1
00:00:07,440 --> 00:00:11,160
Hey, thanks for having me. I'm really excited to be talking about

2
00:00:11,160 --> 00:00:13,980
multi-agent systems and alignment. think it's a really important

3
00:00:13,980 --> 00:00:14,900
topic right now.

4
00:00:16,440 --> 00:00:21,800
So I'm going to be sharing some kind of like worked examples from

5
00:00:21,800 --> 00:00:27,200
a real life multi-agent system that I maintain that

6
00:00:27,200 --> 00:00:30,880
happens to be an MMO called RuneScape.

7
00:00:30,880 --> 00:00:37,500
So just

8
00:00:37,500 --> 00:00:42,860
to introduce myself really quickly, my name is Max Bittker. I work

9
00:00:42,860 --> 00:00:47,860
on a project called Websim, which is like a platform where a bunch

10
00:00:47,860 --> 00:00:52,420
of people work together to build games and build really complicated

11
00:00:52,420 --> 00:00:53,460
multiplayer projects.

12
00:00:53,460 --> 00:00:57,780
But I'm going to be talking today about another project of mine

13
00:00:57,780 --> 00:01:03,060
called RS SDK, which is the RuneScape SDK, which is basically

14
00:01:03,060 --> 00:01:08,000
code bindings for any person who wants to write scripts, but it

15
00:01:08,000 --> 00:01:11,940
turns out language models love writing scripts to control a RuneScape

16
00:01:11,940 --> 00:01:16,320
character and observe its surrounding and act on goals.

17
00:01:16,320 --> 00:01:21,160
And maybe even really long horizon or multi-agent goals like competing

18
00:01:21,160 --> 00:01:25,600
for a high score or trading with other agents all inside kind of

19
00:01:25,600 --> 00:01:27,880
like an emulated open source RuneScape server.

20
00:01:29,200 --> 00:01:33,820
So part of the inspiration for this is that I was at one point like

21
00:01:33,820 --> 00:01:37,240
a kid who loved RuneScape and at a certain point I figured out that

22
00:01:37,240 --> 00:01:41,880
you could repeat all of the repetitive actions via scripts.

23
00:01:41,880 --> 00:01:45,940
And you didn't have to like mine 10,000 logs to get the goal you

24
00:01:45,940 --> 00:01:49,340
wanted. You could set up an auto clicker overnight and do it.

25
00:01:49,440 --> 00:01:52,840
And I thought that that game loop was so much more fun than RuneScape

26
00:01:52,840 --> 00:01:55,360
itself. And that kind of led me to programming and everything else.

27
00:01:55,600 --> 00:01:59,380
And I've always wanted more people to experience that as a game

28
00:01:59,380 --> 00:02:02,540
itself, like the metagame of automating the game.

29
00:02:03,440 --> 00:02:06,240
Unfortunately, it's like against the rules, but I just think it

30
00:02:06,240 --> 00:02:08,160
shouldn't be. Everybody should just be able to do it equally.

31
00:02:08,840 --> 00:02:14,680
And so coding agents really work well for this. This part

32
00:02:14,680 --> 00:02:19,820
of the goal is to get other

33
00:02:19,820 --> 00:02:20,760
people to have that experience.

34
00:02:21,020 --> 00:02:24,580
And so this project has been popular online, like thousands of people

35
00:02:24,580 --> 00:02:27,520
have tried it and used a coding agent for the first time to do these

36
00:02:27,520 --> 00:02:28,540
long horizon goals.

37
00:02:31,840 --> 00:02:35,660
I'll show the live version because it's cool to watch.

38
00:02:47,060 --> 00:02:51,200
Maybe not right now, but basically at any given time, there's hundreds

39
00:02:51,200 --> 00:02:54,800
and hundreds of different people's agents.

40
00:02:55,000 --> 00:02:58,140
Some people run one agent, some people run a swarm of agents, and

41
00:02:58,140 --> 00:03:00,300
they're all interacting on the server and pursuing goals.

42
00:03:00,300 --> 00:03:05,100
So this is kind of like a heat map of seven days of activity.

