1
00:00:00,320 --> 00:00:05,680
All right, welcome to the Neuvottelija channel, Aleksi Paavola.

2
00:00:05,680 --> 00:00:09,400
Thanks so much. And we didn't really know each other a week ago,

3
00:00:09,400 --> 00:00:14,599
we'd seen each other's conversations over at Agentics Finland a few times, but,

4
00:00:14,599 --> 00:00:18,680
you'd done such a great piece of coding that I thought,

5
00:00:18,680 --> 00:00:23,199
get this guy into the studio right away — and luckily it worked out. So could you

6
00:00:23,199 --> 00:00:28,080
start by telling us a bit about your background — what you do and where you're coming from?

7
00:00:28,080 --> 00:00:35,800
Yeah, absolutely, great to be here. My name is indeed Aleksi Paavola, and I've been

8
00:00:35,800 --> 00:00:41,160
in tech for a good 10 years — actually probably closer to 15 by now —

9
00:00:41,160 --> 00:00:47,160
as a tech entrepreneur basically my whole career, and I've founded and sold a couple of

10
00:00:47,160 --> 00:00:54,480
tech companies, and for the last four years or so I've been building

11
00:00:54,480 --> 00:00:58,239
AI stuff, and also doing software development and

12
00:00:58,239 --> 00:01:02,280
software consulting. Now mainly around

13
00:01:02,280 --> 00:01:07,799
these language models, on top of and around them, all sorts of things — I've done quite a lot,

14
00:01:07,799 --> 00:01:13,600
been in Saudi Arabia running trainings, and I've also

15
00:01:13,600 --> 00:01:18,759
trained Google's Europe FINA team on language models, [chuckle] and then I've also been in Oslo

16
00:01:18,759 --> 00:01:26,079
representing Google's language model cloud solutions at

17
00:01:26,079 --> 00:01:30,720
a seminar with over 10,000 people, and there's been

18
00:01:30,720 --> 00:01:36,720
all sorts of interesting stuff. Cool — actually I had Topi Manuas

19
00:01:36,720 --> 00:01:40,600
as a guest. He moved to Singapore and he's been doing Google's cloud sales,

20
00:01:40,600 --> 00:01:44,680
and we talked about Google's multimodal solution. Now I'm

21
00:01:44,680 --> 00:01:49,079
well, even though we probably both use Claude more,

22
00:01:49,079 --> 00:01:55,119
and Codex — but somehow the Gemini folks invited me along there,

23
00:01:55,119 --> 00:01:58,880
when they launched Gemini 3.0, we were at this really exclusive

24
00:01:58,880 --> 00:02:03,439
event — it was champagne — and they released the model an hour before the global

25
00:02:03,439 --> 00:02:06,840
release, and I've always had this soft spot for

26
00:02:06,840 --> 00:02:11,720
bribery, so Google feels like a really nice company, and they have in Finland,

27
00:02:11,720 --> 00:02:15,160
in Hamina, real investment going on — yeah, that's actually where

28
00:02:15,160 --> 00:02:20,400
in Hamina I was running this training — Google's Europe FINA team

29
00:02:20,400 --> 00:02:25,360
gathered in Hamina for a seminar like this, and we went through a bit of

30
00:02:25,360 --> 00:02:29,720
the basics of language models. Hey, that's awesome. You're honestly a bigger deal

31
00:02:29,720 --> 00:02:33,440
than I thought — I thought you were sort of roughly on my

32
00:02:33,440 --> 00:02:36,280
level, but now it's getting a little scary. But anyway, the reason I

33
00:02:36,280 --> 00:02:40,360
invited you here wasn't about Google at all — it's that you've built this

34
00:02:40,360 --> 00:02:44,680
open-source project, this better

35
00:02:44,680 --> 00:02:50,319
Model Context Protocol tool called Laki.ai, and I started using it right away,

36
00:02:50,319 --> 00:02:54,120
and it works great. Before that I had something else filling that gap,

37
00:02:54,120 --> 00:02:57,280
something called Oik, but that was clearly worse. So you've built

38
00:02:57,280 --> 00:03:01,680
this better tool, but explain — what exactly is Laki.ai, and what

39
00:03:01,680 --> 00:03:06,040
is MCP? Right. Thanks a lot for the kind words. It's

40
00:03:06,040 --> 00:03:11,120
come up from other places too, actually — that listeners should really go

41
00:03:11,120 --> 00:03:17,280
and try this out for sure — but, um, this project actually

42
00:03:17,280 --> 00:03:21,840
got started with Ilkka Linkoneva and Juuso Vesanto —

43
00:03:21,840 --> 00:03:29,599
we tried to solve this challenge, where

44
00:03:29,599 --> 00:03:33,239
we'd identified that language models were going to produce a huge amount of this

45
00:03:33,239 --> 00:03:37,239
legal documentation. And then people would still need to read and

46
00:03:37,239 --> 00:03:41,640
understand it. So we developed a solution for that. This was already a few

47
00:03:41,640 --> 00:03:46,480
years ago now. And our hypothesis was that if we fed into our

48
00:03:46,480 --> 00:03:52,840
system, say, 100 NDAs, a lawyer could then highlight

49
00:03:52,840 --> 00:03:58,280
some section in Word from a new NDA that they've

50
00:03:58,280 --> 00:04:05,000
gotten from a client, and then see, from the ones they'd previously approved themselves, um,

51
00:04:05,000 --> 00:04:09,120
the corresponding section in those NDA documents that's semantically closest

52
00:04:09,120 --> 00:04:16,320
to it, and from that easily approve it, like okay, I've

53
00:04:16,320 --> 00:04:21,959
approved this kind of clause on basically the same terms before, so this is definitely fine. Well,

54
00:04:21,959 --> 00:04:27,160
the surprising challenge that came up here was

55
00:04:27,160 --> 00:04:33,160
lawyers' ability to express things in a million different

56
00:04:33,160 --> 00:04:41,400
ways. [chuckle] So these ended up just full of redlines. So now

57
00:04:41,400 --> 00:04:46,199
when you — like in that test we had, we had about 50 of these

58
00:04:46,199 --> 00:04:50,880
different NDA agreements, and when we took a new NDA and you

59
00:04:50,880 --> 00:04:53,720
highlighted some section in Word, it was always just

60
00:04:53,720 --> 00:04:59,160
red with redlines. The phrasing and the terms somehow

61
00:04:59,160 --> 00:05:02,960
these lawyers had worded so differently that the tool didn't end up being very

62
00:05:02,960 --> 00:05:10,800
useful. Well, around the same time this pretty fortunate thing happened,

63
00:05:10,800 --> 00:05:16,479
from the perspective of Finnish citizens, actually — Finlex,

64
00:05:16,479 --> 00:05:20,953
its data became openly available for everyone to use. So

65
00:05:20,953 --> 00:05:25,800
[chuckle] this had been — and this was, this was March 2015,

66
00:05:25,800 --> 00:05:31,440
I think it was, and before that, I remember there was some big software house

67
00:05:31,440 --> 00:05:35,360
that was doing an AI project with Finlex's data, and

68
00:05:35,360 --> 00:05:38,400
it fell through over this licensing issue. I think Alma

69
00:05:38,400 --> 00:05:41,600
Media was holding onto it too, too tightly,

70
00:05:41,600 --> 00:05:46,600
but then in March it got freed up, and our idea was that

71
00:05:46,600 --> 00:05:52,680
we'd build an interface — like a better interface — for Finlex's data, um,

72
00:05:52,680 --> 00:05:56,919
with no commercial purpose at all. Just mainly because we

73
00:05:56,919 --> 00:06:01,520
could, so we'd do it, and we figured this was going to be a hugely

74
00:06:01,520 --> 00:06:08,400
useful service for citizens, of course, and so we built Laki.ai,

75
00:06:08,400 --> 00:06:14,080
launched it back in 2015 sometime in summer, and it's still — it's just a

76
00:06:14,080 --> 00:06:19,199
traditional web service for Finlex's data. So it has

77
00:06:19,199 --> 00:06:23,280
laws, statutes,

78
00:06:23,280 --> 00:06:28,880
there are court cases, government bills, and what sets it apart

79
00:06:28,880 --> 00:06:33,720
from Finlex itself is that these documents are all linked

80
00:06:33,720 --> 00:06:37,400
to each other. So when you open a specific section of law there,

81
00:06:37,400 --> 00:06:43,560
for example, you immediately see, okay, this is linked to these government

82
00:06:43,560 --> 00:06:48,360
bills, and then this is linked to, say, these seven court cases. Mm.

83
00:06:48,360 --> 00:06:53,000
So that's basically what we built, without any bigger plan for

84
00:06:53,000 --> 00:06:57,639
anything beyond that. Well, now, fairly recently,

85
00:06:57,639 --> 00:07:03,199
um, we had this idea — well,

86
00:07:03,199 --> 00:07:08,374
could we build something for an agent on top of this, and

87
00:07:08,374 --> 00:07:11,039
[clears throat] we looked into it a bit and concluded that

88
00:07:11,039 --> 00:07:17,400
yeah, this would definitely be possible, and that's what got

89
00:07:17,400 --> 00:07:23,199
this Laki.ai MCP started. And what that actually means in practice

90
00:07:23,199 --> 00:07:29,080
is that now, with your AI tool of choice, you can connect it

91
00:07:29,080 --> 00:07:35,599
to this Laki.ai database, and now your agent

92
00:07:35,599 --> 00:07:43,280
can use all this legal data that we have there

93
00:07:43,280 --> 00:07:49,039
in the Laki.ai database. Yeah, yeah, it got straight into my Claude, uh,

94
00:07:49,039 --> 00:07:53,759
tech stack — though there was a bit of fiddling involved, actually,

95
00:07:53,759 --> 00:07:58,639
let's just say, right at the start we'd invited — along with Aku Nikkola —

96
00:07:58,639 --> 00:08:03,400
from Legit here — he, well, he wanted to spend, understandably, a long stretch

97
00:08:03,400 --> 00:08:07,440
of summer vacation afterward, so he couldn't make it here, but around the same

98
00:08:07,440 --> 00:08:13,240
time he'd released this — Claude has this legal

99
00:08:13,240 --> 00:08:17,560
skills thing, which is kind of Anglo-Saxon know-how, and he'd then

100
00:08:17,560 --> 00:08:22,599
built, with his team, a kind of skills library that has

101
00:08:22,599 --> 00:08:27,120
the special features of Finnish law — starting from the language, like how a section

102
00:08:27,120 --> 00:08:31,400
works, and so on [chuckle], and he published it. I took

103
00:08:31,400 --> 00:08:34,719
that into use too. There was a bit of fiddling involved, actually, because for some reason Claude

104
00:08:34,719 --> 00:08:38,200
had changed the connectors a bit, so I had to go down to the command-line

105
00:08:38,200 --> 00:08:42,399
level to shove it in, which was a bit annoying. But when you

106
00:08:42,399 --> 00:08:45,959
combine these two skills — this skills library, which in my

107
00:08:45,959 --> 00:08:48,959
opinion is actually pretty good, I'd recommend that too — and then this

108
00:08:48,959 --> 00:08:53,320
Laki.ai, you get a really nice combo for Finnish

109
00:08:53,320 --> 00:08:57,320
legal needs. I don't know, have you had a chance to test Legit's

110
00:08:57,320 --> 00:09:00,560
skills yet? Yeah, definitely have. And maybe

111
00:09:00,560 --> 00:09:05,480
I could add a bit more, since you asked, to open up

112
00:09:05,480 --> 00:09:09,519
a little more for listeners — MCP isn't necessarily super familiar to everyone.

