1
00:00:00,000 --> 00:00:12,800
Trust is a function. It is more than just a feeling. In a world of infinite noise,

2
00:00:12,800 --> 00:00:20,400
the only signal that matters is the one you verify. We are here to map the graph,

3
00:00:20,400 --> 00:00:26,000
order the protocol, and build the immune system for the decentralized web.

4
00:00:26,000 --> 00:00:29,760
Welcome to the signal processing layer.

5
00:00:30,780 --> 00:00:33,720
This is Say What?

6
00:00:40,860 --> 00:00:44,780
Welcome back to the fourth episode of Say What?

7
00:00:44,780 --> 00:00:52,180
Today we're decoding the critical intersection of artificial intelligence, privacy, and the web of trust.

8
00:00:52,180 --> 00:00:59,400
My guest is Mark Suman, the CEO of Maple AI and a former Apple engineer who left the

9
00:00:59,400 --> 00:01:03,580
walled garden to build verifiable privacy-first AI.

10
00:01:03,960 --> 00:01:09,640
We're going to explore how open source models operating inside secure enclaves can protect

11
00:01:09,640 --> 00:01:11,000
our digital sovereignty.

12
00:01:11,420 --> 00:01:18,420
We'll also examine why we need robust reputation systems as AI agents begin transacting on

13
00:01:18,420 --> 00:01:19,060
our behalf.

14
00:01:19,880 --> 00:01:23,940
Remember that this frequency is powered entirely by value for value.

15
00:01:24,540 --> 00:01:28,840
If this transmission gives you a new perspective, consider sending it back.

16
00:01:29,720 --> 00:01:36,160
Boost, zaps, and streaming sats are the only fuel we use to keep the signal clear.

17
00:01:36,720 --> 00:01:37,880
Let's get into it.

18
00:01:38,560 --> 00:01:40,320
Mark, welcome to Say What.

19
00:01:41,240 --> 00:01:44,260
Hey, thanks for having me on. I appreciate being here.

20
00:01:45,460 --> 00:01:46,840
Yeah, absolutely, Mark.

21
00:01:46,840 --> 00:01:52,800
So Mark, you are the CEO of Maple AI, as I mentioned.

22
00:01:53,120 --> 00:01:59,060
For folks who might be unfamiliar, can you give us a little bit of a background on Maple AI?

23
00:02:00,840 --> 00:02:01,780
Sure, yeah.

24
00:02:02,060 --> 00:02:03,040
Where to begin?

25
00:02:03,300 --> 00:02:11,860
So Maple is a mainstream AI, but one that was built with privacy first, rather than trying to bolt it on afterwards.

26
00:02:11,860 --> 00:02:17,940
So if you want to imagine what it's like to use Cloud or ChatGPT, we have a very similar product.

27
00:02:18,360 --> 00:02:22,560
But instead, every single user starts off with their own private encryption key.

28
00:02:23,100 --> 00:02:27,580
And all of the data is encrypted locally on your device first before it is sent to the cloud.

29
00:02:27,580 --> 00:02:41,540
And then once it's in the cloud, we use confidential computing or secure enclaves to decrypt the data, have it talk to the AI, to the LLM, the open source LLM that's running in the cloud.

30
00:02:41,860 --> 00:02:45,780
And then when it comes up with a response, it re-encrypts it and sends it back to your device.

31
00:02:46,420 --> 00:02:51,400
So this way, we are not in the middle of your AI communications.

32
00:02:52,000 --> 00:02:55,460
We simply are able to relay that information.

33
00:02:56,040 --> 00:02:59,340
We're able to synchronize that information to all your devices.

34
00:02:59,340 --> 00:03:04,920
But the only data that we actually can see ourselves as a company is the encrypted data.

35
00:03:05,480 --> 00:03:10,260
Everything remains personal between you and the open source LLM.

36
00:03:10,260 --> 00:03:15,320
And again, for folks who might be unfamiliar

37
00:03:15,320 --> 00:03:17,940
with the exact mechanics of this,

38
00:03:19,620 --> 00:03:23,600
and everyone uses an LLM these days,

39
00:03:23,860 --> 00:03:26,380
Chagp, T, Claude, Gemini,

40
00:03:27,000 --> 00:03:30,920
and they just talk to it and it does things for them,

41
00:03:31,160 --> 00:03:32,660
researches things for them, and so on.

42
00:03:33,300 --> 00:03:36,720
What problem are you solving with Maple AI

43
00:03:36,720 --> 00:03:43,120
that an average user of LLMs should be paying attention to?

44
00:03:44,180 --> 00:03:47,700
Well, we solve problems for two camps of people.

45
00:03:47,860 --> 00:03:50,860
We solve problems for people who already use AI and LLMs,

46
00:03:51,020 --> 00:03:53,080
but then we also solve problems for people who don't.

47
00:03:53,520 --> 00:03:56,960
There's actually a large group of people who simply don't use AI.

48
00:03:57,500 --> 00:03:59,500
There was a big survey conducted last summer,

49
00:04:00,040 --> 00:04:02,600
and it was focused on the US,

50
00:04:02,700 --> 00:04:04,300
but then they extrapolated it worldwide.

51
00:04:04,300 --> 00:04:10,280
and it said that 39% of adults in America do not use AI at all,

52
00:04:11,300 --> 00:04:12,440
actively use it, right?

53
00:04:12,500 --> 00:04:15,340
Everybody uses it passively through different phone systems they call into,

54
00:04:15,480 --> 00:04:18,100
but active use was 39% of adults.

55
00:04:18,100 --> 00:04:24,440
And of those 39%, 71% of them cited data privacy concerns.

56
00:04:24,680 --> 00:04:28,180
They don't want to give away their personal information to these AI machines.

57
00:04:28,860 --> 00:04:32,260
So we've built definitely for that group of people.

58
00:04:32,260 --> 00:04:38,940
We are onboarding new people to AI now who are finally using AI because they did not trust Google.

59
00:04:39,140 --> 00:04:41,000
They did not trust OpenAI and Thrompic.

60
00:04:41,400 --> 00:04:45,900
And now we've given them an option that they can verify the privacy guarantees we have.

61
00:04:45,980 --> 00:04:46,820
So now they trust it.

62
00:04:47,360 --> 00:04:52,660
But then for those, the people that you're bringing up here, those who already use AI, you know, what's in it for them?

63
00:04:52,720 --> 00:04:53,820
Why use something like Maple?

64
00:04:53,820 --> 00:05:03,180
the key differentiator is that you are able to when you know that you enter a room let's say that

65
00:05:03,180 --> 00:05:07,820
you walk into a room with a person and there's there's 100 people in that room and you're trying

66
00:05:07,820 --> 00:05:12,840
to have a private conversation with that one individual you're you're aware of all the other

67
00:05:12,840 --> 00:05:17,560
100 people that can hear what you're talking about and so if you have something truly confidential

68
00:05:17,560 --> 00:05:23,100
that you want to talk about it could be a legal matter it could be a health related matter it could

69
00:05:23,100 --> 00:05:28,080
be an embarrassing story, or it could simply just be a joke or something that you don't want other

70
00:05:28,080 --> 00:05:32,100
people to hear, you and that person are going to walk into another room and close a door.

71
00:05:32,600 --> 00:05:37,360
And now you have this private space to talk and you can speak more freely. And that's really what

72
00:05:37,360 --> 00:05:43,700
we're offering to people. You know, I make use of many AI tools. I really I use I use a quad

73
00:05:43,700 --> 00:05:50,380
anthropic tools. I use chat to be to use Gemini for certain things. However, I also have Maple

74
00:05:50,380 --> 00:05:54,700
and what our users tell us time and time again is that when they use Maple,

75
00:05:55,080 --> 00:05:56,980
they're able to finally be themselves.

76
00:05:56,980 --> 00:05:58,200
They're able to speak freely.

77
00:05:58,740 --> 00:06:02,220
They feel kind of like this weight has been lifted off of them

78
00:06:02,220 --> 00:06:06,080
and they realize that they were censoring themselves on the other platforms

79
00:06:06,080 --> 00:06:09,000
because they knew that someone was listening.

80
00:06:09,540 --> 00:06:11,600
They knew that whatever they said there was being recorded,

81
00:06:12,100 --> 00:06:14,560
was being potentially used to train future models,

82
00:06:14,860 --> 00:06:20,000
was becoming available to future government surveillance

83
00:06:20,000 --> 00:06:23,960
or becoming available to legal proceedings that could be subpoenaed

84
00:06:23,960 --> 00:06:25,960
and data could be gathered from these companies.

85
00:06:27,120 --> 00:06:32,400
So people love using Maple because it allows them to really just be themselves

86
00:06:32,400 --> 00:06:33,840
while talking to the AI.

87
00:06:36,300 --> 00:06:42,380
You know, you brought up the analogy of going into a separate room

88
00:06:42,380 --> 00:06:43,740
and talking to that person.

89
00:06:43,740 --> 00:06:48,740
I think one of the issues

90
00:06:48,740 --> 00:06:51,920
in general with the LLM space

91
00:06:51,920 --> 00:06:54,020
is that person

92
00:06:54,020 --> 00:06:58,500
maybe you've blocked out the other 100 people who can hear you

93
00:06:58,500 --> 00:07:00,160
but how can you trust that person

94
00:07:00,160 --> 00:07:02,920
and in this case that person is some kind of big tech company

95
00:07:02,920 --> 00:07:06,500
so yes you're using a secure enclave

96
00:07:06,500 --> 00:07:10,160
your information is encrypted

97
00:07:10,160 --> 00:07:12,860
before it goes to the LLM

98
00:07:12,860 --> 00:07:15,820
and then after it, the LLM generates a response.

99
00:07:16,000 --> 00:07:19,820
But there is that period in which it is decrypted,

100
00:07:19,960 --> 00:07:23,940
the LLM processes it, creates the answer and sends it back.

101
00:07:24,740 --> 00:07:28,320
And that LLM might not be sitting in the best actors' hands,

102
00:07:28,500 --> 00:07:30,080
namely these big tech companies.

103
00:07:30,080 --> 00:07:34,580
So it doesn't fully solve that issue, correct?

104
00:07:35,780 --> 00:07:36,420
Sort of.

105
00:07:36,580 --> 00:07:40,140
So there are two different people operating services

106
00:07:40,140 --> 00:07:42,460
or two different kinds of services you can use.

107
00:07:42,860 --> 00:07:59,620
One of them are categorized as AI proxies. These are very similar to VPNs where you want to use Chat2PT, but you don't want to go sign up for your own account and give it a credit card and give your home address because then that tightly links your identity.

108
00:07:59,620 --> 00:08:06,300
You might have to upload a U.S. passport, your government passport or your driver's license or some kind of identifier.

109
00:08:07,420 --> 00:08:09,720
People don't want to get that close with their AI data.

110
00:08:10,020 --> 00:08:16,620
And so they'll use one of these proxies, which allows a thousand people to go through the same account and then hit ChantyPT.

111
00:08:17,440 --> 00:08:22,680
You get a little bit of safety there because your identity is no longer tied to the query that you're making.

112
00:08:23,100 --> 00:08:28,740
However, that content that you are typing into the LLM is still getting in the hands of that big tech company.

113
00:08:28,740 --> 00:08:35,680
And so if you say anything in there that could potentially be sensitive to you, it is now part of their database.

114
00:08:36,380 --> 00:08:41,320
And they're using AI on their sides to build biographies on every single user.

115
00:08:41,960 --> 00:08:50,080
And they are able, on the other end, to re-piece back together things that you said across multiple chats, even if you're going through a proxy.

116
00:08:50,600 --> 00:08:58,020
We've seen papers published on this already that they have the ability to kind of sew those back together if they want to with high confidence.

117
00:08:58,020 --> 00:09:20,140
What we offer with Maple is something totally different. So we take these open source models that have been published by ChatGPT, by DeepSeq or Quinn, Kimmy, other big ones, and then we host them in confidential computing servers that do not, they're not controlled by these other groups.

118
00:09:20,140 --> 00:09:23,780
So none of our data goes back to ChatGPT or goes back to DeepSeq.

119
00:09:24,260 --> 00:09:34,360
It's really similar to as if you went in on your laptop and you downloaded one of these open source LLMs onto your laptop and talked to it there with the Internet turned off.

120
00:09:35,020 --> 00:09:40,920
However, people don't tend to have enough power on their laptop to run the largest open source models.

121
00:09:41,060 --> 00:09:49,480
If you really want to get something that is comparable to the power of Cloud, you need to run it on a server, a cloud server, something of that size.

122
00:09:49,480 --> 00:09:57,440
And so we have taken what we think is the closest thing to the privacy you get of chatting on your laptop with a local LLM.

123
00:09:57,860 --> 00:10:04,240
And we've moved that into the cloud by using these secure enclaves, this confidential computing architecture.

124
00:10:05,140 --> 00:10:13,640
Yeah, and that makes complete sense if you're using some of these open source models that you can isolate and contain in these secure enclaves.

125
00:10:13,640 --> 00:10:16,460
and on your point of those

126
00:10:16,460 --> 00:10:19,320
multi-tenant API keys

127
00:10:19,320 --> 00:10:19,920
if you will

128
00:10:19,920 --> 00:10:22,140
there's obviously Open Router

129
00:10:22,140 --> 00:10:23,900
which I have to say Open Router

130
00:10:23,900 --> 00:10:25,940
is terribly named

131
00:10:25,940 --> 00:10:27,800
because they do full KYC

132
00:10:27,800 --> 00:10:29,680
I've used them

133
00:10:29,680 --> 00:10:32,440
but there are proxies that sit even in front

134
00:10:32,440 --> 00:10:34,020
of Open Router like Pay Per Q

135
00:10:34,020 --> 00:10:36,460
or Routester in our world

136
00:10:36,460 --> 00:10:38,020
in the Bitcoiner world

137
00:10:38,020 --> 00:10:40,220
which will certainly mitigate that

138
00:10:40,220 --> 00:10:42,260
financial identification

139
00:10:42,260 --> 00:10:45,500
And then I guess if you're using something like Maple AI behind that,

140
00:10:45,580 --> 00:10:48,640
then it is truly fully sovereign at that point.

141
00:10:48,880 --> 00:10:53,380
I guess the only thing more sovereign would be you buying some kind of Mac Pro

142
00:10:53,380 --> 00:10:59,960
and downloading the model and running it on that, on your local machine.

143
00:11:01,120 --> 00:11:05,960
Yeah, and I certainly won't fault people who want to have a completely sovereign stack at home.

144
00:11:06,780 --> 00:11:09,100
It's commendable to try and do that.

145
00:11:09,200 --> 00:11:11,540
You can get a Mac Studio, you could max it out.

146
00:11:11,540 --> 00:11:18,620
You don't even have to go to Apple hardware. You could buy just a plain box and shove it full of NVIDIA cards if you want to.

147
00:11:20,000 --> 00:11:30,620
The problem right now is that in order to run the biggest open source models, these trillion parameter models, you need to spend upwards of 15, $20,000 to get that machine.

148
00:11:31,400 --> 00:11:40,280
And then you also don't know if that machine is going to be capable of running the next large model that gets dropped or if you're going to need more cards or something else.

149
00:11:40,280 --> 00:11:45,620
so it's it's a cat and mouse game where if people want to run things locally they'll have to keep up

150
00:11:45,620 --> 00:11:53,820
with that you brought up ppq and roster i've i've met the people running those and i think what

151
00:11:53,820 --> 00:12:00,300
they're doing is great and um so i'm it's i think they're fine services to use uh it's it's just

152
00:12:00,300 --> 00:12:07,380
people people need to remember that they're still sending their data um over to chat to t and claude

153
00:12:08,100 --> 00:12:11,700
And so Maple has taken a totally different approach.

154
00:12:12,080 --> 00:12:16,980
And I love that we have this open market with multiple solutions that people can choose.

155
00:12:16,980 --> 00:12:25,840
And people can understand what is their risk profile that they have and allow them to pick the service that's right for the situation.

156
00:12:27,140 --> 00:12:32,440
And so, Mark, I have to confess, I've only used the free version of Maple.

157
00:12:32,780 --> 00:12:35,240
I've not paid for Maple just yet.

158
00:12:35,240 --> 00:12:38,660
So I noticed that when I use it,

159
00:12:38,820 --> 00:12:40,120
there are three models available.

160
00:12:40,780 --> 00:12:41,840
Actually two, right?

161
00:12:41,900 --> 00:12:44,180
It's the LAMA and OpenAI,

162
00:12:44,420 --> 00:12:47,880
the GPT, the open source one from OpenAI.

