1
00:00:00,000 --> 00:00:05,160
John, this is your first experience of me recording from the back porch of my father-in-law's shore house.

2
00:00:05,520 --> 00:00:07,980
First time caller, a long-time listener for something like this.

3
00:00:08,080 --> 00:00:15,500
It's always a key part of the summer for me is when Marty transitions to his shore house and does all the pods on the back porch.

4
00:00:15,920 --> 00:00:17,500
It's a real vibe I look forward to.

5
00:00:17,920 --> 00:00:21,220
It's a great vibe. We've got birds chirping in the background. There's not a cloud in the sky.

6
00:00:21,320 --> 00:00:23,960
It's about 74 degrees, very comfortable, not sweating.

7
00:00:23,960 --> 00:00:32,200
and after the the heat and the smoke that that hit the area last week this is a much a much

8
00:00:32,200 --> 00:00:37,820
need of reprieve from from the weird weather yeah you uh i'm glad i'm glad you you made it through

9
00:00:37,820 --> 00:00:44,060
that that rough rough period there did the smoke make it down your way at all no no we're uh we're

10
00:00:44,060 --> 00:00:49,440
insulated from all of canada's uh shenanigans down in uh the uh lower half of the country

11
00:00:49,440 --> 00:00:52,920
Hey, it turns out forest management control burns kind of important.

12
00:00:52,920 --> 00:00:54,440
You neglect them for a decade.

13
00:00:55,200 --> 00:00:56,880
Bad things happen.

14
00:00:56,880 --> 00:00:59,080
And then again, then it gets blamed on climate change.

15
00:00:59,080 --> 00:01:03,240
And you're forced to to move around your whole life in your portfolio

16
00:01:03,240 --> 00:01:05,440
to make sure that we're investing in in clean climate tech.

17
00:01:05,880 --> 00:01:08,400
I think the narrative was a little less powerful this time

18
00:01:08,880 --> 00:01:09,960
for for one reason or another.

19
00:01:09,960 --> 00:01:13,120
Fortunately, we're past the the ESG days of early 2020s.

20
00:01:13,120 --> 00:01:14,640
But never say never.

21
00:01:14,640 --> 00:01:16,920
Investment cycles are always come back around.

22
00:01:16,920 --> 00:01:22,800
Well, I mean, they may come back around, but as you said, it's out of favor now because war is not good for the environment.

23
00:01:22,960 --> 00:01:26,900
War is back on the table. Iran, Shraddh Hormuz, back on the table, boys.

24
00:01:27,620 --> 00:01:32,600
Yeah, here we are again. All this has happened before, all this will happen again, as they say.

25
00:01:32,600 --> 00:01:46,680
We don't have to hit these too hard. I'm sure anyone who is a professional doom scroller or a professional situation monitor is well aware that where we're sitting right now is definitively less encouraging than it was perhaps a couple months ago.

26
00:01:46,920 --> 00:01:50,920
perhaps not totally surprising that we would see re-escalation maybe you might have thought that

27
00:01:50,920 --> 00:01:55,640
it would have waited until after midterms but i think anyone probably would have assumed that the

28
00:01:55,640 --> 00:02:01,240
the mou that was put across the table a couple months ago was was not super well defined on a

29
00:02:01,240 --> 00:02:05,400
variety of terms which we talked about a couple months back but you know here we are again

30
00:02:05,400 --> 00:02:10,760
straights open straights closed michael scott snip snap gif a schrodinger's cat type situation but

31
00:02:10,760 --> 00:02:17,060
Either way, you know, the punchline here is WTI, Brent up back into the 80s.

32
00:02:17,060 --> 00:02:26,420
But I think even more notably, you've got, you know, European nat gas benchmarks, all getting back into kind of highs of the early, early chaos we were seeing earlier this year.

33
00:02:26,420 --> 00:02:44,220
So that'll be, I think, the biggest constraint on, you know, or the biggest pain point, I guess we would say, is the countries, Europe, Southeast Asia, that are maybe less well supplied on a variety of key energy inputs really getting squeezed here again.

34
00:02:44,400 --> 00:02:47,940
So that'll be the main vulnerability to watch.

35
00:02:48,480 --> 00:02:53,460
Stateside, we have the Strategic Petroleum Reserve falling to 43 days of reserves left, lowest, I think.

36
00:02:54,020 --> 00:02:55,240
83, lowest since 83.

37
00:02:55,240 --> 00:03:02,540
Louis since 83. Yeah. So got done to 46 days in 2023. So I guess we replenished it a little bit

38
00:03:02,540 --> 00:03:08,640
and we've drained it down to its lowest level in over 40 years. So not great, Bob.

39
00:03:08,960 --> 00:03:14,040
No. Well, it might've been a good idea to deal with that before kicking the hornet's nest here,

40
00:03:14,200 --> 00:03:20,620
but here we are. But I think, you know, everyone, everyone knows there's the war is on,

41
00:03:20,660 --> 00:03:24,1000
if we can call it that now. I think maybe the more incremental or like interesting pieces out

42
00:03:25,000 --> 00:03:30,960
of the last week where some headlines just on basically the accelerated urgency to route around

43
00:03:30,960 --> 00:03:36,120
straightforward news by various GCC countries. We've talked about these a little bit in the past

44
00:03:36,120 --> 00:03:40,700
on the pod, but just mentioning analysis out of Goldman this week, kind of suggesting that a

45
00:03:40,700 --> 00:03:48,600
variety of these projects could get up to say 60% of pre-war levels of oil exports in the next two

46
00:03:48,600 --> 00:03:53,460
years being essentially straight proof. And definitely looks like there's a significant

47
00:03:53,460 --> 00:03:58,760
appetite as you would expect to ramp up both investment in these and then also the pace at

48
00:03:58,760 --> 00:04:04,840
which they get completed. Now, this isn't going to help you if you end up with the Houthis over in

49
00:04:04,840 --> 00:04:09,280
the Red Sea causing a lot more trouble and they're Iran aligned. So you're not necessarily out of the

50
00:04:09,280 --> 00:04:12,940
woods just because you built some of these pipeline projects. And of course, they're not

51
00:04:12,940 --> 00:04:17,420
going to materialize tomorrow. It's going to take one or two years to really get meaningful benefit.