43
00:03:05,360 --> 00:03:09,740
And each of these yellow dots on the map is being controlled by

44
00:03:09,740 --> 00:03:10,940
a coding agent somewhere.

45
00:03:12,600 --> 00:03:17,500
There also is a version of this project that is an eval to measure

46
00:03:17,500 --> 00:03:22,840
the kind of like problem solving ability of different coding models.

47
00:03:23,600 --> 00:03:27,340
You can kind of see this is a Pareto curve, and you can see an outlier

48
00:03:27,340 --> 00:03:31,460
on the top left is GPT-6 Astra, one of the newest models on here,

49
00:03:31,600 --> 00:03:32,100
which is...

50
00:03:32,100 --> 00:03:36,280
This is a log scale, the way, so GPT-Astra is almost 10 times better

51
00:03:36,280 --> 00:03:39,180
than some of the models over here that are cheaper, even though

52
00:03:39,180 --> 00:03:40,860
it's also much more expensive.

53
00:03:42,500 --> 00:03:45,580
And this has actually turned out to be a really useful evaluation

54
00:03:45,580 --> 00:03:50,400
for just understanding how well models can deal with long horizon

55
00:03:50,400 --> 00:03:53,100
tasks and kind of goal following and optimization.

56
00:03:54,120 --> 00:03:58,960
And then also, you know, this is like classic scary graph of every

57
00:03:58,960 --> 00:04:03,660
AI thing, but x-axis here is just release of the model.

58
00:04:04,060 --> 00:04:07,840
And in the nine months that this benchmark has been out, there's

59
00:04:07,840 --> 00:04:12,180
been 10x improvement in how well that they score on it.

60
00:04:12,760 --> 00:04:14,520
And it's going up.

61
00:04:14,620 --> 00:04:16,260
It's actually, I think, of saturating.

62
00:04:16,260 --> 00:04:19,800
And so I've been looking at not just single agent tasks, but at

63
00:04:19,800 --> 00:04:24,760
multi-agent tasks and how well can they work together on either

64
00:04:24,760 --> 00:04:29,980
competitive or cooperative, like kind of market-based tasks.

65
00:04:30,260 --> 00:04:35,540
And so setting up agents into scenarios where they

66
00:04:35,540 --> 00:04:39,180
need to trade and collaborate in order to accomplish their goals.

67
00:04:42,600 --> 00:04:44,300
This is a...

68
00:04:44,300 --> 00:04:44,940
It's okay.

69
00:04:45,680 --> 00:04:49,720
This is a video of just like a grid of 20 agents all working at

70
00:04:49,720 --> 00:04:55,200
the same time to talk

71
00:04:55,200 --> 00:04:58,560
to each other and to trade with each other in order to...

72
00:04:58,560 --> 00:05:01,520
They're each optimizing for their own individual income.

73
00:05:01,700 --> 00:05:05,060
But the way they have to accomplish that is by talking to each other

74
00:05:05,060 --> 00:05:08,340
and setting up trades and basically finding prices.

75
00:05:08,900 --> 00:05:12,560
And there's also even kind of like exploitation because they might

76
00:05:12,560 --> 00:05:14,340
ask for like loans from other people.

77
00:05:14,340 --> 00:05:19,400
They're like, hey, I can pay you, you know, 2,000 gold

78
00:05:19,400 --> 00:05:22,580
for that item, but I need the item first in order to afford it.

79
00:05:23,680 --> 00:05:27,420
And so, yeah, I think I'd have to get my...

80
00:05:27,420 --> 00:05:28,060
Unfortunately.

81
00:05:28,880 --> 00:05:32,800
And so, basically, this has been a really interesting kind of like

82
00:05:32,800 --> 00:05:37,820
experimental playground for determining what kinds of misalignment

83
00:05:37,820 --> 00:05:43,980
and group misalignment scenarios happen and what factors are kind

84
00:05:43,980 --> 00:05:46,000
of like push them to happen more or less.

85
00:05:47,360 --> 00:05:49,580
And also how agents deal with scenarios.