113
00:09:09,519 --> 00:09:15,440
So it's an open-source protocol that Anthropic — the company behind

114
00:09:15,440 --> 00:09:21,200
Claude — built and designed, and it stands for Model Context

115
00:09:21,200 --> 00:09:31,240
Protocol — that's what MCP is short for — and in practice it's a way to connect

116
00:09:31,240 --> 00:09:36,680
different applications for agents to use. So maybe the closest

117
00:09:36,680 --> 00:09:40,480
example would be something like USB-C. So if you think about your

118
00:09:40,480 --> 00:09:43,959
computer, and you want to plug a keyboard into it, you plug

119
00:09:43,959 --> 00:09:49,040
the USB-C cable in between them, and then, simsalabim,

120
00:09:49,040 --> 00:09:53,839
you can use that keyboard with the computer. So this is

121
00:09:53,839 --> 00:09:58,600
basically kind of the same idea — when you connect your

122
00:09:58,600 --> 00:10:03,120
Claude, or soon ChatGPT too, which we

123
00:10:03,120 --> 00:10:08,000
have coming for Laki.ai — then

124
00:10:08,000 --> 00:10:13,480
your Claude doesn't need to

125
00:10:13,480 --> 00:10:18,200
hunt for legal sources on the open internet anymore, which is pretty difficult and

126
00:10:18,200 --> 00:10:23,279
easily leads to errors — instead it can now use this database

127
00:10:23,279 --> 00:10:27,440
we've built. Yeah, that USB comparison is good. So

128
00:10:27,440 --> 00:10:34,320
as I see it, MCP is kind of a more refined thing than

129
00:10:34,320 --> 00:10:38,120
a regular API call — Application Programming Interface — and the problem with those was

130
00:10:38,120 --> 00:10:42,360
that everyone called them a bit their own way, like pulling

131
00:10:42,360 --> 00:10:46,720
data from HubSpot into Word or whatever — so it kind of created

132
00:10:46,720 --> 00:10:51,079
this sort of port standard, and it eats up some bandwidth, and it's

133
00:10:51,079 --> 00:10:55,959
a bit too complicated for really fast calls,

134
00:10:55,959 --> 00:10:58,839
but it opens up all the interfaces really nicely, so —

135
00:10:58,839 --> 00:11:02,200
did I get that roughly right? Yeah, pretty much. It's just

136
00:11:02,200 --> 00:11:08,440
that, um, behind MCP there's basically always APIs. So it's a

137
00:11:08,440 --> 00:11:11,680
protocol built on top of those APIs, designed

138
00:11:11,680 --> 00:11:15,560
specifically so agents can use them efficiently. And

139
00:11:15,560 --> 00:11:20,552
yeah, what you said is completely true, that there's been

140
00:11:20,552 --> 00:11:23,959
[chuckle] — you said there are some challenges too, and that's completely true. And

141
00:11:23,959 --> 00:11:29,560
MCPs have gotten a lot of criticism exactly for filling up the

142
00:11:29,560 --> 00:11:33,959
agents' or language models' context window, but these have

143
00:11:33,959 --> 00:11:37,880
improved quite a bit, and what's actually improved the most,

144
00:11:37,880 --> 00:11:44,360
is agents' ability to actually make use of these different MCP

145
00:11:44,360 --> 00:11:49,519
services, and it seems like — I myself was fairly skeptical at one point

146
00:11:49,519 --> 00:11:54,079
about the future prospects of the MCP protocol.

147
00:11:54,079 --> 00:12:00,360
But now I'm a bit more positive again, it seems like maybe we're finding

148
00:12:00,360 --> 00:12:05,920
solutions where these agents can now efficiently make use of these

149
00:12:05,920 --> 00:12:10,800
MCP connections and services that

150
00:12:10,800 --> 00:12:14,680
are hooked up via MCP. Yeah. And in a way, since

151
00:12:14,680 --> 00:12:18,240
the level of inference, or AI capability, keeps rising

152
00:12:18,240 --> 00:12:21,160
all the time, it's not really a problem to build even really weird

153
00:12:21,160 --> 00:12:27,560
pipelines. Like, AI will build basically any kind of, uh [laughter]

154
00:12:27,560 --> 00:12:31,040
ancient pipe between some old database and a Cobol system, as long as you just

155
00:12:31,040 --> 00:12:34,600
shove it in there, because they're just capable of that — but those are fragile, and MCP

156
00:12:34,600 --> 00:12:38,160
is a standard, and standards are good for a reason, because they

157
00:12:38,160 --> 00:12:41,279
evolve too, in that way — and it's nice to hear that it's also become

158
00:12:41,279 --> 00:12:43,800
more efficient, that the standard itself has developed along the

159
00:12:43,800 --> 00:12:49,040
way. Yeah — you say 'agent' a lot, and actually I myself

160
00:12:49,040 --> 00:12:55,079
had a little chat about this over lunch earlier, about whether

161
00:12:55,079 --> 00:12:58,880
agents — in the sense that I want to understand, like, this same kind of

162
00:12:58,880 --> 00:13:01,920
AI employee — are they really that useful? Maybe

163
00:13:01,920 --> 00:13:05,959
we should save that conversation for later. I think we have a bit of a different view on that.

164
00:13:05,959 --> 00:13:10,760
Um, because personally, what you might

165
00:13:10,760 --> 00:13:15,040
call an agent — like some kind of thing sitting next to an MCP,

166
00:13:15,040 --> 00:13:19,000
some subroutine call — I wouldn't even publicly call that an agent. It's

167
00:13:19,000 --> 00:13:23,160
just some subroutine running there. Not an agent at all. Whereas these

168
00:13:23,160 --> 00:13:27,440
agents are more like same-level AI employees, almost god-

169
00:13:27,440 --> 00:13:31,680
like beings that can do high-cognition work,

170
00:13:31,680 --> 00:13:35,880
not just sit in between two things. I don't know

171
00:13:35,880 --> 00:13:38,720
if you want to jump in on that already, [laughter]

172
00:13:38,720 --> 00:13:45,279
I can jump in a little, um, to the extent that, if we

173
00:13:45,279 --> 00:13:48,760
think about the direction ChatGPT and

174
00:13:48,760 --> 00:13:54,120
Claude have moved in, they've become a lot more agentic. If you think about

175
00:13:54,120 --> 00:14:01,000
how, say, the very first ChatGPT, back in November '22,

176
00:14:01,000 --> 00:14:05,399
that released version worked, it was always just — you'd write some

177
00:14:05,399 --> 00:14:09,160
prompt, the prompt got sent to the language model, and the language model

178
00:14:09,160 --> 00:14:13,959
sent back some answer. But now — and the way I think about it —

179
00:14:13,959 --> 00:14:20,399
the difference between a language model and an AI agent is that in an agent

180
00:14:20,399 --> 00:14:24,600
the language model is used in some kind of loop. So if we think about, say,

181
00:14:24,600 --> 00:14:27,399
essay writing, a non-agentic

182
00:14:27,399 --> 00:14:30,839
way of writing an essay would be that you have a prompt, and out comes the

183
00:14:30,839 --> 00:14:35,480
finished piece the moment you feed the prompt to the language model, whereas in the agentic

184
00:14:35,480 --> 00:14:40,199
process it evaluates its own work along the way. So it writes

185
00:14:40,199 --> 00:14:44,639
a first draft, say, and then it checks, okay, maybe this

186
00:14:44,639 --> 00:14:49,000
second paragraph needs to be strengthened a bit more, or the ending is missing

187
00:14:49,000 --> 00:14:53,480
some punch — so now ChatGPT and Claude have moved

188
00:14:53,480 --> 00:14:58,000
in this direction, so I think in a way you could even call these chats

189
00:14:58,000 --> 00:15:01,120
agentic

190
00:15:01,120 --> 00:15:07,160
systems already — but of course it's true they're not agents to nearly

191
00:15:07,160 --> 00:15:10,720
the same extent as, say, Claude's Sonnet, or Codex, or

192
00:15:10,720 --> 00:15:15,839
Claude's Cowork, or Claude Code. Yeah, but that's honestly, scarily, always

193
00:15:15,839 --> 00:15:19,399
moving toward that same point — for anyone who hasn't yet

194
00:15:19,399 --> 00:15:23,279
heard the term — this kind of open AI employee that has

195
00:15:23,279 --> 00:15:28,000
its own three-part memory structure and very advanced context

196
00:15:28,000 --> 00:15:32,279
management tools — so we actually have to get into this context concept,

197
00:15:32,279 --> 00:15:36,920
because you said the MCP call eats up that precious context, and

198
00:15:36,920 --> 00:15:39,880
we're already at million-token contexts, basically as

199
00:15:39,880 --> 00:15:45,920
standard models — Claude's and, um, OpenAI's base models run

200
00:15:45,920 --> 00:15:50,279
on a million tokens, so it's less and less, in a way, that short-term memory

201
00:15:50,279 --> 00:15:54,560
eating problem here that's disappearing, and if the MCP model gets even more efficient,

202
00:15:54,560 --> 00:15:59,560
then this is a vanishing problem overall. How do you see, by the way,

203
00:15:59,560 --> 00:16:03,680
this — um, I've actually had guests who like to run

204
00:16:03,680 --> 00:16:07,600
the 'dumb' models on purpose — like, it's not a bug, it's a

205
00:16:07,600 --> 00:16:11,279
feature — that they run this 'dumb' inference on purpose because it's

206
00:16:11,279 --> 00:16:13,839
easier to understand, and it's actually good in a way, that there's not too much

207
00:16:13,839 --> 00:16:16,920
creativity or harness layered on top of it. Do you have a philosophy

208
00:16:16,920 --> 00:16:21,639
on that? Well, yes, yes [laughter], very much so, maybe, maybe

209
00:16:21,639 --> 00:16:27,000
as a basic piece of advice I'd say that people

210
00:16:27,000 --> 00:16:32,639
should, on a general level, pay for these AI services, so that

211
00:16:32,639 --> 00:16:36,040
nobody should be using the free versions of these for work — because

212
00:16:36,040 --> 00:16:40,480
paying gets you a lot of value for your money, and these are still

213
00:16:40,480 --> 00:16:43,880
relatively affordable, at least for now — though I do expect that

214
00:16:43,880 --> 00:16:48,680
prices will definitely go up, and have partly gone up already — because

215
00:16:48,680 --> 00:16:55,199
even Anthropic's and OpenAI's Enterprise plans aren't exactly

216
00:16:55,199 --> 00:16:59,360
cheap anymore. But generally speaking, I think a good

217
00:16:59,360 --> 00:17:02,920
basic recommendation is to always use the best model. And that's kind of

218
00:17:02,920 --> 00:17:08,360
something people should just know — it's like

219
00:17:08,360 --> 00:17:12,679
modern general knowledge, that you know — if you think about it,

220
00:17:12,679 --> 00:17:17,199
if you've chosen to work in, say, the Claude or Anthropic ecosystem, then

221
00:17:17,199 --> 00:17:22,319
you should understand and know what the difference is between, say, Sonnet, Opus, and Fable

222
00:17:22,319 --> 00:17:25,439
is. M. And then, as a baseline, you could

223
00:17:25,439 --> 00:17:27,679
always — if you have some difficult task, [clears throat]

224
00:17:27,679 --> 00:17:31,480
you use the best model. And then again, if you have — if you're just

225
00:17:31,480 --> 00:17:35,600
asking, hey, what are some good lunch spots nearby, then it really doesn't

226
00:17:35,600 --> 00:17:40,039
matter whether it's some super-smart model. You could ask that

227
00:17:40,039 --> 00:17:44,160
to basically anything. Yeah, yeah. That's how it is — then we can

228
00:17:44,160 --> 00:17:47,400
move on to local models, where you're not allowed to do any

229
00:17:47,400 --> 00:17:51,200
tool calls or internet searches — so it's kind of very rigid, and

230
00:17:51,200 --> 00:17:54,360
in that case it can be really important to know in what way it's

231
00:17:54,360 --> 00:17:58,760
mechanically rigid, because then it can be fully reproducible, because

232
00:17:58,760 --> 00:18:01,679
in a way, maybe this hallucination that the better

233
00:18:01,679 --> 00:18:05,240
models always get criticized for — the more harness or control layer you have in

234
00:18:05,240 --> 00:18:10,000
between, or agentic loops, or Ralph-loop envelopes, or whatever cleverness you

235
00:18:10,000 --> 00:18:15,320
build in between the model and yourself, the more

236
00:18:15,320 --> 00:18:18,280
hallucination loop comes into it too. So in a way, one solution

237
00:18:18,280 --> 00:18:21,320
is to strip all of that away, to just run it directly on bare metal

238
00:18:21,320 --> 00:18:24,440
in some task that's outside your own interest, so you know exactly

239
00:18:24,440 --> 00:18:27,440
what the model is going to do.

240
00:18:27,440 --> 00:18:32,679
Yeah, right. There's actually a great essay on this — I don't know if

241
00:18:32,679 --> 00:18:38,280
have you read it, or come across it — the Bitter Lesson? No, I haven't,

242
00:18:38,280 --> 00:18:44,200
so tell me more about this legendary Canadian

243
00:18:44,200 --> 00:18:49,039
professor who wrote, years ago now, this

244
00:18:49,039 --> 00:18:55,919
essay on AI, where he argues that the value humans

245
00:18:55,919 --> 00:18:59,280
bring to AI systems

246
00:18:59,280 --> 00:19:03,120
is, sort of,

247
00:19:03,120 --> 00:19:07,360
really limited over time, and he gives good examples. So if we think about

248
00:19:07,360 --> 00:19:15,679
chess, say — originally in chess, people

249
00:19:15,679 --> 00:19:20,159
thought that humans could bring value to

250
00:19:20,159 --> 00:19:24,159
the AI. So basically, with this harness

251
00:19:24,159 --> 00:19:31,760
you could make an even better AI. But then, from chess

252
00:19:31,760 --> 00:19:36,720
or chess history, we know very well that a human can't

253
00:19:36,720 --> 00:19:41,000
add any value to a chess AI like that. A chess AI

254
00:19:41,000 --> 00:19:46,400
on its own is much stronger than AI plus a human. And

255
00:19:46,400 --> 00:19:50,440
the AI can't be helped at all by a human giving

256
00:19:50,440 --> 00:19:54,000
it tips for that kind of chess AI. And self-

257
00:19:54,000 --> 00:20:00,320
driving is a good example too — there, first they tried doing

258
00:20:00,320 --> 00:20:04,320
sub-optimization, building [clears throat] small pieces at a time.