52
00:04:17,420 --> 00:04:39,740
But the world really moving kind of in a direction, it seems, of minimizing dependence on and ability to this particular maritime choke point, which interestingly was specifically called out, as we mentioned before, in the national security strategy document last year as a key priority of the administration, you know, minimizing the disruptive impacts of the Stradivariah Hormuz as a choke point.

53
00:04:39,740 --> 00:04:47,120
So, you know, a tinfoil hat theorist might might say that this was maybe an outcome that was that certain parties were looking for.

54
00:04:47,280 --> 00:04:48,960
But we'll see how it actually shapes up.

55
00:04:49,480 --> 00:04:55,900
Yeah. And I'm not sure if you saw the headline this morning out of Yemen, but it looks like the Yemeni's government or the many government, excuse me,

56
00:04:55,960 --> 00:05:00,020
has announced that they're going to put a naval blockade on Saudi Arabia's ports.

57
00:05:01,020 --> 00:05:05,820
I actually I actually missed that one. So you're you're breaking news for everyone here.

58
00:05:06,620 --> 00:05:06,900
Yes.

59
00:05:07,220 --> 00:05:12,760
So if you're thinking about what's going to happen when these pipelines do make it to the Red Sea, as you mentioned, the Houthis.

60
00:05:12,920 --> 00:05:17,180
I mean, Yemen, I think, loosely lying there with some Houthi outshoots in the country.

61
00:05:17,600 --> 00:05:20,980
Looks like they're already beginning to signal like, hey, you're not going to get away with this.

62
00:05:21,060 --> 00:05:22,460
We have something to say, too.

63
00:05:22,460 --> 00:05:25,720
So chaos in the Middle East oil markets.

64
00:05:25,720 --> 00:05:41,920
And as this chaos is going, it looks like the U.S. is trying to carry favor with some of the countries in the area, and the latest of which is Iraq, which the U.S. and Iraq have a pretty muddy history in the 21st century.

65
00:05:42,040 --> 00:05:47,740
But it looks like the Iraqi prime minister will be visiting Washington to talk about oil and gas deals.

66
00:05:48,440 --> 00:05:55,060
Yeah. I mean, you talk about cyclicality, right? You think about the 90s and H.W. Bush shaking hands with Saddam Hussein.

67
00:05:55,560 --> 00:06:03,160
15, 20 years later, he gets summarily deposed and we do some nation building and it's a chaotic scene all around.

68
00:06:03,280 --> 00:06:10,860
And then 20 years later, here we are again, striking deals and building up partnerships that will benefit allegedly U.S. companies like Chevron.

69
00:06:10,860 --> 00:06:26,360
So, you know, just I think these are these are self-explanatory, but I think it's as it relates to a U.S. that's looking to flex its might more and more in energy markets internationally and become more and more of a dominant player across petroleum markets.

70
00:06:26,360 --> 00:06:30,920
clearly one to watch here, humorous enough that it's all coming back around to being a partner

71
00:06:30,920 --> 00:06:38,060
with Iraq. But just the latest kind of puzzle piece, plug me in here to the US's clear desire

72
00:06:38,060 --> 00:06:42,480
to be more and more influential and puts hand in the till more and more in global oil markets.

73
00:06:43,220 --> 00:06:48,860
Yes. And I mean, we've been covering this extensively, which is a lot of the focus on

74
00:06:48,860 --> 00:06:55,180
the straighter communities may be a focus that's not in the right area. It's what's happening

75
00:06:55,180 --> 00:06:57,500
on the edges as the

76
00:06:57,500 --> 00:06:59,240
Strait of Hormuz shuts down

77
00:06:59,240 --> 00:07:01,020
and how is everything being

78
00:07:01,020 --> 00:07:03,360
redirected and re-architected on the

79
00:07:03,360 --> 00:07:05,300
back end and helium, which

80
00:07:05,300 --> 00:07:07,260
as we now

81
00:07:07,260 --> 00:07:09,040
know after becoming experts on

82
00:07:09,040 --> 00:07:11,560
Middle East oil and gas strategy

83
00:07:11,560 --> 00:07:13,540
is a key input

84
00:07:13,540 --> 00:07:15,160
in fertilizer and other

85
00:07:15,160 --> 00:07:17,000
sort of other

86
00:07:17,000 --> 00:07:19,120
semis

87
00:07:19,120 --> 00:07:21,340
semis. It's good for

88
00:07:21,340 --> 00:07:23,520
helium is very important. That's what I'm trying to say

89
00:07:23,520 --> 00:07:28,560
back deck fog is hitting, but it looks like the US is going to be a massive

90
00:07:28,560 --> 00:07:31,920
benefactor in the helium markets because of all this.

91
00:07:31,920 --> 00:07:36,320
So look, if you're looking at the chart that we have on the screen right now,

92
00:07:36,320 --> 00:07:39,120
it's looking, it says US emerges as helium winner.