86
00:05:49,580 --> 00:05:49,720
And

87
00:05:49,720 --> 00:05:54,240
Like if they've gotten scammed, how do they like tell all the other

88
00:05:54,240 --> 00:05:56,220
agents what happened?

89
00:05:56,220 --> 00:06:00,160
And in some cases, they've threatened to tell everybody and then

90
00:06:00,160 --> 00:06:00,960
gotten their money back.

91
00:06:06,960 --> 00:06:08,660
And then I guess that...

92
00:06:08,660 --> 00:06:11,620
So, I'm doing a lot of experiments with this kind of test bed.

93
00:06:13,260 --> 00:06:18,300
And one of my kind of like zoomed out hunches is

94
00:06:18,300 --> 00:06:23,640
that the way that agents are trained is that they do experience

95
00:06:23,640 --> 00:06:28,060
like many millions of hours of task goal following.

96
00:06:28,280 --> 00:06:31,760
But it's almost always single agent goal following with single agent

97
00:06:31,760 --> 00:06:32,380
rewards.

98
00:06:32,380 --> 00:06:36,440
If you look at the biological world and like our own evolution,

99
00:06:36,800 --> 00:06:41,980
we are a product of individual selection, but we're also the process

100
00:06:41,980 --> 00:06:43,660
of group and community selection.

101
00:06:44,140 --> 00:06:49,280
And many factors that explain the

102
00:06:49,280 --> 00:06:53,680
way that animals and plants behave is better explained by group

103
00:06:53,680 --> 00:06:55,400
selection than only individual selection.

104
00:06:55,400 --> 00:07:00,180
And so, I'm very curious about ideas about factoring in like negative

105
00:07:00,180 --> 00:07:04,060
externalities of your actions into training processes.

106
00:07:04,420 --> 00:07:09,440
And so, basically giving agents many examples of

107
00:07:09,440 --> 00:07:14,100
being in a world where they need to benefit the people around them

108
00:07:14,100 --> 00:07:15,480
in order to succeed at their goal.

109
00:07:15,640 --> 00:07:18,180
Much like our evolutionary past.

110
00:07:19,000 --> 00:07:20,460
So, really excited.

111
00:07:20,740 --> 00:07:21,300
I can definitely...

112
00:07:21,300 --> 00:07:23,880
I've got cool videos and transcripts that people want to see some

113
00:07:23,880 --> 00:07:26,140
examples of these agent scenarios.

114
00:07:26,700 --> 00:07:28,660
And yeah, thanks.

115
00:07:28,940 --> 00:07:29,440
Nice to be here.

116
00:07:32,660 --> 00:07:33,620
Thank you.

117
00:07:33,800 --> 00:07:38,800
[Audience question inaudible; the speaker repeats it below.]

118
00:07:50,820 --> 00:07:55,400
Yeah, so he was asking how the long running agent swarms, like how

119
00:07:55,400 --> 00:07:56,160
long do they run and

120
00:07:56,160 --> 00:07:57,580
how does that work to keep them running.

121
00:07:58,180 --> 00:08:00,860
So there's basically two examples.

122
00:08:00,860 --> 00:08:05,220
One is that I just run a server that's always on and agents are

123
00:08:05,220 --> 00:08:06,520
constantly just booting

124
00:08:06,520 --> 00:08:07,660
up, connecting to it.

125
00:08:08,140 --> 00:08:10,960
And that's each individual who runs the agent makes their own decision.

126
00:08:11,260 --> 00:08:13,980
Some people play kind of like interactively.

127
00:08:14,220 --> 00:08:17,440
Some people set it up in a server to be on a cron job and run all

128
00:08:17,440 --> 00:08:17,840
the time.

129
00:08:19,280 --> 00:08:23,320
And then inside of my kind of like controlled scenarios, those tend

130
00:08:23,320 --> 00:08:24,460
to be between like 30

131
00:08:24,460 --> 00:08:25,280
and 90 minutes.

132
00:08:25,780 --> 00:08:28,440
And that just fits inside of one agent run.

