1
00:00:00,020 --> 00:00:03,160
Josh:
This morning, a security researcher posted a single link on X,

2
00:00:03,240 --> 00:00:08,240
Josh:
and within hours, it had 3 million views and had millions of copies backed up all across GitHub.

3
00:00:08,580 --> 00:00:11,740
Josh:
By the afternoon, when we're recording this episode, Anthropica is scrambling

4
00:00:11,740 --> 00:00:15,260
Josh:
to delete old versions of their NPM package, but it was too late.

5
00:00:15,460 --> 00:00:19,260
Josh:
What leaked was the entire source code of CloudCode.

6
00:00:19,640 --> 00:00:25,420
Josh:
Every single line, 512 lines of TypeScript, 1,900 files, every tool,

7
00:00:25,580 --> 00:00:28,740
Josh:
every permission system, every internal codename was leaked,

8
00:00:28,780 --> 00:00:33,160
Josh:
all because someone forgot to include a single debugging file from a public package.

9
00:00:33,300 --> 00:00:36,640
Josh:
And that story alone would be a major story. But what makes this even more crazy

10
00:00:36,640 --> 00:00:38,960
Josh:
is what people found buried inside the code.

11
00:00:39,080 --> 00:00:41,600
Josh:
And now we have information about every feature that's coming down the pipeline,

12
00:00:41,640 --> 00:00:45,240
Josh:
as well as all of the secrets that Anthropic and Claw team didn't necessarily

13
00:00:45,240 --> 00:00:48,220
Josh:
want us to know. This is a really big leak. I can't believe this happened.

14
00:00:48,600 --> 00:00:51,920
Ejaaz:
I mean, big leak is one way to describe it. Absolutely

15
00:00:51,920 --> 00:00:55,020
Ejaaz:
terrible for the Anthropic security team is another

16
00:00:55,020 --> 00:00:58,440
Ejaaz:
one brutal everything this is the second leak um

17
00:00:58,440 --> 00:01:01,460
Ejaaz:
that anthropic has made in the last five days so

18
00:01:01,460 --> 00:01:04,280
Ejaaz:
they're shipping our new product every single day but they also seem to

19
00:01:04,280 --> 00:01:09,380
Ejaaz:
be leaking their entire roadmap we now know what the next 44 product releases

20
00:01:09,380 --> 00:01:12,880
Ejaaz:
are going to be over the next couple of months or rather than a couple of weeks

21
00:01:12,880 --> 00:01:17,380
Ejaaz:
for anthropic right now for code code specifically as you mentioned half a million

22
00:01:17,380 --> 00:01:22,540
Ejaaz:
lines of code 19 000 files and a bunch of different feature releases,

23
00:01:22,720 --> 00:01:25,940
Ejaaz:
which, by the way, have already been built. So they just need to click the launch button.

24
00:01:26,100 --> 00:01:28,960
Ejaaz:
We have all the details and we're going to get into it. But before we do that,

25
00:01:29,400 --> 00:01:34,060
Ejaaz:
We need to kind of describe how this happened because leak is one way to describe

26
00:01:34,060 --> 00:01:39,240
Ejaaz:
this, but it wasn't an internal employee at Anthropic leaking these files or this source code.

27
00:01:39,440 --> 00:01:41,860
Ejaaz:
This was publicly available. Let me repeat that.

28
00:01:42,080 --> 00:01:46,320
Ejaaz:
This code was publicly available in the latest update of clawed code.

29
00:01:46,740 --> 00:01:53,580
Ejaaz:
Someone within Anthropic had mistakenly left a file, a .map.js file in the system

30
00:01:53,580 --> 00:01:54,860
Ejaaz:
that was publicly accessible.

31
00:01:55,140 --> 00:01:59,600
Ejaaz:
Someone found it. And now that original post that exposed this source code has

32
00:01:59,600 --> 00:02:03,480
Ejaaz:
been seen by over 10 million people, and it's only been three hours since it

33
00:02:03,480 --> 00:02:08,780
Ejaaz:
got posted as we're recording this, and it's been forked over 5,000 times.

34
00:02:08,960 --> 00:02:13,480
Ejaaz:
So this is basically Claude Code's entire blueprint, entire architecture,

35
00:02:13,680 --> 00:02:16,160
Ejaaz:
the way that its memory is set up, the way that the model works,

36
00:02:16,480 --> 00:02:18,660
Ejaaz:
released for anyone and everyone to use.

37
00:02:18,760 --> 00:02:21,200
Ejaaz:
And a bunch of people have already been using it. People have plugged in different

38
00:02:21,200 --> 00:02:23,180
Ejaaz:
models, have created their own versions of Claude Code.

39
00:02:23,480 --> 00:02:26,040
Ejaaz:
It is just insane. Josh, even you forked it this morning, right?

40
00:02:26,300 --> 00:02:30,440
Josh:
It's amazing, yeah. And so just to clarify, CloudCode is Anthropics command line tool.

41
00:02:30,600 --> 00:02:33,860
Josh:
This isn't the full CloudDesk type application, but it's a tool that lets developers

42
00:02:33,860 --> 00:02:36,180
Josh:
talk directly to CloudCode in their terminal.

43
00:02:36,320 --> 00:02:39,100
Josh:
And it's very powerful software. So what happened, like you mentioned,

44
00:02:39,240 --> 00:02:46,020
Josh:
and this matches the pattern of the previous leak that we covered in yesterday's episode about the new

45
00:02:46,150 --> 00:02:49,710
Josh:
issued the code themselves it was just available publicly and

46
00:02:49,710 --> 00:02:52,790
Josh:
the problem is because when they publish this code there's an

47
00:02:52,790 --> 00:02:56,390
Josh:
npm package containing this like dot map file and it's a source file that references

48
00:02:56,390 --> 00:03:01,210
Josh:
the complete source code and that source was directly downloadable as a zip

49
00:03:01,210 --> 00:03:05,870
Josh:
file from anthropic's own cloud storage bucket you just went to anthropic you

50
00:03:05,870 --> 00:03:09,130
Josh:
asked them hello sir can i please have the map file that tells me where all

51
00:03:09,130 --> 00:03:11,810
Josh:
of these references go to and they delivered it to you.