93
00:07:39,120 --> 00:07:59,852
We looking at the import of noble gases from top three sources in 2026 percentage of total So US for looking at 2022 and to japan was hovering around 25 30 it be around 80 this year south korea it was below 20 in 2022 and is approaching 60 this year

94
00:07:59,852 --> 00:08:06,972
and then taiwan similarly below 20 and looking like it's going to be just at 60 of their helium

95
00:08:06,972 --> 00:08:11,052
imports will be from the US. So big winner in the helium markets, the US is.

96
00:08:12,092 --> 00:08:16,332
Yeah. And if you look at kind of the trend here, if you're watching on video, like this was

97
00:08:16,332 --> 00:08:19,932
already kind of in play and in progress over the last few years. So it's not just

98
00:08:21,132 --> 00:08:25,852
the disruption that we've seen in the Gulf, but clearly there's been a jump this year,

99
00:08:25,852 --> 00:08:33,292
a meaningful jump, especially for Korea and Taiwan across for the US as a source of helium imports

100
00:08:33,292 --> 00:08:40,092
for these countries and you know it's the the the red bar here of china getting squeezed as a as a

101
00:08:40,092 --> 00:08:45,452
player that was also that trend was was playing out even before this but i just thought it was a

102
00:08:45,452 --> 00:08:50,972
good chart to to reiterate it came out of i believe nikai this this week something we've been talking

103
00:08:50,972 --> 00:08:57,692
about again over the last few months just the the wins on the margin that the us is is seeing a lot

104
00:08:57,692 --> 00:09:02,892
of kind of key key inputs that where it's becoming more and more of a relevant player an exporter

105
00:09:02,892 --> 00:09:14,012
And I think, again, just cuts against the idea that there's no, you know, there's no edge, there's no strategy, nothing behind what the U.S. is doing other than kind of Trump flailing blindly in the Gulf.

106
00:09:14,172 --> 00:09:17,512
Not to say that everything has been executed perfectly or without cost.

107
00:09:17,692 --> 00:09:22,172
And, you know, we don't have the chart on here, but there was a really good paper out of, I forget the organization.

108
00:09:22,292 --> 00:09:30,352
It's one of the big, basically, energy monitoring and commodity monitoring organizations, kind of their annual paper on the state of play on critical minerals.

109
00:09:30,352 --> 00:09:38,672
So the flip side of this chart is, you know, unsurprisingly, China remains well ahead of anybody on kind of refined capacity for rare earths.

110
00:09:38,672 --> 00:09:45,552
I think that explains, you know, unsurprisingly why you've seen all the headlines we flagged over the last few weeks on the U.S.

111
00:09:45,712 --> 00:09:58,692
doing a lot more on the industrial policy side to build out dual sourcing capabilities and do partnerships with nations like Brazil to further bolster U.S. rare earth capacity.

112
00:09:58,692 --> 00:10:10,752
So it's pros and cons on both sides, again, like we've been saying, but there's definitely clear benefit that you're seeing to the U.S. from this disruption kind of one way or another.

113
00:10:10,752 --> 00:10:28,192
Yes. And then we transition to how the Fed is trying to operate and navigate in this world, not only this world, but in a world of their own in which they're trying to drastically overhaul, get away from forward guidance, add more data resources that can react in real time.

114
00:10:28,192 --> 00:10:34,432
And we've got some statements from the Fed this week pertaining to inflation.

115
00:10:35,012 --> 00:10:47,952
And I think they're really, again, trying to position away from forward guidance and saying they're going to be more reactive with lifetime data saying we can't get stuck in the cognitive dissonance of fighting the last war on inflation.

116
00:10:48,792 --> 00:10:53,832
And a couple of Fed governors coming out and say they believe inflation has peaked and saying it pretty confidently.

117
00:10:53,832 --> 00:10:58,492
So we've got the Fed beginning to position its new regime here.

118
00:10:59,152 --> 00:11:06,012
Yeah, which is never exactly what you want to see the Fed trying to call it inflection one way or another, especially on inflation.

119
00:11:06,232 --> 00:11:11,872
You know, it brings back the PTSD memories of inflation is transitory from anyone who lived through 20 and 21.

120
00:11:12,392 --> 00:11:15,712
But I just thought it was an interesting set of comments that came out last week.

121
00:11:15,712 --> 00:11:24,512
The top one is from new Fed chair Kevin Warsh in front of Congress testifying that, you know, the if we do things right, the inflation surge the last five years is a thing of the past.

122
00:11:24,512 --> 00:11:28,912
And he talked a little about AI as well in in that in those remarks.

123
00:11:28,912 --> 00:11:39,792
And in the past, if you've looked, he's been very on the train of AI productivity and enabling kind of a deflationary impulse that can counteract some of the more inflationary forces that we've seen over the last few years.

124
00:11:39,792 --> 00:11:44,672
and not wanting to get in the way of that and hamstring that and really allow that to flourish

125
00:11:44,672 --> 00:11:49,512
and to allow for that kind of deflationary impact. And then Fed Governor Waller

126
00:11:49,512 --> 00:11:55,172
telling people that we shouldn't be, as he says, fighting the last war, just kind of looking

127
00:11:55,172 --> 00:11:57,952
backward and doing what we should have done five years ago in a totally different set of

128
00:11:57,952 --> 00:12:02,652
circumstances. And then John Williams saying, like you said, putting in a top for rightly or wrongly

129
00:12:02,652 --> 00:12:06,832
on inflation. Collectively, I just think that's an interesting talk track from all these guys that

130
00:12:06,832 --> 00:12:10,152
Um, allegedly we're moving away from forward guidance.