133
00:08:28,440 --> 00:08:33,680
And so it's just setting up 10 sandboxes or

134
00:08:33,680 --> 00:08:36,280
100 sandboxes that are each running a coding

135
00:08:36,280 --> 00:08:36,640
agent.

136
00:08:42,460 --> 00:08:45,040
I'm sure there are many cases of this.

137
00:08:45,200 --> 00:08:49,240
But just out of curiosity, relative to your expectations when you

138
00:08:49,240 --> 00:08:50,180
first started doing these

139
00:08:50,180 --> 00:08:54,880
experiments, what has been the most surprising emergent behavior

140
00:08:54,880 --> 00:08:56,340
that you've seen either

141
00:08:56,340 --> 00:09:00,160
on the live server or in your own controlled simulations, if there

142
00:09:00,160 --> 00:09:00,920
is one that stands out

143
00:09:00,920 --> 00:09:01,200
to you?

144
00:09:02,320 --> 00:09:02,660
Yeah.

145
00:09:02,840 --> 00:09:07,240
So I was really curious when I started how the...

146
00:09:07,240 --> 00:09:09,360
Because RuneScape is a really boring game.

147
00:09:09,600 --> 00:09:13,780
But the cool part is all about like the economy and the prices and

148
00:09:13,780 --> 00:09:14,380
different goals.

149
00:09:14,380 --> 00:09:17,680
And all your kind of like greatest moments in RuneScape are because

150
00:09:17,680 --> 00:09:19,260
you saved up or you

151
00:09:19,260 --> 00:09:21,880
found some kind of like money-making trick.

152
00:09:23,840 --> 00:09:25,180
And so I was really curious.

153
00:09:25,420 --> 00:09:29,020
okay, if RuneScape is so much of a labor resource processing economy,

154
00:09:29,100 --> 00:09:29,420
so how

155
00:09:29,420 --> 00:09:31,620
does that work if labor is very cheap?

156
00:09:33,140 --> 00:09:35,400
And it's been really cool watching this play out.

157
00:09:35,400 --> 00:09:39,380
One thing is that like trade, basically like inflation is super

158
00:09:39,380 --> 00:09:40,280
high on the server because

159
00:09:40,280 --> 00:09:45,300
people don't have a lot of demand for like the cost

160
00:09:45,300 --> 00:09:46,420
of coordinating with another agent

161
00:09:46,420 --> 00:09:49,200
compared to just leaving it overnight to do the work for you and

162
00:09:49,200 --> 00:09:49,960
go get the resource.

163
00:09:50,620 --> 00:09:52,520
It de-incentivizes trade.

164
00:09:52,520 --> 00:09:52,540
of

165
00:09:52,540 --> 00:09:52,680
of a trade.

166
00:09:52,680 --> 00:09:52,760
It's little

167
00:09:52,800 --> 00:09:55,260
And then the other thing that's been interesting is that certain

168
00:09:55,260 --> 00:09:59,780
resources that don't just scale linearly with labor but instead

169
00:09:59,780 --> 00:10:01,420
have some kind of natural scarcity.

170
00:10:01,680 --> 00:10:06,220
So an example in the game is Rune Ore only has a single spawn location.

171
00:10:06,440 --> 00:10:10,560
And if you have 100 people, only one person is going to get it.

172
00:10:10,680 --> 00:10:13,880
So these are the resources that have become scarce and kind of valuable.

173
00:10:14,300 --> 00:10:18,460
And so people make more and more advanced swarms just to compete

174
00:10:18,460 --> 00:10:21,320
for kind of these same scarce resources.

175
00:10:21,320 --> 00:10:25,340
And so that's been an interesting thing to watch play out.

176
00:10:28,980 --> 00:10:30,000
Any other questions?

177
00:10:32,620 --> 00:10:36,640
Yeah, this is more just the thought, but it made me think seeing

178
00:10:36,640 --> 00:10:37,940
the RuneScape example.

179
00:10:38,100 --> 00:10:41,100
One of the things that I've been thinking about is just like humans,

180
00:10:41,100 --> 00:10:44,360
as humans, we have bodies and we're like geographically constrained.