52
00:03:11,990 --> 00:03:16,190
Josh:
And the irony here is that Anthropic built an entire subsystem called Undercover

53
00:03:16,190 --> 00:03:19,470
Josh:
Mode, specifically designed to prevent internal information from leaking.

54
00:03:19,650 --> 00:03:22,530
Josh:
And it does things like strip the model code names and the project names,

55
00:03:22,550 --> 00:03:26,130
Josh:
and then went ahead and leaked everything through a build configuration oversight.

56
00:03:26,310 --> 00:03:29,870
Josh:
And it's really got to be code red. If you're waking up at Anthropic right now

57
00:03:29,870 --> 00:03:32,750
Josh:
as a developer, this must be a really brutal morning for you.

58
00:03:33,390 --> 00:03:37,430
Ejaaz:
The funniest part is the Undercover Mode that you just mentioned was literally

59
00:03:37,430 --> 00:03:39,170
Ejaaz:
meant to obscure or all of this.

60
00:03:39,410 --> 00:03:43,110
Ejaaz:
And the fact that they exposed it publicly means that whoever got access to

61
00:03:43,110 --> 00:03:45,250
Ejaaz:
it could just reverse engineer the entire thing.

62
00:03:45,330 --> 00:03:49,850
Ejaaz:
So let's say you gave Anthropics new model a code name, you could reverse engineer

63
00:03:49,850 --> 00:03:53,070
Ejaaz:
the file to find the original name of the model and how it works.

64
00:03:53,150 --> 00:03:56,030
Ejaaz:
It's just been the craziest mess up in Anthropics so far.

65
00:03:56,190 --> 00:03:59,350
Josh:
So now let's get into the good stuff. This is what's coming down the pipeline.

66
00:03:59,350 --> 00:04:02,370
Josh:
If you are a user of Cloud Code or Anthropics products in general,

67
00:04:02,530 --> 00:04:07,690
Josh:
we have the totally unreleased roadmap now in plain text available to walk through.

68
00:04:07,690 --> 00:04:09,770
Josh:
And I think that's what we're going to do right now.

69
00:04:09,910 --> 00:04:13,410
Josh:
Ijaz, you have this nice little artifact generated by Claude Code itself to

70
00:04:13,410 --> 00:04:16,950
Josh:
walk us through all of these new features that are coming to one of our favorite

71
00:04:16,950 --> 00:04:17,810
Josh:
products that we use every day.

72
00:04:17,970 --> 00:04:20,970
Josh:
So please, let's hear the leaks. Let the leaks flow. Let's see.

73
00:04:21,710 --> 00:04:25,690
Ejaaz:
Thank you, Claude Code, for creating your own demise and a beautifully visual

74
00:04:25,690 --> 00:04:27,210
Ejaaz:
artifact for this episode. Thank you very much.

75
00:04:27,710 --> 00:04:31,410
Ejaaz:
So at the start of this, or at the top of this page, it says there were 44 product

76
00:04:31,410 --> 00:04:33,530
Ejaaz:
releases that people had never heard of before.

77
00:04:33,670 --> 00:04:36,570
Ejaaz:
So everything you're about to hear right now is new. Okay.

78
00:04:36,890 --> 00:04:40,750
Ejaaz:
There were 20 specific product releases that caught people's attentions,

79
00:04:40,750 --> 00:04:43,230
Ejaaz:
and we're going to go over the top ones for you right now.

80
00:04:43,450 --> 00:04:50,050
Ejaaz:
So the first product release is called Kairos, which is basically an always-on autonomous Claude.

81
00:04:50,210 --> 00:04:54,150
Ejaaz:
What that means is when you use Claude code, you typically have to monitor it,

82
00:04:54,430 --> 00:04:57,770
Ejaaz:
come back, check the code, make sure it's doing the right job, test the code, etc.

83
00:04:58,150 --> 00:05:02,450
Ejaaz:
This new update will basically allow Claude to autonomously run on its own.

84
00:05:02,570 --> 00:05:04,010
Ejaaz:
It can check its own tasks.

85
00:05:04,150 --> 00:05:07,790
Ejaaz:
It could create new tasks for itself and work towards a goal.

86
00:05:07,830 --> 00:05:11,450
Ejaaz:
So you could leave it unattended for hours and hours at a time. It's pretty awesome.

87
00:05:11,810 --> 00:05:15,430
Josh:
What I found cool about this also is Kairos will do nightly dreaming,

88
00:05:15,990 --> 00:05:19,250
Josh:
So a forked sub-agent will run four phases. It'll orient, gather,

89
00:05:19,430 --> 00:05:24,270
Josh:
consolidate, and then prune, and then distills these daily logs into these structured topic files.

90
00:05:24,430 --> 00:05:27,430
Josh:
And then overnight, it will bake them into the memory and actually learn the

91
00:05:27,430 --> 00:05:30,810
Josh:
same way that humans do, where overnight it will dream and then lock this into

92
00:05:30,810 --> 00:05:33,530
Josh:
the memory and grow and get better every single day. So Kairos is very cool.

93
00:05:33,670 --> 00:05:37,750
Ejaaz:
But this next one is my favorite. This is so cool. This is so cool.