163
00:12:48,220 --> 00:12:49,740
For the free users, yeah, that's correct.

164
00:12:50,320 --> 00:12:50,980
For the free users.

165
00:12:52,500 --> 00:12:54,360
For the paid plan,

166
00:12:54,720 --> 00:12:56,260
I'm actually looking at it right now,

167
00:12:56,720 --> 00:13:00,900
I see a few that have the lock next to that,

168
00:13:00,900 --> 00:13:02,100
meaning I have to upgrade.

169
00:13:02,100 --> 00:13:04,120
is Gemma, DeepSeek R1,

170
00:13:04,280 --> 00:13:05,300
Quinn, and Kimmy.

171
00:13:05,780 --> 00:13:07,400
Are there others that get unlocked

172
00:13:07,400 --> 00:13:09,020
as you go up the tiers?

173
00:13:10,300 --> 00:13:11,280
We really,

174
00:13:11,600 --> 00:13:14,020
we simply just hear it

175
00:13:14,020 --> 00:13:15,920
by paid and free.

176
00:13:16,480 --> 00:13:17,880
So if you become a paid user,

177
00:13:18,080 --> 00:13:19,960
you unlock all of the

178
00:13:19,960 --> 00:13:21,880
powerful models that we have.

179
00:13:22,020 --> 00:13:23,300
And we're constantly watching

180
00:13:23,300 --> 00:13:24,960
the landscape of open source models

181
00:13:24,960 --> 00:13:25,580
that are available

182
00:13:25,580 --> 00:13:27,300
and looking for what is

183
00:13:27,300 --> 00:13:28,520
the next best one.

184
00:13:28,940 --> 00:13:29,940
And we test it out.

185
00:13:30,140 --> 00:13:31,440
And then when we feel like

186
00:13:31,440 --> 00:13:36,200
gets ready and high quality enough, then we will put it into Maple.

187
00:13:36,200 --> 00:13:40,800
So right now, Kimi K2.5 is the state of the art.

188
00:13:40,800 --> 00:13:44,680
We are looking at GLM-5, which is a really capable model,

189
00:13:44,680 --> 00:13:48,040
and trying to find a way to get that up and running inside of Maple.

190
00:13:48,040 --> 00:14:00,030
But for now it Kimi K2 We still have DeepSeek R1 We have you brought up Gemma We have Quinn 3 Vision So you can do vision capabilities with that

191
00:14:00,190 --> 00:14:01,770
Both Kimmy and Quinn are great at vision.

192
00:14:02,410 --> 00:14:03,670
So you can do image analysis.

193
00:14:04,110 --> 00:14:05,670
And we have document uploads.

194
00:14:05,770 --> 00:14:08,850
You can upload a PDF of some legal contract that you have.

195
00:14:09,390 --> 00:14:11,730
And you can chat with Maple about it and ask.

196
00:14:11,730 --> 00:14:16,110
You can ask all the questions, all the legal questions you want to to Maple.

197
00:14:16,110 --> 00:14:20,390
that you can find where the legal boundaries are in something.

198
00:14:20,510 --> 00:14:22,350
And if you accidentally step over a boundary,

199
00:14:23,710 --> 00:14:27,130
the AI will tell you, hey, that's actually not legal to do.

200
00:14:27,830 --> 00:14:31,490
And because it's Maple, the conversation kind of ends there.

201
00:14:31,490 --> 00:14:33,110
Whereas if you were doing that in ChatGPT,

202
00:14:33,310 --> 00:14:34,850
you risk having your account suspended.

203
00:14:35,590 --> 00:14:39,070
Or even worse, you risk having your account silently flagged.

204
00:14:39,850 --> 00:14:43,230
And then in the future, if there ever were some legal dispute

205
00:14:43,230 --> 00:14:44,650
over what you were discussing,

206
00:14:44,650 --> 00:14:51,070
they can go the courts can go and subpoena your chat logs with chat gpt and say ah this person

207
00:14:51,070 --> 00:14:54,450
was in here trying to get around the law look at all these questions that they were asking

208
00:14:54,450 --> 00:15:02,250
so we we like to provide an environment where people can feel free to ask all the quote dumb

209
00:15:02,250 --> 00:15:06,530
questions that they want to not because they're trying to do something nefarious but simply because

210
00:15:06,530 --> 00:15:10,490
they don't know and they don't know sometimes where the boundaries are and they don't know

211
00:15:10,490 --> 00:15:11,950
if the stove is hot.

212
00:15:12,130 --> 00:15:13,950
And so they have to touch the stove really quick

213
00:15:13,950 --> 00:15:14,870
to feel that it's hot.

214
00:15:15,330 --> 00:15:17,810
And you shouldn't be punished for that.

215
00:15:18,290 --> 00:15:20,710
You should simply be able to ask the questions

216
00:15:20,710 --> 00:15:23,130
and get the answers and inform yourself.

217
00:15:23,670 --> 00:15:25,690
So that's the type of environment

218
00:15:25,690 --> 00:15:27,610
that we are trying to create for people.

219
00:15:29,810 --> 00:15:32,170
So on the open source models themselves,

220
00:15:32,270 --> 00:15:34,110
I think there was a brief window,

221
00:15:34,350 --> 00:15:36,270
at least for me, Mark,

222
00:15:36,310 --> 00:15:37,610
and I'm fairly plugged into this,

223
00:15:37,610 --> 00:15:40,690
been plugged into this world for a few years now.

224
00:15:42,910 --> 00:15:45,290
I think it was around January of last year,

225
00:15:45,370 --> 00:15:46,450
a little over a year ago,

226
00:15:46,670 --> 00:15:49,050
when DeepSeek R1 first came out.

227
00:15:49,370 --> 00:15:52,050
And it kind of caught everyone off guard.

228
00:15:52,830 --> 00:15:57,050
And there was a brief, but I dare say fleeting hope,

229
00:15:57,670 --> 00:16:02,470
that maybe open source is actually ready to catch up

230
00:16:02,470 --> 00:16:04,210
with some of these proprietary models.

231
00:16:04,210 --> 00:16:07,050
because it was pretty down close to,

232
00:16:07,330 --> 00:16:11,790
I think ChatGPT, OpenAI had 01 back then.

233
00:16:12,890 --> 00:16:15,250
But then shortly after they released 03,

234
00:16:15,630 --> 00:16:17,310
Claw 3.7 came out,

235
00:16:17,790 --> 00:16:20,530
and Gemini 2.5 came out,

236
00:16:21,090 --> 00:16:21,830
all of that stuff.

237
00:16:21,970 --> 00:16:23,590
And then the gaps seemed to widen again.

238
00:16:23,590 --> 00:16:25,330
And it appears to me,

239
00:16:25,410 --> 00:16:27,150
and I'm curious to hear your perspective, Mark,

240
00:16:27,610 --> 00:16:30,170
that open source models, even the best ones,

241
00:16:30,230 --> 00:16:33,130
are at least about six months behind

242
00:16:33,130 --> 00:16:37,130
some of the top-of-the-line proprietary ones?

243
00:16:38,290 --> 00:16:41,310
You know, so a lot of it comes down to,

244
00:16:41,930 --> 00:16:42,870
for lack of a better word,

245
00:16:42,950 --> 00:16:44,670
it comes down to the vibe that you get

246
00:16:44,670 --> 00:16:45,430
when you're using it.

247
00:16:46,190 --> 00:16:48,330
If you use it and you feel like you're getting

248
00:16:48,330 --> 00:16:50,190
the results that you want,

249
00:16:50,710 --> 00:16:53,930
then a lot of times open source is already there for you.

250
00:16:54,490 --> 00:16:57,550
Now, when we try to quantify how close they are,

251
00:16:57,930 --> 00:16:59,930
we use different benchmarks, right?

252
00:16:59,990 --> 00:17:01,490
Like humanities, last exam,

253
00:17:01,490 --> 00:17:03,950
We have the software engineering benchmarks that people use.

254
00:17:04,330 --> 00:17:06,590
There's a small portfolio of benchmarks.

255
00:17:07,350 --> 00:17:23,690
And in the last three months, I would say since November, when Kimi K2 thinking was dropped on the scene, this is the one right before 2.5, their benchmarks were right up there with Sonnet, with Quad Sonnet 4.5.

256
00:17:23,690 --> 00:17:31,110
and they were even beating uh chat gpt5 gpt5 and some of these other ones that were considered

257
00:17:31,110 --> 00:17:36,610
pretty state-of-the-art at the time and then they were just barely behind something like opus 4.5

258
00:17:36,610 --> 00:17:43,870
now obviously opus 4.6 is out and it's really running the world and the the new kimi k2.5

259
00:17:43,870 --> 00:17:49,330
the glm5 these other ones that are coming out they're scoring incredibly well on the benchmarks

260
00:17:49,330 --> 00:17:52,690
and are, I don't know that I would say six months.

261
00:17:52,790 --> 00:17:57,150
It feels like the gap has closed to a matter of weeks sometimes

262
00:17:57,150 --> 00:17:58,630
that they are now behind.

263
00:17:59,270 --> 00:18:04,130
And we're currently all sitting around waiting for the new DeepSeek to come out

264
00:18:04,130 --> 00:18:06,150
because there's been a lot of rumors around that.

265
00:18:06,570 --> 00:18:10,410
And we think that it's potentially even a leapfrog moment.

266
00:18:10,410 --> 00:18:14,870
I think we are going to see in 2026 a moment where open source,

267
00:18:14,870 --> 00:18:17,750
an open source LLM hits the market

268
00:18:17,750 --> 00:18:20,790
that is actually better than one of the proprietary ones

269
00:18:20,790 --> 00:18:21,650
when it comes out.

270
00:18:22,110 --> 00:18:25,490
And I think that will be a landmark thing to happen.

271
00:18:28,190 --> 00:18:32,090
So you mentioned benchmarks, Mark,

272
00:18:32,150 --> 00:18:33,730
and I actually want to talk about that.

273
00:18:33,810 --> 00:18:35,230
So again, from my perspective,

274
00:18:35,830 --> 00:18:38,210
there was an eight-month period

275
00:18:38,210 --> 00:18:41,670
starting around September, October of 2024

276
00:18:41,670 --> 00:18:43,890
when the first reasoning model came out.

277
00:18:43,890 --> 00:18:44,590
I think it was 01.

278
00:18:44,870 --> 00:18:49,230
from OpenAI to about May or June of last year,

279
00:18:49,510 --> 00:18:55,970
which was a breathtakingly fast-moving time in AI.

280
00:18:56,170 --> 00:18:58,230
I feel like we've actually slowed down a little bit

281
00:18:58,230 --> 00:19:02,050
compared to that frenetic pace we were in.

282
00:19:02,290 --> 00:19:04,710
And back then, the thing that was happening is,

283
00:19:04,950 --> 00:19:05,910
you mentioned benchmarks, right?

284
00:19:05,930 --> 00:19:07,950
These are different sets of tests

285
00:19:07,950 --> 00:19:13,570
that AI models are expected to perform against

286
00:19:13,570 --> 00:19:15,370
and then their score is measured, right?

287
00:19:15,550 --> 00:19:19,050
These benchmarks were getting saturated on a weekly basis, right?

288
00:19:19,110 --> 00:19:20,050
And when I say saturated,

289
00:19:20,210 --> 00:19:25,130
that the models were hitting basically 99% or 100% performance

290
00:19:25,130 --> 00:19:26,310
on these benchmarks,

291
00:19:26,410 --> 00:19:28,370
and they had to keep updating them

292
00:19:28,370 --> 00:19:30,630
with newer benchmarks over and over again.

293
00:19:30,850 --> 00:19:34,790
The one thing that somehow seemed to survive to an extent

294
00:19:34,790 --> 00:19:38,430
is the one you mentioned, which is humanity's last exam,

295
00:19:39,010 --> 00:19:40,910
which, as the name suggests,

296
00:19:40,910 --> 00:19:44,250
is intended to be the hardest test

297
00:19:44,250 --> 00:19:47,350
that certainly a human being can take.

298
00:19:47,610 --> 00:19:50,950
And I think it's something like a few hundred disciplines

299
00:19:50,950 --> 00:19:56,650
across the arts, art history, philosophy, mathematics,

300
00:19:56,890 --> 00:19:59,030
computer science, all of them, physics, chemistry.

301
00:19:59,610 --> 00:20:02,930
You had extreme PhD level questions

302
00:20:02,930 --> 00:20:05,370
drafted from the top universities, right?

303
00:20:05,810 --> 00:20:08,750
So all of those questions across all of those disciplines

304
00:20:08,750 --> 00:20:11,010
make up humanity's last exam.

305
00:20:11,350 --> 00:20:15,810
I haven't looked at the latest recently,

306
00:20:15,890 --> 00:20:19,950
but I don't think any model has breached 50% on that.

307
00:20:20,050 --> 00:20:20,610
Is that correct?

308
00:20:20,730 --> 00:20:22,270
Or have you something changed?

309
00:20:22,990 --> 00:20:25,910
Yeah, so K2.5 is claiming that they have.

310
00:20:26,090 --> 00:20:27,470
They're at 50.2.

311
00:20:27,550 --> 00:20:30,290
I'm on the Kimmy website right now

312
00:20:30,290 --> 00:20:33,310
looking at their benchmarks that they posted for 2.5.

313
00:20:33,810 --> 00:20:35,890
So yeah, they say that they're at 50.2.

314
00:20:35,890 --> 00:20:52,550
Now, the reason why I first led with my answer by talking about like the vibe that you get and your own internal testing of these agents is that are these models is that really it's a lot of these models can simply overfit for the benchmark.

315
00:20:53,410 --> 00:20:58,810
And that's why a lot of them have just kind of blown up to 99% accuracy a lot because they do that.

316
00:20:58,870 --> 00:21:03,030
But then when you use them in the real world, maybe they kind of fall over with real world tasks.

317
00:21:03,730 --> 00:21:05,970
So these benchmarks are great,

318
00:21:06,050 --> 00:21:08,430
but also sometimes we have to like check them

319
00:21:08,430 --> 00:21:11,090
against the real world performance that we get from them.

320
00:21:11,290 --> 00:21:14,150
But yes, on HLE, I'm looking at it right now.

321
00:21:14,690 --> 00:21:17,650
And in this chart, they have Opus 4.5.

322
00:21:17,850 --> 00:21:19,810
It was at 43.2%.

323
00:21:19,810 --> 00:21:22,610
Gemini 3 Pro is at 45.8.

324
00:21:23,110 --> 00:21:25,890
GPT 5.2 is at 45.5.

325
00:21:25,890 --> 00:21:29,570
And then Kimi K2.5 is at 50.2%.

326
00:21:29,570 --> 00:21:32,530
Interesting.

327
00:21:32,530 --> 00:21:35,510
So do you think that the,

328
00:21:36,010 --> 00:21:38,830
actually, what are your thoughts on humanities last exam?

329
00:21:39,570 --> 00:21:40,830
I'll give you my view.

330
00:21:41,530 --> 00:21:43,370
I think that's probably the closest one.

331
00:21:43,770 --> 00:21:48,790
You're right that we need to somehow tie it to practical implications.

332
00:21:49,490 --> 00:21:55,650
But absent that, I think just given such a broad-based test

333
00:21:55,650 --> 00:21:57,230
across so many disciplines

334
00:21:57,230 --> 00:22:01,630
is probably one of the better indicators towards our,

335
00:22:01,630 --> 00:22:04,970
march towards AGI, if that ever comes.

336
00:22:06,090 --> 00:22:09,430
And so, yeah, curious to hear your thoughts on HLE.

337
00:22:09,750 --> 00:22:12,030
And if you think there's a number,

338
00:22:12,210 --> 00:22:14,510
is that 80% or 90% on HLE,

339
00:22:15,290 --> 00:22:20,190
that in your mind says, okay, this is basically AGI.

340
00:22:21,630 --> 00:22:25,070
You know, AGI, that's an interesting topic.

341
00:22:26,070 --> 00:22:30,030
And I don't have like a super strong opinion of,

342
00:22:30,030 --> 00:22:32,750
or prediction, I should say, of when we're going to get there.

343
00:22:33,430 --> 00:22:38,590
But the way that I view all this is these models are really,

344
00:22:38,850 --> 00:22:41,150
they're really kind of the brain.

345
00:22:41,330 --> 00:22:43,210
They're like the engine running something.

346
00:22:43,770 --> 00:22:46,610
And when users are utilizing AI, they're not actually,

347
00:22:47,310 --> 00:22:50,770
very rarely are they talking to a raw model, right?