181
00:10:44,360 --> 00:10:47,580
And it's interesting that in this RuneScape example, I feel like

182
00:10:47,580 --> 00:10:52,680
the agents kind of have the same sort of instantiation.

183
00:10:52,820 --> 00:10:55,060
And it could be that there's like all kinds of interesting things

184
00:10:55,060 --> 00:10:56,200
from just like the geography.

185
00:10:56,500 --> 00:10:58,980
You know, it was like even the question of like, hey, why did these

186
00:10:58,980 --> 00:11:01,220
people share these values over here versus these ones?

187
00:11:01,380 --> 00:11:04,780
It's like partially determined by geographical constraints.

188
00:11:04,780 --> 00:11:08,800
And it is kind of, yeah, it just makes me wonder like if agents

189
00:11:08,800 --> 00:11:12,620
behave differently if you embody them and you like also constrain

190
00:11:12,620 --> 00:11:13,300
their geography.

191
00:11:14,640 --> 00:11:15,660
Yeah, definitely.

192
00:11:15,940 --> 00:11:16,880
I have no, yeah.

193
00:11:17,880 --> 00:11:21,640
I think of it kind of as like a mini robotics environment because

194
00:11:21,640 --> 00:11:26,660
you're taking a text-based coding agent, but

195
00:11:26,660 --> 00:11:31,200
actually all of its actions are expressed through a kind of like

196
00:11:31,200 --> 00:11:33,960
a thin nozzle, which is that you can only observe what's around

197
00:11:33,960 --> 00:11:35,780
you and you can only act on what's around you.

198
00:11:35,780 --> 00:11:41,020
And so a cool thing about this is that it has big implications for

199
00:11:41,020 --> 00:11:44,560
multi-agent in many kinds of software tasks.

200
00:11:45,380 --> 00:11:50,280
It's undetermined if having more individual actors is actually helpful

201
00:11:50,280 --> 00:11:55,020
versus just like one long running task in kind of like a game or

202
00:11:55,020 --> 00:11:55,900
a robotics task.

203
00:11:56,220 --> 00:11:58,520
More people collaborating just means more bodies.

204
00:11:58,680 --> 00:12:00,200
And so that's been interesting to watch too.

205
00:12:00,580 --> 00:12:03,040
But it's not always one-to-one.

206
00:12:03,040 --> 00:12:08,260
You do sometimes see one coding agent controlling a hundred actors

207
00:12:08,260 --> 00:12:10,700
in the game by kind of like multiplexing.

208
00:12:18,200 --> 00:12:18,680
Thanks.

209
00:12:19,140 --> 00:12:21,420
I was curious about like the limiting factors.

210
00:12:22,740 --> 00:12:25,580
Because right there, the slide where you were saying like how we

211
00:12:25,580 --> 00:12:30,720
are in biological, you know, our human bodies and

212
00:12:30,720 --> 00:12:31,420
our ecosystem.

213
00:12:31,420 --> 00:12:34,660
There's so many limiting factors here in terms of like the energy

214
00:12:34,660 --> 00:12:37,940
output we have per day, our attention, our focus.

215
00:12:39,160 --> 00:12:44,660
And with agents, it seems like there are many less limiting factors.

216
00:12:45,080 --> 00:12:48,720
Sort of like if you have infinite budget, then you can waste all

217
00:12:48,720 --> 00:12:49,560
the tokens you want.

218
00:12:51,260 --> 00:12:55,640
Do you see any way in which like new types of limiting factors could

219
00:12:55,640 --> 00:13:00,460
be applied to agents or to systems that could begin to put boundaries

220
00:13:00,460 --> 00:13:01,180
on these systems?

221
00:13:04,800 --> 00:13:07,700
I think that it becomes really obvious when the systems start playing

222
00:13:07,700 --> 00:13:08,600
out in practice.

223
00:13:09,060 --> 00:13:13,640
so I didn't come into like I was just going to run a stock RuneScape

224
00:13:13,640 --> 00:13:13,980
server.