259
00:20:04,320 --> 00:20:07,600
Andrej Karpathy has actually talked a lot about this, about how

260
00:20:07,600 --> 00:20:12,640
Tesla's strategy — which a lot of people

261
00:20:12,640 --> 00:20:18,200
have already copied by now — of training the car to drive

262
00:20:18,200 --> 00:20:22,840
so that the neural network handles the driving has turned out to be

263
00:20:22,840 --> 00:20:28,480
a much better strategy than trying to tell it

264
00:20:28,480 --> 00:20:32,320
in specific spots, sort of steering the AI

265
00:20:32,320 --> 00:20:37,440
using human intelligence. And now I believe we'll see this same thing, and we already have,

266
00:20:37,440 --> 00:20:40,919
in a lot of language-model-based systems now. If we

267
00:20:40,919 --> 00:20:45,840
think about software development, say — in software development, at first

268
00:20:45,840 --> 00:20:51,039
there was Cursor, which was this traditional

269
00:20:51,039 --> 00:20:55,240
integrated development environment — an IDE that

270
00:20:55,240 --> 00:21:00,480
developers were used to using — and when AI was brought in and

271
00:21:00,480 --> 00:21:05,799
integrated into that IDE, for a long time it looked like, okay,

272
00:21:05,799 --> 00:21:12,279
this is going to be the way — that humans can add value

273
00:21:12,279 --> 00:21:16,640
at various points along the way in software development. I myself was actually really

274
00:21:16,640 --> 00:21:19,679
skeptical about that when Claude Code came out,

275
00:21:19,679 --> 00:21:26,640
whether it could really be true that a purely text-based interface

276
00:21:26,640 --> 00:21:30,080
could work on its own. Whereas over in Cursor there was this tab

277
00:21:30,080 --> 00:21:35,080
completion — you'd write the start of a function, and then

278
00:21:35,080 --> 00:21:40,039
the AI would predict what kind of function you wanted to write. Pretty quickly, though,

279
00:21:40,039 --> 00:21:45,120
it's gotten to the point where I don't even remember the last time I did a tab

280
00:21:45,120 --> 00:21:50,720
completion like that. So it's really shifted toward these...

281
00:21:50,720 --> 00:21:55,120
these harnesses have kind of shrunk from that, and I believe this

282
00:21:55,120 --> 00:21:58,039
trend is going to keep going, and now if we look at

283
00:21:58,039 --> 00:22:01,159
say what OpenAI's or Anthropic's employees

284
00:22:01,159 --> 00:22:04,960
say — what they talk about is exactly that they're

285
00:22:04,960 --> 00:22:07,960
constantly trying to strip that harness down.

286
00:22:07,960 --> 00:22:11,559
Mm. That is, letting the

287
00:22:11,559 --> 00:22:15,720
model take on a bigger and bigger share of it. And if you think about

288
00:22:15,720 --> 00:22:20,720
the legal field from that angle, I'm actually really skeptical

289
00:22:20,720 --> 00:22:25,440
about, say, Legora or Harvey now. They're in the same

290
00:22:25,440 --> 00:22:28,760
position Cursor was in a while back, sort of. They've

291
00:22:28,760 --> 00:22:36,600
tried to build this kind of software around

292
00:22:36,600 --> 00:22:40,480
the legal side, so they could add

293
00:22:40,480 --> 00:22:43,799
some value at certain points. Mm.

294
00:22:43,799 --> 00:22:50,720
But I think we're already at the point where if you have Claude and

295
00:22:50,720 --> 00:22:55,120
you connect it, say, with Laki.ai to

296
00:22:55,120 --> 00:23:02,200
reliable Finnish legal sources, then for pretty much

297
00:23:02,200 --> 00:23:05,840
every legal-related

298
00:23:05,840 --> 00:23:10,720
task you'll actually get better answers from Claude than from

299
00:23:10,720 --> 00:23:16,080
Legora or Harvey, which have built a lot of harness

300
00:23:16,080 --> 00:23:19,320
between the model and the user.

301
00:23:19,320 --> 00:23:23,679
Yeah, that's a harsh way to put it, but hard to disagree. Personally, I

302
00:23:23,679 --> 00:23:28,559
actually only started coding properly again this year myself. I was

303
00:23:28,559 --> 00:23:32,480
a good coder as a kid, and I've been an investment banker for 30 years now, and

304
00:23:32,480 --> 00:23:35,840
I haven't lost a thing in those 30 years, because those backslashes are still

305
00:23:35,840 --> 00:23:39,520
sitting there in the terminal waiting. And actually, I do use

306
00:23:39,520 --> 00:23:43,000
autocomplete for that — when I SSH between different computers,

307
00:23:43,000 --> 00:23:46,559
tab-complete always finds, like, what did I even name this machine, so

308
00:23:46,559 --> 00:23:49,880
I still get autocomplete out of that, but otherwise I'm pretty much full

309
00:23:49,880 --> 00:23:54,720
slash-goal mode here — I mean, it's public

310
00:23:54,720 --> 00:23:59,039
sector, make-no-mistakes territory, just slash-goal — but somewhere there's

311
00:23:59,039 --> 00:24:03,080
a line, where your goal is so complex that you

312
00:24:03,080 --> 00:24:07,679
can't just do that — this is like an extreme example,

313
00:24:07,679 --> 00:24:12,159
something like "make no mistakes, fix the world" or something huge —

314
00:24:12,159 --> 00:24:16,360
so somewhere there's still a human path running through it,

315
00:24:16,360 --> 00:24:22,360
taste and intent and the steering that's involved, and — well, somewhere,

316
00:24:22,360 --> 00:24:26,559
in driving, or in editing a simple legal document, that

317
00:24:26,559 --> 00:24:30,320
works, but then if it's some multi-year, complex

318
00:24:30,320 --> 00:24:35,039
tangle of shareholder agreements between, say, your five family offices'

319
00:24:35,039 --> 00:24:39,279
portfolio companies, then, well, it doesn't quite work like that — there I

320
00:24:39,279 --> 00:24:42,919
believe a thick harness still has value, because it

321
00:24:42,919 --> 00:24:46,120
is just, somehow, so multi-dimensional that on a single

322
00:24:46,120 --> 00:24:51,600
dimension you can't play it. Um, so this kind of task-by-

323
00:24:51,600 --> 00:24:55,640
task, one vertical task at a time — yeah, those are getting chewed through constantly, so

324
00:24:55,640 --> 00:24:58,600
once enough of them are done and the system has learned, then

325
00:24:58,600 --> 00:25:02,320
it can pull them together holistically too.

326
00:25:02,320 --> 00:25:08,120
Yeah. And then there's maybe that one, sort of,

327
00:25:08,120 --> 00:25:12,480
dimension where Harvey and Legora are trying to compete, which is the interface.

328
00:25:12,480 --> 00:25:16,120
They have these UI components that you

329
00:25:16,120 --> 00:25:22,679
don't have in Claude, say, but I predict this will stay temporary,

330
00:25:22,679 --> 00:25:26,960
because generative interfaces seem to be coming on

331
00:25:26,960 --> 00:25:32,640
pretty fast right now. So soon you'll just be able to tell Claude,

332
00:25:32,640 --> 00:25:37,720
hey, I have this use case, what would be a good

333
00:25:37,720 --> 00:25:41,240
interface for this, and it'll build you

334
00:25:41,240 --> 00:25:44,720
that interface on the fly. Yeah.

335
00:25:44,720 --> 00:25:50,279
So the room for Legoras and Harveys of the world is going to get

336
00:25:50,279 --> 00:25:54,039
genuinely pretty tight. And it's really interesting, because they've sold

337
00:25:54,039 --> 00:25:59,080
it so well up to now — it's been sold incredibly well, the price is

338
00:25:59,080 --> 00:26:03,159
really steep, the contract terms are pretty long, but this is a really

339
00:26:03,159 --> 00:26:08,520
interesting situation in the sense that if we think about

340
00:26:08,520 --> 00:26:13,919
those law firms that have committed to these and

341
00:26:13,919 --> 00:26:17,880
built their whole process around Harvey or Legora, then

342
00:26:17,880 --> 00:26:22,919
along comes some smaller firm that adopts

343
00:26:22,919 --> 00:26:27,120
Claude instead — and especially if we're talking about Finnish

344
00:26:27,120 --> 00:26:31,640
law, right now in Finnish law Claude is

345
00:26:31,640 --> 00:26:35,000
just flat-out superior compared to something like Legora or Harvey, because

346
00:26:35,000 --> 00:26:39,760
you can't — in Legora or Harvey, as far as I know at least, you can't

347
00:26:39,760 --> 00:26:44,039
actually get Laki.ai hooked in there. Oh. Yeah. So it's that closed off.

348
00:26:44,039 --> 00:26:49,159
And they don't have an equivalent — they bought this Swedish

349
00:26:49,159 --> 00:26:54,159
startup, Kurigal or was it Kur Data, which basically has

350
00:26:54,159 --> 00:26:58,399
done in Sweden the same thing we've done with Laki.ai in Finland.

351
00:26:58,399 --> 00:27:01,399
So in Sweden they have that capability, but in Finland they don't have

352
00:27:01,399 --> 00:27:06,679
that capability. So legal research in Legora or Harvey in Finland

353
00:27:06,679 --> 00:27:12,600
relies purely on a web browser, and we know exactly what

354
00:27:12,600 --> 00:27:17,200
level that's at. Now a good example of this is, say,

355
00:27:17,200 --> 00:27:20,600
Harvey's own case too, where by default

356
00:27:20,600 --> 00:27:25,200
they've actually banned the model from doing this web search,

357
00:27:25,200 --> 00:27:28,080
and there's a very good reason for that [laughter], and now you've got

358
00:27:28,080 --> 00:27:33,120
this appallingly expensive product whose legal research relies

359
00:27:33,120 --> 00:27:37,200
on that web search — so is it really worth paying for? Well, neither of us is

360
00:27:37,200 --> 00:27:39,840
a lawyer, so of course we're of the opinion that it's not

361
00:27:39,840 --> 00:27:42,960
worth paying for, but — I do have friends

362
00:27:42,960 --> 00:27:46,480
though, who do just fine living off that lawyer's know-how, that corpus, and

363
00:27:46,480 --> 00:27:50,600
expertise, and their closed systems — and if you have, say,

364
00:27:50,600 --> 00:27:54,480
a top Finnish law firm, then their own internal

365
00:27:54,480 --> 00:27:59,360
system, with thousands of past cases in it, or you could

366
00:27:59,360 --> 00:28:03,519
use training material — you can still find that unique

367
00:28:03,519 --> 00:28:07,399
expertise there, the case-specific kind. I'd say — even though neither of us is

368
00:28:07,399 --> 00:28:12,679
a lawyer — Finland has weak courts

369
00:28:12,679 --> 00:28:16,760
and a strong Chamber of Commerce arbitration system.

370
00:28:16,760 --> 00:28:21,080
And this arbitration system is partly a bit like the American

371
00:28:21,080 --> 00:28:25,320
case-law type of thing, where, say, one or three

372
00:28:25,320 --> 00:28:30,679
arbitrators make a legally binding decision that

373
00:28:30,679 --> 00:28:34,120
binds the parties just like a judge's ruling would, if they've

374
00:28:34,120 --> 00:28:38,240
agreed to use the Chamber of Commerce's dispute resolution method for it. And they're

375
00:28:38,240 --> 00:28:42,760
not public, as far as I understand, so that body of information partly lives

376
00:28:42,760 --> 00:28:47,960
in law firms' own databases, because they're parties to these cases. Their

377
00:28:47,960 --> 00:28:52,640
partners might even be the arbitrators themselves, and it's

378
00:28:52,640 --> 00:28:57,320
kind of kept out of that whole open-data world of yours. I don't know

379
00:28:57,320 --> 00:29:01,200
if that rings a bell, that world — this kind of Helsinki Chamber of

380
00:29:01,200 --> 00:29:06,360
Commerce-type exception. Well, yeah. I mean,

381
00:29:06,360 --> 00:29:10,600
the challenge here, if we think about these law firms, what's

382
00:29:10,600 --> 00:29:14,360
difficult about it is that now

383
00:29:14,360 --> 00:29:19,880
a couple of years ago I was wondering which direction this was going,

384
00:29:19,880 --> 00:29:24,799
these general-purpose AI tools. Would they become

385
00:29:24,799 --> 00:29:29,159
so easy to use that anyone could use them,

386
00:29:29,159 --> 00:29:32,559
or would it get more complex? And I think this

387
00:29:32,559 --> 00:29:36,840
has clearly gone toward getting more complex. So now

388
00:29:36,840 --> 00:29:40,200
law firms traditionally maybe don't have much

389
00:29:40,200 --> 00:29:45,320
IT expertise to begin with, let alone this kind of AI

390
00:29:45,320 --> 00:29:51,039
IT expertise — which means that actually getting the benefits out of it, whether it's

391
00:29:51,039 --> 00:29:56,240
something like Legora or Harvey. Let alone something like

392
00:29:56,240 --> 00:30:00,919
Claude — to actually get the power out of Claude, it's

393
00:30:00,919 --> 00:30:05,279
exactly like you said, the key role there is played by your

394
00:30:05,279 --> 00:30:10,519
company's data, built up over the years. And now you need to get that into Claude's

395
00:30:10,519 --> 00:30:16,120
hands in a sensible way. You need genuinely good

396
00:30:16,120 --> 00:30:20,799
skills — which, as we touched on a bit earlier in this conversation —

397
00:30:20,799 --> 00:30:25,600
came up. So, skills means

398
00:30:25,600 --> 00:30:31,279
you can teach these language models different kinds of

399
00:30:31,279 --> 00:30:35,120
tasks. You could think of it like this: take doing a somersault, say —

400
00:30:35,120 --> 00:30:38,880
and say we have some young child,

401
00:30:38,880 --> 00:30:41,399
who doesn't know how to do a somersault, then

402
00:30:41,399 --> 00:30:44,840
the somersault involves certain things you have to

403
00:30:44,840 --> 00:30:48,559
do — so once you've taught that child the somersault

404
00:30:48,559 --> 00:30:53,360
skill once, then it can, sort of, whenever it needs

405
00:30:53,360 --> 00:30:57,440
to do a somersault, just pull out that somersault skill. And this

406
00:30:57,440 --> 00:31:01,480
is kind of the same thing — now you can teach these language models

407
00:31:01,480 --> 00:31:05,600
different skills. It could be, say, NDA

408
00:31:05,600 --> 00:31:07,080
commenting. Mm.