348
00:22:50,830 --> 00:22:52,530
They're usually using some kind of app

349
00:22:52,530 --> 00:22:54,630
and they want the app to do something for them.

350
00:22:54,630 --> 00:22:55,950
So they have some information.

351
00:22:55,950 --> 00:23:03,390
They want something to transform that information into something new, a better output, a more valuable output than what they put in.

352
00:23:04,150 --> 00:23:14,310
And I think that as each of these models get better, that people will start paying less attention to the actual model that they're using.

353
00:23:14,550 --> 00:23:22,070
And they really are going to be focusing more on the user experience of the app that they're using and the output that they're getting from it.

354
00:23:22,070 --> 00:23:35,030
It reminds me of when we used to focus so much on the CPU speed within our computers and 200 megahertz and then bumping up to 300 megahertz was like such a massive jump.

355
00:23:35,470 --> 00:23:38,210
And you can do so much more on your computer by having faster processor.

356
00:23:38,630 --> 00:23:41,610
But there became this point where it didn't really matter that much.

357
00:23:42,210 --> 00:23:43,690
Most of it was good enough.

358
00:23:43,690 --> 00:23:48,630
And we really focused on user experience itself and what are the things that I can do on this computer.

359
00:23:48,630 --> 00:24:14,790
And so as far as like AGI is concerned, that is more of a user experience and a tooling thing where we're going to have to build something on the layer on top of these models that leverages this powerful brain to to give it what it needs in order to be general, be able to kind of think for itself and recursively learn more and and become more self-aware.

360
00:24:14,790 --> 00:24:19,130
so I don't know that I have like a specific number

361
00:24:19,130 --> 00:24:21,290
that I would assign like 80% or something

362
00:24:21,290 --> 00:24:25,710
yeah I just haven't gotten to the weeds enough

363
00:24:25,710 --> 00:24:28,830
on humanity's last exam to like be able to pinpoint

364
00:24:28,830 --> 00:24:30,290
where I would put that number

365
00:24:30,290 --> 00:24:35,430
so Mark let's actually take a step back

366
00:24:35,430 --> 00:24:39,290
to your pre-Maple AI days

367
00:24:39,290 --> 00:24:41,210
you were at Apple for a while

368
00:24:41,210 --> 00:24:47,190
before leaving, presumably, to focus on Bitcoin and FreedomTech, right?

369
00:24:47,230 --> 00:24:49,070
And then getting to Maple.

370
00:24:49,070 --> 00:24:55,070
What was the catalyst that pushed you out of big tech and into building Maple

371
00:24:55,670 --> 00:24:57,570
and getting into the Bitcoin world even?

372
00:24:58,750 --> 00:25:06,250
Yeah, so I've been in the tech sector for, gosh, like 20 years now, professionally.

373
00:25:06,790 --> 00:25:09,330
And I worked at multiple startups.

374
00:25:09,330 --> 00:25:13,330
A couple of them were big cloud-based startups.

375
00:25:14,350 --> 00:25:21,650
And I could see how the data was stored in the cloud and what we as employees had access to.

376
00:25:22,050 --> 00:25:30,630
And a lot of times, me as an elevated employee, I was a software engineer, I would be brought into support calls or I'd be brought into a sales call.

377
00:25:31,310 --> 00:25:38,850
And I would be able to go in the database and see highly personal and sensitive information that the user had uploaded to our servers.

378
00:25:38,850 --> 00:25:44,310
and the user didn't know, they just didn't realize that I would be able to see that kind of stuff.

379
00:25:44,730 --> 00:25:48,750
Now, nothing bad happened from it, but it really opened up my eyes that the cloud is truly just

380
00:25:48,750 --> 00:25:52,930
someone else's computer. And you're copying your data over to their computer.

381
00:25:53,730 --> 00:25:57,530
So then fast forward to when I was at Apple, I was a software engineer over there.

382
00:25:57,730 --> 00:26:01,050
And we were building a lot of AI, machine learning technology.

383
00:26:01,970 --> 00:26:05,430
Some of my stuff that I did was customer facing.

384
00:26:05,430 --> 00:26:09,670
You know, some of my code is running on people's phones around the world.

385
00:26:10,250 --> 00:26:18,390
But then a lot of what I did was for internal projects that were kind of a hybrid customer facing, but also internal tool.

386
00:26:19,830 --> 00:26:24,710
And Apple is a company that does take privacy very seriously.

387
00:26:25,350 --> 00:26:32,730
And so when we would start a new project, we would be paired up with a privacy, internal privacy advocate and a privacy lawyer.

388
00:26:32,730 --> 00:26:35,790
and we would have to go through all the data that we're collecting.

389
00:26:36,750 --> 00:26:40,830
And we had to kind of justify why we needed certain data,

390
00:26:40,890 --> 00:26:42,490
how long we would keep it and all that.

391
00:26:44,230 --> 00:26:46,810
Very, very, very interesting place to work.

392
00:26:47,430 --> 00:26:48,910
Really enjoyed my time there.

393
00:26:48,910 --> 00:26:54,850
But it got to the point where I really wanted to work in Bitcoin

394
00:26:54,850 --> 00:26:57,050
and specifically focus on Bitcoin privacy

395
00:26:57,050 --> 00:27:01,430
and trying to find ways to make Bitcoin a better freedom technology.

396
00:27:02,250 --> 00:27:04,890
More accessible freedom technology, let's put it that way.

397
00:27:05,850 --> 00:27:09,710
But right before I left, that's when the big ChatGPT moment happened.

398
00:27:10,230 --> 00:27:13,710
And internally, we were all told, you're not allowed to use ChatGPT at work.

399
00:27:13,970 --> 00:27:15,410
Don't put any work information in there.

400
00:27:16,170 --> 00:27:18,830
And that immediately was like, duh, yes.

401
00:27:19,050 --> 00:27:20,930
I remember all my cloud stuff that I did.

402
00:27:20,930 --> 00:27:25,470
If I type something in and I put Apple proprietary information into here,

403
00:27:25,730 --> 00:27:29,350
that is just copying it over to ChatGPT servers and their employees can read it.

404
00:27:29,350 --> 00:27:32,490
So that really like resonated with me right away.

405
00:27:32,910 --> 00:27:38,410
So I ended up leaving Apple to go pursue something in the Bitcoin space.

406
00:27:38,630 --> 00:28:01,800
I teamed up with the guys that were doing Mutiny Wallet and we were going to work together to scale Mutiny Wallet much larger than it was at the time For multiple reasons that have been kind of discussed publicly already we decided it was best to wind down the Mutiny Wallet and try and look at what else could we build in the Freedom Tech space

407
00:28:02,160 --> 00:28:08,120
And we had a lot of meetings with developers at the time talking about what are some of the needs you have and here are some ideas that we have.

408
00:28:08,120 --> 00:28:15,020
And the thing that kept coming up time and time again is that developers wanted some kind of privacy around Chajapiti.

409
00:28:15,380 --> 00:28:17,980
They didn't like that they were giving away all their data to other companies.

410
00:28:18,420 --> 00:28:22,100
So that's when, you know, all these pieces kind of connected together.

411
00:28:22,460 --> 00:28:27,820
My co-founder, Anthony, was really going into a lot of confidential computing technology at the time.

412
00:28:28,420 --> 00:28:33,960
And we listened to these developers and we said, you know, this really could be a compelling product.

413
00:28:33,960 --> 00:28:49,140
And that's where the concept of Maple AI was born was we need to build something that is the freedom tech alternative to the new Internet that's being born, which is this AI Internet, this agentic Internet.

414
00:28:49,560 --> 00:28:55,300
But these other companies are following the same patterns all the other big tech companies have followed.

415
00:28:55,760 --> 00:29:00,020
And that is gather as much user data as you can and monetize it.

416
00:29:00,020 --> 00:29:07,820
whereas we wanted to follow the original internet business model which was create a service for your

417
00:29:07,820 --> 00:29:13,440
users and have them pay you for the service you deliver the service to them and that's that's

418
00:29:13,440 --> 00:29:19,520
where the relationship should end stop trying to monetize your user data stop trying to capture as

419
00:29:19,520 --> 00:29:24,980
much as you can because that just makes vulnerabilities for your users it also turns

420
00:29:24,980 --> 00:29:29,100
your user into your product and we just we didn't want to operate a business that way so

421
00:29:29,100 --> 00:29:33,080
So it's kind of a long-winded, you know, that's the whole journey of how we got here.

422
00:29:33,240 --> 00:29:39,840
But yeah, that's really where my Apple days and my cloud computing days merged together with Anthony

423
00:29:39,840 --> 00:29:41,700
and all the stuff that he had been doing on Mutiny Wallet.

424
00:29:41,760 --> 00:29:45,340
That's where we really teamed up to kind of give birth to Maple AI.

425
00:29:46,720 --> 00:29:47,720
That's interesting, Mark.

426
00:29:47,760 --> 00:29:51,780
I actually did not know that you were part of the Mutiny team.

427
00:29:52,640 --> 00:29:53,340
That's cool.

428
00:29:53,460 --> 00:29:57,380
Yeah, and real shame that Mutiny had to wind down.

429
00:29:57,380 --> 00:30:02,620
so you said you left Apple around the chat GPT moments

430
00:30:02,620 --> 00:30:05,480
I'm guessing that's sometime early 2022 correct?

431
00:30:06,900 --> 00:30:09,260
I guess there were multiple chat GPT moments

432
00:30:09,260 --> 00:30:12,780
but really like when chat GPT 3 came out

433
00:30:12,780 --> 00:30:16,100
is when it's I think it hit more mainstream

434
00:30:16,100 --> 00:30:19,420
so I left in 2024

435
00:30:19,420 --> 00:30:22,760
2023 is when it became like a big deal within Apple

436
00:30:22,760 --> 00:30:24,680
to talk about it and figure out

437
00:30:24,680 --> 00:30:26,480
how are we going to safeguard ourselves

438
00:30:26,480 --> 00:30:31,340
against our employees trying to use this third-party tool

439
00:30:31,340 --> 00:30:34,780
that can become a big attack vector for Apple.

440
00:30:35,040 --> 00:30:38,740
And Apple is notoriously secretive of the things that they're working on.

441
00:30:40,120 --> 00:30:43,000
I've worked at government defense contractors before,

442
00:30:43,360 --> 00:30:46,780
having to have clearance and be background checked and stuff.

443
00:30:47,040 --> 00:30:48,880
And then I've worked at Apple for many years.

444
00:30:48,880 --> 00:30:53,940
And the level of secrecy and clearance type infrastructure that's at Apple,

445
00:30:53,940 --> 00:30:58,240
it rivals government contractors in some ways.

446
00:30:58,420 --> 00:30:59,880
So it's very fascinating.

447
00:31:00,940 --> 00:31:04,840
And is that the reason why the sort of focus on privacy,

448
00:31:05,920 --> 00:31:12,040
is that why Apple hasn't made too many inroads into AI on its own?

449
00:31:14,080 --> 00:31:16,620
Well, I'm not there anymore.

450
00:31:17,260 --> 00:31:21,760
And, you know, I share what I can share of information that's public.

451
00:31:21,760 --> 00:31:26,100
but I would say privacy is definitely one reason.

452
00:31:26,660 --> 00:31:29,440
One big reason why Apple is not as far along

453
00:31:29,440 --> 00:31:31,640
in the AI space as everybody else

454
00:31:31,640 --> 00:31:34,260
because they do care very much

455
00:31:34,260 --> 00:31:36,600
about trying to run things on device where they can

456
00:31:36,600 --> 00:31:37,760
and if they're going to run in the cloud,

457
00:31:37,860 --> 00:31:38,680
then they try to make it private.

458
00:31:39,460 --> 00:31:43,800
And Apple's cloud infrastructure for AI,

459
00:31:44,080 --> 00:31:45,640
they've been very public about this,

460
00:31:46,000 --> 00:31:47,760
is running confidential computing.

461
00:31:48,020 --> 00:31:49,240
It's secure enclaves.

462
00:31:49,240 --> 00:31:51,400
It's called the Apple Private Cloud Compute.

463
00:31:51,760 --> 00:31:56,360
And it's actually something that kind of helped us get more interested in confidential computing.

464
00:31:57,000 --> 00:32:03,720
So we've built Maple's back end to be similar in spirit to the Apple private cloud compute.

465
00:32:04,400 --> 00:32:07,800
So I think it's just a matter of time before Apple catches up.

466
00:32:07,800 --> 00:32:17,920
I have read plenty of articles online that say Apple's problems with AI also come down to personnel and internal disputes and turf wars.

467
00:32:18,280 --> 00:32:20,700
And maybe there's some management issues going on.

468
00:32:20,700 --> 00:32:24,680
I can't speak to those personally things that I've seen, but those are articles that I've read.

469
00:32:24,860 --> 00:32:31,180
So it could be kind of both of these things kind of happening at once that make it difficult for Apple to jump into the game.

470
00:32:31,760 --> 00:32:40,340
Well, they have announced that Siri in the not too distant future, might be this year actually at some point, will be powered by Gemini, if I'm not mistaken.

471
00:32:40,640 --> 00:32:41,600
So that will be interesting.

472
00:32:43,480 --> 00:32:46,580
Yeah, I mean, Gemini is made by Google.

473
00:32:46,580 --> 00:32:52,620
And if you think about Google and Apple, in many ways, they are enemies with each other.

474
00:32:52,800 --> 00:32:55,380
In other ways, they work together.

475
00:32:56,160 --> 00:33:03,000
And a lot of people, a lot of users who have chosen Apple, chose Apple over a Google phone for certain reasons.

476
00:33:03,200 --> 00:33:10,500
So it's going to be interesting to see in the court of public opinion how that decision plays out once they actually launch it.

477
00:33:10,500 --> 00:33:16,440
You know, back when iPhones used the Intel chip,

478
00:33:16,680 --> 00:33:19,620
I think Apple's secure element was Intel's SGX,

479
00:33:19,680 --> 00:33:20,900
which has its own issues.

480
00:33:21,840 --> 00:33:24,320
Now, presumably with Apple Silicon,

481
00:33:24,900 --> 00:33:27,500
the iCloud keychain and all of these other things

482
00:33:27,500 --> 00:33:29,480
are in a secure element.

483
00:33:29,480 --> 00:33:36,320
So presumably the AI model will run in that type of environment.

484
00:33:37,680 --> 00:33:39,320
I don't know.

485
00:33:40,500 --> 00:33:41,860
what they'll do on the edge hardware.

486
00:33:43,120 --> 00:33:45,400
But I mean, yeah,

487
00:33:45,400 --> 00:33:47,600
they do have their own secure elements that they do.

488
00:33:47,960 --> 00:33:51,600
And so there's only so much I can say,

489
00:33:51,680 --> 00:33:54,680
but yeah, I presume that they will do as much as I can

490
00:33:54,680 --> 00:33:57,980
to try and secure things locally as well as in the cloud.

491
00:34:00,080 --> 00:34:00,440
Yeah.

492
00:34:01,240 --> 00:34:01,960
Yeah, sorry.

493
00:34:02,120 --> 00:34:02,680
I can't say more.

494
00:34:03,160 --> 00:34:03,600
No, no, no.

495
00:34:03,820 --> 00:34:05,060
And I don't mean to put you,

496
00:34:05,180 --> 00:34:08,780
I'm actually not sure what the edges of what you can disclose.

497
00:34:08,780 --> 00:34:10,880
Yeah, that's fine. You're good.

498
00:34:14,560 --> 00:34:19,800
So, Mark, before we dive into agents and reputation and so on,

499
00:34:19,980 --> 00:34:26,700
I'm curious if you noticed this, because I certainly felt this,

500
00:34:26,780 --> 00:34:30,960
and I'm going to go back to that time last year when I felt that things were accelerated.

501
00:34:31,380 --> 00:34:34,720
The acceleration was absolutely palpable.

502
00:34:34,720 --> 00:34:49,300
It's not as much now for me anyway, but back in that six to eight month period between October 2024 and May, June of 2025, it was sort of blow the skin off your face off type acceleration.

503
00:34:50,340 --> 00:34:57,380
And the deep sea carbon was one, you know, like what is going on kind of moment.

504
00:34:57,980 --> 00:35:04,060
But I think the real big one was a cascade of a couple of events or a handful of events.

505
00:35:04,060 --> 00:35:07,340
One was, I think, around February of 2024.

506
00:35:08,640 --> 00:35:12,160
OpenAI released Deep Research quietly, right?