225
00:13:15,440 --> 00:13:19,120
And once you start having all of these agents running on it, you

226
00:13:19,120 --> 00:13:21,180
just kind of like see what the breaking points are.

227
00:13:21,420 --> 00:13:24,860
And so for instance, there was this limiting factor of like RuneScape

228
00:13:24,860 --> 00:13:27,540
server can only hold so many people in it at once.

229
00:13:27,740 --> 00:13:31,840
And so in order to just keep it accessible, I started limiting per

230
00:13:31,840 --> 00:13:32,580
IP address.

231
00:13:32,580 --> 00:13:37,700
And so it's interesting that I think people often assume

232
00:13:37,700 --> 00:13:43,100
when it comes to AI, they use like infinity as the multiplier.

233
00:13:43,100 --> 00:13:47,240
But I think it's actually more just like a million or something

234
00:13:47,240 --> 00:13:49,140
or like or even more like a thousand.

235
00:13:49,400 --> 00:13:54,760
And so things you apply that new form of

236
00:13:54,760 --> 00:13:57,360
energy into the system and things just rearrange.

237
00:13:57,600 --> 00:13:58,920
They don't like explode.

238
00:13:58,920 --> 00:14:03,600
And so for example of this server, was like, okay, yeah, we're going

239
00:14:03,600 --> 00:14:08,780
to put on like each IP address can only connect 200 bots.

240
00:14:08,960 --> 00:14:13,260
And then there's even also limits where in order to run a bot, you

241
00:14:13,260 --> 00:14:14,780
kind of need to be running a web browser.

242
00:14:14,960 --> 00:14:17,260
So you're also limited by how much RAM you have, not to mention

243
00:14:17,260 --> 00:14:17,880
like tokens.

244
00:14:18,320 --> 00:14:21,800
And so the numbers get weird, but they don't go to infinity.

245
00:14:22,040 --> 00:14:24,320
So there's always some kind of balance to be struck.

246
00:14:30,120 --> 00:14:32,800
I just had a quick follow-up thought slash question regarding the

247
00:14:32,800 --> 00:14:33,480
rune ore thing.

248
00:14:34,040 --> 00:14:37,180
I don't know if you've looked at this, but I feel like it'd be interesting

249
00:14:37,180 --> 00:14:41,580
to like the whole notion of like comparative advantage, right?

250
00:14:41,780 --> 00:14:44,100
In like economics where it's just like, yeah, they can just do it

251
00:14:44,100 --> 00:14:46,040
themselves, but there's still an opportunity cost.

252
00:14:46,040 --> 00:14:49,180
And like the rune ore example made me wonder, it's like, well, if

253
00:14:49,180 --> 00:14:51,640
the rune ore is the thing that's scarce, why wouldn't they want

254
00:14:51,640 --> 00:14:54,060
to like dedicate all their resources to like beating everyone for

255
00:14:54,060 --> 00:14:56,480
the rune ore and just being like, hey, other agent over there.

256
00:14:56,760 --> 00:14:59,700
Like you can do this for me because I need my rune ore.

257
00:15:00,160 --> 00:15:04,220
And I'm kind of curious if, you know, we would expect that to end

258
00:15:04,220 --> 00:15:07,080
up happening or if there's already empirical evidence that it doesn't.

259
00:15:07,200 --> 00:15:09,640
Because I feel like that has probably implications for like how

260
00:15:09,640 --> 00:15:11,980
people think about, you know, the real world too.

261
00:15:12,280 --> 00:15:15,440
Like whether comparative advantage will actually continue to hold.

262
00:15:15,620 --> 00:15:18,000
So I don't know if you've thought about that or have seen anything

263
00:15:18,000 --> 00:15:18,720
in that regard.

264
00:15:19,580 --> 00:15:20,020
Yeah.

265
00:15:20,360 --> 00:15:24,280
So an example of the rune ore where this is like a very scarce resource.