131
00:12:10,152 --> 00:12:14,112
So I don't know if this is intended to be a forward guidance light, or if, you know,

132
00:12:14,112 --> 00:12:19,032
maybe the worst doesn't totally have a handle on how much the rest of his team goes out

133
00:12:19,032 --> 00:12:20,392
and talks to the media.

134
00:12:20,652 --> 00:12:25,712
But in any case, the trend line here is not suggestive of, if a governor is trying to

135
00:12:25,712 --> 00:12:29,392
tell you that they're going to have to get more hawkish and that their, their key focus

136
00:12:29,392 --> 00:12:31,072
is on getting, you know, CPI down.

137
00:12:31,432 --> 00:12:35,492
Which is interesting given that all of these comments were happening in the midst of, you

138
00:12:35,492 --> 00:12:39,972
know, WTI ripping back over 80 and all the, you know, the energy input spikes that we talked about

139
00:12:39,972 --> 00:12:45,492
as uncertainty flared up to, you know, a multi-month high in the Gulf again, against that

140
00:12:45,492 --> 00:12:48,512
backdrop, you've got all these guys telling you, well, you know, inflation is probably like, it's

141
00:12:48,512 --> 00:12:52,252
important, but let's not fight the last war. I think it's peaked. I think we can do enough to

142
00:12:52,252 --> 00:12:56,472
make it a thing of the past. Like just an interesting dissonance between those two that

143
00:12:56,472 --> 00:13:03,892
is, I think, indicative of how the, how the institution is, is shading toward positioning

144
00:13:03,892 --> 00:13:07,672
itself. You know, there's a lot of internal division and strife apparently among all the

145
00:13:07,672 --> 00:13:12,752
different Fed governors and people with a vote here. But I think this is putting into the zeitgeist

146
00:13:12,752 --> 00:13:17,632
and expectation for incrementally, you know, incremental dovishness versus incremental

147
00:13:17,632 --> 00:13:24,052
hawkishness. Yeah, it seems I mean, and then you factor in the inflation prints of last week, too,

148
00:13:24,152 --> 00:13:32,492
which surprised to the downside. And you could see if the war is blowing up, obviously interest

149
00:13:32,492 --> 00:13:37,812
expenses going up, I think this is going to be the biggest annual military spin that we've ever

150
00:13:37,812 --> 00:13:41,792
seen. At least that's what Trump and the administration are asking for. So the fiscal

151
00:13:41,792 --> 00:13:45,972
side is continuing to completely blow out. And I find it hard that they're going to be able to

152
00:13:45,972 --> 00:13:51,772
hike rates into that. Yeah. I mean, we've talked about it a bunch here, but the flexibility isn't

153
00:13:51,772 --> 00:13:56,352
tremendous here, particularly if you want to enable everything you just talked about

154
00:13:56,352 --> 00:14:02,432
and enable Worsh's key priority of not stepping on the AI build out. And perhaps that

155
00:14:02,432 --> 00:14:04,952
leads us nicely into this next topic.

156
00:14:05,472 --> 00:14:06,512
Yes, which is Kimmy.

157
00:14:06,592 --> 00:14:07,612
Kimmy's so hot right now.

158
00:14:08,012 --> 00:14:08,892
Kimmy, come back.

159
00:14:09,612 --> 00:14:11,652
Any other fool could see.

160
00:14:12,292 --> 00:14:13,472
Yeah, I mean, this is...

161
00:14:13,472 --> 00:14:15,832
It's an important component of every episode.

162
00:14:16,272 --> 00:14:18,912
This is a shot across the bow of the frontier labs

163
00:14:18,912 --> 00:14:19,832
in the U.S.

164
00:14:19,932 --> 00:14:21,972
Kimmy K3 came out

165
00:14:21,972 --> 00:14:24,092
surpassing many of the

166
00:14:24,092 --> 00:14:26,572
U.S. frontier models on some benchmarks.

167
00:14:26,572 --> 00:14:28,232
And I think

168
00:14:28,232 --> 00:14:30,532
at least reading the tea leaves of people

169
00:14:30,532 --> 00:14:32,092
who are using Kimmy K3

170
00:14:32,092 --> 00:14:38,452
and comparing it to Fable 5 and Chachi BT 5.6 Sol and Terra,

171
00:14:39,112 --> 00:14:43,952
it is a legitimate contender for top Frontier model out there.

172
00:14:44,132 --> 00:14:47,712
Obviously, it is a Chinese open weight model.

173
00:14:48,312 --> 00:14:52,012
And I think one of the interesting things here is,

174
00:14:52,092 --> 00:14:53,472
I think we might have discussed it last week,

175
00:14:53,512 --> 00:14:57,772
but if we didn't, Dylan Patel from Semi Analysis wrote a piece earlier this year

176
00:14:57,772 --> 00:15:04,212
that said the Frontier Labs have a massive lead and will be able to defend that lead

177
00:15:04,212 --> 00:15:09,632
because of the relative lack of access to GPUs that the Chinese market has.