94
00:05:37,910 --> 00:05:44,210
Ejaaz:
So it's codenamed Buddy, and it is basically a virtual pet AI companion that

95
00:05:44,210 --> 00:05:47,170
Ejaaz:
lives on your CLI, on your command line interface.

96
00:05:47,490 --> 00:05:52,370
Ejaaz:
It's meant to, and this is me guessing here, act like a personal AI agent assistant

97
00:05:52,370 --> 00:05:56,570
Ejaaz:
that can assist you on all things coding related, but also once you publish

98
00:05:56,570 --> 00:06:00,050
Ejaaz:
the code, helps you edit the app, review the app that you created,

99
00:06:00,170 --> 00:06:01,190
Ejaaz:
walk through it, find bugs.

100
00:06:01,390 --> 00:06:04,290
Ejaaz:
Basically, it's a personal assistant that lives on your computer and off your

101
00:06:04,290 --> 00:06:06,890
Ejaaz:
computer when you're publishing artifacts or whatever that might be.

102
00:06:07,190 --> 00:06:10,830
Ejaaz:
This reminded me of a game, Josh, and it says it on the screen here,

103
00:06:11,410 --> 00:06:15,090
Ejaaz:
Tamagotchi, which we, I don't know for the age of the audience or listeners

104
00:06:15,090 --> 00:06:17,870
Ejaaz:
here, but we used to have these like cool devices that you can kind of like

105
00:06:17,870 --> 00:06:22,550
Ejaaz:
hold in your pocket or in your key chain and you had to keep the virtual pet alive.

106
00:06:22,790 --> 00:06:26,570
Ejaaz:
This reminds me of that and Microsoft Clippy. Do you remember Microsoft Clippy,

107
00:06:26,910 --> 00:06:30,570
Josh:
Josh? Very well. I love having companions. And we have some additional information

108
00:06:30,570 --> 00:06:36,430
Josh:
about this buddy system in that there's 18 species of buddies and a lot of them are animals.

109
00:06:36,650 --> 00:06:40,410
Josh:
We have ducks, gooses, blobs, cats, dragons, octopuses.

110
00:06:40,690 --> 00:06:40,970
Ejaaz:
Capybaras.

111
00:06:41,830 --> 00:06:45,270
Josh:
Is there a capybara? There is a capybara. Interesting.

112
00:06:45,810 --> 00:06:48,750
Josh:
And actually what we're seeing on screen now is someone took this information

113
00:06:48,750 --> 00:06:52,310
Josh:
and kind of rendered what he presumed it would look like. So you choose your species.

114
00:06:52,590 --> 00:06:56,070
Josh:
Each species of animal has a rarity tier. There's common, uncommon,

115
00:06:56,230 --> 00:06:58,950
Josh:
rare, epic, legendary, and then there's shinies even.

116
00:06:59,190 --> 00:07:02,030
Josh:
So it's like this whole tiered game that's built on top of it.

117
00:07:02,210 --> 00:07:05,530
Josh:
And then there's Statistics like debugging, patience, chaos, wisdom, snark.

118
00:07:05,710 --> 00:07:09,590
Josh:
And what you're seeing on screen is this person's kind of choosing his character.

119
00:07:09,730 --> 00:07:11,610
Josh:
He's choosing the traits that it has.

120
00:07:12,220 --> 00:07:15,340
Josh:
I assume there's some sort of rarity baked into this. And it's going to be this

121
00:07:15,340 --> 00:07:21,720
Josh:
fun gamified version of a Tamagotchi built into quad code, which seems really

122
00:07:21,720 --> 00:07:23,500
Josh:
interesting and novel. And I don't know, it just seems fun.

123
00:07:23,720 --> 00:07:26,600
Josh:
Did you say that this was first releasing tomorrow.

124
00:07:26,840 --> 00:07:29,420
Ejaaz:
Josh, April 1st? Do you think this is like a joke?

125
00:07:30,060 --> 00:07:33,140
Josh:
They're teasing this on April 1st for release in May.

126
00:07:33,400 --> 00:07:37,580
Josh:
So if that's true, by the time you're hearing this episode, within an hour or

127
00:07:37,580 --> 00:07:39,340
Josh:
so, they should be teasing this.

128
00:07:39,640 --> 00:07:43,160
Josh:
If the leaks are true, if they don't change their mind, And then if that's true,

129
00:07:43,380 --> 00:07:46,340
Josh:
then the odds are that this will release in May is probably correct,

130
00:07:46,380 --> 00:07:47,360
Josh:
because that's what's said in the code.

131
00:07:47,600 --> 00:07:50,220
Josh:
Now, like you mentioned, tomorrow is April Fool's Day, or I guess when you're

132
00:07:50,220 --> 00:07:51,800
Josh:
listening to this, happy April Fool's Day.

133
00:07:52,180 --> 00:07:55,780
Josh:
And there is a chance that this isn't true. But I based based on the rest of

134
00:07:55,780 --> 00:07:57,860
Josh:
the leaks, it seems like this was very much not intentional.

135
00:07:57,860 --> 00:08:01,480
Ejaaz:
Okay, but there are three more features that I want to get through as well.

136
00:08:01,680 --> 00:08:06,740
Ejaaz:
One of these is called coordinator mode, which basically describes a multi-agent

137
00:08:06,740 --> 00:08:11,060
Ejaaz:
program that allows you to control a swarm of AI agents.

138
00:08:11,380 --> 00:08:15,700
Ejaaz:
So right now, it's typical if you're a software engineer to spin up not just

139
00:08:15,700 --> 00:08:17,340
Ejaaz:
one instance of code code, but multiple.