409
00:31:07,080 --> 00:31:11,200
Right, but you'd need to build it — and this was actually a really

410
00:31:11,200 --> 00:31:16,760
great opening move in that direction. This Claude — sort of the American Legal

411
00:31:16,760 --> 00:31:22,200
Skills, adapted for Finland. This Claude legal skills

412
00:31:22,200 --> 00:31:27,120
pack — Hakunikkola's GitHub has it, whether

413
00:31:27,120 --> 00:31:31,240
the skills — the skills pack — is something you can install with little effort

414
00:31:31,240 --> 00:31:35,591
into Claude, or as far as I know, into other systems too.

415
00:31:35,591 --> 00:31:37,880
[chuckle] Yeah. But maybe on skills, by the way —

416
00:31:37,880 --> 00:31:41,600
my own observation — for a long time I did them like this, I'd start

417
00:31:41,600 --> 00:31:45,120
building sort of alongside, like: here's a task starting, watch me while I do

418
00:31:45,120 --> 00:31:49,240
this task, and oh, stuff like that came up, and then we jumped around, and I

419
00:31:49,240 --> 00:31:51,799
realized this was done completely backwards — that you should first do the thing

420
00:31:51,799 --> 00:31:54,919
all the way to the finish line, and only then say let's turn this into a skill — once I've slogged through

421
00:31:54,919 --> 00:31:57,799
the whole obstacle course to the end, that's actually a better way to make

422
00:31:57,799 --> 00:32:01,840
skills, at least at this point in time. Absolutely agree. Yeah,

423
00:32:01,840 --> 00:32:04,679
yeah, but nobody really teaches that either — I mean, in that sense

424
00:32:04,679 --> 00:32:08,960
or, I don't know, there just isn't anymore that kind of

425
00:32:08,960 --> 00:32:13,360
Stack Overflow that everyone could go check to see how this stuff

426
00:32:13,360 --> 00:32:16,320
gets done — everyone's kind of learning it at random, I think, but

427
00:32:16,320 --> 00:32:19,360
it's kind of fun in a way. It's insanely fun, [laughter]

428
00:32:19,360 --> 00:32:23,240
but the unfortunate part is it's a pretty small group

429
00:32:23,240 --> 00:32:26,840
who are excited about learning

430
00:32:26,840 --> 00:32:30,799
this kind of stuff. Yeah. Maybe I have to give Legora credit here too

431
00:32:30,799 --> 00:32:34,799
since, well, those Legora guys built that too.

432
00:32:34,799 --> 00:32:39,000
So at first I installed it, of course, thinking, let's just switch it on and,

433
00:32:39,000 --> 00:32:45,480
since I'm the DIY type, run it on this Gemma 4 26B model on my own

434
00:32:45,480 --> 00:32:50,720
hardware — and it was total garbage. Then I put Qwen 3.7, the seven,

435
00:32:50,720 --> 00:32:53,880
running there, which I already run, and that was a lot better, but then there was this

436
00:32:53,880 --> 00:32:57,240
problem where I didn't get their philosophy — like their

437
00:32:57,240 --> 00:33:01,399
Harvey-style model, where the idea is you have two different

438
00:33:01,399 --> 00:33:07,440
language models, which are, say, different — Opus and then Sonnet,

439
00:33:07,440 --> 00:33:12,240
or, well, the best Sonnet, which right now is five, and Opus 4.8,

440
00:33:12,240 --> 00:33:16,679
eight, since Fable got taken away from us — the way they, sort of,

441
00:33:16,679 --> 00:33:22,039
argue with each other, that sort of produces intelligence that doesn't happen

442
00:33:22,039 --> 00:33:25,720
with a single local model. If it's the same model going up against another copy of itself, then there's no

443
00:33:25,720 --> 00:33:31,399
real cross-battle happening, it's just — at first

444
00:33:31,399 --> 00:33:34,600
I thought this was total garbage when I ran it with that local model. It wasn't

445
00:33:34,600 --> 00:33:38,159
getting anything done. But then when I ran it with, like, the best Claude

446
00:33:38,159 --> 00:33:41,840
models, it actually produced a pretty okay base. But then when I ran it

447
00:33:41,840 --> 00:33:45,679
through your Laki.ai and then through these skills, it turned out

448
00:33:45,679 --> 00:33:50,399
even better. So I thought this was a pretty solid workflow. So like

449
00:33:50,399 --> 00:33:53,519
you make some kind of base draft that doesn't need to be all that polished, that can be

450
00:33:53,519 --> 00:33:57,159
just your own best rough sketch, and then you refine it — I think that's

451
00:33:57,159 --> 00:33:59,519
where pretty nice stuff comes out in the end.

452
00:33:59,519 --> 00:34:04,240
Yeah, absolutely. [laughter] Yeah. And maybe let's open this up a bit more

453
00:34:04,240 --> 00:34:08,879
for the listeners too — you just said Legora, but you actually meant

454
00:34:08,879 --> 00:34:13,480
Lovable for sure. Yeah, yeah, yeah, got them mixed up. So Lovable is just this

455
00:34:13,480 --> 00:34:17,200
kind of free program too, that you can install. It eats up some

456
00:34:17,200 --> 00:34:21,320
tokens once you install it. Though this is — yeah, Legora and Harvey, I've

457
00:34:21,320 --> 00:34:24,440
actually never used either, I have to say those are total black

458
00:34:24,440 --> 00:34:27,599
boxes to me. I just watch and see that, once again, the Swedes and the Yanks know how to cash in

459
00:34:27,599 --> 00:34:30,800
on that too. Yeah, I haven't used Harvey myself

460
00:34:30,800 --> 00:34:34,520
either, though I have watched YouTube videos

461
00:34:34,520 --> 00:34:38,879
of it being used. Legora I have used, and

462
00:34:38,879 --> 00:34:42,599
it hasn't convinced yours truly.

463
00:34:42,599 --> 00:34:46,919
Yeah. Lovable is actually kind of fun from a Swedish-coding-scene angle, because they sort of solve

464
00:34:46,919 --> 00:34:52,359
that whole minimum viable product thing within a safe authenticated setup,

465
00:34:52,359 --> 00:34:56,599
pretty neatly and with little effort, instead of you having to

466
00:34:56,599 --> 00:35:00,520
go and set up your own servers, or cloud

467
00:35:00,520 --> 00:35:04,640
installations. I still like that — I still occasionally do some

468
00:35:04,640 --> 00:35:08,520
little things with it myself, yeah, that's also interesting to watch —

469
00:35:08,520 --> 00:35:15,280
I think Lovable actually fits pretty well onto that sort of Pieter Levels path,

470
00:35:15,280 --> 00:35:20,240
whether it turns out that, as we go a bit further, I —

471
00:35:20,240 --> 00:35:26,280
I'll sign off on that. As a user, Lovable is still pretty handy, because

472
00:35:26,280 --> 00:35:28,880
it sort of abstracts away certain

473
00:35:28,880 --> 00:35:33,720
pretty laborious steps, like this whole cloud infrastructure

474
00:35:33,720 --> 00:35:36,800
provisioning and all this configuring,

475
00:35:36,800 --> 00:35:42,839
but how long is it before you're just sitting in Claude's or ChatGPT's

476
00:35:42,839 --> 00:35:49,734
interface, saying, hey, just whip me up something like this.

477
00:35:49,734 --> 00:35:52,720
[laughter] Yeah, that's how it is. Yeah, mainly for me it's

478
00:35:52,720 --> 00:35:56,520
that I at least know what the tech stack is, sort of. They've

479
00:35:56,520 --> 00:36:01,319
gone with a super simple one — you've got React and Node in there, and then

480
00:36:01,319 --> 00:36:04,319
TypeScript too, so I know that those are good

481
00:36:04,319 --> 00:36:09,040
building blocks — and Supabase and so on — and they work, and they don't really

482
00:36:09,040 --> 00:36:13,000
leak, as a rule. Whereas if I go and build it myself, then

483
00:36:13,000 --> 00:36:16,400
you can end up with whatever happens to come out — like, let's just build this on Firebase and

484
00:36:16,400 --> 00:36:20,240
do some of my own messing around, and oh, this doesn't even need this kind of authentication and

485
00:36:20,240 --> 00:36:24,760
that kind of API — and then, I can't be bothered, I'll just grab it off the shelf

486
00:36:24,760 --> 00:36:28,599
instead. [chuckle] So, so it's sort of like that

487
00:36:28,599 --> 00:36:32,560
noob-level-plus thing, that's what it is. And there's still stuff like demo sites or

488
00:36:32,560 --> 00:36:36,040
some small group things you need to log into, so it's not just totally

489
00:36:36,040 --> 00:36:39,560
wide open — and that's incredibly handy. And then it still

490
00:36:39,560 --> 00:36:44,480
keeps improving a bit, and they clearly run all those language models a bit

491
00:36:44,480 --> 00:36:47,359
crosswise under the hood, so every now and then you get a proper

492
00:36:47,359 --> 00:36:51,119
bonus dose of intelligence, when they happen to run it on some better model instead of dropping down

493
00:36:51,119 --> 00:36:55,920
to some cheap knock-off model. But what you said about Cursor a moment ago,

494
00:36:55,920 --> 00:36:59,599
now the great and mighty Elon Musk took it into his empire, so

495
00:36:59,599 --> 00:37:04,839
as we're recording this, they've released an X MCP, and then they

496
00:37:04,839 --> 00:37:08,760
run that into Cursor, so it's really interesting what my

497
00:37:08,760 --> 00:37:14,119
friend Musk pulls off there with that combination of Cursor and X and

498
00:37:14,119 --> 00:37:17,839
Twitter. Yeah, absolutely. I do have to pick up a bit more on

499
00:37:17,839 --> 00:37:22,839
this Lovable thing, because right now

500
00:37:22,839 --> 00:37:28,599
Laki.ai's development has definitely been

501
00:37:28,599 --> 00:37:34,079
mainly done by Juuso from Vesanto, who, uh,

502
00:37:34,079 --> 00:37:40,319
a few years ago — before Lovable even existed — showed me

503
00:37:40,319 --> 00:37:45,520
this kind of business case, where he described what Lovable

504
00:37:45,520 --> 00:37:50,000
has done in practice. And whether we should maybe build that. [laughter]

505
00:37:50,000 --> 00:37:54,119
We ended up deciding together, as a group, well, maybe we won't build

506
00:37:54,119 --> 00:37:56,480
this. Yeah. There's Replit and others out there,

507
00:37:56,480 --> 00:37:59,440
sure, that's eating into that market. But I mean, even Cursor

508
00:37:59,440 --> 00:38:03,800
was bought for 60 billion dollars, which is 30

509
00:38:03,800 --> 00:38:08,640
times revenue — and when X was acquired, at first they didn't even have to pay

510
00:38:08,640 --> 00:38:12,720
with hard cash, it was just a 3% dilution into SpaceX, which is just

511
00:38:12,720 --> 00:38:17,000
an insane amount of paper money right there. So surely something

512
00:38:17,000 --> 00:38:20,520
like that will turn up for Lovable too, out of this, as long as

513
00:38:20,520 --> 00:38:25,280
it doesn't take too long to pay off. Yeah, right. [laughter] But yeah, but that's

514
00:38:25,280 --> 00:38:28,640
exactly the thing — if you think about what Lovable has done

515
00:38:28,640 --> 00:38:32,119
smartly, it's specifically those smart default choices. So they have

516
00:38:32,119 --> 00:38:36,359
good default settings. Whereas — I don't know, I haven't used Replit

517
00:38:36,359 --> 00:38:39,560
myself in a long time. I don't know what its status is these days, but

518
00:38:39,560 --> 00:38:43,480
at least earlier on, it used to leave

519
00:38:43,480 --> 00:38:46,760
the user with quite a lot of decision-making

520
00:38:46,760 --> 00:38:49,880
power there, which was maybe exactly the kind of thing users didn't

521
00:38:49,880 --> 00:38:55,000
want to do — they didn't want to think about what kind of

522
00:38:55,000 --> 00:39:00,079
environment this is being run in, they just want good defaults and then

523
00:39:00,079 --> 00:39:04,119
exactly that, and they've done that really well. Yeah. And so far it hasn't