178
00:15:09,772 --> 00:15:16,892
He was making the case that they won't be able to train the models in the way that Anthropic and OpenAI can do so here in the States

179
00:15:16,892 --> 00:15:24,312
because they have access to all NVIDIA's chips in China due to export controls, tariffs, and other reasons,

180
00:15:24,312 --> 00:15:25,992
doesn't have access to those ships

181
00:15:25,992 --> 00:15:28,332
and therefore will not be able to train models

182
00:15:28,332 --> 00:15:43,224
comparable with U Frontier models But it can be launched It is comparable It is competitive And I guess that begs the question are they getting access to the GPUs Are they finding ways to train these models way more

183
00:15:43,224 --> 00:15:50,564
efficiently with less tokens? And what does this mean for the future of AI, particularly as it

184
00:15:50,564 --> 00:15:56,164
pertains to this juxtaposition of open weight models versus closed source frontier models here

185
00:15:56,164 --> 00:16:01,964
in the United States, which are cozying up with the government to try to prevent the layman from

186
00:16:01,964 --> 00:16:06,784
getting access to the full power of these AI tools. And I don't know if you have them on the

187
00:16:06,784 --> 00:16:11,504
list, but I'm sure we'll talk about them. But there were two headlines that came out over the

188
00:16:11,504 --> 00:16:16,144
last five days that really highlight this one pertaining to a dear friend of ours, Calais,

189
00:16:16,764 --> 00:16:20,584
who's a developer of Cashew, getting a shout out from David Sachs last night.

190
00:16:21,164 --> 00:16:25,704
Again, you're breaking news here. I missed that one. We should definitely talk about that

191
00:16:25,704 --> 00:16:27,744
Um, cause that's, uh, that's big stuff.

192
00:16:27,744 --> 00:16:32,724
Uh, if you've been following Bitcoin for a little while, but yeah, I think, I mean,

193
00:16:32,724 --> 00:16:36,304
there's a ton to say here, you know, we're not gonna, you don't have time to get into

194
00:16:36,304 --> 00:16:41,384
drill into every possible thing, but I mean, there's, there's a lot of nuance within the

195
00:16:41,384 --> 00:16:45,624
chart like this and within this whole, this whole topic, this immediately kicked off,

196
00:16:45,664 --> 00:16:48,584
you know, all of the debates about, well, are they, are they benchmark maxing?

197
00:16:48,704 --> 00:16:52,344
And it's not actually as, you know, uh, it's not as actually as comparable in real world

198
00:16:52,344 --> 00:16:52,624
use.

199
00:16:52,624 --> 00:16:57,764
We've talked about kind of that dynamic with open source versus true kind of frontier models in the past.

200
00:16:58,244 --> 00:17:01,524
I think I haven't played with Kimi K3 enough to have an opinion on that.

201
00:17:01,524 --> 00:17:08,904
Definitely in the past, I've been less impressed for like at least agentic deployments by the open source models relative to everything else on the frontier.

202
00:17:09,304 --> 00:17:14,384
And I think you've generally felt similarly, but, you know, maybe we'll come back in a week or two on analysis there.

203
00:17:14,524 --> 00:17:20,164
But there's also the question of, well, is like you kind of alluded to, are there shadow GPU clusters in like Singapore?

204
00:17:20,164 --> 00:17:26,064
There are all sorts of, you know, shell games that can be played to, you know, mask who the actual end user is.

205
00:17:26,184 --> 00:17:31,264
So, like, has China actually been getting, you know, more access to true frontier accelerators?

206
00:17:31,664 --> 00:17:38,484
TPD, are they using distillation that both the labs and the U.S. government could eventually, you know, crack down on more effectively?

207
00:17:38,744 --> 00:17:39,984
You know, that's that's potential, too.

208
00:17:40,344 --> 00:17:45,144
I think there's good arguments to be had by guys like Gavin Baker.

209
00:17:45,144 --> 00:17:46,964
We'll have a tweet from him here in a second.

210
00:17:46,964 --> 00:18:00,364
But, you know, some good analysis that you put out just on the relative cost of Kemi K3 versus, say, Fable or Sol on a true per task basis, because K3 appears to be much more token hungry.

211
00:18:00,544 --> 00:18:08,104
And so if you need a lot more tries effectively per task completion, do you actually get real savings in commercial deployments or not?

212
00:18:08,104 --> 00:18:20,104
I think that's, you know, very much TBD. All this is to say, like, I don't think that the massive kind of doomer headlines or the the the desire that I see from a lot of people, weirdly, especially like Bitcoiners.

213
00:18:20,764 --> 00:18:27,244
I think there's like this this desire to like this desperation to call a top on the quote unquote AI bubble and to say, like, this is it.

214
00:18:27,344 --> 00:18:30,784
It's all going to be, you know, one shot by open source models.

215
00:18:30,984 --> 00:18:36,044
You know, so I think maybe there's like an underlying desire to tell people like sell all your semis and buy Bitcoin.

216
00:18:36,044 --> 00:18:41,304
so maybe that's a piece of the the motivation but everything i'm kind of getting at here is that i

217
00:18:41,304 --> 00:18:46,524
think all simplistic readings on kind of what this this headline means 48 hours in or you know

218
00:18:46,524 --> 00:18:52,904
maybe quite overdone and need a little reconsideration but you know i think there's that's

219
00:18:52,904 --> 00:18:57,524
that's all like one one piece of it um i have i have a sense that you you want to jump in here

220
00:18:57,524 --> 00:19:03,284
on hugging face so i'm gonna pause and pause my rant well yeah the calling tops are like saying

221
00:19:03,284 --> 00:19:06,844
that this is the death of US closed-source frontier models.

222
00:19:06,944 --> 00:19:07,844
I think that's way overblown.

223
00:19:07,944 --> 00:19:11,484
I think the more important signal here is the kneecapping of these models

224
00:19:11,484 --> 00:19:15,984
and the sort of regulatory encroachment that the Trump administration is taking.