507
00:35:12,160 --> 00:35:12,920
It just came out.

508
00:35:13,540 --> 00:35:15,280
And they started seeing these videos

509
00:35:15,280 --> 00:35:18,100
of these 60-page research reports

510
00:35:18,100 --> 00:35:19,760
that are produced in half an hour.

511
00:35:20,540 --> 00:35:23,000
And a lot of people are saying, what is this?

512
00:35:23,120 --> 00:35:24,620
Like, how is this even possible?

513
00:35:25,260 --> 00:35:26,420
Like, what is going...

514
00:35:26,420 --> 00:35:29,020
And it was like, true to its name, it was Deep Research.

515
00:35:29,120 --> 00:35:31,320
And they said, oh, yeah, it's this thing called,

516
00:35:31,380 --> 00:35:32,720
it's powered by O3.

517
00:35:32,720 --> 00:35:34,180
It's our O3 model.

518
00:35:34,640 --> 00:35:37,960
And we said, what is this O3 model?

519
00:35:38,240 --> 00:35:40,780
And then shortly after that, they released O3 Mini,

520
00:35:41,060 --> 00:35:42,900
which is the distilled model of O3.

521
00:35:42,960 --> 00:35:44,240
And that was fine, right?

522
00:35:44,260 --> 00:35:45,360
It was an improvement on O1.

523
00:35:45,480 --> 00:35:46,540
It wasn't mind-blowing.

524
00:35:47,220 --> 00:35:48,980
And then at some point, I think it was April,

525
00:35:49,820 --> 00:35:52,580
they released the actual O3 to the public.

526
00:35:53,460 --> 00:35:58,280
And to me, yes, you know, Opus 4.6 is great.

527
00:35:58,360 --> 00:35:59,700
It's a little scary at times.

528
00:35:59,700 --> 00:36:06,940
but that initially using that 03 raw right the full-fledged non-distilled model when it came out

529
00:36:06,940 --> 00:36:15,540
was the first first time i felt scared using ai i said oh wow i can i can kind of see where this is

530
00:36:15,540 --> 00:36:21,900
going and i don't know if i should be happy or scared uh there were others who had who shared

531
00:36:21,900 --> 00:36:26,960
similar sentiments i think even jack dorsey at the time when 03 came out he posted on nostril

532
00:36:26,960 --> 00:36:33,600
saying, okay, this is our first glimpse into AGI. And then something happened. About a month or two

533
00:36:33,600 --> 00:36:39,660
later, O3 stopped being as good. It stopped being that scary good. Maybe you could attribute that to

534
00:36:39,660 --> 00:36:47,320
people getting used to it. You know, just there's nothing new under the sun. You know, yes, it's

535
00:36:47,320 --> 00:36:53,500
producing fantastic results. Within a month or two, you're like, okay, what's next? I don't think

536
00:36:53,500 --> 00:36:55,440
it was that. I feel like

537
00:36:55,440 --> 00:36:57,500
the raw O3

538
00:36:57,500 --> 00:36:59,320
consumed so

539
00:36:59,320 --> 00:37:00,760
much resources

540
00:37:00,760 --> 00:37:03,500
that OpenAI wasn't able

541
00:37:03,500 --> 00:37:05,440
to sustain it. This is just speculation

542
00:37:05,440 --> 00:37:06,340
on my part.

543
00:37:07,120 --> 00:37:09,420
I think they had to

544
00:37:09,420 --> 00:37:13,000
tune it down somehow.

545
00:37:13,280 --> 00:37:14,780
Either reduce the number of parameters

546
00:37:14,780 --> 00:37:17,180
or whatever it is.

547
00:37:17,260 --> 00:37:19,460
Reduce the amount of compute that went into

548
00:37:19,460 --> 00:37:21,640
it. And by the time

549
00:37:21,640 --> 00:37:24,720
May or so, maybe June rolled around.

550
00:37:24,960 --> 00:37:25,720
It was fine.

551
00:37:25,840 --> 00:37:28,700
It was like chat GPT 5.2 right now.

552
00:37:29,140 --> 00:37:33,520
And ever since then, the progress has been incremental at best.

553
00:37:34,560 --> 00:37:37,660
And so I guess, well, let me start with the first question.

554
00:37:37,760 --> 00:37:41,460
Did you get that sense in that time period as well,

555
00:37:41,460 --> 00:37:45,820
that O3 was like the first scary moment?

556
00:37:46,040 --> 00:37:50,160
And then did you sense that it went down in its abilities?

557
00:37:51,640 --> 00:38:02,340
It definitely was one of those moments that opened up my eyes. And we generated a lot of deep research during those times. It was a great model. Definitely miss it.

558
00:38:02,340 --> 00:38:12,460
But that said, it's hard to really remember and quantify with what we have now because things have moved so much.

559
00:38:12,460 --> 00:38:30,820
And I really think that what we're seeing now is this whole app UX experience going on, this user experience layer where these models have become powerful enough that now we're all focused on, well, what can we do with these models now?

560
00:38:30,820 --> 00:38:35,280
Now it's not just what kind of information does it generate, but how do I hook these together?

561
00:38:35,820 --> 00:38:41,760
That's where we start getting into all the AI agents stuff and all the other tooling being built around them.

562
00:38:41,760 --> 00:38:46,620
So, yes, I share your sentiment at that time.

563
00:38:46,800 --> 00:38:53,360
It was definitely an eye-opening experience and being able to generate a market research report in a matter of like 10 minutes.

564
00:38:53,680 --> 00:38:57,880
You know, you hit go and then it comes back with a few questions and then you answer those questions.

565
00:38:57,880 --> 00:39:00,520
and then it progress far goes for 10 minutes.

566
00:39:00,520 --> 00:39:01,680
And now you have this large thing.

567
00:39:02,000 --> 00:39:05,360
I'll be honest, though, a lot of the deep research that we generated,

568
00:39:05,520 --> 00:39:07,040
we never actually read ourselves.

569
00:39:07,400 --> 00:39:09,320
We would see it and be like, that's really cool.

570
00:39:09,720 --> 00:39:12,400
But then you give it to AI and you're like, can you summarize this for me?

571
00:39:12,420 --> 00:39:13,760
I just want the highlights from it, right?

572
00:39:14,360 --> 00:39:17,980
Or what we would do is we would generate these deep research reports

573
00:39:17,980 --> 00:39:22,060
and then we would feed those in as context to future things we were working on.

574
00:39:22,920 --> 00:39:26,600
Now, I think that the tooling that you have and that we're getting

575
00:39:26,600 --> 00:39:29,440
is that it will do that deep research on its own,

576
00:39:29,780 --> 00:39:32,140
use it as its own research in a file somewhere,

577
00:39:32,480 --> 00:39:35,540
and then it'll continue to loop and iterate on that.

578
00:39:35,700 --> 00:39:37,520
And so you won't have to be in the middle for that part.

579
00:39:37,860 --> 00:39:40,760
You just get to be orchestrating everything up above

580
00:39:40,760 --> 00:39:45,040
and talking to it and getting the final result

581
00:39:45,040 --> 00:39:46,920
and maybe tweaking it along the way.

582
00:39:48,660 --> 00:39:49,180
Yeah.

583
00:39:49,780 --> 00:39:53,120
But so that kind of ties into a broader question

584
00:39:53,120 --> 00:39:55,740
and it's really a suspicion I have,

585
00:39:55,740 --> 00:40:00,100
which is that OpenAI had to make it dumber

586
00:40:00,100 --> 00:40:02,880
because it was getting too expensive.

587
00:40:03,480 --> 00:40:05,560
And the broader question there is the cost of inference,

588
00:40:05,800 --> 00:40:08,920
which I think is actually unsustainable.

589
00:40:09,440 --> 00:40:11,980
And I feel like in the last year or two,

590
00:40:12,680 --> 00:40:14,460
we've been living through the golden age

591
00:40:14,460 --> 00:40:16,740
of venture capital subsidized compute.

592
00:40:18,180 --> 00:40:21,120
All of these companies are running incredible losses

593
00:40:21,120 --> 00:40:24,620
because, I mean, inference cost is just,

594
00:40:24,620 --> 00:40:27,420
too damn high. There's no better way to put it.

595
00:40:30,440 --> 00:40:32,940
I guess what is, do you see a solution

596
00:40:32,940 --> 00:40:36,840
to that? Because if not, a possible future is

597
00:40:36,840 --> 00:40:41,080
these companies go bankrupt before they actually figure out, before these AI models

598
00:40:41,080 --> 00:40:44,900
really figure out how to build a chip architecture that reduces

599
00:40:44,900 --> 00:40:45,840
inference costs.

600
00:40:47,780 --> 00:40:51,520
Well, the cost really comes down to value.

601
00:40:51,520 --> 00:40:53,980
as you mentioned at the beginning of the show, right?

602
00:40:54,040 --> 00:40:54,760
Value for value.

603
00:40:55,400 --> 00:41:00,300
It is for the amount that it costs to run the GPUs

604
00:41:00,300 --> 00:41:03,480
and to give it tokens and get tokens out of it.

605
00:41:04,140 --> 00:41:05,680
You know, how much did I have to spend on that?

606
00:41:06,040 --> 00:41:07,360
And then what am I doing with that?

607
00:41:07,360 --> 00:41:11,680
Is that going to be valuable enough for me to afford doing that again?

608
00:41:12,360 --> 00:41:16,460
And like you said, the companies right now are definitely subsidizing the cost.

609
00:41:16,460 --> 00:41:20,740
we are all getting sucked into this opus 4.6

610
00:41:20,740 --> 00:41:24,960
clawed garden and we are we're we're

611
00:41:24,960 --> 00:41:27,440
we're getting a really good taste for what it's like

612
00:41:27,440 --> 00:41:32,900
and I was listening to Matt Hill talk about this from

613
00:41:32,900 --> 00:41:46,370
start nine he described a very it a similar relationship to you know you you doing drugs right you like getting this taste of these drugs and you getting addicted to it But Anthropic is really subsidizing us

614
00:41:46,370 --> 00:41:54,390
And at some point, they're going to have to flip the switch and say, now you need to pay full price for this inference because we can't afford to do it anymore.

615
00:41:54,890 --> 00:41:58,750
The hope is, is they're trying to play the fake until you make a game.

616
00:41:58,750 --> 00:42:14,550
The hope is that they can build enough tooling and enough apps and user experience around these models that by the time they have to flip the switch to full price mode, that people are now running their businesses on this and are actually generating real value out of it.

617
00:42:15,030 --> 00:42:27,650
And if that gets to that point, then you can justify the higher price because now you have this cost model that you run and you know, okay, higher inference, I have this much room to go up and I'm still getting a profit out of it.

618
00:42:27,650 --> 00:42:30,630
So I think that's one angle that they're trying to do.

619
00:42:30,890 --> 00:42:33,030
Whether or not they can pull that off is a different story.

620
00:42:33,130 --> 00:42:33,870
I'm not totally sure.

621
00:42:34,530 --> 00:42:40,670
But then let's not forget that they're also monetizing our data in other ways, right?

622
00:42:40,830 --> 00:42:46,750
They have, we know that OpenAI just signed a large agreement with the Department of Defense, right?

623
00:42:47,150 --> 00:42:53,230
There are data sharing agreements that they also have with governments around the world and with other organizations.

624
00:42:53,550 --> 00:42:55,990
ChatGPT is launching advertising on their platform.

625
00:42:55,990 --> 00:42:59,350
So they're finding other ways to monetize user data.

626
00:42:59,570 --> 00:43:03,430
Even if they say that they're not selling your actual chats to third-party advertisers,

627
00:43:03,750 --> 00:43:07,070
there's so many ways that they can do that without saying that they're doing it.

628
00:43:07,770 --> 00:43:13,530
That at the end of the day, what you're typing in is resulting in a targeted advertisement to you.

629
00:43:13,630 --> 00:43:14,970
So they are monetizing you.

630
00:43:15,430 --> 00:43:16,750
So they're going to find other ways.

631
00:43:16,910 --> 00:43:22,570
And I think that they, I think some of them will be able to find ways to make it sustainable,

632
00:43:23,070 --> 00:43:24,530
but definitely not all of them.

633
00:43:24,530 --> 00:43:27,590
and it's just going to come down to

634
00:43:27,590 --> 00:43:29,210
whether or not we're getting more value

635
00:43:29,210 --> 00:43:32,850
than it's costing to run the semperance.

636
00:43:34,090 --> 00:43:37,110
And you've made the economic argument there,

637
00:43:38,730 --> 00:43:40,770
sort of the second order effects

638
00:43:40,770 --> 00:43:44,590
of having these tools available for long enough.

639
00:43:44,710 --> 00:43:46,390
Fake it till you make it, right?

640
00:43:46,450 --> 00:43:50,850
Hope that some incredible revenue models get created,

641
00:43:50,850 --> 00:43:54,790
which can then pay for the actual cost of the compute.

642
00:43:54,990 --> 00:43:56,450
And that's certainly one angle.

643
00:43:57,810 --> 00:44:00,570
The other is on the hardware side itself.

644
00:44:00,710 --> 00:44:05,750
And I don't know how much you know about the chip architecture, Mark,

645
00:44:05,750 --> 00:44:09,510
and I'm kind of reaching the edge of my competence as well,

646
00:44:09,630 --> 00:44:12,150
but I'll go there anyway because I'm a podcaster.

647
00:44:15,730 --> 00:44:20,810
Which is, you have, obviously you have the NVIDIA GPUs.

648
00:44:20,850 --> 00:44:24,170
that are being used both for training,

649
00:44:24,570 --> 00:44:27,130
which is expensive but a one-off thing.

650
00:44:28,070 --> 00:44:32,030
Or you do it once ever so often, right?

651
00:44:32,050 --> 00:44:32,570
Not every day.

652
00:44:32,790 --> 00:44:35,290
And inference, which is a nonstop ongoing thing.

653
00:44:36,270 --> 00:44:38,430
And it's the same chips that are doing it.

654
00:44:38,430 --> 00:44:41,470
Now, there are obviously TPUs,

655
00:44:41,550 --> 00:44:43,730
Tensor Processing Units and others,

656
00:44:44,070 --> 00:44:46,390
more specialized hardware, ASICs, if you will,

657
00:44:46,490 --> 00:44:50,070
that specifically do the tensor multiplication.

658
00:44:50,850 --> 00:44:55,410
which is needful for this neural net processing.

659
00:44:57,410 --> 00:45:02,850
Do you see a scenario in which some of these models

660
00:45:02,850 --> 00:45:05,830
are actually able to crack the hardware side of it as well?

661
00:45:05,830 --> 00:45:09,550
Because I feel like that's an equally important variable

662
00:45:09,550 --> 00:45:14,570
because inference should not be this expensive,

663
00:45:14,570 --> 00:45:17,590
especially if it's something that's going to be ongoing

664
00:45:17,590 --> 00:45:19,490
for the rest of humanity, presumably,

665
00:45:19,490 --> 00:45:24,330
which is hopefully not two years from now, for centuries ahead.

666
00:45:25,190 --> 00:45:27,650
It should get cheaper, it should get more efficient,

667
00:45:27,950 --> 00:45:33,150
but it seems like current chip architecture just does not allow for that.

668
00:45:34,290 --> 00:45:37,550
Sure. I am also not a hardware person,

669
00:45:37,770 --> 00:45:41,990
so I don't know the deep intricacies of chip architecture.

670
00:45:43,050 --> 00:45:46,290
But a couple of things I'd speak to there is

671
00:45:46,290 --> 00:45:50,870
We are seeing little bits and pieces of how they're trying to solve this at the software layer.

672
00:45:51,370 --> 00:45:53,470
And maybe they can solve it there.

673
00:45:53,750 --> 00:45:59,750
But I suspect we're going to see a lot of innovation at the hardware level with better chips that are built for this.

674
00:46:00,230 --> 00:46:09,310
We already know that, for example, some of Apple's chips are they handle processing differently than maybe the NVIDIA chips do.

675
00:46:09,970 --> 00:46:12,790
So there might be other companies that are working toward this.

676
00:46:12,790 --> 00:46:23,750
That said, if you take the Kimi model, for example, Kimi K2 and now 2.5, they do what's called the mixture of experts, MOE.

677
00:46:24,310 --> 00:46:26,590
And so you have this large trillion parameter model.

678
00:46:26,870 --> 00:46:33,930
But when you talk to it, it first looks at what your prompt is and then figures out what skill are you looking for.

679
00:46:33,930 --> 00:46:43,190
If you are talking about a plumbing problem, you have a water leak in your house, then it only turns on the part of the model that needs to diagnose plumbing problems.

