Ejaaz:
pick them up now that you know what's going on or are you just kind of on the

351
00:20:49,910 --> 00:20:53,130
Ejaaz:
side of like yeah this is happening and it's cheap for me to use and I can run

352
00:20:53,130 --> 00:20:55,910
Ejaaz:
it privately at home. Maybe it doesn't matter. I don't know. Let us know.

353
00:20:56,050 --> 00:20:59,290
Josh:
We didn't talk about that. Are you now more or less inclined to use these models?

354
00:20:59,930 --> 00:21:04,430
Ejaaz:
Dude, I, okay, I'm just gonna be very honest. I'm still gonna use these models

355
00:21:04,430 --> 00:21:07,790
Ejaaz:
because I'm not exactly convinced, even though I understand where Anthropik's

356
00:21:07,790 --> 00:21:10,270
Ejaaz:
coming from, that distillation is such a bad thing.

357
00:21:10,550 --> 00:21:14,130
Ejaaz:
I think they need to figure out a way to prevent people from distilling them.

358
00:21:14,330 --> 00:21:17,690
Ejaaz:
If you can access it via an API, you've got a security issue, not a national threat.

359
00:21:17,950 --> 00:21:21,610
Josh:
Yeah, I think I'm probably in the same boat where like I will continue to experiment

360
00:21:21,610 --> 00:21:24,050
Josh:
with SeedDance because C Dance is so much better than everything else.

361
00:21:24,370 --> 00:21:28,190
Josh:
And I'll just use the best products at the time. And I hope that the American

362
00:21:28,190 --> 00:21:30,710
Josh:
companies continue to provide the best products.

363
00:21:31,110 --> 00:21:34,350
Josh:
And yeah, I guess that concludes today's episode. So if you did enjoy,

364
00:21:34,490 --> 00:21:37,990
Josh:
please don't forget to share it with your friends. That's a big way to help us grow.

365
00:21:38,770 --> 00:21:42,010
Josh:
Liking, subscribing, commenting. If you're listening to this podcast,

366
00:21:42,330 --> 00:21:43,610
Josh:
rating five stars goes a long way.

367
00:21:43,890 --> 00:21:47,290
Josh:
And yeah, we have a amazing sub stack that comes out twice a week that you can

368
00:21:47,290 --> 00:21:50,130
Josh:
also subscribe to. Everything is linked down below in the description.

369
00:21:50,430 --> 00:21:52,570
Josh:
And Ejaz, unless you got anything else, So I think that's it for today.

370
00:21:52,770 --> 00:21:56,330
Ejaaz:
No, that's it. So we'll see you guys on the next one. I'll see you folks. See you guys.