225
00:19:16,144 --> 00:19:19,664
And I think that's what David Sachs was pointing out in his tweet last night.

226
00:19:19,724 --> 00:19:23,244
But just to sort of preface this the right way, to your point, like who's to know?

227
00:19:23,624 --> 00:19:27,784
If they're benchmark maxing, if they have these shadow GPUs,

228
00:19:27,784 --> 00:19:31,484
if they're waging distillation training attacks, I don't know if they're attacks.

229
00:19:31,484 --> 00:19:33,424
however you would describe them.

230
00:19:33,444 --> 00:19:35,344
But I think that misses the point.

231
00:19:35,524 --> 00:19:37,524
And I think the two headlines were the Hugging Face story

232
00:19:37,524 --> 00:19:39,444
and then Calais, he basically wrote a tweet,

233
00:19:39,564 --> 00:19:41,504
I think Friday or Saturday, basically saying,

234
00:19:41,924 --> 00:19:47,284
I found 17 or 12 critical security vulnerabilities

235
00:19:47,284 --> 00:19:50,204
in a couple of the projects I'm maintaining with Kemi K3.

236
00:19:50,484 --> 00:19:55,784
I tried to do them with Fable 5 and with GPT 5.6 sold,

237
00:19:55,984 --> 00:20:00,284
but due to the sort of curbing of those models,

238
00:20:00,284 --> 00:20:03,384
particularly when it comes to working on code bases.

239
00:20:04,164 --> 00:20:07,704
So for those who are unaware, Anthropic famously will let you use Fable 5,

240
00:20:07,884 --> 00:20:12,704
but if you're doing anything that is security sensitive in a code base,

241
00:20:12,824 --> 00:20:18,864
they will divert you to Opus 4.8, to a lesser model, to a model that's not as performant.

242
00:20:19,704 --> 00:20:23,784
And I think this would be a signal that KimiK3 is comparable, at least to an extent,

243
00:20:23,964 --> 00:20:29,324
but two cases over the weekend, Hugging Face, they had, I mean, this is just a crazy story.

244
00:20:29,324 --> 00:20:35,804
crazy story generally and Kimmy K3 is just like a small detail of it but they had in the what

245
00:20:35,804 --> 00:20:41,164
they're describing as the first sort of autonomous AI attack on their code base they got a code

246
00:20:41,164 --> 00:20:48,904
injection in one of their databases and they had an autonomous AI attacker execute more than 17,000

247
00:20:48,904 --> 00:20:55,304
events over the course of a weekend they they found the the attack and they started protecting

248
00:20:55,304 --> 00:21:01,304
against it. But again, they were trying to use Fable 5 and chat GBT 5.6 sold to basically

249
00:21:01,304 --> 00:21:08,424
wage a defensive attack, a security audit of their systems of what was being infiltrated.

250
00:21:08,904 --> 00:21:14,004
And Fable 5 and GBT 5.6 said, you can't do this. You're not allowed to work on this type of

251
00:21:14,004 --> 00:21:20,244
this type of task with with these models. And they were forced to download Kimmy K3 in a local

252
00:21:20,244 --> 00:21:25,004
environment and run it and run the security audit with it against their systems. And it was able to

253
00:21:25,004 --> 00:21:31,484
identify with Kimmy K3 what was wrong. Similarly with Kali, I don't know even say which project it

254
00:21:31,484 --> 00:21:36,324
was specifically. It could be Cashew, it could be BitChack, it could be Klawi. He's running many

255
00:21:36,324 --> 00:21:41,844
things, but similarly was trying to do a security audit of his systems and was using Fable 5,

256
00:21:42,064 --> 00:21:47,344
GBT 5.6. They said, hey, you're not allowed to do this task with these models. And he was forced

257
00:21:47,344 --> 00:21:55,304
use Kimmy K3 and was able to find 12 security flaws in the code base that he was auditing at

258
00:21:55,304 --> 00:22:02,184
that time. So this isn't really a debate about whether or not which frontier model is better

259
00:22:02,184 --> 00:22:07,304
than the other. It's what is the nature of being able to use these models? How are we going to be

260
00:22:07,304 --> 00:22:12,584
able to use these models? And I think the open weight Chinese model is really proving that what

261
00:22:12,584 --> 00:22:16,424
we're doing here in the United States is not going to work long term. You can't cuck the models. You

262
00:22:16,424 --> 00:22:20,224
can't prevent people from using them or you're going to get out competed and whether or not

263
00:22:20,224 --> 00:22:25,864
Kimmy K3 is actually as performant or more performant than Fable 5 or ChatGPT 5.6 sold

264
00:22:25,864 --> 00:22:31,024
right now I think is irrelevant it's like if like people are going to want to use these tools in

265
00:22:31,024 --> 00:22:36,084
particular ways and if the frontier models here in the U.S. don't let you do that because of

266
00:22:36,084 --> 00:22:41,984
fears of of the models breaking containment or doing things that the government doesn't like

267
00:22:41,984 --> 00:22:45,204
it doesn't matter if you have an open source competitor that's going to let you do it you're

268
00:22:45,204 --> 00:22:51,284
going to fall back to that, even if it's not as performant. Yeah, I mean, I think this is a great

269
00:22:51,284 --> 00:22:57,724
maybe segue into that, the Gavin Baker tweet, and the other thing on that page, which is basically

270
00:22:57,724 --> 00:23:03,964
just getting to like your old friend, Jevin's paradox, you know, everyone, it's become this

271
00:23:03,964 --> 00:23:07,304
Jevin's paradox was the thing that I think big corners were talking about, you know, well before

272
00:23:07,304 --> 00:23:13,584
it became kind of a mainstream talking point among everyone in the world. But in any case,

273
00:23:13,584 --> 00:23:16,004
just, just, we're not gonna read everything on here.