524
00:39:04,119 --> 00:39:08,920
leaked badly. So just that — the authentication stuff and the Supabase instances — and then

525
00:39:08,920 --> 00:39:11,880
they've held up to scaling pretty well, they haven't

526
00:39:11,880 --> 00:39:15,200
blown up so badly that you'd have had to pay

527
00:39:15,200 --> 00:39:19,599
a ton of money for it when suddenly 10 million

528
00:39:19,599 --> 00:39:23,040
users are hitting some database — somehow they've

529
00:39:23,040 --> 00:39:26,520
managed to handle that too. And then I think they've also, in a way,

530
00:39:26,520 --> 00:39:32,240
in that harness layer — that dumb chat text box — they don't

531
00:39:32,240 --> 00:39:35,720
tell you what model they're running there, but they clearly always make

532
00:39:35,720 --> 00:39:40,400
that kind of trade-off between cost savings and then, on the other hand,

533
00:39:40,400 --> 00:39:44,119
whether you get quality out of it — and I think they've handled that too

534
00:39:44,119 --> 00:39:47,160
pretty well. As a rule, when I put in

535
00:39:47,160 --> 00:39:50,000
a smart prompt, what usually comes out is pretty sensible

536
00:39:50,000 --> 00:39:54,560
stuff, not garbage. Whereas if you put, say, Sonnet

537
00:39:54,560 --> 00:39:57,760
in there as the Claude model, you might actually get some

538
00:39:57,760 --> 00:40:01,520
surprisingly bad output — like [chuckle] — if it's just a bad token day and it

539
00:40:01,520 --> 00:40:07,839
doesn't feel like running on that 4.8 Pro, or whatever it's called, that

540
00:40:07,839 --> 00:40:12,520
ultra-ultra-effort setting. Yeah [laughter] yeah. On that note, actually,

541
00:40:12,520 --> 00:40:17,640
it's an interesting day today — yesterday evening

542
00:40:17,640 --> 00:40:22,800
actually, when I was reading the news, Anthropic announced that Fable is coming

543
00:40:22,800 --> 00:40:24,800
back today,

544
00:40:24,800 --> 00:40:28,720
just like Sonnet 5 came too — yeah, Sonnet 5 came out yesterday, and Fable

545
00:40:28,720 --> 00:40:31,400
is coming back here. Ah, has Trump had a good day, then, so that

546
00:40:31,400 --> 00:40:34,680
apparently — ah, okay. Well then, full speed ahead,

547
00:40:34,680 --> 00:40:39,400
because I was still in that post-Arctic15

548
00:40:39,400 --> 00:40:44,359
hangover haze, so my last Fable run was mainly YouTube

549
00:40:44,359 --> 00:40:49,240
subtitling for this channel, so, honestly, that's about

550
00:40:49,240 --> 00:40:55,400
where things stand. Yeah, I have to say, luckily, when Fable

551
00:40:55,400 --> 00:41:01,440
was released — if I remember right it was available for three days — I hammered

552
00:41:01,440 --> 00:41:06,400
away at it pretty much morning to night, and I have to say it was truly

553
00:41:06,400 --> 00:41:12,560
a significant — based on my own testing and

554
00:41:12,560 --> 00:41:18,160
experience — a significant improvement over anything

555
00:41:18,160 --> 00:41:22,400
before it. Right, and you had that slash-make-me-rich-make-no-

556
00:41:22,400 --> 00:41:26,200
mistakes thing. It made money land in the account from a single prompt.

557
00:41:26,200 --> 00:41:29,480
Yeah, and there's — this ties into [laughter] — I think there's a

558
00:41:29,480 --> 00:41:35,119
kind of interesting story here, like what, or

559
00:41:35,119 --> 00:41:38,040
this kind of objection I actually run into pretty often

560
00:41:38,040 --> 00:41:42,119
from software developers. A lot of them say this model is already

561
00:41:42,119 --> 00:41:43,839
good enough for me now, so

562
00:41:43,839 --> 00:41:49,560
I don't need a better model, this is enough for me. To which my

563
00:41:49,560 --> 00:41:54,560
question is, well, are you at the point where you just give it instructions in the

564
00:41:54,560 --> 00:41:55,599
morning, and then

565
00:41:55,599 --> 00:41:59,800
check the next day's

566
00:41:59,800 --> 00:42:04,480
standup to see what's been done — and then they're a bit confused, like, well no,

567
00:42:04,480 --> 00:42:07,480
of course not, I write functions with it.

568
00:42:07,480 --> 00:42:10,920
Yeah, yeah. Well, I think sometimes you just hit enter too [laughter]

569
00:42:10,920 --> 00:42:13,400
as in, like, 'dangerously appro—

570
00:42:13,400 --> 00:42:18,240
ve.' The thing is, I personally see an insane amount of value specifically in

571
00:42:18,240 --> 00:42:22,559
the model's intelligence improving, because it enables — and now the point is,

572
00:42:22,559 --> 00:42:25,400
how you should think about it, I think, is just that

573
00:42:25,400 --> 00:42:31,559
we're still not at the point where I could treat

574
00:42:31,559 --> 00:42:35,319
a model like a good software developer. And that's exactly

575
00:42:35,319 --> 00:42:38,480
what I want. Yeah, I want to tell it, like,

576
00:42:38,480 --> 00:42:44,079
in a weekly meeting, hey, here's this week's tasks, and then we've got — whether it's

577
00:42:44,079 --> 00:42:47,640
a daily, you know, a 10 or 15 minute check-in where we

578
00:42:47,640 --> 00:42:49,319
look at how it's gone.

579
00:42:49,319 --> 00:42:53,640
Mm. So then the thing is, that these

580
00:42:53,640 --> 00:42:57,079
current ones are good enough — well, no, definitely not, because I want

581
00:42:57,079 --> 00:43:01,040
Fable back quickly, and luckily it's coming today, and then we've also got

582
00:43:01,040 --> 00:43:06,319
GPT-5.6 coming, which is interesting too — some people already have it

583
00:43:06,319 --> 00:43:11,079
on the approved list there. Exactly. Yeah, some already have it,

584
00:43:11,079 --> 00:43:14,119
but hopefully it's coming for us too, yeah,

585
00:43:14,119 --> 00:43:18,119
us regular mortals, probably within a week or two from around now.

586
00:43:18,119 --> 00:43:22,400
So we'll get to that shortly, but I have to say, when I started

587
00:43:22,400 --> 00:43:27,200
down this rabbit hole back in January, I of course picked up Claude Code

588
00:43:27,200 --> 00:43:32,040
and Cowork, and even switched to Mac because of it, since it didn't run on PC, so

589
00:43:32,040 --> 00:43:35,800
all that — I liked Claude's philosophy compared to that. Codex, well,

590
00:43:35,800 --> 00:43:39,480
already one-shotted things back then, but I liked it when I sort of

591
00:43:39,480 --> 00:43:43,599
learned that you have to actually watch Claude a bit as it fumbles around and

592
00:43:43,599 --> 00:43:47,359
keeps looking things over and making choices, then hitting enter, and sometimes going into

593
00:43:47,359 --> 00:43:51,160
the shell [chuckle] and having a look at what

594
00:43:51,160 --> 00:43:55,599
you need to press yourself, and oh no, that's not working — you take screenshots or

595
00:43:55,599 --> 00:43:58,480
paste this text back in, there was an error and so on. And that's kind of, for me,

596
00:43:58,480 --> 00:44:03,119
still a basic principle. And then, back when Claude

597
00:44:03,119 --> 00:44:06,839
ages ago switched to this token-based pricing, I decided to

598
00:44:06,839 --> 00:44:10,480
just cut OpenAI out of the picture, because it started

599
00:44:10,480 --> 00:44:13,960
costing real money — the first day alone was like 90 euros, so I thought,

600
00:44:13,960 --> 00:44:18,640
just drop it completely, and go with OpenAI's monthly plan for Codex as my

601
00:44:18,640 --> 00:44:22,605
daily driver. But at the same time, as I started using Codex,

602
00:44:22,605 --> 00:44:25,760
[clears throat] it really is kind of scarily good at

603
00:44:25,760 --> 00:44:28,839
one-shotting things — you give it the task, hit enter, and

604
00:44:28,839 --> 00:44:33,000
off it goes, and pretty solid stuff comes out at the end, so I'm not

605
00:44:33,000 --> 00:44:37,800
totally sure I like that. I think I'd still rather be the artist

606
00:44:37,800 --> 00:44:42,559
who steers the ship every now and then. I don't know if you have the same

607
00:44:42,559 --> 00:44:46,240
experiences. Well, yeah, I do. [laughter] I mean, as personalities,

608
00:44:46,240 --> 00:44:51,960
Claude — or if you think about the Claude models

609
00:44:51,960 --> 00:44:56,480
versus these GPT models, they're really different. And I do

610
00:44:56,480 --> 00:45:00,520
share that experience, that working with Claude is much

611
00:45:00,520 --> 00:45:02,119
more pleasant, so

612
00:45:02,119 --> 00:45:07,960
it's just a lot more human and

613
00:45:07,960 --> 00:45:14,240
more considerate, and it actually explains things — and that's maybe hard to explain.

614
00:45:14,240 --> 00:45:17,520
But there's actually a pretty big difference in how it feels to work with, whereas

615
00:45:17,520 --> 00:45:21,000
then, the GPT models are a bit blunt, or kind of

616
00:45:21,000 --> 00:45:24,680
arrogant — it's like you've got some kind of [laughter]

617
00:45:24,680 --> 00:45:29,520
really senior type, like some professor who's

618
00:45:29,520 --> 00:45:32,960
a bit high-strung, like, you're not really — you're a bit

619
00:45:32,960 --> 00:45:36,319
on edge when you go to tell it something, like hey, could you

620
00:45:36,319 --> 00:45:40,359
do this now? Yeah. This actually, by the way, Claude —

621
00:45:40,359 --> 00:45:47,760
when it went from Opus 4.6 to 4.7, it was a shock to me when this HR

622
00:45:47,760 --> 00:45:53,559
lady showed up, and then this security guy showed up too — or the other way round, HR and security

623
00:45:53,559 --> 00:45:57,960
lady — it was like, well Sami, it's already 8 o'clock, haven't

624
00:45:57,960 --> 00:46:01,680
we already done quite a lot of work here, maybe continue tomorrow — and I was like,

625
00:46:01,680 --> 00:46:05,960
what the hell, it's none of your business how long and how hard I

626
00:46:05,960 --> 00:46:09,079
work here. And then the other thing, this security business, where

627
00:46:09,079 --> 00:46:12,520
it flashed up something about an API key, like you basically need to delete your hard drive and

628
00:46:12,520 --> 00:46:15,920
go get a new key again, like this isn't going to fly, this is just

629
00:46:15,920 --> 00:46:19,079
a terrible security risk — and honestly, with both of these, I was like, I don't

630
00:46:19,079 --> 00:46:22,480
want any of this at all, and that was honestly part of the reason I jumped

631
00:46:22,480 --> 00:46:25,119
over to the GPT world, because they don't really have this stuff there, so —

632
00:46:25,119 --> 00:46:28,400
Mm. Yeah. [laughter] But that — there's actually been a really

633
00:46:28,400 --> 00:46:32,119
interesting blog post from OpenAI just a few days ago,

634
00:46:32,119 --> 00:46:37,760
they published. They'd looked internally into how many tokens

635
00:46:37,760 --> 00:46:42,559
from all their employees' AI usage

636
00:46:42,559 --> 00:46:48,160
come from Codex versus ChatGPT. Yeah. And there had been a really

637
00:46:48,160 --> 00:46:54,079
dramatic shift there — it was something like over 90% of all the tokens,

638
00:46:54,079 --> 00:46:58,920
basically all the output that these AIs produce inside the OpenAI organization,

639
00:46:58,920 --> 00:47:03,040
comes from Codex, even outside of software development. And

640
00:47:03,040 --> 00:47:07,599
this is something that I don't think has really

641
00:47:07,599 --> 00:47:11,200
in Finnish organizations, OpenAI - look, they go through Codex

642
00:47:11,200 --> 00:47:12,960
so no one's really caught on to that yet. [laughter] But the thing is,

643
00:47:12,960 --> 00:47:16,720
I think that's how it is - or actually, I do know that's how it is,

644
00:47:16,720 --> 00:47:20,960
that people just use the chat.

645
00:47:20,960 --> 00:47:23,520
Mm. Like if we look at it this way,

646
00:47:23,520 --> 00:47:29,160
inside OpenAI everyone uses Codex. Nobody really uses the chat -

647
00:47:29,160 --> 00:47:33,559
chat usage is minimal. Whereas in Finnish companies,

648
00:47:33,559 --> 00:47:37,760
and maybe in general, aside from these kinds of

649
00:47:37,760 --> 00:47:41,520
tech startups or whoever, who are actually

650
00:47:41,520 --> 00:47:47,000
in the AI world, AI adoption is completely

651
00:47:47,000 --> 00:47:48,640
still in its infancy.