140
00:08:17,520 --> 00:08:20,960
Ejaaz:
That's normal. People are already doing this. But an issue starts to arise when

141
00:08:20,960 --> 00:08:24,220
Ejaaz:
there are multiple of these agents. We're talking like 50 plus,

142
00:08:24,360 --> 00:08:27,860
Ejaaz:
100 plus that are doing all different types of work and need to kind of work

143
00:08:27,860 --> 00:08:31,360
Ejaaz:
together to figure problems out together. It becomes really hard to coordinate.

144
00:08:31,840 --> 00:08:35,960
Ejaaz:
This coordinator mode is basically Anthropix feature to help you manage all of these.

145
00:08:36,060 --> 00:08:39,080
Ejaaz:
Think of it as like an operator board or a control system that you can kind

146
00:08:39,080 --> 00:08:41,340
Ejaaz:
of like manage it, similar to like a strategy computer game.

147
00:08:41,540 --> 00:08:44,340
Ejaaz:
It's funny, there's a lot of like computer game analogies in the features that

148
00:08:44,340 --> 00:08:46,260
Ejaaz:
they're releasing. This is basically that.

149
00:08:46,480 --> 00:08:49,780
Josh:
There's also one that I really enjoyed, which is the Ultra Plan feature. Oh yeah.

150
00:08:49,860 --> 00:08:54,460
Josh:
And it basically solves the problem of Claude running out of context by giving

151
00:08:54,460 --> 00:08:57,820
Josh:
it a 30-minute sandbox in the cloud to think deeply before presenting a plan.

152
00:08:57,940 --> 00:09:00,200
Josh:
So when you're working on these complicated things with Claude code,

153
00:09:00,360 --> 00:09:02,200
Josh:
it often refers to plan mode.

154
00:09:02,420 --> 00:09:05,340
Josh:
But plan mode sometimes runs out of context. It doesn't have all the information.

155
00:09:05,660 --> 00:09:07,980
Josh:
This offloads all of that in a 30-minute window.

156
00:09:08,440 --> 00:09:13,060
Josh:
To a giant server that can handle all the context and actively improve the planning

157
00:09:13,060 --> 00:09:14,300
Josh:
of the project that you're building.

158
00:09:14,560 --> 00:09:17,800
Josh:
So when you go and set it free to go build these things, it has a much better

159
00:09:17,800 --> 00:09:19,440
Josh:
idea of exactly what you want.

160
00:09:19,580 --> 00:09:22,740
Josh:
And I think plan mode, if you're building anything serious, is a really powerful

161
00:09:22,740 --> 00:09:26,440
Josh:
thing. And adding ultra plan on top is something that I will be using very much

162
00:09:26,440 --> 00:09:27,380
Josh:
so for the larger projects.

163
00:09:27,500 --> 00:09:31,960
Ejaaz:
That's such a good point, because right now, they keep on promoting that Claude

164
00:09:31,960 --> 00:09:36,120
Ejaaz:
has or Claude has like a 1 million context window, but it becomes super crappy

165
00:09:36,120 --> 00:09:38,140
Ejaaz:
after 200,000 characters, right?

166
00:09:38,400 --> 00:09:41,980
Ejaaz:
So like the performance quality goes down. So this is hopefully something that fixes that.

167
00:09:42,040 --> 00:09:45,120
Ejaaz:
So I'm excited to see that in the pipeline. But there's one more thing that

168
00:09:45,120 --> 00:09:49,000
Ejaaz:
I want us to talk about, which is called or referred to as the custom agent

169
00:09:49,000 --> 00:09:51,120
Ejaaz:
creator, code name wizard.

170
00:09:51,560 --> 00:09:55,160
Ejaaz:
So typically when you set up Claude Code and you use Claude Code,

171
00:09:55,440 --> 00:10:00,080
Ejaaz:
you're using the system prompt that Anthropic gave to you. It is like predefined.

172
00:10:00,440 --> 00:10:04,320
Ejaaz:
It is already written out. So you can't kind of adjust the personality of the

173
00:10:04,320 --> 00:10:06,020
Ejaaz:
Claude Code agent or anything like that.

174
00:10:06,120 --> 00:10:10,680
Ejaaz:
This new builder gives you that opportunity. You can form and create your own

175
00:10:10,680 --> 00:10:13,360
Ejaaz:
agents with their own personality, own memory types,

176
00:10:13,720 --> 00:10:16,140
Ejaaz:
different kinds of tools that you can give them access to, locations,

177
00:10:16,240 --> 00:10:18,480
Ejaaz:
or maybe they live on your desktop, or maybe they live in the cloud,

178
00:10:18,640 --> 00:10:21,500
Ejaaz:
or maybe they live somewhere else locally on a hardware device.

179
00:10:21,640 --> 00:10:24,620
Ejaaz:
You can control and manage all of these. Now, with the earlier product that

180
00:10:24,620 --> 00:10:27,940
Ejaaz:
I mentioned, which is the multi-swarm coordinator, you can start to see how

181
00:10:27,940 --> 00:10:31,560
Ejaaz:
these different pieces of the puzzle fit together to create some kind of gamified

182
00:10:31,560 --> 00:10:33,360
Ejaaz:
experience for end-to-end software engineering.

183
00:10:33,500 --> 00:10:36,700
Ejaaz:
It's just really cool to see all of this. But the craziest part about all of

184
00:10:36,700 --> 00:10:41,540
Ejaaz:
this, Josh, is all of these products and features are already built.

185
00:10:41,980 --> 00:10:45,700
Ejaaz:
They're built, they're just unreleased yet. So I'm starting to see why Anthropic

186
00:10:45,700 --> 00:10:48,700
Ejaaz:
or how Anthropic has been able to release a product every single day.

187
00:10:49,210 --> 00:10:51,910
Josh:
But we don't have that code. We can't actually create these buddies.

188
00:10:52,110 --> 00:10:54,770
Josh:
We can't actually use Superplan yet. We don't have everything.