274
00:23:16,104 --> 00:23:32,636
I would highly recommend going to read um Gavin tweet and and it his pen tweet and everything else that been on this timeline over the past couple of days On the right is just an example of this playing out where basically you got the CEO of Databricks outlining essentially this dynamic of as on a per unit

275
00:23:32,636 --> 00:23:39,436
basis, this resource effectively gets cheaper. There is that much more or even more a compensating

276
00:23:39,436 --> 00:23:43,476
amount of demand for volume for that resource. Like it's, you know, same as it ever was,

277
00:23:43,476 --> 00:23:48,956
same dynamic that we've seen play out across commodity markets, you know, over the history

278
00:23:48,956 --> 00:23:54,216
of capitalism. And so I think if you split out like what, who is this good for? Who is this bad

279
00:23:54,216 --> 00:23:59,636
for in the broader AI complex? And I think this relates ultimately to not just companies, but also

280
00:23:59,636 --> 00:24:04,436
the US and China. There's like the inference piece and the training piece. And I think this slide is

281
00:24:04,436 --> 00:24:09,156
all about how, you know, the inference piece is going to, is going to be just fine and likely

282
00:24:09,156 --> 00:24:12,916
continue to explode. I think, you know, I think open source models, especially like a diverse

283
00:24:12,916 --> 00:24:19,576
ecosystem of open source models is, you know, giga bullish for most definitely pieces of compute

284
00:24:19,576 --> 00:24:24,456
broadly, different pieces of semiconductors and, you know, the power suppliers that will ultimately

285
00:24:24,456 --> 00:24:28,096
be required to stand up enough capacity to take advantage of all that. I think it's one of the

286
00:24:28,096 --> 00:24:33,316
reasons you see NVIDIA promoting their own open source models so aggressively to kind of avoid the,

287
00:24:33,376 --> 00:24:38,556
you know, the monopsony or the oligopsony of just having a few, you know, close frontier players in

288
00:24:38,556 --> 00:24:42,496
the space dominating the whole, the whole of everything. But that's the inference side. I think

289
00:24:42,496 --> 00:24:47,316
maybe the more interesting strategically like the derivative question is more on the training side,

290
00:24:47,316 --> 00:24:53,716
which you see people kind of wring their hands once again over this last weekend on the degree

291
00:24:53,716 --> 00:24:58,556
to which like training will be financeable at all. These, you know, these massive and increasingly

292
00:24:58,556 --> 00:25:02,316
expensive training runs of tens of billions or like hundreds of billions of dollars over time

293
00:25:02,316 --> 00:25:06,756
to get to the next frontier generation. How can you really underwrite that if, you know,

294
00:25:06,796 --> 00:25:11,156
the ROI is going to be meaningfully undercut by, you know, whether it's a distillation attack or a

295
00:25:11,156 --> 00:25:15,176
fast follower or however you want to think about it. But if, if open source can come in and, and

296
00:25:15,176 --> 00:25:20,436
constantly just ride on the coattails of the, the gains that are made in training of these new

297
00:25:20,436 --> 00:25:24,356
models, then, you know, how are you even going to get a further, further model gains over time?

298
00:25:24,356 --> 00:25:29,476
Because there won't be like an incentive to continue spending on that. And I think this is

299
00:25:29,476 --> 00:25:35,436
like, to your point on open AI and, and Thropic in different ways, cozying up to, cozying after the

300
00:25:35,436 --> 00:25:39,896
government. Like, I think it just kind of gets to a point of like, I don't necessarily even think

301
00:25:39,896 --> 00:25:42,856
that that's right. And that's how it's going to work out, especially if you think that the frontier

302
00:25:42,856 --> 00:25:47,896
labs ultimate strategy is to verticalize everything and to stand up kind of, you know, their own

303
00:25:47,896 --> 00:25:52,896
accounting and law firms and all these different professional services they could go attack as like,

304
00:25:52,896 --> 00:25:56,856
you know, vertical infrastructure on what they build. You know, I think you can very easily

305
00:25:56,856 --> 00:26:01,476
justify meaningful ROI, you know, on that basis, but like, let's even just game it out and say that

306
00:26:01,476 --> 00:26:06,156
that's actually what happens and that you have a meaningful kind of headwind to training spend.

307
00:26:06,156 --> 00:26:12,476
for that economic reason. I think you just have to ask yourself, well, as the tweet on the left

308
00:26:12,476 --> 00:26:17,796
here suggests, K3 shows that scaling laws are continuing, right? Effectively, larger models

309
00:26:17,796 --> 00:26:21,956
are still giving you concomitantly better performance and throwing more compute at

310
00:26:21,956 --> 00:26:27,196
training, throwing more resources at it, growing the size gives you the ability to get better

311
00:26:27,196 --> 00:26:31,796
models. I think if scaling laws generally are still holding and there's still meaningfully more

312
00:26:31,796 --> 00:26:38,896
juice to be squeezed out of massive training runs. And there's still technically like progress that

313
00:26:38,896 --> 00:26:47,036
can be made. It kind of doesn't matter like if the private ROIC is there because neither China

314
00:26:47,036 --> 00:26:50,756
nor the US nor anyone else who's trying to play in this game is going to be able to say,

315
00:26:51,116 --> 00:26:55,576
is going to be able to just concede and let the other side kind of get to it first and own this

316
00:26:55,576 --> 00:26:59,776
incredibly important technology. And they're not going to, as you see on the right here,

317
00:26:59,776 --> 00:27:06,216
of these projections of spend from AI and Anthropic over, especially OpenAI, over the

318
00:27:06,216 --> 00:27:10,176
next five years, we're talking about hundreds of billions of dollars in incremental spend.