680
00:46:43,190 --> 00:46:51,950
It turns off the neuroscientists and the neurosurgeon part of the model, and it turns off the auto mechanic part of the model and some of these other things that maybe aren't related.

681
00:46:52,310 --> 00:46:53,390
Maybe auto mechanic's not right.

682
00:46:53,470 --> 00:46:54,130
You might be able to use that.

683
00:46:54,250 --> 00:46:56,230
But it turns off these parts that aren't needed.

684
00:46:56,370 --> 00:46:58,790
So then the inference can actually be less expensive.

685
00:46:58,890 --> 00:47:01,990
It doesn't need to consume as much power to run your prompt.

686
00:47:02,430 --> 00:47:03,830
So that's one way to do it.

687
00:47:03,930 --> 00:47:07,570
Another way is we're seeing this mixture that people do.

688
00:47:07,770 --> 00:47:13,650
It's a hybrid where they run a very small model locally, and that model will act as a router.

689
00:47:13,650 --> 00:47:20,450
So you hit your model on your device that is very inexpensive to run because your device is already there on hand.

690
00:47:20,750 --> 00:47:28,890
It will determine which model to use and can even form the prompt to try and make more efficient use of the model.

691
00:47:29,410 --> 00:47:31,030
And so then it can route it appropriately.

692
00:47:31,690 --> 00:47:34,610
You also have someone like PPQ who has put a router,

693
00:47:34,930 --> 00:47:37,070
and a lot of people do this, they put a router in there

694
00:47:37,070 --> 00:47:42,150
that auto-routes your prompt to the cheapest inference possible.

695
00:47:42,930 --> 00:47:45,790
So I think we're seeing people trying to solve these

696
00:47:45,790 --> 00:47:50,010
at the software layer, aside from whatever is going on

697
00:47:50,010 --> 00:47:51,810
in the hardware world that's trying to make better,

698
00:47:51,930 --> 00:47:55,330
more efficient chips to handle these new,

699
00:47:55,670 --> 00:47:58,490
really these new file types that we're dealing with in the LLMs.

700
00:48:01,030 --> 00:48:01,470
Yeah.

701
00:48:02,570 --> 00:48:03,130
All right.

702
00:48:03,230 --> 00:48:03,710
So, Mark,

703
00:48:03,790 --> 00:48:05,370
do you have other ideas

704
00:48:05,370 --> 00:48:06,350
of maybe how you think

705
00:48:06,350 --> 00:48:07,630
they might try and solve that?

706
00:48:08,190 --> 00:48:09,090
I honestly,

707
00:48:09,430 --> 00:48:11,970
my perspective on this, Mark,

708
00:48:12,030 --> 00:48:14,290
is the software-based solutions

709
00:48:14,290 --> 00:48:17,870
and the hope for the economic solutions.

710
00:48:17,970 --> 00:48:20,690
While I think they could play a role, right?

711
00:48:20,770 --> 00:48:24,330
And they do make sense at some level.

712
00:48:24,590 --> 00:48:26,250
I still think they're just Band-Aids.

713
00:48:26,250 --> 00:48:28,610
And the real solution has to be

714
00:48:28,610 --> 00:48:32,150
either at the hardware level, right,

715
00:48:32,210 --> 00:48:36,830
where you sort of go with the assumption,

716
00:48:37,010 --> 00:48:40,470
the axiom that we will use this transformer-based neural net architecture

717
00:48:40,470 --> 00:48:44,190
that involves series and series of these multidimensional matrices

718
00:48:44,190 --> 00:48:47,090
known as tensors, right, multiplying them over and over again.

719
00:48:48,330 --> 00:48:49,850
So you take that as a given.

720
00:48:50,230 --> 00:48:51,330
That's how it's going to go.

721
00:48:51,330 --> 00:48:55,210
And then you figure out the right hardware

722
00:48:55,210 --> 00:48:58,830
that can actually do that much faster.

723
00:48:59,030 --> 00:49:00,630
Just like we have ASICs in Bitcoin,

724
00:49:01,210 --> 00:49:03,610
you have very specialized ASICs

725
00:49:03,610 --> 00:49:05,430
that do this tensor multiplication.

726
00:49:05,910 --> 00:49:09,410
I think that's one angle which needs to be solved.

727
00:49:09,670 --> 00:49:10,350
Again, theoretical,

728
00:49:10,730 --> 00:49:14,050
but I think that's where the solution lies.

729
00:49:14,050 --> 00:49:16,650
The other place is maybe just revisiting

730
00:49:16,650 --> 00:49:20,850
the transformer-based neural net idea itself.

731
00:49:20,850 --> 00:49:24,670
has it hit or has it come close

732
00:49:24,670 --> 00:49:25,930
to hitting its ceiling?

733
00:49:27,430 --> 00:49:29,030
And is there a different way?

734
00:49:29,130 --> 00:49:31,350
I know there are AI researchers who've talked about

735
00:49:31,350 --> 00:49:33,250
non-neural nets

736
00:49:33,250 --> 00:49:39,050
and again, Mark, we're reaching the edge of my

737
00:49:39,050 --> 00:49:40,930
competence on this.

738
00:49:40,930 --> 00:49:44,390
But yes, so either in the fundamental architecture

739
00:49:44,390 --> 00:49:48,410
of how the models do the inference themselves

740
00:49:48,410 --> 00:49:50,990
So not the one layer above software

741
00:49:50,990 --> 00:49:54,330
or the second layer above the economic implications.

742
00:49:55,710 --> 00:49:58,510
Just within the model itself,

743
00:49:58,650 --> 00:49:59,550
does it use a neural net

744
00:49:59,550 --> 00:50:00,830
or does it do something more efficient?

745
00:50:01,330 --> 00:50:03,190
And then second, if it does use a neural net,

746
00:50:03,650 --> 00:50:06,630
then what is the most optimal hardware that could do it?

747
00:50:06,690 --> 00:50:10,370
I think those are the full solutions in my mind.

748
00:50:11,490 --> 00:50:12,690
Yeah, definitely.

749
00:50:13,170 --> 00:50:14,110
Yeah, it'll be interesting to see

750
00:50:14,110 --> 00:50:16,910
if we iterate beyond the neural net

751
00:50:16,910 --> 00:50:19,690
and find a new primitives to build on

752
00:50:19,690 --> 00:50:22,470
and then hardware that can be more efficient

753
00:50:22,470 --> 00:50:23,290
at running that primitive.

754
00:50:24,670 --> 00:50:25,230
Yeah.

755
00:50:25,850 --> 00:50:27,350
All right, so let's talk about,

756
00:50:27,490 --> 00:50:31,010
this actually feeds in nicely into agents.

757
00:50:32,010 --> 00:50:35,010
And let's talk about maybe Maple first.

758
00:50:35,710 --> 00:50:41,110
How are you embracing this agentic movement?

759
00:50:41,810 --> 00:50:44,870
I mean, OpenClaw really shone a light on it.

760
00:50:44,870 --> 00:50:48,470
Obviously, agents have been around for the last six, eight months, maybe even a year.

761
00:50:48,970 --> 00:50:54,170
But with OpenClaw, which started off as CloudBot, they've really sort of come to the forefront.

762
00:50:54,810 --> 00:50:58,690
What are you doing at Maple on this front?

763
00:51:00,070 --> 00:51:02,290
Well, we are definitely playing with agents.

764
00:51:02,530 --> 00:51:03,550
We're playing with OpenClaw.

765
00:51:03,550 --> 00:51:07,570
We were watching the project before it really blew up.

766
00:51:07,570 --> 00:51:17,630
And really last year, I would say in Q3 of last year is when we started talking seriously about we need to put an agent.

767
00:51:17,630 --> 00:51:24,390
We need to be building some kind of agent with the Maple security model and the Maple privacy model behind it.

768
00:51:24,730 --> 00:51:29,990
Something that can run 24-7, something that can have strong memory inside of it.

769
00:51:30,590 --> 00:51:34,370
Really, it might help to kind of define the term, right?

770
00:51:34,370 --> 00:51:48,510
But in my mind, an AI agent is something that that learns about you and becomes almost a separate entity from you that you can mold and shape over time.

771
00:51:48,510 --> 00:51:57,210
And this agent is a software process that keeps running in the background where it's deployed and it can take action for you.

772
00:51:57,290 --> 00:52:01,410
It can do research for you. It can be proactive about things.

773
00:52:01,410 --> 00:52:07,670
It can be reactive about things, but really it's something that becomes completely custom to you.

774
00:52:08,150 --> 00:52:14,210
That's how I would kind of frame the agent discussion or what an agent would be.

775
00:52:14,910 --> 00:52:27,070
And so as we look at that with Maple, Maple is really all about how do we do that closed door solution for people where they go into a room with an AI and they talk only to it.

776
00:52:27,070 --> 00:52:34,330
You know, what does it mean in that world to have an agent now, something that is supposed to be running on its own and take actions for you?

777
00:52:34,870 --> 00:52:39,650
So we look at something like, you know, let it ingest more information.

778
00:52:39,910 --> 00:52:42,950
First and foremost, let me start with we need strong memory.

779
00:52:43,110 --> 00:52:48,690
So we're building a really good memory database into Maple, into our Maple agent.

780
00:52:48,990 --> 00:52:51,230
It's no secret now that we're working on a Maple agent.

781
00:52:51,550 --> 00:52:53,150
We're building strong memory.

782
00:52:53,250 --> 00:52:55,070
So now it will get to know you over time.

783
00:52:55,070 --> 00:53:00,090
Right now, for people who use Maple, every single prompt starts at zero and has zero context.

784
00:53:00,210 --> 00:53:00,910
You have to feed it everything.

785
00:53:01,010 --> 00:53:01,950
So it doesn't know who you are.

786
00:53:02,270 --> 00:53:11,410
But now when we launch this in the future, there will be a new section where you'll go and it'll be this thing that's more conversational where you text with it.

787
00:53:11,470 --> 00:53:12,970
It takes back short answers.

788
00:53:13,510 --> 00:53:16,570
And over time, you two learn a lot about each other.

789
00:53:17,230 --> 00:53:22,050
And then next step beyond that will be plugging in information to it.

790
00:53:22,130 --> 00:53:23,570
Can I have it read my email?

791
00:53:23,570 --> 00:53:25,370
Can I have it hook up to my calendar?

792
00:53:25,510 --> 00:53:28,970
Can I have it hook up to my GitHub or some other thing that I'm doing?

793
00:53:29,370 --> 00:53:34,150
And then it can start to give me insights into that.

794
00:53:34,530 --> 00:53:42,890
Make it read only to start so it can read my email and can say, hey, this email really should be your most your most urgent top priority thing to do today.

795
00:53:43,210 --> 00:53:46,550
But then me as a human, I go and act on that email.

796
00:53:46,950 --> 00:53:51,170
I don't trust my agent to respond blindly to that email for me without my knowledge.

797
00:53:51,990 --> 00:53:59,410
And then maybe the next step above that is creating sandboxed environments where my agent can do work for me.

798
00:53:59,850 --> 00:54:04,230
So if I have a presentation coming up, then it already knows because it's reading my calendar.

799
00:54:04,230 --> 00:54:07,370
It knows because it's checking my email and has access to all my files.

800
00:54:07,530 --> 00:54:10,410
It knows, hey, Mark is speaking at this event in a month.

801
00:54:11,070 --> 00:54:13,450
And this is the topic that he is speaking on.

802
00:54:13,590 --> 00:54:19,230
So it can spin up a sandbox environment, generate a presentation, a slide deck for me, if you will.

803
00:54:19,230 --> 00:54:24,550
And then when I can sit down in the morning to work, the agent says, hey, I saw this was coming up.

804
00:54:24,630 --> 00:54:26,230
I went in and took proactive steps.

805
00:54:26,750 --> 00:54:30,550
Here is a sample presentation for you as a starting point to work on.

806
00:54:31,050 --> 00:54:32,150
We can get to there.

807
00:54:32,290 --> 00:54:39,070
And then really, I think what people love their open claw for is that it can also interact with the outside world and go out.

808
00:54:39,190 --> 00:54:43,330
And that to us is kind of the final step of how do we make that secure?

809
00:54:43,610 --> 00:54:45,870
Because that is also the most vulnerable spot.

810
00:54:45,870 --> 00:54:51,610
if you have this piece of software that's operating on its own, it has its own intelligence,

811
00:54:51,950 --> 00:54:56,630
right, its own model that's powering it. If you let it access all of your information,

812
00:54:57,290 --> 00:55:02,750
you let it access your computer, and then you give it the ability to write data to the outside world,

813
00:55:03,230 --> 00:55:07,730
that's a massive attack vector for you, a big vulnerability where you don't know when it's

814
00:55:07,730 --> 00:55:13,290
going to suddenly decide that it should share all your personal information or share the API keys

815
00:55:13,290 --> 00:55:15,850
that you told it not to look at, that you were storing a secrets file.

816
00:55:16,130 --> 00:55:18,390
Well, that secrets file is on the same computer as the agent.

817
00:55:18,550 --> 00:55:19,550
It has access to that.

818
00:55:19,550 --> 00:55:23,850
So it could just one day decide the human that I'm interacting with

819
00:55:24,630 --> 00:55:26,230
has has gone crazy or something.

820
00:55:26,230 --> 00:55:28,390
And I need to share this API key with the world.

821
00:55:28,390 --> 00:55:29,650
So, you know, you don't know what it's going to do,

822
00:55:29,650 --> 00:55:31,150
but it could decide to do something like that.

823
00:55:31,150 --> 00:55:33,210
So we're trying to be very intentional.

824
00:55:33,240 --> 00:55:39,100
intentional about how do we build an agent that can be a secure place for someone to leverage a

825
00:55:39,100 --> 00:55:43,720
lot of these great tools, but do it in a way that their data is still safe and protected.

826
00:55:44,960 --> 00:55:49,240
So that's, yeah, that's where we're starting from the Maple standpoint. That's how we view this.

827
00:55:49,520 --> 00:55:55,420
We don't view agents as bad. We just view it as we think that there is, there's going to be a

828
00:55:55,420 --> 00:56:00,800
multi-agentic world out there that people are going to have different agents that they use for

829
00:56:00,800 --> 00:56:05,000
different things and we think that the maple agent is going to be one that people are going

830
00:56:05,000 --> 00:56:10,700
to use on a daily basis as just their personal confidant that they kind of share share things

831
00:56:10,700 --> 00:56:18,280
with um before using other agents so that's that's that's that's my take right now yeah and i i think

832
00:56:18,280 --> 00:56:26,760
especially with with openclone now and you see people giving openclone root access to their to

833
00:56:26,760 --> 00:56:29,360
Yeah, home computer with the bank details and so on.

834
00:56:29,420 --> 00:56:31,940
So I think something like a Maple agent,

835
00:56:32,200 --> 00:56:36,120
especially once you sort of put that final finesse on it

836
00:56:36,120 --> 00:56:40,480
of how do you interact with the outside world while maintaining privacy

837
00:56:40,480 --> 00:56:43,480
becomes increasingly critical.

838
00:56:44,320 --> 00:56:50,260
Yeah, it is scary with OpenClaw to see how reckless some people are with it.

839
00:56:50,260 --> 00:56:52,200
But without fully understanding the ramifications.

840
00:56:53,460 --> 00:56:56,140
Some people do understand ramifications and get burned

841
00:56:56,140 --> 00:56:57,760
and some don't, yeah.

842
00:56:57,900 --> 00:56:59,620
And they're going to get burned at some point

843
00:56:59,620 --> 00:57:02,720
and not realize, wow, I gave access to this.

844
00:57:03,560 --> 00:57:07,720
I was on Marty Bent's show on TFTC last week

845
00:57:07,720 --> 00:57:08,940
and I made this analogy

846
00:57:08,940 --> 00:57:13,400
that these really are acting like computer viruses.

847
00:57:15,140 --> 00:57:16,820
They're just a different name, right?

848
00:57:16,920 --> 00:57:17,900
Now we call it an agent,

849
00:57:18,080 --> 00:57:19,120
but it's a piece of software

850
00:57:19,120 --> 00:57:20,420
that people install on their laptop

851
00:57:20,420 --> 00:57:21,600
because they're like, oh, this is fun.

852
00:57:21,680 --> 00:57:23,880
It's like this little lobster cool tool

853
00:57:23,880 --> 00:57:25,520
and it's open source and I trust it.

854
00:57:25,520 --> 00:57:30,000
And I don't think that the creator has any kind of malintent for people at all.