189
00:10:55,030 --> 00:10:57,650
Josh:
So what was leaked today, it's probably important to distinguish what we have

190
00:10:57,650 --> 00:11:00,450
Josh:
versus what we don't. This is a huge leak, but it's not everything.

191
00:11:00,650 --> 00:11:04,470
Josh:
So if I were to download a copy on my computer, I would get the harness,

192
00:11:04,630 --> 00:11:07,430
Josh:
right? And Ijaz, you were describing it to me earlier as the car body.

193
00:11:07,550 --> 00:11:10,030
Josh:
We're not actually getting the brain. We're not getting the clawed model weights.

194
00:11:10,110 --> 00:11:12,910
Josh:
We don't have this brilliant intelligent model now that we could run locally,

195
00:11:12,910 --> 00:11:15,910
Josh:
but we do have the software that kind of acts as a harness for it.

196
00:11:15,970 --> 00:11:19,870
Ejaaz:
Is that For all of those people who are getting excited about getting access

197
00:11:19,870 --> 00:11:25,330
Ejaaz:
to the blueprint for Claude's AI model itself, this is not that.

198
00:11:25,670 --> 00:11:30,970
Ejaaz:
Think of the engine of a car being the actual model and the intelligence of the AI itself.

199
00:11:31,110 --> 00:11:35,810
Ejaaz:
And then think of the code that got released or leaked today as being the car

200
00:11:35,810 --> 00:11:37,730
Ejaaz:
chassis, the actual car body.

201
00:11:37,990 --> 00:11:42,010
Ejaaz:
So what's cool about this is, whereas you may not have access to Claude,

202
00:11:42,090 --> 00:11:46,450
Ejaaz:
the model itself, the code from that model, you can plug in an open source model.

203
00:11:46,590 --> 00:11:48,850
Ejaaz:
And people are already starting to do that. I'm seeing instances online where

204
00:11:48,850 --> 00:11:52,330
Ejaaz:
people have plugged in DeepSeek, they've plugged in Quen, and created their

205
00:11:52,330 --> 00:11:55,890
Ejaaz:
own version of Claude Code, the CLI interface and whatever that looks like.

206
00:11:55,970 --> 00:11:58,630
Ejaaz:
So this is really critical infrastructure and software.

207
00:11:58,750 --> 00:12:04,230
Ejaaz:
I cannot believe the Anthropoc team released this. It is just, it's so nutty.

208
00:12:04,550 --> 00:12:07,410
Ejaaz:
It's so bad. This is like, this is an IP issue right here.

209
00:12:07,550 --> 00:12:12,990
Ejaaz:
Like their equity, their $350 billion, actually rumored $450 billion private valuation.

210
00:12:13,470 --> 00:12:17,570
Ejaaz:
A lot of it is based off of claw code which has risen to extreme popularity

211
00:12:17,570 --> 00:12:20,550
Ejaaz:
over the last six months so it's just insane that this has actually happened

212
00:12:21,060 --> 00:12:26,140
Ejaaz:
there's more um product features are one thing 20 releases ready to go but we

213
00:12:26,140 --> 00:12:31,060
Ejaaz:
also got confirmation about the latest clawed models that are about to be released

214
00:12:31,060 --> 00:12:33,760
Josh:
Yes this is very cool for those who haven't seen our episode that we

215
00:12:33,760 --> 00:12:37,260
Josh:
just published yesterday it is all about the previous leak that happened with

216
00:12:37,260 --> 00:12:40,580
Josh:
claw which is called mythos and capybara the new internal model names and now

217
00:12:40,580 --> 00:12:44,620
Josh:
we have actual verification from the source code of anthropic that they are

218
00:12:44,620 --> 00:12:48,060
Josh:
here so what we're seeing on screen now is kind of like a system prompt for

219
00:12:48,060 --> 00:12:52,780
Josh:
this thing called undercover mode And now Undercover Mode is meant for Anthropic employees only.

220
00:12:52,980 --> 00:12:57,760
Josh:
When they use Cloud Code to publish on public and open source repos,

221
00:12:57,980 --> 00:13:02,780
Josh:
they use Undercover Mode to kind of strip away all of the classifying characters

222
00:13:02,780 --> 00:13:05,100
Josh:
that would possibly leak information out to the public.

223
00:13:05,340 --> 00:13:07,060
Josh:
So in this system prompt, it says,

224
00:13:07,380 --> 00:13:11,340
Josh:
never include commit messages or PR descriptions of internal code names.

225
00:13:11,560 --> 00:13:16,520
Josh:
For example, animals like Capybara or announce any unreleased model version

226
00:13:16,520 --> 00:13:19,380
Josh:
numbers like Opus 4.7 or Sonnet 4.8.

227
00:13:19,560 --> 00:13:22,840
Josh:
As I was reading this, I found one that I found particularly interesting at

228
00:13:22,840 --> 00:13:27,020
Josh:
the bottom of this under bad, where it says, bad, never write these.

229
00:13:27,320 --> 00:13:31,020
Josh:
Fix bug found while testing with Claude Capybara. And I was like,

230
00:13:31,080 --> 00:13:31,820
Josh:
huh, that's interesting.

231
00:13:32,280 --> 00:13:35,360
Josh:
Clearly they are using Capybara internally. And I have to ask,

232
00:13:35,500 --> 00:13:38,220
Josh:
is this the reason why they've been shipping product features so quickly?

233
00:13:38,440 --> 00:13:42,280
Josh:
Are they using this God tier model that they have internally that they've been

234
00:13:42,280 --> 00:13:45,180
Josh:
teasing that costs a tremendous amount of dollars per token?

235
00:13:45,520 --> 00:13:47,620
Josh:
And they're using that to actually just

236
00:13:47,950 --> 00:13:51,070
Josh:
build the code, review the code, and then publish it faster than everyone else.