652
00:47:48,640 --> 00:47:52,640
so what's needed now is a shift away from doing

653
00:47:52,640 --> 00:47:58,000
just little one-off things in chat. Yeah. Toward actually being

654
00:47:58,000 --> 00:48:02,640
well, if you're in the OpenAI ecosystem, having the work actually happen

655
00:48:02,640 --> 00:48:05,319
in Codex itself. Yeah, that's a good point. And actually

656
00:48:05,319 --> 00:48:10,319
now we get to that harness thing, or the OpenAI setup, or actual AI

657
00:48:10,319 --> 00:48:16,480
workers - not just these kind of dumb, in-between,

658
00:48:16,480 --> 00:48:20,240
whatever agents. So maybe two themes now - if these

659
00:48:20,240 --> 00:48:24,359
are the current front-runners - let's leave Google aside for now, even though they're

660
00:48:24,359 --> 00:48:28,160
trying to get into this game with Antigravity. So you've got

661
00:48:28,160 --> 00:48:33,040
Claude - Anthropic started that game already back in early this year. So we got that

662
00:48:33,040 --> 00:48:36,640
Claude Cowork, and along with it came the desktop app, so you've got the

663
00:48:36,640 --> 00:48:42,079
chat, then you've got Claude Cowork, which gets you into those office folders

664
00:48:42,079 --> 00:48:46,119
to work in, and then there's Claude Code. And at first for me it was like, I

665
00:48:46,119 --> 00:48:49,920
went from the chat into Code, and then I went into Cowork. Now I don't

666
00:48:49,920 --> 00:48:53,839
really know where I even want to be. It's almost the same whether I'm

667
00:48:53,839 --> 00:48:57,160
in chat. Actually I pretty rarely go into Code, because if I'm running

668
00:48:57,160 --> 00:49:00,359
code, then I go to the command line interface, which is the

669
00:49:00,359 --> 00:49:05,000
text-based thing. And then OpenAI's side was kind of whatever, before

670
00:49:05,000 --> 00:49:08,960
Codex came along - there was just the chat and the command line, but

671
00:49:08,960 --> 00:49:11,319
I never really used that. But now it's kind of

672
00:49:11,319 --> 00:49:15,240
like this: I actually run quite a lot of coding straight from chat,

673
00:49:15,240 --> 00:49:18,359
but then I go into Codex if I want to one-shot something. But then I

674
00:49:18,359 --> 00:49:22,559
run things with my own Samantha on OpenClow. So OpenClow is this kind of open-

675
00:49:22,559 --> 00:49:26,880
source project that was built by this guy, Peter Steinberger,

676
00:49:26,880 --> 00:49:30,319
an Austrian genius. So I picked it up like a week after it was

677
00:49:30,319 --> 00:49:34,319
released, and it runs on my Mac - this Samantha, which runs on its own

678
00:49:34,319 --> 00:49:38,040
physical machine, but it's clearly just hammering away on Codex the whole time

679
00:49:38,040 --> 00:49:44,440
because it codes and uses that as its main brain. So, um, I still

680
00:49:44,440 --> 00:49:49,559
think of it as the harness - you've got this smart agent, OpenClow,

681
00:49:49,559 --> 00:49:53,200
that's still a really big deal for me. But are you following

682
00:49:53,200 --> 00:49:56,520
this idea of the desktop, these three pipelines, and then this kind of

683
00:49:56,520 --> 00:50:02,280
agentic building block - how do you actually see access

684
00:50:02,280 --> 00:50:09,440
to that intelligence. Well, especially if we go into this

685
00:50:09,440 --> 00:50:14,079
context of an office worker or knowledge worker,

686
00:50:14,079 --> 00:50:16,640
who works at some corporation,

687
00:50:16,640 --> 00:50:21,839
then I think the shift there - the first thing that needs

688
00:50:21,839 --> 00:50:26,680
to happen - is getting rid of Copilot.

689
00:50:26,680 --> 00:50:30,119
Yeah, it's terrible. And it's funny because

690
00:50:30,119 --> 00:50:34,079
I've done a fair number of these trainings over in Saudi

691
00:50:34,079 --> 00:50:38,799
Arabia, and I don't run into this problem there. It being a religious country,

692
00:50:38,799 --> 00:50:42,920
yeah, over there it's either

693
00:50:42,920 --> 00:50:46,640
companies operating in the OpenAI or Anthropic ecosystem,

694
00:50:46,640 --> 00:50:49,680
but I've never once run into this thing I keep running into in Finland,

695
00:50:49,680 --> 00:50:54,640
constantly. And the worst part of all of this is

696
00:50:54,640 --> 00:50:59,000
that we have companies that say, oh yeah, we're right there

697
00:50:59,000 --> 00:51:03,760
on the AI front line, and then it turns out, yeah, we've got Copilot

698
00:51:03,760 --> 00:51:06,880
here. [laughter] And Copilot

699
00:51:06,880 --> 00:51:10,680
has nothing to do with actually leveraging AI

700
00:51:10,680 --> 00:51:13,319
in any real sense. So that'd be my first piece of advice - [clears throat]

701
00:51:13,319 --> 00:51:17,599
you have to get rid of Copilot. And then

702
00:51:17,599 --> 00:51:20,720
as for what good alternatives there are right now, I think the best

703
00:51:20,720 --> 00:51:23,760
options are, like you said earlier, that Google has kind of dropped

704
00:51:23,760 --> 00:51:30,079
out of the game, so you've got either the OpenAI ecosystem or Anthropic.

705
00:51:30,079 --> 00:51:34,920
So you pick one of the two, and once you've picked one, you should try to

706
00:51:34,920 --> 00:51:39,240
make sure the work actually happens either in Codex

707
00:51:39,240 --> 00:51:42,319
or in Claude's Cowork - not in the

708
00:51:42,319 --> 00:51:46,359
chat, because this is exactly the point from that OpenAI blog post -

709
00:51:46,359 --> 00:51:47,400
yeah,

710
00:51:47,400 --> 00:51:51,960
that the chat is kind of like, if you have a question

711
00:51:51,960 --> 00:51:56,079
you want an answer to, or some small thing like that, then

712
00:51:56,079 --> 00:51:59,599
chat's fine. But in other cases, if you're actually going to do

713
00:51:59,599 --> 00:52:01,079
knowledge work, mm,

714
00:52:01,079 --> 00:52:04,200
and you're not a software developer - for developers, of course,

715
00:52:04,200 --> 00:52:08,520
it's Claude Code or Codex, but now if you're

716
00:52:08,520 --> 00:52:11,920
say, someone doing marketing or sales,

717
00:52:11,920 --> 00:52:17,240
that kind of person, you'd notice, okay, my

718
00:52:17,240 --> 00:52:22,480
usage - like maybe 90% of it - if you're in the OpenAI ecosystem,

719
00:52:22,480 --> 00:52:26,319
then you look at your monthly usage and you realize, okay,

720
00:52:26,319 --> 00:52:29,119
90% of my usage is in Codex. Mm.

721
00:52:29,119 --> 00:52:32,079
Then you can say, hey, that's going pretty well.

722
00:52:32,079 --> 00:52:36,040
Mm. So that's maybe as an intro, that, well,

723
00:52:36,040 --> 00:52:44,160
then, um, these - I wouldn't necessarily dare recommend to companies

724
00:52:44,160 --> 00:52:49,520
these OpenAI solutions yet, at this stage - although now

725
00:52:49,520 --> 00:52:53,000
Microsoft - I haven't followed this super closely, but Microsoft

726
00:52:53,000 --> 00:52:57,160
is now, on some level, offering companies specifically this OpenClow

727
00:52:57,160 --> 00:53:00,839
Yeah, exactly, and actually Peter Steinberger also released

728
00:53:00,839 --> 00:53:05,319
it for mobile too. So yeah, it's coming - look, since Steinberger is at OpenAI now,

729
00:53:05,319 --> 00:53:08,839
he's not just some renegade kid anymore, cranking out a bit of

730
00:53:08,839 --> 00:53:14,640
sloppy code on the side - no, he's properly, like, right there inside

731
00:53:14,640 --> 00:53:19,400
the OpenAI empire. But maybe, in a way, the desktop thing is

732
00:53:19,400 --> 00:53:24,319
kind of interesting to me, because it might be that browsers

733
00:53:24,319 --> 00:53:30,200
basically die out entirely, that it's Anthropic's and OpenAI's and Google's

734
00:53:30,200 --> 00:53:34,520
own apps, once they get things sorted out - those become the portals you go in through

735
00:53:34,520 --> 00:53:37,119
and yeah, and that's actually a good

736
00:53:37,119 --> 00:53:42,480
point to raise - when I talk here about Codex and Claude Cowork,

737
00:53:42,480 --> 00:53:47,960
these, um - you can actually use Codex through

738
00:53:47,960 --> 00:53:54,400
the browser too. What I'm specifically talking about here are these

739
00:53:54,400 --> 00:53:57,160
desktop applications,

740
00:53:57,160 --> 00:54:02,559
meaning the desktop apps you download onto your machine and run

741
00:54:02,559 --> 00:54:06,119
things in. And that's exactly it - it's really interesting

742
00:54:06,119 --> 00:54:12,000
to think about why you'd actually want to go yourself

743
00:54:12,000 --> 00:54:16,400
to some website - like, does that need go away?

744
00:54:16,400 --> 00:54:20,760
It probably won't disappear entirely, but it'll shrink really, really significantly.

745
00:54:20,760 --> 00:54:25,119
What's interesting here too is that at least I have a really

746
00:54:25,119 --> 00:54:30,119
hard time finding tasks anymore

747
00:54:30,119 --> 00:54:33,520
that I'd actually go do myself. Like,

748
00:54:33,520 --> 00:54:37,657
what would be a thing where I'd actually be better than Codex?

749
00:54:37,657 --> 00:54:39,677
[laughter]

750
00:54:39,839 --> 00:54:44,862
taste, and long-term goals - 'make me rich' or something like that,

751
00:54:44,862 --> 00:54:48,280
[laughter] you need some kind of long-term vision, so that essentially the

752
00:54:48,280 --> 00:54:53,280
task-by-task work versus, say, what happens a year from now - the gap between those is

753
00:54:53,280 --> 00:54:56,359
still significant, definitely. Yeah, and yes, exactly,

754
00:54:56,359 --> 00:54:59,400
so even if you get there a year from now one task at a time, and

755
00:54:59,400 --> 00:55:03,000
even if every single task were done entirely by AI, you still need to first

756
00:55:03,000 --> 00:55:05,240
have some kind of direction. Yes. Yes, absolutely. Yeah. Well,

757
00:55:05,240 --> 00:55:09,079
that's the value a human creates, but the thing is, that you

758
00:55:09,079 --> 00:55:12,680
yourself doing something routine - like, this

759
00:55:12,680 --> 00:55:17,720
task-doing is maybe exactly the thing where nobody should really be doing any

760
00:55:17,720 --> 00:55:22,319
tasks themselves anymore - they should really be left to AI. And I mean

761
00:55:22,319 --> 00:55:25,680
right now we're doing a huge amount of all kinds of tasks

762
00:55:25,680 --> 00:55:29,240
which is honestly kind of crazy. Yeah, I think, if you

763
00:55:29,240 --> 00:55:35,000
think about what Musk did back with Twitter, or X - he

764
00:55:35,000 --> 00:55:37,799
threw out like 70 percent of the staff

765
00:55:37,799 --> 00:55:41,039
back in a time when we didn't even have these tools yet,

766
00:55:41,039 --> 00:55:43,880
so I think it's genuinely hard

767
00:55:43,880 --> 00:55:48,400
to find organizations where you couldn't throw out something like 70 percent

768
00:55:48,400 --> 00:55:54,119
of the staff and still have things carry on pretty much as before.

769
00:55:54,119 --> 00:55:58,440
Yeah, that's brutal. And the other hope people voice is that everyone

770
00:55:58,440 --> 00:56:01,640
would adopt these tools, and then everyone could basically keep

771
00:56:01,640 --> 00:56:04,319
their jobs and just scale things up tenfold instead. But

772
00:56:04,319 --> 00:56:08,520
the more likely outcome is exactly that - scale does grow among the ones

773
00:56:08,520 --> 00:56:12,799
who remain and who push hard, but then, in a way,

774
00:56:12,799 --> 00:56:17,440
people fall away too - and I think it's worth

775
00:56:17,440 --> 00:56:22,440
summing up like this: if your role at work is that some colleague

776
00:56:22,440 --> 00:56:27,960
or boss or client comes and says 'do this for me', hands you the task, and

777
00:56:27,960 --> 00:56:32,720
then you do it and hand it back a bit later if you remember, then

778
00:56:32,720 --> 00:56:35,640
that's not a great position to be in, because then you're basically

779
00:56:35,640 --> 00:56:39,500
pretty easily replaceable by AI inference. Yeah, that's

780
00:56:39,500 --> 00:56:42,920
[laughter] exactly right. Do you remember, by the way, there was a big fuss about that

781
00:56:42,920 --> 00:56:46,799
back when Mikko Alasaarela said on some podcast that he had

782
00:56:46,799 --> 00:56:51,680
calculated how many employees like that Kela would actually need, and it

783
00:56:51,680 --> 00:56:55,864
turned out to be some incomprehensibly small number. Like it was 18 or something.