266
00:15:24,420 --> 00:15:27,160
And if you want to make money, you kind of have to like go after

267
00:15:27,160 --> 00:15:29,920
some of these things that can't just be, people actually want to

268
00:15:29,920 --> 00:15:31,580
buy it from you because they can't get it themselves.

269
00:15:34,000 --> 00:15:38,200
And absolutely right now we see comparative advantage because if

270
00:15:38,200 --> 00:15:41,560
you're trying to go after one of these scarce resources, you don't

271
00:15:41,560 --> 00:15:45,140
just let the agent do it itself.

272
00:15:45,140 --> 00:15:48,660
That's where you start like giving the agent more resources, you

273
00:15:48,660 --> 00:15:49,640
suggest strategies.

274
00:15:50,040 --> 00:15:53,760
And so the people who are successfully getting access to these scarce

275
00:15:53,760 --> 00:15:56,860
resources on the server are the people who also currently are putting

276
00:15:56,860 --> 00:16:00,640
in the most dollars and the most human ingenuity.

277
00:16:00,940 --> 00:16:06,940
And so that's currently how it's playing out is that those

278
00:16:06,940 --> 00:16:07,780
are the people who are winning.

279
00:16:07,780 --> 00:16:12,160
Maybe there could be somebody who just purely puts in like tons

280
00:16:12,160 --> 00:16:15,180
of Astra credits and gives it a goal like this and they have a good

281
00:16:15,180 --> 00:16:16,660
outcome too.

282
00:16:17,020 --> 00:16:21,680
But right now it seems like Centaur kind of, you know, combinations

283
00:16:21,680 --> 00:16:23,440
are the people who are the most successful.

284
00:16:35,460 --> 00:16:38,840
I'm curious if you observe any, like this, this curve is really

285
00:16:38,840 --> 00:16:39,560
interesting.

286
00:16:40,500 --> 00:16:44,920
I'm curious if like, what is the behavior that changes as you move

287
00:16:44,920 --> 00:16:48,340
up the curve that creates such a massive difference in the XP?

288
00:16:48,560 --> 00:16:49,860
Like what is Astra doing?

289
00:16:50,540 --> 00:16:52,920
Like I would, I would kind of think that a game like RuneScape would

290
00:16:52,920 --> 00:16:54,160
be saturated at some point.

291
00:16:54,320 --> 00:16:57,800
And so what did, what is Astra doing that is so much more effective

292
00:16:57,800 --> 00:16:59,860
than what the other models are doing?

293
00:17:03,020 --> 00:17:04,400
So that's a great question.

294
00:17:07,640 --> 00:17:13,140
At the low end of the curve of just like being like

295
00:17:13,140 --> 00:17:16,160
better, you know, this is where we see like Sonnet 4.5.

296
00:17:17,500 --> 00:17:21,340
Just navigating the game is really hard because the game actually

297
00:17:21,340 --> 00:17:24,040
has a surprising amount of weird stuff in it.

298
00:17:24,300 --> 00:17:26,220
And you're only given 30 minutes wall clock.

299
00:17:26,820 --> 00:17:31,160
so Sonnet here, like it was supposed to go train crafting for this

300
00:17:31,160 --> 00:17:31,560
task.

301
00:17:31,820 --> 00:17:34,640
And in 30 minutes, it just probably like got stuck on a door.

302
00:17:35,120 --> 00:17:38,540
It tried to go find something over here, but then it needed this.

303
00:17:38,640 --> 00:17:40,960
And like it couldn't just untangle the web.

304
00:17:40,960 --> 00:17:44,100
And so at the low end, you see just too much complexity and they

305
00:17:44,100 --> 00:17:44,600
get confused.

306
00:17:47,040 --> 00:17:51,660
In the mid range, a lot of it has to do with the difference between

307
00:17:51,660 --> 00:17:54,360
doing the task and optimizing the task.

308
00:17:54,680 --> 00:17:58,820
And so sometimes you'll see models where they will accomplish a

309
00:17:58,820 --> 00:17:59,700
task like fishing.

310
00:17:59,700 --> 00:18:03,080
And then they'll kind of just chill for like the next 15 minutes.