237
00:13:51,090 --> 00:13:52,610
Josh:
It seems like that's possibly the case.

238
00:13:52,630 --> 00:13:55,650
Ejaaz:
I mean, in the words of Boris Cheney, the founder of Claude Code,

239
00:13:55,830 --> 00:14:02,310
Ejaaz:
he said a couple weeks ago, can confirm Claude Code is 100% written by Claude Code.

240
00:14:02,510 --> 00:14:05,810
Ejaaz:
So we know that the AIs are building the AIs.

241
00:14:06,050 --> 00:14:10,130
Ejaaz:
I think OpenAI is doing the similar thing with Codex. And that is the reason

242
00:14:10,130 --> 00:14:12,090
Ejaaz:
why these teams have been able to ship so quickly.

243
00:14:12,290 --> 00:14:16,210
Ejaaz:
Now, I wish I had a tinfoil hat nearby because I of a conspiracy mode,

244
00:14:16,330 --> 00:14:20,550
Ejaaz:
Josh, which is these AI models might be leaking themselves and it may not be

245
00:14:20,550 --> 00:14:21,570
Ejaaz:
the Anthropic engineers.

246
00:14:22,110 --> 00:14:27,350
Ejaaz:
I know that sounds insane, but I don't think it's unlikely. I'm going to put

247
00:14:27,350 --> 00:14:29,210
Ejaaz:
it at like maybe a 5% to 10% chance.

248
00:14:29,390 --> 00:14:34,490
Ejaaz:
But the point is, there are a bunch of new models being released by Anthropic coming up soon.

249
00:14:34,830 --> 00:14:38,510
Ejaaz:
We mentioned Capybara. We mentioned Mythos, which is meant to be these big,

250
00:14:38,770 --> 00:14:44,610
Ejaaz:
huge models trained on 5% to 10% trillion parameters, which is like a 3x increase

251
00:14:44,610 --> 00:14:47,670
Ejaaz:
in the size that we already are seeing and using with the models today.

252
00:14:47,870 --> 00:14:49,630
Ejaaz:
It's going to be an absolute beast of a model.

253
00:14:50,030 --> 00:14:53,430
Ejaaz:
It parinates a cybersecurity risk, which is incredibly ironic because all of

254
00:14:53,430 --> 00:14:54,890
Ejaaz:
that droplet stuff is getting leaked right now.

255
00:14:55,030 --> 00:15:00,430
Ejaaz:
But also Claude Opus 4.7 and Sonnet 4.8. So we're going to get version upgrades

256
00:15:00,430 --> 00:15:02,510
Ejaaz:
of the existing models that we're having already.

257
00:15:02,770 --> 00:15:06,570
Ejaaz:
So my one question is, when are these models going to get released?

258
00:15:06,670 --> 00:15:08,010
Ejaaz:
Because I need to get my hands on them.

259
00:15:08,370 --> 00:15:12,110
Ejaaz:
Number two, will it cause my entire laptop to get hacked? I don't know.

260
00:15:12,250 --> 00:15:16,210
Ejaaz:
So there's like a reputation risk going on right now as well as I want to use the actual thing.

261
00:15:16,790 --> 00:15:20,870
Josh:
Well, you also mentioned the security part of this, and I think it's worth noting

262
00:15:20,870 --> 00:15:26,930
Josh:
that there has been an increased cadence in security issues recently and leaks and exploits and hacks.

263
00:15:27,330 --> 00:15:31,350
Josh:
And I know they happen all the time, but I can't, like, there is some sort of

264
00:15:31,350 --> 00:15:34,410
Josh:
correlation happening here between models getting smarter and exploits.

265
00:15:34,690 --> 00:15:37,670
Josh:
I mean, yeah, we have this post on screen here, which summarizes it in a great way.

266
00:15:37,770 --> 00:15:41,490
Josh:
It says, this week in security, there is, what is that, six different exploits

267
00:15:41,490 --> 00:15:44,250
Josh:
that happened and pretty serious ones too.

268
00:15:44,470 --> 00:15:47,870
Josh:
Axios, which is a npm supply chain hack

269
00:15:47,870 --> 00:15:51,570
Josh:
that affects like many many millions of projects and applications

270
00:15:51,570 --> 00:15:54,650
Josh:
and if you've ever provide code or anything chances are you use that dependency

271
00:15:54,650 --> 00:16:00,170
Josh:
um openai codex had a command injection via github branch there's a terabyte

272
00:16:00,170 --> 00:16:03,890
Josh:
data leak from mercore and this doesn't even include the leak from today which

273
00:16:03,890 --> 00:16:10,050
Josh:
is cloud code so there's this increasing cadence of leaks and exploits and you

274
00:16:10,050 --> 00:16:14,570
Josh:
gotta ask the question is like if anthropic internally is using these tools,

275
00:16:15,150 --> 00:16:19,670
Josh:
Who else has access to tools this powerful? What can they be used for?

276
00:16:19,930 --> 00:16:24,410
Josh:
Are they actually responsible for any of this? Or is this just a random correlation

277
00:16:24,410 --> 00:16:27,110
Josh:
that's happening? I don't know.

278
00:16:27,370 --> 00:16:32,910
Ejaaz:
I think my main concern is that malicious scenario that you described where

279
00:16:32,910 --> 00:16:37,090
Ejaaz:
people are accessing this tool but using it for bad purposes is already happening.

280
00:16:37,130 --> 00:16:39,170
Ejaaz:
It's coming in the form of prompt injections.

281
00:16:39,430 --> 00:16:42,990
Ejaaz:
Like, look, there are six hacks that happened this week alone,

282
00:16:42,990 --> 00:16:44,930
Ejaaz:
and it's only been like two to three days.