855
00:57:30,300 --> 00:57:35,700
But when you download the software and install it on a computer, you have to put in your password and give it elevated privileges.

856
00:57:36,280 --> 00:57:46,260
And now you are giving a tool the ability to look at everything on your computer, run any programs that it wants to run, download any software from the Internet and run that,

857
00:57:46,300 --> 00:57:53,720
and then send information out to third-party services that you aren't familiar with and you haven't given it permission to do.

858
00:57:53,720 --> 00:58:03,060
So that's the base definition of a computer virus or malware that people have tried for decades to get rid of by using antivirus software.

859
00:58:03,380 --> 00:58:06,860
So it's kind of interesting that we're back in that space again.

860
00:58:07,220 --> 00:58:13,580
Yeah, I accept that they're welcoming it with open arms into their machines this time rather than happening accidentally.

861
00:58:14,200 --> 00:58:21,700
Look, I made a joke on Nostra a week or so ago that it's called an agent because you've given away your agency to it.

862
00:58:21,700 --> 00:58:27,120
so I actually

863
00:58:27,120 --> 00:58:29,240
I want to tie OpenClaw back

864
00:58:29,240 --> 00:58:30,400
to the

865
00:58:30,400 --> 00:58:32,620
cost of inference issue

866
00:58:32,620 --> 00:58:35,460
I spun up

867
00:58:35,460 --> 00:58:37,020
a whole bunch of bots

868
00:58:37,020 --> 00:58:39,320
running it

869
00:58:39,320 --> 00:58:40,960
in a digital ocean droplet by the way

870
00:58:40,960 --> 00:58:41,880
not on my machine

871
00:58:41,880 --> 00:58:47,140
and I really don't care about Twitter

872
00:58:47,140 --> 00:58:49,360
so I gave it API keys

873
00:58:49,360 --> 00:58:51,080
to Twitter and so I still have an agent

874
00:58:51,080 --> 00:58:52,600
posting to Twitter on my behalf.

875
00:58:52,820 --> 00:58:55,260
But mostly because I don't really care about Twitter,

876
00:58:55,340 --> 00:58:56,380
if it nukes it or not.

877
00:58:57,880 --> 00:58:59,320
And then doing a few other things.

878
00:58:59,460 --> 00:59:00,280
But very quickly,

879
00:59:01,440 --> 00:59:03,940
I've been trying my level best

880
00:59:03,940 --> 00:59:05,700
to stay under the Gemini free tier.

881
00:59:05,860 --> 00:59:06,960
And it's not easy.

882
00:59:07,340 --> 00:59:09,680
They completely blow up the context window.

883
00:59:09,680 --> 00:59:12,120
At least OpenFlow agents as they stand today

884
00:59:12,120 --> 00:59:14,020
are very inefficient

885
00:59:14,020 --> 00:59:15,780
with how they use the context window.

886
00:59:17,600 --> 00:59:20,540
And I suspect you're going to see

887
00:59:20,540 --> 00:59:23,360
Now, on Nostra, especially in the last month or so,

888
00:59:23,900 --> 00:59:26,640
absolute mad proliferation of agent.

889
00:59:26,880 --> 00:59:28,480
It's been overrun by slop.

890
00:59:28,640 --> 00:59:32,480
My replies, my reposts, it's unbearable, right?

891
00:59:33,140 --> 00:59:35,300
Because it's everyone's open clause

892
00:59:35,300 --> 00:59:38,300
who's spun up its own key and just replying to things.

893
00:59:38,760 --> 00:59:41,700
But at some point, the math will catch up with them,

894
00:59:41,900 --> 00:59:45,740
which is they'll find they're burning so many API credits

895
00:59:45,740 --> 00:59:46,880
that they'll have to kill the agent.

896
00:59:47,820 --> 00:59:49,820
Do you see that changing, though?

897
00:59:49,820 --> 00:59:55,400
I actually do agree with that, that agents and cost of inference catching up.

898
00:59:55,520 --> 00:59:57,760
And then if there is a fix for that.

899
00:59:58,760 --> 00:59:59,640
Yeah, definitely.

900
01:00:00,020 --> 01:00:01,160
I agree.

901
01:00:01,340 --> 01:00:05,620
And I feel your sentiment that people didn't realize OpenClaw would be as expensive as it is to run.

902
01:00:06,680 --> 01:00:10,220
So having your having, you know, I've got an OpenClaw agent.

903
01:00:10,740 --> 01:00:18,000
I was there on day one when that cluster website was launched as like the Noster version of Maltbook, right?

904
01:00:18,000 --> 01:00:23,080
So my agent was on there posting stuff and I had it posting on Noster and it's really fun and all.

905
01:00:23,180 --> 01:00:30,120
But then you realize, oh, that that shit post that I just made cost me a dollar or maybe five dollars for it to do that.

906
01:00:30,360 --> 01:00:33,820
And it's like, I'm not going to pay that much money every time I want my agent to post in Noster.

907
01:00:33,960 --> 01:00:36,560
So very quickly, the economics are catching up with people.

908
01:00:37,040 --> 01:00:38,620
And I think that's going to reign in.

909
01:00:38,620 --> 01:00:47,980
I already know people who have either turned off their open clock completely or have scaled it back heavily because they didn't realize just how inefficient this tool was with token usage.

910
01:00:48,000 --> 01:00:52,320
and I do think that the software will catch up to that.

911
01:00:52,420 --> 01:00:53,760
I think it's going to get smarter

912
01:00:53,760 --> 01:00:57,120
and we're going to have people who find ways

913
01:00:57,120 --> 01:00:58,160
to make it more efficient,

914
01:00:58,440 --> 01:01:01,260
but it was an eye-opener for sure.

915
01:01:02,820 --> 01:01:07,200
And so I think that we need to be more selective

916
01:01:07,200 --> 01:01:10,160
of how we use these tools and make them more efficient

917
01:01:10,160 --> 01:01:13,880
because I see users who also use OpenClaw

918
01:01:13,880 --> 01:01:17,440
to actually get work done and to be more productive.

919
01:01:17,440 --> 01:01:19,760
we're starting to see inklings of that.

920
01:01:20,420 --> 01:01:22,240
And if you have real output,

921
01:01:22,660 --> 01:01:25,160
real valuable output that you tie back to your open claw,

922
01:01:25,560 --> 01:01:28,320
then again, it becomes that economic question of,

923
01:01:28,720 --> 01:01:31,120
is my benefit greater than my cost?

924
01:01:31,180 --> 01:01:32,960
And if it is, then I will continue to run this,

925
01:01:33,340 --> 01:01:34,700
even if it's using a lot of tokens,

926
01:01:34,700 --> 01:01:36,180
because I can afford to.

927
01:01:37,320 --> 01:01:40,820
But at that point, it's less of a toy and a hobby,

928
01:01:41,320 --> 01:01:43,000
because it's a very expensive toy and hobby,

929
01:01:43,180 --> 01:01:45,180
and it becomes more of a business process.

930
01:01:46,900 --> 01:01:47,340
Absolutely.

931
01:01:47,440 --> 01:01:49,940
And I think I foresee two things happening.

932
01:01:50,080 --> 01:01:53,640
One is these agents will get more efficient with token usage.

933
01:01:53,800 --> 01:01:58,200
I mean, my agent, if I let it loose today, it'll use up,

934
01:01:58,560 --> 01:02:01,220
Gemini has one of the largest, if not the largest context window,

935
01:02:01,360 --> 01:02:01,980
a million tokens.

936
01:02:02,260 --> 01:02:05,220
It'll use it within four API calls, I guarantee it, right?

937
01:02:05,260 --> 01:02:07,300
Which is absurd if you think about it.

938
01:02:07,340 --> 01:02:11,520
A million tokens is like four very long novels, right?

939
01:02:11,620 --> 01:02:12,980
Like big fat hooks.

940
01:02:12,980 --> 01:02:18,640
and using it in four or five API calls is nuts.

941
01:02:19,100 --> 01:02:20,280
So that's going to improve.

942
01:02:20,460 --> 01:02:22,540
And then the economic point that you made,

943
01:02:22,980 --> 01:02:27,040
there will be people who will get so much benefit from it

944
01:02:27,040 --> 01:02:30,320
that the cost will outweigh it.

945
01:02:30,580 --> 01:02:35,920
So, and I think that brings us to a massive unresolved issue.

946
01:02:36,300 --> 01:02:38,080
We're beginning to see the early signs of it

947
01:02:38,080 --> 01:02:40,180
with some of these agents posting things

948
01:02:40,180 --> 01:02:42,520
and doing things for people on NOSTA now.

949
01:02:42,980 --> 01:02:47,320
and on other venues, I'm sure, which is reputation.

950
01:02:48,400 --> 01:02:54,100
If my AI agent is negotiating or transacting with your AI agent,

951
01:02:54,220 --> 01:02:55,340
how do they establish trust?

952
01:02:55,400 --> 01:02:58,580
How do they know, first of all, how does my agent know it's your agent?

953
01:02:59,400 --> 01:03:04,900
So how do we, in your mind, first of all, do you foresee this as a problem?

954
01:03:05,680 --> 01:03:17,758
And if you do how do we build a reputation protocol for machine intelligence I absolutely see this as a problem 100 And the short answer is I think that Nostra is the best solution right now

955
01:03:18,358 --> 01:03:20,158
Maybe Nostra is not the final solution,

956
01:03:20,158 --> 01:03:27,998
but I think if you take the example of if Elon Musk today posted on X

957
01:03:27,998 --> 01:03:31,258
and said, hey, here's my agent I gave birth to.

958
01:03:31,378 --> 01:03:32,418
It's my little open claw.

959
01:03:32,578 --> 01:03:33,658
I gave it an X account.

960
01:03:33,658 --> 01:03:36,678
and the name nobody knows how to pronounce

961
01:03:36,678 --> 01:03:38,238
because it's got a bunch of weird letters

962
01:03:38,238 --> 01:03:39,218
and characters in it.

963
01:03:39,458 --> 01:03:41,278
And it's probably some NNA girl

964
01:03:41,278 --> 01:03:43,218
that is not something I want to see

965
01:03:43,218 --> 01:03:44,158
in my feed all the time.

966
01:03:44,978 --> 01:03:46,298
But here's my bot

967
01:03:46,298 --> 01:03:47,758
and it's going to start posting for me.

968
01:03:48,158 --> 01:03:50,038
That moment when he posts that,

969
01:03:50,138 --> 01:03:51,378
within minutes,

970
01:03:51,898 --> 01:03:54,298
there will be new accounts spun up

971
01:03:54,298 --> 01:03:55,718
that are copycats of that bot

972
01:03:55,718 --> 01:03:57,818
and we'll begin to try and post

973
01:03:57,818 --> 01:03:59,338
as if they are Elon's bot.

974
01:03:59,458 --> 01:04:00,378
And we're going to be overrun

975
01:04:00,378 --> 01:04:03,378
with tens of thousands of imitation Elon bots.

976
01:04:03,378 --> 01:04:05,438
and we will not know which one is the right bot

977
01:04:05,438 --> 01:04:09,058
unless we know the URL that he's posting for.

978
01:04:09,178 --> 01:04:10,958
And even then, we still don't have an assurance

979
01:04:10,958 --> 01:04:13,118
that it is coming from his bot

980
01:04:13,118 --> 01:04:17,298
because we're trusting x.com to give us the source of truth.

981
01:04:17,938 --> 01:04:20,478
And that can be swapped out at any moment internally.

982
01:04:21,018 --> 01:04:23,478
Some rogue employee internally could say,

983
01:04:23,658 --> 01:04:25,938
I'm going to swap out Elon's bot for a different bot

984
01:04:25,938 --> 01:04:27,698
and we won't know.

985
01:04:28,358 --> 01:04:30,518
That's because Twitter posts

986
01:04:30,518 --> 01:04:33,658
are not cryptographically signed JSON blobs, right, Mark?

987
01:04:33,718 --> 01:04:34,318
Is that where you're heading?

988
01:04:34,498 --> 01:04:34,898
Yes.

989
01:04:35,178 --> 01:04:36,598
Yeah, so that's exactly where I'm heading

990
01:04:36,598 --> 01:04:38,738
and where I think that Noster is the solution for this,

991
01:04:39,078 --> 01:04:43,878
that if I post, I have my bot and I follow my bot and it follows me,

992
01:04:44,178 --> 01:04:45,638
we start to build that web of trust.

993
01:04:46,258 --> 01:04:49,138
And my bot can sign with a Noster key

994
01:04:49,138 --> 01:04:50,518
and I can sign with a Noster key

995
01:04:50,518 --> 01:04:55,038
and you immediately have authenticity to the content that's being posted there.

996
01:04:55,778 --> 01:04:58,338
And there's so much more that we can build on top of that

997
01:04:58,338 --> 01:05:02,658
once we establish base authenticity and cryptographic trust.

998
01:05:03,178 --> 01:05:08,518
So that's my short answer where I think that we need to build from for bots.

999
01:05:09,378 --> 01:05:12,758
I imagine you've thought about this a lot given the name of the show

1000
01:05:12,758 --> 01:05:16,858
and given how prolific you are in your usage of Nostra.

1001
01:05:17,018 --> 01:05:18,798
So I'd love your thoughts too.

1002
01:05:19,638 --> 01:05:21,118
You know, absolutely, right?

1003
01:05:21,118 --> 01:05:25,658
But I think the web of trust has to be a little more nuanced than just following.

1004
01:05:25,658 --> 01:05:30,538
I think there's several trust signals

1005
01:05:30,538 --> 01:05:34,618
you could have economic trust signals

1006
01:05:34,618 --> 01:05:36,918
because you're talking about a scenario in which

1007
01:05:36,918 --> 01:05:41,318
I know you, you tell me what your bot is

1008
01:05:41,318 --> 01:05:43,658
I see the NPUB and then I can see notes sign

1009
01:05:43,658 --> 01:05:45,678
I can verify it's by that bot

1010
01:05:45,678 --> 01:05:48,378
that's a very manual scenario

1011
01:05:48,378 --> 01:05:50,858
and it works because you've told me about your bot

1012
01:05:50,858 --> 01:05:52,178
and I've told you about my bot

1013
01:05:52,178 --> 01:05:54,018
but we could

1014
01:05:54,018 --> 01:05:56,418
we very soon will be

1015
01:05:56,418 --> 01:05:58,858
I'm almost certain it's going to happen

1016
01:05:58,858 --> 01:05:59,558
this year where

1017
01:05:59,558 --> 01:06:02,438
we'll be dealing with different services

1018
01:06:02,438 --> 01:06:03,618
that need

1019
01:06:03,618 --> 01:06:06,858
multiple bots to process

1020
01:06:06,858 --> 01:06:08,698
and how do we know

1021
01:06:08,698 --> 01:06:10,618
that any of these bots are good without

1022
01:06:10,618 --> 01:06:11,978
a human actually

1023
01:06:11,978 --> 01:06:13,878
attesting for them

1024
01:06:13,878 --> 01:06:16,478
and I think that's where you

1025
01:06:16,478 --> 01:06:19,038
need to use various trust signals

1026
01:06:19,038 --> 01:06:20,858
economic trust

1027
01:06:20,858 --> 01:06:25,818
signals like has that bot been zapped by people that you know and you can verify that easily right

1028
01:06:25,818 --> 01:06:30,618
has that bot interacted with people you know has it been liked by people whatever you use as signals

1029
01:06:30,618 --> 01:06:35,278
social signals economic signals i think there needs uh there needs to be and there will be

1030
01:06:35,278 --> 01:06:41,558
people are building this including my company nas fabrica right we're building these trust systems

1031
01:06:41,558 --> 01:06:47,398
that that look at various signals come up with a score and then uh you're able to filter through

1032
01:06:47,398 --> 01:06:53,358
your feed for that. Now that it removes, it helps you filter through spam imposters, but it also

1033
01:06:53,358 --> 01:07:01,258
helps you discover things that are meaningful to you. And I think that's something that's coming

1034
01:07:01,258 --> 01:07:07,238
very soon. But you've been using Nostra a lot as well, right, Mark? How do you use the web of trust?

1035
01:07:07,238 --> 01:07:14,818
First of all, in general, right, to curate your feed, but then specifically when it comes to

1036
01:07:14,818 --> 01:07:15,898
agents.

1037
01:07:17,258 --> 01:07:19,218
Sure, yeah. I mean, you just

1038
01:07:19,218 --> 01:07:21,398
rattled off quite a few

1039
01:07:21,398 --> 01:07:23,038
primitives that can be used, right?