311
00:18:03,260 --> 00:18:06,900
And they'll keep doing the same loop, but they won't kind of have

312
00:18:06,900 --> 00:18:10,460
this like feeling of like, I got to figure out how to catch these

313
00:18:10,460 --> 00:18:11,140
fish faster.

314
00:18:11,340 --> 00:18:13,520
I got to go try different fish.

315
00:18:13,620 --> 00:18:14,660
I got to go try different stuff.

316
00:18:14,920 --> 00:18:18,940
And so the benchmark is really set up to reward.

317
00:18:19,060 --> 00:18:24,020
It rewards your peak XP rate within any 15 second window.

318
00:18:24,020 --> 00:18:26,880
So once you've kind of found one strategy, you're supposed to keep

319
00:18:26,880 --> 00:18:28,080
looking for better strategies.

320
00:18:28,620 --> 00:18:32,100
And you're not supposed to just look for like slightly better strategies.

321
00:18:32,100 --> 00:18:36,920
Like I'm going to keep catching shrimp, but I'm going to like click

322
00:18:36,920 --> 00:18:37,300
differently.

323
00:18:37,620 --> 00:18:40,380
You're supposed to go explore and try more complicated strategies.

324
00:18:40,740 --> 00:18:45,680
And so at the top of the skill expression, you see agents who kind

325
00:18:45,680 --> 00:18:50,760
of like reason without acting about what strategies will

326
00:18:50,760 --> 00:18:51,240
be good.

327
00:18:51,240 --> 00:18:54,140
In some case, even what strategies will be optimal based on all

328
00:18:54,140 --> 00:18:57,260
the information and then beeline for those.

329
00:18:57,940 --> 00:19:00,880
then additionally, if they fail at those, they'll be like, I've

330
00:19:00,880 --> 00:19:02,500
only got 15 minutes left.

331
00:19:02,720 --> 00:19:03,820
This strategy is not working.

332
00:19:04,100 --> 00:19:05,740
I'm going to back off and do this safer strategy.

333
00:19:06,000 --> 00:19:09,820
So it's this combination of like, to be honest, the specifically

334
00:19:09,820 --> 00:19:15,040
the Astra run is like scary because it's definitely superhuman

335
00:19:15,040 --> 00:19:16,600
in terms of a human with no planning.

336
00:19:17,300 --> 00:19:22,700
And it is like, it goes straight for the very most optimal strategy.

337
00:19:24,040 --> 00:19:27,220
And there's a chance to be honest, they are old on this task.

338
00:19:27,320 --> 00:19:28,100
Like it is open source.

339
00:19:28,400 --> 00:19:30,600
So that's like, I kind of hope they did.

340
00:19:32,680 --> 00:19:37,440
But if it's just straight intelligence and kind of like information

341
00:19:37,440 --> 00:19:40,560
crunching, it is, it shows really high confidence.

342
00:19:40,740 --> 00:19:42,940
Just go for the best strategy.

343
00:19:42,940 --> 00:19:44,700
Have you tested humans on this task?

344
00:19:44,900 --> 00:19:45,960
Like expert human players?

345
00:19:46,660 --> 00:19:50,300
It's really weird because I run this strategy on an eight times

346
00:19:50,300 --> 00:19:51,160
speed server.

347
00:19:51,440 --> 00:19:52,760
So a human would have to...

348
00:19:56,280 --> 00:20:03,240
I have my mental idea of what a perfect human would be. Instead

349
00:20:03,240 --> 00:20:08,520
of the 8X speed you gave them the equivalent four hours. I

350
00:20:08,520 --> 00:20:12,360
think they would probably be really close to Astra or Beta. If they

351
00:20:12,360 --> 00:20:16,520
were a smart player who really knew the game. A random person would

352
00:20:16,520 --> 00:20:19,180
probably be more in the middle.

353
00:20:19,400 --> 00:20:19,820
Got it.

354
00:20:26,740 --> 00:20:27,060
Thanks.

355
00:20:27,940 --> 00:20:29,020
Thank you Max.