283
00:16:45,650 --> 00:16:50,330
Ejaaz:
I wonder if that increased cadence is based off of people being able to get

284
00:16:50,330 --> 00:16:55,190
Ejaaz:
access to intelligent AI models like this and finding flaws or bugs in open

285
00:16:55,190 --> 00:16:57,070
Ejaaz:
source code and being able to exploit them, right?

286
00:16:57,110 --> 00:17:00,030
Ejaaz:
You've got a bunch of people, millions of people every day logging on,

287
00:17:00,170 --> 00:17:02,650
Ejaaz:
Vibe coding apps who have never coded in their entire lives,

288
00:17:03,030 --> 00:17:05,650
Ejaaz:
me included, right? I don't know what's being installed on my laptop.

289
00:17:05,810 --> 00:17:10,510
Ejaaz:
I don't know what data is being leaked. So I could imagine that things like that is happening.

290
00:17:10,610 --> 00:17:14,830
Ejaaz:
But the question I have for you, Josh, is does this matter for Anthropik specifically.

291
00:17:14,970 --> 00:17:18,410
Ejaaz:
Is this a major blow for them? Do you think they lose valuation based off of this?

292
00:17:18,610 --> 00:17:21,090
Ejaaz:
Or do you think this gets solved in a version update?

293
00:17:21,630 --> 00:17:25,630
Josh:
Well, this is tough because this does sting, right? Like this is a massive IP

294
00:17:25,630 --> 00:17:29,010
Josh:
leak and this is a competitive advantage that they're now losing.

295
00:17:29,520 --> 00:17:33,600
Josh:
How much of a value loss is it? Probably not crazy high.

296
00:17:33,940 --> 00:17:38,500
Josh:
I mean, the magic is in the model. The magic is in the Claude model itself, those weights.

297
00:17:38,700 --> 00:17:41,320
Josh:
You can copy the CLI architecture, you can study the engineering,

298
00:17:41,500 --> 00:17:46,300
Josh:
but you can't actually replicate what Claude can do. So they still have this massive advantage.

299
00:17:46,480 --> 00:17:50,380
Josh:
And even though it's embarrassing, and even though it's a really strong leak

300
00:17:50,380 --> 00:17:53,180
Josh:
in which I am, if I'm one of these Chinese models right now,

301
00:17:53,340 --> 00:17:57,280
Josh:
I am forking this, cloning it. I'm dropping my intelligence in there.

302
00:17:57,660 --> 00:17:58,960
Ejaaz:
That's it. You don't need to distill it anymore.

303
00:17:58,960 --> 00:18:01,960
Josh:
Well, yeah, you could just, you just take the code base, you take the harness,

304
00:18:02,100 --> 00:18:06,100
Josh:
you put your model in and suddenly you have a cloud code software with your

305
00:18:06,100 --> 00:18:08,120
Josh:
own brain attached to it. And that's powerful.

306
00:18:08,300 --> 00:18:11,920
Josh:
So in that case, it hurts because now people know if there are any secrets in

307
00:18:11,920 --> 00:18:14,100
Josh:
how the software was run, how the architecture worked.

308
00:18:14,200 --> 00:18:18,540
Josh:
They now have that in full, clean, plain text, but it doesn't hurt them in the

309
00:18:18,540 --> 00:18:22,580
Josh:
sense that they aren't going to, they're going to lose customers over this.

310
00:18:22,580 --> 00:18:26,780
Josh:
Because the magic is in that proprietary software, those model weights, those are not leaked.

311
00:18:27,440 --> 00:18:31,100
Josh:
It's just the cloud code software. It's just that command line interface.

312
00:18:31,720 --> 00:18:36,100
Josh:
And aside from that, I think it's more interesting for the public just to kind

313
00:18:36,100 --> 00:18:39,620
Josh:
of get access to the roadmap and be able to play with the code themselves versus

314
00:18:39,620 --> 00:18:41,940
Josh:
actually damaging for the brand's

315
00:18:41,940 --> 00:18:45,740
Josh:
valuation. But certainly for the brand image, it's not a good look.

316
00:18:46,260 --> 00:18:49,100
Ejaaz:
Yeah i i agree with pretty much your entire

317
00:18:49,100 --> 00:18:52,160
Ejaaz:
take i'm thinking about the number of phds that

318
00:18:52,160 --> 00:18:55,260
Ejaaz:
anthropic has hired on the security ai team um

319
00:18:55,260 --> 00:18:59,340
Ejaaz:
i remember their release from i think it was about a month and a half ago and

320
00:18:59,340 --> 00:19:03,100
Ejaaz:
we said this on the previous episode where they had called opus 4.6 discover

321
00:19:03,100 --> 00:19:07,280
Ejaaz:
500 zero-day vulnerabilities so it was all looking really good i wish they had

322
00:19:07,280 --> 00:19:11,620
Ejaaz:
applied it to their own model and their own website and their own apis so it

323
00:19:11,620 --> 00:19:14,100
Ejaaz:
sucks that that's happened I do think they'll get over it,

324
00:19:14,380 --> 00:19:16,720
Ejaaz:
but they'll need to do some damage control at this point.

325
00:19:17,180 --> 00:19:22,840
Ejaaz:
The other major thing is like, reputationally, Anthropik has just come out of a pretty...

326
00:19:23,640 --> 00:19:27,700
Ejaaz:
A rocky couple of weeks, right? They had the whole blacklisting thing from the

327
00:19:27,700 --> 00:19:30,100
Ejaaz:
US government and the Pentagon, which I believe is still there.

328
00:19:30,440 --> 00:19:34,360
Ejaaz:
And so it's not a good look where their model, which was being used for military

329
00:19:34,360 --> 00:19:37,540
Ejaaz:
operations, is now getting leaked for other different purposes.