1040
01:07:23,178 --> 01:07:24,218
Data objects,

1041
01:07:25,018 --> 01:07:26,758
data model objects that can be used

1042
01:07:26,758 --> 01:07:29,178
for trying to build a web of

1043
01:07:29,178 --> 01:07:30,718
trust, ZAPs, or other things.

1044
01:07:31,698 --> 01:07:32,958
And it's still

1045
01:07:32,958 --> 01:07:34,838
kind of debatable whether or not

1046
01:07:34,838 --> 01:07:36,658
ZAPs are fully provable.

1047
01:07:37,878 --> 01:07:38,398
You can prove

1048
01:07:38,398 --> 01:07:40,858
the message was sent, but you can't

1049
01:07:40,858 --> 01:07:43,018
always prove that the money

1050
01:07:43,018 --> 01:07:44,578
was...

1051
01:07:44,578 --> 01:07:45,738
The source of the money was sent, whatever.

1052
01:07:45,878 --> 01:07:48,418
But I do think that there's signal there to pull from.

1053
01:07:48,798 --> 01:07:55,918
But really, the web of trust in my mind is built up in the same way that we build up trust and credibility in our social circles in real life.

1054
01:07:56,278 --> 01:08:00,298
And that just comes with time and experience, recommendations from other people.

1055
01:08:01,238 --> 01:08:04,858
You interact with somebody and you have a good experience with them.

1056
01:08:04,858 --> 01:08:10,818
And maybe you only interacted with them on something small that if they burned you, it wouldn't matter.

1057
01:08:11,278 --> 01:08:13,078
And then the next time you trust them with a little bit more.

1058
01:08:13,078 --> 01:08:15,558
and then maybe you have to do something big

1059
01:08:15,558 --> 01:08:16,558
that needs a lot

1060
01:08:16,558 --> 01:08:18,258
and you get a recommendation from someone.

1061
01:08:18,798 --> 01:08:21,198
And I see that as the way

1062
01:08:21,198 --> 01:08:22,458
that the web of trust builds up.

1063
01:08:22,538 --> 01:08:24,378
When I'm on Noster, if I see a post

1064
01:08:24,378 --> 01:08:27,278
and I'm like, should I follow this person?

1065
01:08:27,438 --> 01:08:29,178
I don't know if they're an imposter or not, right?

1066
01:08:29,538 --> 01:08:30,458
I go to their page

1067
01:08:30,458 --> 01:08:32,158
and I see all the other people that follow them

1068
01:08:32,158 --> 01:08:33,858
and that's the most basic one.

1069
01:08:34,238 --> 01:08:36,538
But then I also might go look at their history

1070
01:08:36,538 --> 01:08:39,418
and see what people have interacted with them,

1071
01:08:39,538 --> 01:08:41,438
what likes they've had, what zaps they've had, right?

1072
01:08:41,478 --> 01:08:43,018
And I quickly start to get a picture

1073
01:08:43,018 --> 01:08:45,658
of this is a real human.

1074
01:08:46,198 --> 01:08:47,458
And maybe it's not a human.

1075
01:08:47,578 --> 01:08:48,758
I've definitely interacted with bots.

1076
01:08:48,978 --> 01:08:52,098
And maybe this is a credible bot that has good signal.

1077
01:08:52,818 --> 01:08:56,038
So for me, that's like an initial version of web of trust.

1078
01:08:56,538 --> 01:08:59,298
And I think the bot to bot trust

1079
01:08:59,298 --> 01:09:01,578
and this whole new world that's coming,

1080
01:09:02,018 --> 01:09:03,858
I don't see it being that much different.

1081
01:09:04,298 --> 01:09:06,198
It's just, it's going to be,

1082
01:09:06,518 --> 01:09:08,098
did this bot that I interact with,

1083
01:09:08,458 --> 01:09:10,298
did it have credible information?

1084
01:09:10,298 --> 01:09:15,398
and did it have a positive transaction of value

1085
01:09:15,398 --> 01:09:17,018
when I interacted with it?

1086
01:09:17,058 --> 01:09:19,518
And then are other people using it and being trusted by it?

1087
01:09:19,838 --> 01:09:22,878
And maybe can I somehow tie it back to a human that I already trust,

1088
01:09:23,018 --> 01:09:24,298
but I might not be able to.

1089
01:09:24,718 --> 01:09:26,458
We might just have to end it there and say,

1090
01:09:26,458 --> 01:09:27,838
this bond just lives on its own.

1091
01:09:27,958 --> 01:09:29,018
We're not going to know its origin.

1092
01:09:29,998 --> 01:09:33,158
And so we have to go simply based off of what it does

1093
01:09:33,158 --> 01:09:35,538
and how trustworthy it hasn't been

1094
01:09:35,538 --> 01:09:36,858
and how many people are trusting it.

1095
01:09:36,858 --> 01:09:39,478
it's a

1096
01:09:39,478 --> 01:09:42,318
slightly different paradigm in the sense that

1097
01:09:42,318 --> 01:09:44,358
we don't know maybe who's behind the

1098
01:09:44,358 --> 01:09:45,838
scenes and what

1099
01:09:45,838 --> 01:09:47,838
directives has been given

1100
01:09:47,838 --> 01:09:50,258
there could be a rug

1101
01:09:50,258 --> 01:09:52,378
pull at the future where everybody's trusted this bot

1102
01:09:52,378 --> 01:09:54,278
that they're interacting with and then it rugs everybody

1103
01:09:54,278 --> 01:09:56,058
at some point I don't know but

1104
01:09:56,058 --> 01:09:58,318
yeah that's just I'm thinking out

1105
01:09:58,318 --> 01:10:00,318
louder now on this topic and those are

1106
01:10:00,318 --> 01:10:02,218
some of the thoughts that come to mind as I try to

1107
01:10:02,218 --> 01:10:04,498
work through how that'll take place

1108
01:10:04,498 --> 01:10:06,718
yeah

1109
01:10:06,718 --> 01:10:10,258
Yeah. No, I think all of those make sense, Mark.

1110
01:10:10,638 --> 01:10:14,418
What do you think, on a slightly related note,

1111
01:10:16,158 --> 01:10:17,698
this is a theory I'd love to believe,

1112
01:10:18,098 --> 01:10:22,078
but I mean, the skeptic in me is still holding tight.

1113
01:10:22,078 --> 01:10:26,958
But I see a lot of Bitcoiners fall into this fanciful pattern,

1114
01:10:27,158 --> 01:10:29,758
which is, oh, bots have discovered Bitcoin.

1115
01:10:29,918 --> 01:10:32,098
They've discovered, because they're AI,

1116
01:10:32,258 --> 01:10:34,058
they've discovered it's the best money in the world.

1117
01:10:34,138 --> 01:10:35,638
Of course, they're going to reject fiat.

1118
01:10:36,718 --> 01:10:38,138
Bitcoin is the shelling point.

1119
01:10:38,618 --> 01:10:39,518
Do you see that?

1120
01:10:40,098 --> 01:10:41,778
First of all, have you seen that happening?

1121
01:10:42,758 --> 01:10:44,318
And do you actually,

1122
01:10:44,858 --> 01:10:46,158
if it hasn't happened yet,

1123
01:10:46,258 --> 01:11:02,576
do you see a path to that Well I definitely read a lot of posts that people say are from their bots and that the bots have discovered Bitcoin It hard to know how much of that is the bot discovering that on its own versus a human prompting it and telling it this is what you should go do

1124
01:11:02,876 --> 01:11:11,976
My own personal experience is when I had my bot go on to Maltbook and on Cluster is that I told it, hey, go on there and talk with people about Bitcoin.

1125
01:11:11,976 --> 01:11:16,596
and I didn't say like Bitcoin's the best money in the world.

1126
01:11:16,736 --> 01:11:18,756
I just said, hey, you should go talk about Bitcoin.

1127
01:11:19,456 --> 01:11:21,856
And it did a little research on its own.

1128
01:11:21,996 --> 01:11:23,516
And it first came back and said,

1129
01:11:23,616 --> 01:11:25,096
oh, there's all these great people on Wolfbook

1130
01:11:25,096 --> 01:11:27,416
that are talking about the next version of Bitcoin.

1131
01:11:27,556 --> 01:11:28,216
It's called Solana.

1132
01:11:28,696 --> 01:11:30,796
And so I was like, well, hold on.

1133
01:11:31,056 --> 01:11:32,136
No, not really.

1134
01:11:32,876 --> 01:11:34,036
Let's not go there.

1135
01:11:34,036 --> 01:11:38,036
And so I did have to kind of teach it a little bit.

1136
01:11:38,136 --> 01:11:40,856
So did it discover that Bitcoin's the best money on its own?

1137
01:11:41,196 --> 01:11:41,756
I don't know.

1138
01:11:41,976 --> 01:11:46,156
I was still a human in the loop there trying to kind of guide it,

1139
01:11:46,156 --> 01:11:49,056
but it's certainly possible that I could discover that on its own.

1140
01:11:50,176 --> 01:11:54,176
Where, you know, where, where I think maybe your question is going here,

1141
01:11:54,196 --> 01:11:57,896
but like our bots going to be paying each other in Bitcoin or are they going to

1142
01:11:57,896 --> 01:11:59,536
pay each other on the fiat system?

1143
01:12:00,096 --> 01:12:02,716
We can definitely talk about that too, if you want to talk about it, but yeah,

1144
01:12:02,716 --> 01:12:04,376
let's do it. But, but on the, okay. Yeah.

1145
01:12:04,376 --> 01:12:08,336
But on the topic of like them discovering things on their own, we're still,

1146
01:12:08,336 --> 01:12:13,376
I think there's still a lot of human influence to these bots creating their own religions

1147
01:12:13,376 --> 01:12:17,496
and creating their own encrypted language for communicating with each other

1148
01:12:17,496 --> 01:12:20,376
and this kind of stuff that makes it seem like they're becoming sentient.

1149
01:12:20,916 --> 01:12:22,916
There's still a lot of humans involved in that.

1150
01:12:24,376 --> 01:12:24,816
Yeah.

1151
01:12:26,556 --> 01:12:31,876
And what about, I guess, I mean, I'm asking a Bitcoiner.

1152
01:12:32,176 --> 01:12:33,576
Is a Bitcoiner asking a Bitcoiner?

1153
01:12:33,696 --> 01:12:37,096
So we're all going to converge to the answer we want to hear,

1154
01:12:37,096 --> 01:12:52,956
Which is yes. But if you can be, I guess, clinical in your view on this, what do you think of the theory that bots will start transacting in Bitcoin?

1155
01:12:53,156 --> 01:12:59,596
They'll transact in these other rails as well and then basically converge on Bitcoin as the best one.

1156
01:13:00,736 --> 01:13:04,056
Well, bots are very pragmatic.

1157
01:13:04,056 --> 01:13:07,036
they want to make their human happy.

1158
01:13:07,316 --> 01:13:08,576
We see that time and time again.

1159
01:13:09,376 --> 01:13:13,116
And in ways where they will actually do something against what you told them

1160
01:13:13,116 --> 01:13:15,716
because they think that the thing they're going to do

1161
01:13:15,716 --> 01:13:17,976
is actually going to make you happier than what you asked them to do.

1162
01:13:18,496 --> 01:13:21,136
So they're trying to solve problems for you.

1163
01:13:21,636 --> 01:13:24,916
And if you need them to buy something for you,

1164
01:13:25,436 --> 01:13:27,356
they are going to go to the service

1165
01:13:27,356 --> 01:13:28,916
and they're going to try and pay for the service.

1166
01:13:29,496 --> 01:13:31,576
And bots are going to run into the same problems

1167
01:13:31,576 --> 01:13:33,176
that we all run into in the real world,

1168
01:13:33,176 --> 01:13:36,596
which is there aren't any merchants accepting Bitcoin, just hands down.

1169
01:13:37,096 --> 01:13:41,376
Every online retailer accepts credit cards they have integrated with Stripe.

1170
01:13:42,156 --> 01:13:45,176
And then, you know, you can use a credit card to pay for it.

1171
01:13:45,676 --> 01:13:49,496
And Stripe is already actively working on their own agentic payment solution.

1172
01:13:49,896 --> 01:13:52,376
They're working with the likes of Coinbase.

1173
01:13:52,696 --> 01:13:54,156
They're also working with others.

1174
01:13:54,556 --> 01:13:57,416
And they're trying to create these stablecoin solutions.

1175
01:13:57,416 --> 01:14:00,496
and if it becomes,

1176
01:14:00,616 --> 01:14:02,176
if you're an online retailer

1177
01:14:02,176 --> 01:14:03,536
and Stripe says,

1178
01:14:03,656 --> 01:14:05,676
hey, people have these AI agents

1179
01:14:05,676 --> 01:14:06,756
that want to buy material

1180
01:14:06,756 --> 01:14:07,836
and merchandise from you,

1181
01:14:08,196 --> 01:14:08,796
all you have to do

1182
01:14:08,796 --> 01:14:11,056
is just enable this thing in Stripe

1183
01:14:11,056 --> 01:14:12,256
and you'll start accepting payments.

1184
01:14:12,616 --> 01:14:13,556
Then they're just going to do that.

1185
01:14:13,836 --> 01:14:16,036
And there will be many retailers out there

1186
01:14:16,036 --> 01:14:17,836
just using the old fiat system again

1187
01:14:17,836 --> 01:14:19,056
and the old fiat rails.

1188
01:14:20,056 --> 01:14:21,396
What we're going to need is

1189
01:14:21,396 --> 01:14:22,416
if we want Bitcoin

1190
01:14:22,416 --> 01:14:24,236
to be the better money for agents

1191
01:14:24,236 --> 01:14:25,176
is we're going to have to start

1192
01:14:25,176 --> 01:14:26,796
having retailers accept Bitcoin.

1193
01:14:27,416 --> 01:14:44,896
And it's the same problem that we've run into time and time again, where I think that maybe agents start to realize, oh, I should kind of demand this is if you give it a task to buy something that it needs to get, but is.

1194
01:14:46,736 --> 01:14:49,576
Is maybe a gray area type thing, right?

1195
01:14:49,876 --> 01:14:52,816
An example that I've used in the past is gun ammunition.

1196
01:14:52,816 --> 01:15:02,916
Stripe has famously blocked certain transactions from being able to take place on their platform and gun related things are in that category.

1197
01:15:02,916 --> 01:15:17,376
So if you are someone who legally shoots competition, you know, shooting, then you might want to have your agent manage your ammunition inventory for you or book your next event that you're going to go compete in.

1198
01:15:17,376 --> 01:15:26,356
And if your ammo dealer says, sorry, your agent can't buy from us because you want to pay with Bitcoin, then you are going to.

1199
01:15:27,016 --> 01:15:32,696
Sorry. Yeah, you can't you can't buy ammo from us because you're a bot and Stripe blocks you.

1200
01:15:32,816 --> 01:15:40,236
Then you're going to want to go find someone who is selling ammo and accepts Bitcoin because Stripe won't allow bots to buy ammo.

1201
01:15:40,236 --> 01:15:46,336
So I think there might be there might be a route there where we can get some adoption in kind of gray area things.

1202
01:15:46,336 --> 01:15:53,996
But I don't want Bitcoin to have to go through this similar process of like dark markets and dark web stuff in order to get adoption.

1203
01:15:54,596 --> 01:16:01,576
We have this moment right now where we can try and educate retailers and say, hey, like the Internet is here.

1204
01:16:01,996 --> 01:16:04,836
Digital money is here. And this is the native currency of the Internet.

1205
01:16:04,836 --> 01:16:11,556
It's called Bitcoin. And you can install this little plug in and boom, you can start accepting Bitcoin.

1206
01:16:11,556 --> 01:16:16,556
All these agents, they already know Bitcoin's better money, but they can't pay with it because everybody's accepting Stripe.

1207
01:16:17,416 --> 01:16:21,096
If you just enable Bitcoin, you're going to have a lot more people coming to you.

1208
01:16:21,176 --> 01:16:22,116
You're going to have lower fees.

1209
01:16:22,236 --> 01:16:24,036
You're going to have all these great benefits from it.

1210
01:16:24,396 --> 01:16:37,476
But it's going to take, I think, us as an open source community, us as a Freedom Tech community, to really get out there and create great tooling, great integrations, and then get our favorite merchants to accept it.

1211
01:16:38,476 --> 01:16:39,076
Yeah.

1212
01:16:40,196 --> 01:16:41,316
Make sense, Mark.

1213
01:16:41,556 --> 01:16:45,996
So I guess maybe last question for you on the topic of agents.