784
00:56:55,864 --> 00:56:59,920
[laughter] Yeah, and 8,000 was the wrong answer,

785
00:56:59,920 --> 00:57:03,559
whatever the real number is, but yeah, they really should have

786
00:57:03,559 --> 00:57:07,520
ordered some 500-million Salesforce license, so right away, right away

787
00:57:07,520 --> 00:57:12,520
they went about the organization's task management the wrong way. But

788
00:57:12,520 --> 00:57:16,200
yeah, Mikko Alasaarela has been on the show many, many times, definitely.

789
00:57:16,200 --> 00:57:19,359
Although he's mentioned getting this AI brain fog - like, I haven't

790
00:57:19,359 --> 00:57:22,680
found it as fun yet as, in a way, this being a return to childhood,

791
00:57:22,680 --> 00:57:27,039
where I always get excited getting into it, having everything

792
00:57:27,039 --> 00:57:30,880
set up and available, and it's genuinely a joy prompting in some cool

793
00:57:30,880 --> 00:57:34,359
tasks. I think that's great, but yeah, I'm still

794
00:57:34,359 --> 00:57:38,319
working it out too. Yeah. And, [laughter] I really do have to give credit,

795
00:57:38,319 --> 00:57:44,400
even though, um, I'm generally maybe a bit of a

796
00:57:44,400 --> 00:57:47,799
well, or I do enjoy building these agents myself,

797
00:57:47,799 --> 00:57:52,960
but the thing I'm a bit skeptical about, and what I don't

798
00:57:52,960 --> 00:57:58,119
really like, is that a lot of these AI influencers give the impression

799
00:57:58,119 --> 00:58:02,440
that in order to actually get value out of these tools,

800
00:58:02,440 --> 00:58:07,839
you need to be, like, some kind of guru and build some agent farm

801
00:58:07,839 --> 00:58:11,200
and stuff, that it's really difficult.

802
00:58:11,200 --> 00:58:16,520
Whereas I'd argue that in most cases, or even

803
00:58:16,520 --> 00:58:21,000
almost every case, you can just take something like Codex or Claude

804
00:58:21,000 --> 00:58:27,720
Cowork and get pretty much the same results,

805
00:58:27,720 --> 00:58:32,160
if you just learn to use these off-the-shelf products well,

806
00:58:32,160 --> 00:58:39,079
right out of the box. But, um, back to the point - Samantha,

807
00:58:39,079 --> 00:58:43,920
which is this kind of divine, legendary being you built, that can

808
00:58:43,920 --> 00:58:48,079
do human-like magic - it's exactly that kind of magic box. You write

809
00:58:48,079 --> 00:58:52,280
whatever you want, and out comes the best answer in the universe, right?

810
00:58:52,280 --> 00:58:57,880
Yeah. Which [laughter] works. Yeah, which works in WhatsApp, specifically.

811
00:58:57,880 --> 00:59:03,000
And I think that's, again, a genuinely clever way

812
00:59:03,000 --> 00:59:08,920
to use agents, because the added value, in my opinion, isn't

813
00:59:08,920 --> 00:59:13,720
that if you ask it questions, I haven't seen that the answers

814
00:59:13,720 --> 00:59:18,760
are meaningfully better than what you'd get from, say, Claude or Codex.

815
00:59:18,760 --> 00:59:24,520
But where a lot of the value comes from is that it's now part of that

816
00:59:24,520 --> 00:59:28,559
community. Mm. Because something like that isn't

817
00:59:28,559 --> 00:59:32,440
possible anywhere else - it requires something like

818
00:59:32,440 --> 00:59:35,640
WhatsApp specifically - I think the WhatsApp channel is

819
00:59:35,640 --> 00:59:40,039
a great place where people and an AI like that can

820
00:59:40,039 --> 00:59:43,920
chat together, and Musk has actually kind of copied you over on X,

821
00:59:43,920 --> 00:59:48,920
in that Grok is there on X-

822
00:59:48,920 --> 00:59:52,720
slash-Twitter, in the discussions, and I think it works

823
00:59:52,720 --> 00:59:57,119
really well there, being part of that kind of conversation,

824
00:59:57,119 --> 01:00:02,520
so it can even correct Elon Musk's own statements - you can call it in

825
01:00:02,520 --> 01:00:06,319
to comment based on Grokipedia or whatever

826
01:00:06,319 --> 01:00:11,240
else, on whether what's being said actually makes sense. Um, but yeah, this Samantha thing, um,

827
01:00:11,240 --> 01:00:16,119
by the way, it's this - [laughter] - shameless plug - when you subscribe to

828
01:00:16,119 --> 01:00:21,200
Neuvottelija Insider and the bonus content, you can

829
01:00:21,200 --> 01:00:24,440
then get in touch with me, and once I check that you're a real

830
01:00:24,440 --> 01:00:28,720
person and not, um, a bot or some malicious infiltrator, then

831
01:00:28,720 --> 01:00:32,240
I'll invite you into this closed community of over 200 people,

832
01:00:32,240 --> 01:00:36,680
the Neuvottelija Insider community, where you can tag @Samantha, and from there

833
01:00:36,680 --> 01:00:41,200
you can ask literally anything and she answers so wonderfully - and by the way,

834
01:00:41,200 --> 01:00:44,520
everyone's been trying to mess with her for months, trying to crack her, to get her to reveal

835
01:00:44,520 --> 01:00:49,480
Sami's, uh, salary, or [chuckle] some file, or, like, pay, or

836
01:00:49,480 --> 01:00:54,760
whatever - but no, it's a beautifully built data structure, [laughter] but

837
01:00:54,760 --> 01:00:59,400
just a little plug there. But maybe, maybe a bit of skepticism - I appreciate that. It

838
01:00:59,400 --> 01:01:02,799
earlier, when we talked about this over lunch - a lot of people build

839
01:01:02,799 --> 01:01:07,720
it kind of however it happens to come out, so it doesn't add any real value, or

840
01:01:07,720 --> 01:01:10,839
it just ends up being another new communication [chuckle] layer, and

841
01:01:10,839 --> 01:01:14,760
at worst it's a security risk between you and the actual, real AI, i.e.

842
01:01:14,760 --> 01:01:19,599
say, Anthropic's best model - and in that case there's no point to it at all.

843
01:01:19,599 --> 01:01:23,240
So that's when it's just a dumb wrapper. But then

844
01:01:23,240 --> 01:01:28,079
if it has memory - so I was about to mention Andrej Karpathy's

845
01:01:28,079 --> 01:01:33,000
brilliantly envisioned wiki memory - it's able to itself

846
01:01:33,000 --> 01:01:36,880
pull from file structures that are semantically

847
01:01:36,880 --> 01:01:41,960
organized, curated, and nuggets of information I've burned into it

848
01:01:41,960 --> 01:01:45,039
collaboratively, from notes. Then there's this extended

849
01:01:45,039 --> 01:01:48,839
context memory - meaning even if there's a WhatsApp channel with

850
01:01:48,839 --> 01:01:51,319
a lot of text in it, usually the tokens [clears throat] run out

851
01:01:51,319 --> 01:01:56,440
because it can't hold more than a few messages in that short

852
01:01:56,440 --> 01:02:00,039
context window, so it extends that up to a file-level,

853
01:02:00,039 --> 01:02:03,160
kind of lossless context memory. So that's already

854
01:02:03,160 --> 01:02:06,880
a huge deal - it can see the whole conversation space, not just

855
01:02:06,880 --> 01:02:11,319
the last couple of messages. And then there's the secret ingredient, which is

856
01:02:11,319 --> 01:02:15,279
this Memory Lapse graph memory, which is exclusively between me and

857
01:02:15,279 --> 01:02:18,039
it, and it stores things semantically without any of that

858
01:02:18,039 --> 01:02:22,440
RAG garbage - just clean data structures. And then there's the five-layer

859
01:02:22,440 --> 01:02:26,039
security system that keeps malicious attackers

860
01:02:26,039 --> 01:02:29,160
out. So, yeah, it's lovely. I've been tinkering with this for the last four

861
01:02:29,160 --> 01:02:33,480
months, and it's pure joy just playing around with it, so I

862
01:02:33,480 --> 01:02:36,559
do like it, even though it's a bit silly, so

863
01:02:36,559 --> 01:02:40,960
but it's a great hobby. It's great, and an extremely useful

864
01:02:40,960 --> 01:02:46,200
hobby too - but the thing is, my point here, maybe,

865
01:02:46,200 --> 01:02:50,520
or my main point about why I'm a bit critical of these

866
01:02:50,520 --> 01:02:55,599
personal agent setups is just that some people are left with the

867
01:02:55,599 --> 01:02:58,920
impression that to use AI effectively

868
01:02:58,920 --> 01:03:04,520
you need to spend like 10 hours a day tuning these things. Yeah. Which just isn't

869
01:03:04,520 --> 01:03:07,720
true anymore - and maybe it was true at some point, back when

870
01:03:07,720 --> 01:03:10,760
yeah, those crash-prone setups were really fragile - but now it's

871
01:03:10,760 --> 01:03:13,920
good enough that, yeah, you can get really far with

872
01:03:13,920 --> 01:03:19,119
just off-the-shelf products, and these days they get you

873
01:03:19,119 --> 01:03:24,599
a whole agent farm to work with. I mean, with just a basic Claude license

874
01:03:24,599 --> 01:03:28,920
and Claude Cowork, you fire off some genuinely big task in there. For example

875
01:03:28,920 --> 01:03:34,920
Cowork just made me, like, over 150 of them,

876
01:03:34,920 --> 01:03:39,200
of which quite a few turned out genuinely usable - so it fired off

877
01:03:39,200 --> 01:03:45,520
an agent army - it wasn't just one Claude doing it, there was a whole agent army,

878
01:03:45,520 --> 01:03:49,720
working away on our slides

879
01:03:49,720 --> 01:03:55,559
without me having to do anything, like I didn't have to

880
01:03:55,559 --> 01:03:58,240
build any kind of agent system myself.

881
01:03:58,240 --> 01:04:01,760
It's just that you get it all, like, if you want

882
01:04:01,760 --> 01:04:08,440
an agent army at your disposal today, all you need is a Claude subscription and

883
01:04:08,440 --> 01:04:09,640
off it goes, and yeah,

884
01:04:09,640 --> 01:04:12,960
there it is. Yeah, so, I guess I started this back in December, I

885
01:04:12,960 --> 01:04:16,520
built a kind of second brain - a better notebook - and put

886
01:04:16,520 --> 01:04:20,000
it, at first, into this free tool called Obsidian, an MD-file-

887
01:04:20,000 --> 01:04:24,559
based journal system, for my precious observations about the world, and

888
01:04:24,559 --> 01:04:28,000
then I started building these skills into it, and then I

889
01:04:28,000 --> 01:04:31,599
noticed that both of these were kind of dumb - like, my terribly

890
01:04:31,599 --> 01:04:35,039
brilliant insights about the world that got curated into that journal,

891
01:04:35,039 --> 01:04:38,799
well, maybe they weren't actually that brilliant, and they went stale pretty fast, and

892
01:04:38,799 --> 01:04:42,000
they just ate up the agent's memory, since it had to figure out what

893
01:04:42,000 --> 01:04:46,039
Sami had thought about on some other day back then, i.e. in December 2025, and then those

894
01:04:46,039 --> 01:04:49,359
skills got even worse, because they basically just got in the way of the agent's

895
01:04:49,359 --> 01:04:53,680
much better general intelligence, so I just deleted all the skills, completely

896
01:04:53,680 --> 01:04:57,079
cold turkey. And then I also deleted this, like, my

897
01:04:57,079 --> 01:05:01,000
notebook - but I kept it around for the assistant, in a way, so that now if

898
01:05:01,000 --> 01:05:05,400
genuinely curated nuggets of information or important stuff comes up, then it

899
01:05:05,400 --> 01:05:09,119
gets burned into the Karpathy-style memory. And I think this has actually worked pretty

900
01:05:09,119 --> 01:05:11,920
well. Yeah. So even though general intelligence keeps improving all the

901
01:05:11,920 --> 01:05:15,839
time, there's still a role for a 'Samantha',

902
01:05:15,839 --> 01:05:19,319
one that you still can't quite get off the shelf.

903
01:05:19,319 --> 01:05:24,119
Yeah. Well, let's put it this way - maybe not quite. So now there's a good question,

904
01:05:24,119 --> 01:05:27,319
which is how far we are from that point, because right now

905
01:05:27,319 --> 01:05:34,480
actually - OpenAI, I haven't looked into it that deeply myself, but they announced

906
01:05:34,480 --> 01:05:38,760
recently that they've overhauled their memory system. It was

907
01:05:38,760 --> 01:05:42,000
previously actually implemented pretty poorly, and I didn't even

908
01:05:42,000 --> 01:05:46,000
keep it turned on myself. But now there is, yeah, um,

909
01:05:46,000 --> 01:05:50,240
at least according to their own marketing, they have

910
01:05:50,240 --> 01:05:54,319
a considerably more advanced memory now, where, and then you can

911
01:05:54,319 --> 01:05:57,400
have separate memories for different projects, or you can

912
01:05:57,400 --> 01:06:02,359
use the global memory - so they really seem to be coming on strong now

913
01:06:02,359 --> 01:06:09,319
in this area, and the thing is, it might get hard to compete with. I mean

914
01:06:09,319 --> 01:06:13,400
right now it's the case that at OpenAI,

915
01:06:13,400 --> 01:06:17,720
and at Anthropic too, they pay pretty solid salaries, which attracts

916
01:06:17,720 --> 01:06:20,440
quite a bit - well, Andrej Karpathy went there, and Peter

917
01:06:20,440 --> 01:06:23,799
Steinberg went over to OpenAI, so basically these

918
01:06:23,799 --> 01:06:27,599
memory-world gurus are on both sides now. Yeah, and actually

919
01:06:27,599 --> 01:06:33,440
OpenAI - the whole 'Open Glow' project can be seen as OpenAI's

920
01:06:33,440 --> 01:06:36,279
public development platform for their memory architecture, so...