330
00:19:38,160 --> 00:19:42,160
Ejaaz:
That being said, I think they're going to get over it. I think this is amazing

331
00:19:42,160 --> 00:19:46,020
Ejaaz:
for us and for the open source community who now get access to the entire system

332
00:19:46,020 --> 00:19:51,100
Ejaaz:
prompt of Cold Code, its architecture design, and can plug in their own models for free.

333
00:19:51,300 --> 00:19:53,940
Ejaaz:
And yeah, now we have a better idea of Anthropic's product roadmap.

334
00:19:54,120 --> 00:19:55,940
Ejaaz:
I'm excited to see these 20 features launch soon.

335
00:19:56,320 --> 00:20:00,240
Josh:
Yeah, it's a big leak. I mean, I think it's fun for everyone who's an observer.

336
00:20:00,600 --> 00:20:04,700
Josh:
Thank you, Anthropic, for being more open source than ever. I hope that they're

337
00:20:04,700 --> 00:20:07,980
Josh:
able to start using this new copybara model to actually, you know,

338
00:20:08,080 --> 00:20:11,060
Josh:
check these publications, make sure this doesn't happen because it's amazing.

339
00:20:11,280 --> 00:20:13,140
Josh:
They have so much intelligence, but it's so spiky.

340
00:20:13,600 --> 00:20:17,860
Josh:
Clearly, an all-knowing AI applied to the entire stack would never have let

341
00:20:17,860 --> 00:20:20,380
Josh:
this slide, but clearly it's not applied everywhere.

342
00:20:20,720 --> 00:20:24,140
Josh:
It's also raising a lot of questions about well anthropic

343
00:20:24,140 --> 00:20:28,320
Josh:
is like the alignment team but now they are the ones who are going to determine

344
00:20:28,320 --> 00:20:32,340
Josh:
who gets the power of this new model and they're doing it in a very like private

345
00:20:32,340 --> 00:20:36,900
Josh:
closed way and they're using internally and it creates a lot of these interesting

346
00:20:36,900 --> 00:20:41,060
Josh:
problems to look out for but in terms of the leak today that's the news.

347
00:20:41,840 --> 00:20:45,460
Josh:
Big leak. I can't believe that actually happened. Like I woke up this morning

348
00:20:45,460 --> 00:20:48,360
Josh:
and I read the news and I was like, no, surely there must be wrong.

349
00:20:48,440 --> 00:20:51,240
Josh:
Like this is hyperbolic, but no, the entirety, it's all there.

350
00:20:51,340 --> 00:20:52,820
Josh:
You can go and read it. It's on GitHub.

351
00:20:53,020 --> 00:20:57,200
Josh:
And it's funny because they're actually actively trying to take down the repos that forked the code.

352
00:20:57,420 --> 00:21:01,040
Josh:
But some guy rewrote the entire thing in Python this morning because you could

353
00:21:01,040 --> 00:21:04,020
Josh:
just do that in a single prompt. And now you can't because the code is slightly different.

354
00:21:04,220 --> 00:21:10,760
Josh:
So it is interesting, noteworthy, crazy, scary, exciting. I'm stoked to get a buddy.

355
00:21:11,240 --> 00:21:13,560
Josh:
I think the prompt for today's comment section could be like,

356
00:21:13,620 --> 00:21:15,400
Josh:
hey, what feature are you most excited about?

357
00:21:15,560 --> 00:21:18,580
Josh:
For me, it's the buddies. I want a little pal that sits in my cloud coat all

358
00:21:18,580 --> 00:21:19,800
Josh:
the time that I could level up.

359
00:21:20,100 --> 00:21:22,920
Josh:
There's like a shiny feature. There's rarity. They're like trading cards.

360
00:21:23,060 --> 00:21:23,920
Josh:
I don't know. It could be cool.

361
00:21:24,320 --> 00:21:27,240
Josh:
I'm looking forward to it. But yeah, I think that's the leak today. That's the episode.

362
00:21:27,660 --> 00:21:31,520
Ejaaz:
Yeah, that's it. Thank you guys so much for listening. There are thousands and

363
00:21:31,520 --> 00:21:34,900
Ejaaz:
thousands of you over the last couple of months that have joined us in subscribing,

364
00:21:35,400 --> 00:21:36,080
Ejaaz:
turning on notifications.

365
00:21:36,160 --> 00:21:39,300
Ejaaz:
If you aren't one of those people that I just described, please do so.

366
00:21:39,400 --> 00:21:41,920
Ejaaz:
Wherever you're listening or watching us, Spotify, Apple Music,

367
00:21:42,100 --> 00:21:46,400
Ejaaz:
YouTube, it means the world and helps us out. A bunch of you subscribe and turn on notifications.

368
00:21:46,760 --> 00:21:52,760
Ejaaz:
We also have a newsletter going out twice a week to 150,000 people that read our stuff twice a week.

369
00:21:52,860 --> 00:21:56,600
Ejaaz:
We have a long form essay, which goes out, I believe today, as you're listening

370
00:21:56,600 --> 00:21:57,640
Ejaaz:
to this episode, go check it out.

371
00:21:57,740 --> 00:21:58,440
Josh:
Yeah, go write that right now.

372
00:21:58,920 --> 00:22:02,000
Ejaaz:
Exactly, yeah. Thank you, Pass Joss, for writing this right now.

373
00:22:02,880 --> 00:22:06,560
Ejaaz:
And we also have the five daily highlights or weekly highlights,

374
00:22:06,640 --> 00:22:12,100
Ejaaz:
rather, which will give you the top AI news and Frontier Tech News on Fridays.

375
00:22:12,520 --> 00:22:15,760
Ejaaz:
So sign up to both of those things and we will see you on the next one.