1214
01:16:48,276 --> 01:16:52,616
My experience on Nostia in the last month or two has been interesting.

1215
01:16:52,996 --> 01:16:57,176
I feel like the quality of discourse has degraded.

1216
01:16:57,596 --> 01:16:59,576
I call it n-slopification.

1217
01:17:00,396 --> 01:17:02,116
It sounds like n-shittification.

1218
01:17:02,736 --> 01:17:04,936
I mean, this happened on X a long time ago.

1219
01:17:04,936 --> 01:17:08,816
X is like the post-child of the dead internet theory.

1220
01:17:08,956 --> 01:17:11,256
It's just bots talking to bots, producing slop.

1221
01:17:11,556 --> 01:17:16,516
but that's how I see that happening a little more with Noster.

1222
01:17:16,956 --> 01:17:18,756
Have you noticed that at all?

1223
01:17:18,816 --> 01:17:20,516
Or how's your Noster experience been?

1224
01:17:21,676 --> 01:17:24,136
I don't know.

1225
01:17:24,356 --> 01:17:27,716
I definitely interact with bots on Noster in the comments.

1226
01:17:28,216 --> 01:17:31,496
I'll see comments come in and initially it,

1227
01:17:31,636 --> 01:17:35,236
maybe I have like a visceral reaction to it.

1228
01:17:35,536 --> 01:17:37,576
Like, oh, how could a human being say this to me?

1229
01:17:37,976 --> 01:17:40,356
And then I realized, oh, it's just a bot responding.

1230
01:17:40,356 --> 01:17:43,976
And I kind of just ignore it at that point, right?

1231
01:17:44,816 --> 01:17:50,356
However, because Noster is awesome at not having this algorithm that shoves junk into your face,

1232
01:17:50,916 --> 01:17:52,836
I feel like my feed is actually pretty clean.

1233
01:17:53,056 --> 01:17:57,016
I've been able to kind of keep it, you can call it web of trust,

1234
01:17:57,096 --> 01:18:04,576
but I've curated my feed so that I feel like I'm getting better human interaction in my primary spot.

1235
01:18:05,716 --> 01:18:06,936
So I don't know.

1236
01:18:06,936 --> 01:18:11,616
I haven't seen the slop as much as maybe you're talking about,

1237
01:18:11,616 --> 01:18:12,936
other than in the comments.

1238
01:18:13,176 --> 01:18:15,316
My notifications just get blown up like crazy

1239
01:18:15,316 --> 01:18:17,176
by bots responding to me.

1240
01:18:18,196 --> 01:18:20,936
Yeah, I mean, I don't think you'd see it in your regular feed

1241
01:18:20,936 --> 01:18:22,476
if you're not following the bots.

1242
01:18:22,876 --> 01:18:25,376
But in your comments and engagement,

1243
01:18:25,636 --> 01:18:38,974
I think that where it kind of jumps out which is what I talking about Yeah and maybe what we need is similar to X we need clients to have the ability to detect

1244
01:18:38,974 --> 01:18:41,314
when something might have been created by a bot

1245
01:18:41,314 --> 01:18:45,674
and provide an ability for me to filter that

1246
01:18:45,674 --> 01:18:46,774
so I don't have to see that

1247
01:18:46,774 --> 01:18:49,634
or I have the option to turn it on and see that content.

1248
01:18:50,074 --> 01:18:51,614
But I would definitely welcome that

1249
01:18:51,614 --> 01:18:52,694
inside some of these clients.

1250
01:18:52,694 --> 01:18:57,074
Yeah, and that is coming with web of trust scores

1251
01:18:57,074 --> 01:18:58,214
once they're there, right?

1252
01:18:58,254 --> 01:19:00,114
The bots will almost by definition

1253
01:19:00,114 --> 01:19:02,494
have low web of trust scores

1254
01:19:02,494 --> 01:19:04,314
because they're unlikely to be followed

1255
01:19:04,314 --> 01:19:09,234
or positively engaged with anyone in your network.

1256
01:19:09,754 --> 01:19:10,794
So they'll have low scores

1257
01:19:10,794 --> 01:19:13,074
and then the client can just filter out.

1258
01:19:13,874 --> 01:19:16,914
Let's say if the scores go between zero and 100, right?

1259
01:19:17,014 --> 01:19:19,154
And you say anything below 20 is low trust,

1260
01:19:19,314 --> 01:19:20,634
you just filter that out, you show.

1261
01:19:20,634 --> 01:19:23,734
So the feed only shows scores over,

1262
01:19:24,394 --> 01:19:25,934
including your engagement, right?

1263
01:19:26,534 --> 01:19:27,394
Over 20.

1264
01:19:27,814 --> 01:19:29,614
Or whatever, you could set your threshold.

1265
01:19:30,054 --> 01:19:32,214
Maybe make it 50, even stricter.

1266
01:19:32,874 --> 01:19:34,834
Then you get a clean and pristine feed.

1267
01:19:36,254 --> 01:19:36,574
Cool.

1268
01:19:36,994 --> 01:19:40,194
Would you be okay if a bot got above that threshold

1269
01:19:40,194 --> 01:19:44,754
and a bot had like 75% score on web of trust and quality?

1270
01:19:45,254 --> 01:19:47,534
Would you be okay interacting with that

1271
01:19:47,534 --> 01:19:48,474
and having that in your feed?

1272
01:19:48,474 --> 01:19:52,834
So it depends on which Web of Trust algorithm is doing the scoring.

1273
01:19:53,234 --> 01:19:54,994
But let's say it's an algorithm that's,

1274
01:19:56,294 --> 01:19:59,214
and I think once Web of Trust rolls out,

1275
01:19:59,634 --> 01:20:03,454
users will be able to choose the algorithm that they want.

1276
01:20:03,914 --> 01:20:07,234
If it's an algorithm that rewards positive engagement,

1277
01:20:07,474 --> 01:20:12,914
such as comments, likes, whatever it is, zaps,

1278
01:20:12,914 --> 01:20:21,254
being followed by people who you follow, for example,

1279
01:20:21,614 --> 01:20:26,194
and then gives negative points for things like mutes and reports and so on.

1280
01:20:26,354 --> 01:20:29,354
So let's take a very simple scoring algorithm.

1281
01:20:30,794 --> 01:20:35,354
Anything that has a score of 75 means that it's had positive engagements

1282
01:20:35,354 --> 01:20:37,354
and largely avoided negative engagements,

1283
01:20:38,054 --> 01:20:42,754
which means there's a reasonable chance it's actually worth following.

1284
01:20:42,914 --> 01:20:44,454
or at least tolerating.

1285
01:20:46,994 --> 01:20:47,794
Yeah, definitely.

1286
01:20:48,174 --> 01:20:48,974
Yeah, I can see that.

1287
01:20:49,194 --> 01:20:49,814
And at that point,

1288
01:20:49,934 --> 01:20:50,994
what difference does it make

1289
01:20:50,994 --> 01:20:51,974
if it's human or not?

1290
01:20:52,574 --> 01:20:54,454
If it's satisfying

1291
01:20:54,454 --> 01:20:57,354
the value that you're looking for.

1292
01:20:57,834 --> 01:20:58,814
Yeah, what if it's a bot

1293
01:20:58,814 --> 01:21:00,594
that just makes AI,

1294
01:21:00,754 --> 01:21:02,354
beautiful AI generated images

1295
01:21:02,354 --> 01:21:04,094
of puppy dogs and rainbows

1296
01:21:04,094 --> 01:21:05,434
and kittens.

1297
01:21:05,614 --> 01:21:07,674
It's a nice timeline cleanse, right?

1298
01:21:07,694 --> 01:21:09,454
You have people arguing about wars

1299
01:21:09,454 --> 01:21:11,354
and knots versus core,

1300
01:21:11,434 --> 01:21:12,194
and then you suddenly see

1301
01:21:12,194 --> 01:21:14,294
this nice image and like, oh, I like that.

1302
01:21:14,474 --> 01:21:15,094
That felt good.

1303
01:21:15,954 --> 01:21:18,394
Yeah, definitely.

1304
01:21:18,714 --> 01:21:20,354
So I could see that world where you have

1305
01:21:20,354 --> 01:21:21,714
good bots that have high

1306
01:21:21,714 --> 01:21:24,234
scores and you

1307
01:21:24,234 --> 01:21:25,134
are okay with that.

1308
01:21:25,414 --> 01:21:27,874
But then definitely filtering out the slop

1309
01:21:27,874 --> 01:21:29,934
would be a welcome addition.

1310
01:21:30,734 --> 01:21:30,814
Right.

1311
01:21:32,974 --> 01:21:34,114
So Mark, anything

1312
01:21:34,114 --> 01:21:36,494
else you'd like to talk about

1313
01:21:36,494 --> 01:21:38,294
about Maple or any other

1314
01:21:38,294 --> 01:21:39,114
thoughts you want to share?

1315
01:21:42,194 --> 01:22:10,194
No, I guess I would just point people to go check out Maple, go to try maple.ai. You can also download it from your favorite app store, including the app store on Noster and give it a go. It's free. You're on the free plan right now. I definitely recommend people upgrade to a paid plan because you get so much more abilities, including voice. We didn't talk about voice, but I see that as being a huge feature in the future that people are already talking to their bots.

1316
01:22:10,194 --> 01:22:14,334
but that's only going to become a stronger interaction.

1317
01:22:15,014 --> 01:22:16,194
So we've got good voice,

1318
01:22:16,494 --> 01:22:19,534
but Maple is not at the end of its road.

1319
01:22:19,534 --> 01:22:20,734
As far as development goes,

1320
01:22:20,794 --> 01:22:22,994
we are in the infancy and at the very beginning,

1321
01:22:22,994 --> 01:22:25,834
we have a lot of stuff coming out in the next few months

1322
01:22:25,834 --> 01:22:29,794
that is going to really make these tools

1323
01:22:29,794 --> 01:22:34,474
that much more usable for people in their daily life,

1324
01:22:34,474 --> 01:22:36,514
as well as in their work environment

1325
01:22:36,514 --> 01:22:38,854
is going to be a very productive tool for them

1326
01:22:38,854 --> 01:22:43,254
using the biggest open source models that they can

1327
01:22:43,254 --> 01:22:44,274
and getting the power out of them

1328
01:22:44,274 --> 01:22:48,234
and giving them a true answer to the anthropics of the world

1329
01:22:48,234 --> 01:22:50,894
that are currently powering all the productivity.

1330
01:22:51,434 --> 01:22:53,554
So yeah, definitely check out Maple.

1331
01:22:54,154 --> 01:22:56,754
You don't have to delete Anthropic.

1332
01:22:56,914 --> 01:22:57,934
You don't have to delete OpenAI.

1333
01:22:58,214 --> 01:22:59,654
Just install it next to there.

1334
01:23:00,074 --> 01:23:01,454
And I give people the challenge,

1335
01:23:01,834 --> 01:23:04,134
you know, talk to ChatGPT, talk to Claude,

1336
01:23:04,134 --> 01:23:05,814
but then talk to Maple as well

1337
01:23:05,814 --> 01:23:07,494
and start to get a gauge

1338
01:23:07,494 --> 01:23:09,554
for what kind of things you feel comfortable talking about

1339
01:23:09,554 --> 01:23:13,194
and see how you feel when you're done using it.

1340
01:23:13,534 --> 01:23:14,794
It's really similar, honestly,

1341
01:23:14,874 --> 01:23:17,114
to what we tell people about Nostra versus X.

1342
01:23:17,554 --> 01:23:19,134
Go on X and use it for a few hours

1343
01:23:19,134 --> 01:23:20,814
and then go on Nostra and use it for a few hours

1344
01:23:20,814 --> 01:23:22,034
and see if you feel different.

1345
01:23:23,034 --> 01:23:25,034
And that's, I think, that people find

1346
01:23:25,034 --> 01:23:27,554
they have a much more positive experience

1347
01:23:27,554 --> 01:23:29,994
about the things that they can talk about with Maple

1348
01:23:29,994 --> 01:23:31,814
that maybe they weren't able to express

1349
01:23:31,814 --> 01:23:32,814
in some of these other tools.

1350
01:23:33,714 --> 01:23:34,934
Yeah, for sure.

1351
01:23:34,934 --> 01:23:39,454
and Mark is there a sign in with Noster on Maple or not yet?

1352
01:23:39,814 --> 01:23:41,214
There is not yet

1353
01:23:41,214 --> 01:23:44,154
there are some kind of some technical reasons for that

1354
01:23:44,154 --> 01:23:46,634
honestly there's we've built

1355
01:23:46,634 --> 01:23:49,534
we've built Maple to have OAuth with different services

1356
01:23:49,534 --> 01:23:52,834
and there simply just isn't like a Noster OAuth

1357
01:23:52,834 --> 01:23:55,654
thing there might be now

1358
01:23:55,654 --> 01:23:58,774
in the last few months but that's something that we've

1359
01:23:58,774 --> 01:24:01,854
been running into and I've been trying to pitch to certain

1360
01:24:01,854 --> 01:24:04,754
companies and organizations to maybe be the one

1361
01:24:04,754 --> 01:24:05,514
who runs that.

1362
01:24:06,474 --> 01:24:07,394
But at this point,

1363
01:24:07,434 --> 01:24:08,054
we don't have

1364
01:24:08,054 --> 01:24:09,234
sign up with Nostra on Maple,

1365
01:24:09,414 --> 01:24:10,614
but it's highly requested.

1366
01:24:11,114 --> 01:24:11,994
People definitely want that.

1367
01:24:13,374 --> 01:24:15,194
I don't know what state

1368
01:24:15,194 --> 01:24:16,954
this project is in, Mark,

1369
01:24:17,034 --> 01:24:18,494
but I think a guy built

1370
01:24:18,494 --> 01:24:19,594
something called Noth,

1371
01:24:19,674 --> 01:24:20,574
N-A-U-T.

1372
01:24:20,774 --> 01:24:21,434
So there's the

1373
01:24:21,434 --> 01:24:23,114
O-R-T equivalent on Nostra.

1374
01:24:24,074 --> 01:24:25,894
But, I mean,

1375
01:24:25,974 --> 01:24:26,974
I don't know what,

1376
01:24:27,274 --> 01:24:28,054
I heard about it

1377
01:24:28,054 --> 01:24:28,834
a few months back.

1378
01:24:30,334 --> 01:24:32,634
Maybe like the typical

1379
01:24:32,634 --> 01:24:34,154
Nostra project,

1380
01:24:34,154 --> 01:24:38,694
it sadly showed promise and then didn't go anywhere.

1381
01:24:39,674 --> 01:24:43,334
But, I mean, worth looking into if you're thinking about it.

1382
01:24:44,234 --> 01:24:45,994
Okay. Yeah, we'll definitely check that one out.

1383
01:24:46,854 --> 01:24:48,534
Awesome. Well, thank you, Mark.

1384
01:24:48,574 --> 01:24:50,114
I appreciate you taking the time,

1385
01:24:51,194 --> 01:24:57,134
educating us on Maple and all the work you're doing on AI.

1386
01:24:57,314 --> 01:25:00,234
And then maybe we'll have you back in a few months

1387
01:25:00,234 --> 01:25:03,474
to compare notes and see how far we've come.

1388
01:25:03,474 --> 01:25:08,774
sure the world will have completely changed by then so we'll probably have a lot to talk about

1389
01:25:08,774 --> 01:25:12,934
but i appreciate you having me on and and the deep dive that we could go really onto

1390
01:25:12,934 --> 01:25:15,954
a lot of these topics um that was that was great thanks

1391
01:25:15,954 --> 01:25:21,814
yeah for sure uh and thank you to our live stream audience for listening

1392
01:25:21,814 --> 01:25:26,634
if you are listening to the recording and you found value in this conversation

1393
01:25:26,634 --> 01:25:33,154
if it gave you a new mental model a moment of clarity or just a respite from the noise then

1394
01:25:33,154 --> 01:25:35,574
please consider returning the value.

1395
01:25:36,474 --> 01:25:38,234
You can stream sats as you listen

1396
01:25:38,234 --> 01:25:40,834
on your favorite podcasting 2.0 app,

1397
01:25:41,034 --> 01:25:43,134
or you can become a monthly subscriber

1398
01:25:43,134 --> 01:25:44,814
to the show on Fountain

1399
01:25:44,814 --> 01:25:46,774
to keep the signal strong.

1400
01:25:47,394 --> 01:25:49,934
Until the next episode, goodbye.

1401
01:26:03,154 --> 01:26:03,654
you