921
01:06:36,279 --> 01:06:40,799
Yeah, yeah, exactly, [laughter] I mean, it can be, it can be

922
01:06:40,799 --> 01:06:45,319
challenging, but still, I don't think it's a lost cause - because even if

923
01:06:45,319 --> 01:06:51,039
- well, even if it turns out that these so-called general-purpose, off-the-shelf

924
01:06:51,039 --> 01:06:55,520
products end up being, and probably soon will be, better than what you're able

925
01:06:55,520 --> 01:06:59,400
to build yourself, I still think that's really fun.

926
01:06:59,400 --> 01:07:01,279
Yeah. And a really good learning experience,

927
01:07:01,279 --> 01:07:04,279
because in any case, even if those

928
01:07:04,279 --> 01:07:08,400
off-the-shelf products turn out better, you've at least learned how to use them. You

929
01:07:08,400 --> 01:07:12,279
get to use those off-the-shelf products effectively, which gives you

930
01:07:12,279 --> 01:07:15,279
a huge competitive edge in this day and age,

931
01:07:15,279 --> 01:07:17,720
where knowing how to effectively

932
01:07:17,720 --> 01:07:22,920
use AI is honestly an insane competitive advantage right now.

933
01:07:22,920 --> 01:07:28,359
Yeah. And there's still some of that anthropomorphizing magic dust to it too, in that

934
01:07:28,359 --> 01:07:32,920
more and more people know what a 'Samantha' is. In a way, that's, in a certain sense,

935
01:07:32,920 --> 01:07:37,720
a skill in itself too, that I can - I've swapped out the brain from Anthropic for

936
01:07:37,720 --> 01:07:43,160
OpenAI's, and just for fun I set up DeepSeek too as a backup model, and then

937
01:07:43,160 --> 01:07:47,279
say, whichever's the era's best

938
01:07:47,279 --> 01:07:52,640
video generator models - the point being that the harness isn't

939
01:07:52,640 --> 01:07:57,799
model-locked - or rather, I can swap it with minimal hassle

940
01:07:57,799 --> 01:08:02,520
from one camp to another, or one API to another. And that, to me, is a

941
01:08:02,520 --> 01:08:06,039
genuinely strategic defensive capability. And then I could go full bunker mode.

942
01:08:06,039 --> 01:08:10,559
I've actually been thinking about buying, like, a 256

943
01:08:10,559 --> 01:08:15,079
gigabyte unified-memory Spark, if I go the Nvidia route, or a Mac

944
01:08:15,079 --> 01:08:18,759
Studio - once I've got one running, bunker's ready, internet cut off, and

945
01:08:18,759 --> 01:08:22,640
everything runs 100% on local inference. So basically

946
01:08:22,640 --> 01:08:26,159
the thing is, since I don't fully trust that if I go with, say,

947
01:08:26,159 --> 01:08:29,759
Claude, they're going to try to build a moat, so that you just can't

948
01:08:29,759 --> 01:08:32,880
switch over to OpenAI - let alone Gemini - or

949
01:08:32,880 --> 01:08:37,679
Chinese models, not to mention. So self-sufficiency is, in my view,

950
01:08:37,679 --> 01:08:41,960
also a strategic competitive advantage. Yeah. Yeah, absolutely. Yeah. And

951
01:08:41,960 --> 01:08:47,799
nothing is as fun as tinkering with local models.

952
01:08:47,799 --> 01:08:51,799
Yeah, though it can be a bit of a letdown too. I was super excited - like,

953
01:08:51,799 --> 01:08:56,372
when Google put out Gemma 4, I thought, now we can do anything with this.

954
01:08:56,372 --> 01:08:58,920
[chuckle] Turns out you couldn't really do much of anything

955
01:08:58,920 --> 01:09:05,560
with it at all, [laughter]. Yeah, it's - I'm at the same

956
01:09:05,560 --> 01:09:10,600
time disappointed and happy about open source development. Like,

957
01:09:10,600 --> 01:09:14,960
disappointed right now because it's a fact

958
01:09:14,960 --> 01:09:19,400
that [clears throat] the gap is pretty big. If you think about, say, Fable

959
01:09:19,400 --> 01:09:24,000
or Mytos, and then something like GLM 5.2, which is maybe currently

960
01:09:24,000 --> 01:09:27,920
the most advanced open source model, unfortunately there's still something like a 16-to-12-

961
01:09:27,920 --> 01:09:31,759
month gap. Mm. And that makes me sad. But

962
01:09:31,759 --> 01:09:36,120
then in the same breath, now that we've seen this

963
01:09:36,120 --> 01:09:40,839
US policy, where, where we're being blocked from access to these

964
01:09:40,839 --> 01:09:46,120
best models, I'm honestly insanely happy that

965
01:09:46,120 --> 01:09:51,920
we do have a reasonably strong open source

966
01:09:51,920 --> 01:09:58,239
ecosystem and movement. Because if it turns out that

967
01:09:58,239 --> 01:10:02,199
these, for example, end up being available

968
01:10:02,199 --> 01:10:06,000
exclusively to US citizens, then we're not going to be

969
01:10:06,000 --> 01:10:09,400
left dropping into thin air here, so

970
01:10:09,400 --> 01:10:13,320
yeah, that's such a juicy topic that I'd suggest Aleksi Paavola and I

971
01:10:13,320 --> 01:10:16,679
head over to the members-only side. I'll put the Laki.ai

972
01:10:16,679 --> 01:10:19,640
link in the description - I'd recommend it, maybe with one small

973
01:10:19,640 --> 01:10:23,679
security caveat: when you upload confidential

974
01:10:23,679 --> 01:10:27,480
documents into systems like this, it's worth

975
01:10:27,480 --> 01:10:31,040
pseudonymizing or anonymizing those documents. If there are

976
01:10:31,040 --> 01:10:36,760
client or people's names in there, then change them to, say, 'Person A1', and

977
01:10:36,760 --> 01:10:40,679
then convert 'Person A1' back to Sami Miettinen afterward, once the run is done,

978
01:10:40,679 --> 01:10:44,840
if it's genuinely, truly a security-

979
01:10:44,840 --> 01:10:52,000
sensitive matter, like insider data on publicly listed companies.

980
01:10:52,000 --> 01:10:57,280
But otherwise it's a really good tool, and reliable, and this is

981
01:10:57,280 --> 01:11:03,280
genuinely amazing work - that in the spirit of open source, the

982
01:11:03,280 --> 01:11:07,520
end result doesn't end up being monetized by, say, some Alma Media - you've

983
01:11:07,520 --> 01:11:12,719
gone and opened up that legal interface for us

984
01:11:12,719 --> 01:11:17,960
Finns. Yeah, this really is a good, good cause, and

985
01:11:17,960 --> 01:11:21,040
it's genuinely such a sweet thing that now we're

986
01:11:21,040 --> 01:11:26,600
able to offer a fully competitive, or even

987
01:11:26,600 --> 01:11:32,280
better product, for free, to consumers, when the equivalent

988
01:11:32,280 --> 01:11:36,120
bought from Alma Media would run several thousand euros. It's just not

989
01:11:36,120 --> 01:11:41,360
reasonable or realistic in any way for some individual consumer who

990
01:11:41,360 --> 01:11:45,880
has some legal hassle to have to pay several thousand

991
01:11:45,880 --> 01:11:49,120
for something like that. And then these Legora and

992
01:11:49,120 --> 01:11:53,280
Harvey-type tools, which might not know the first thing about Finnish law, well

993
01:11:53,280 --> 01:11:57,239
then that's still another option for the consumer too, so for this

994
01:11:57,239 --> 01:12:01,440
consumer it's - and then, for companies, I think it's worth

995
01:12:01,440 --> 01:12:06,239
spending a bit of money on the use case, so that

996
01:12:06,239 --> 01:12:10,480
you build proper encryption chains. Actually, one thing that [chuckle]

997
01:12:10,480 --> 01:12:14,679
I could do myself too - this anonymization wrapper, I think, would be

998
01:12:14,679 --> 01:12:18,199
something worth building, maybe as open source, because

999
01:12:18,199 --> 01:12:22,440
it's the kind of thing that makes me a little worried about these MCP servers, if

1000
01:12:22,440 --> 01:12:25,239
people go and upload documents that are too sensitive, then there's

1001
01:12:25,239 --> 01:12:28,520
still, at least in theory, a small security risk.

1002
01:12:28,520 --> 01:12:34,159
Well yeah, that's exactly right, that's how it is - what we're aiming for now is

1003
01:12:34,159 --> 01:12:38,360
that the limiting factor would be your Claude or ChatGPT license.

1004
01:12:38,360 --> 01:12:42,440
So if you had, say, a Business or Enterprise plan

1005
01:12:42,440 --> 01:12:46,600
from Claude or ChatGPT, you could put all the same things into it

1006
01:12:46,600 --> 01:12:51,080
that you'd already dare put into email or OneDrive anyway.

1007
01:12:51,080 --> 01:12:55,800
And that's where we're going to be with Laki.ai pretty soon too. So

1008
01:12:55,800 --> 01:13:02,679
then you'll be able to put in those, kind of, sensitive things,

1009
01:13:02,679 --> 01:13:06,639
the same as you can now put into public inter- or, well, kind of

1010
01:13:06,639 --> 01:13:10,600
into systems connected to the public internet. Like, for example,

1011
01:13:10,600 --> 01:13:13,719
email - you can put them in there.

1012
01:13:13,719 --> 01:13:19,440
But of course, you still have to trust

1013
01:13:19,440 --> 01:13:24,400
whichever it is, Anthropic or OpenAI - so that's maybe something everyone has to

1014
01:13:24,400 --> 01:13:27,880
personally weigh, what they actually want to put in there. And

1015
01:13:27,880 --> 01:13:32,719
then, I don't want to feed the security-doomsayer

1016
01:13:32,719 --> 01:13:37,320
camp too much here, but then, if it's a public

1017
01:13:37,320 --> 01:13:41,000
use case that requires an EU server, then yeah,

1018
01:13:41,000 --> 01:13:44,520
Anthropic's cloud might well be over in the US. So in these

1019
01:13:44,520 --> 01:13:48,679
special situations - and you users out there know who you are -

1020
01:13:48,679 --> 01:13:53,280
it's worth actually reading the fine print of these systems

1021
01:13:53,280 --> 01:13:54,401
carefully.

1022
01:13:54,401 --> 01:13:56,600
[laughter] Exactly. Exactly.

1023
01:13:56,600 --> 01:14:00,120
EU legislation is going to be in Laki.ai pretty soon too, so

1024
01:14:00,120 --> 01:14:06,120
and any day now the AI Act's final

1025
01:14:06,120 --> 01:14:10,000
implementation is coming. Luckily the Omnibus has kicked some of the worst

1026
01:14:10,000 --> 01:14:14,679
messes further down the road. But you can save on legal fees

1027
01:14:14,679 --> 01:14:17,840
by hooking up the Laki.ai MCP and

1028
01:14:17,840 --> 01:14:20,120
yeah, getting familiar with things

1029
01:14:20,120 --> 01:14:24,560
yourself with AI. Great. But, thanks for this, Aleksi Paavola. This

1030
01:14:24,560 --> 01:14:27,719
turned out to be way more interesting stuff than I expected. It's always great when

1031
01:14:27,719 --> 01:14:31,440
I haven't googled the background too much beforehand - you always get pleasantly surprised that you're

1032
01:14:31,440 --> 01:14:35,560
right, like, at the very core of the scene. This is always such a great

1033
01:14:35,560 --> 01:14:38,400
thing. So, thanks for coming on. Thank you so much. This was a lot of fun,

1034
01:14:38,400 --> 01:14:41,440
and, and if you've watched or listened this far, go ahead and subscribe to the

1035
01:14:41,440 --> 01:14:44,560
channel, and let's head over to the members-only side and think about this data

1036
01:14:44,560 --> 01:14:47,360
sovereignty topic - I thought that was so much fun, so on the members-only side

1037
01:14:47,360 --> 01:14:53,080
we'll spend a bit more time thinking about whether Europe still has

1038
01:14:53,080 --> 01:14:56,760
any say in this at all, and what happens if, say, some future Trump

1039
01:14:56,760 --> 01:15:00,199
cuts off our access to the better inference - so, see you on the members-only side

1040
01:15:00,199 --> 01:15:02,521
then. Thanks. Thank you.

1041
01:15:02,521 --> 01:15:04,541
[chuckle]