319
00:27:10,556 --> 00:27:13,636
And this is very likely not even counting all the derivative spend that needs to come

320
00:27:13,636 --> 00:27:18,536
in from the power and infrastructure side, which is very clearly critical to the way

321
00:27:18,536 --> 00:27:21,356
that the Trump administration is thinking about re-industrialization.

322
00:27:21,756 --> 00:27:25,856
It's critical to the way that Kevin Warsh is thinking about managing a monetary policy

323
00:27:25,856 --> 00:27:32,076
and inflation is having this, having to spend, having to spend here, building this out and then

324
00:27:32,076 --> 00:27:36,956
getting return on that. I think what I'm basically getting to is like, if even if you think that

325
00:27:36,956 --> 00:27:41,196
that's what happens, the training just gets kicked to the sovereign level, right? Like,

326
00:27:41,656 --> 00:27:47,296
you're in no way do I think that the US is ever going to look at this and say, well, we could

327
00:27:47,296 --> 00:27:53,416
continue to dramatically expand model capabilities, if we just kept spending on, you know, these big

328
00:27:53,416 --> 00:27:57,776
training runs, but it's hard to finance because open source, you know, makes it tougher to capture

329
00:27:57,776 --> 00:28:01,976
those returns. And so we're just going to concede on that. I think like what you're looking at here

330
00:28:01,976 --> 00:28:06,836
on the right side is even if you're super bearish on open sources impact on frontier labs, what

331
00:28:06,836 --> 00:28:10,136
you're looking at is just another line item on the federal budget, right? Like someone's going

332
00:28:10,136 --> 00:28:13,696
to be spending that money, right? Whether it's open AI and drop back in there, you know, their

333
00:28:13,696 --> 00:28:18,976
investors or, you know, private financiers or like the government, I think like that curve is going

334
00:28:18,976 --> 00:28:24,596
to move that way regardless, as long as scaling laws hold up. Yeah. And it'd be remiss of us not

335
00:28:24,596 --> 00:28:31,296
to mention, but I think a whole new sort of twist to this juxtaposition of the superpowers in the

336
00:28:31,296 --> 00:28:35,976
US and China and this AI race entered the conversation last Thursday night when President

337
00:28:35,976 --> 00:28:44,076
Trump gave the speech on Chinese meddling in the 2020 election, which is probably a strategic

338
00:28:44,076 --> 00:28:45,576
It's interesting timing, right?

339
00:28:45,796 --> 00:28:51,036
A strategic announcement based off of everything going on right now in the AI world.

340
00:28:51,576 --> 00:28:55,236
Yeah, it's an interesting time for sure to bring that up out of nowhere.

341
00:28:57,936 --> 00:29:04,876
Speaking of interesting times, interesting times in Bitcoin hovering around $64,000 up from the lows of the prior week right now.

342
00:29:04,876 --> 00:29:19,836
And our good friend Alex Thorne and his team at Galaxy Research are just putting together a chart to see what indicators have historically triggered a bottom in Bitcoin have been hit already.

343
00:29:19,976 --> 00:29:24,976
And so we've had four that have hit, two that are approaching, and about seven that are not yet there.

344
00:29:25,196 --> 00:29:25,596
Yeah.

345
00:29:26,256 --> 00:29:31,596
You know, this is, if this is wrong, you know, this is, you can put it on Alex and Galaxy and don't blame us.

346
00:29:31,596 --> 00:29:39,476
But just an interesting kind of data point or set of data points and mile marker on where we are in Bitcoin's cycles, such as it may be.

347
00:29:40,296 --> 00:29:49,776
And I think it certainly looks like if you if you kind of just eyeball this closer, much closer to the bottom than the top, although not fully there yet.

348
00:29:49,776 --> 00:30:01,556
And I just think interesting to as the market has completely moved on from Bitcoin, not in the meta at all, not in the zeitgeist, bouncing around this very kind of tight consolidation range in the low 60s.

349
00:30:01,936 --> 00:30:18,076
If you compare that to everything else we've just talked about with war, defense budget, the Fed's current positioning on inflation, the potential need if you're super bearish on frontier labs for the government to effectively pick up that tab and backstop them, as has been hinted, if that's your belief.

350
00:30:18,076 --> 00:30:25,796
then it's a really interesting time, I think, to consider all those topics against the backdrop of, you know, what Bitcoin's currently doing.

351
00:30:25,956 --> 00:30:36,036
And this last headline that we have from BPI on the Bitcoin, the strategic Bitcoin reserve, Armabil, you know, moving, finally moving to committee, you know, for the first time.

352
00:30:36,436 --> 00:30:40,896
Interesting set of data points here as we look at Bitcoin's breadcrumbs against everything else.

353
00:30:41,116 --> 00:30:44,376
Yes. And I actually was in D.C. last week recorded with Connor Brown.

354
00:30:44,376 --> 00:30:46,976
so if you want the inside look

355
00:30:46,976 --> 00:30:49,416
from the Bitcoin Policy Institute's

356
00:30:49,416 --> 00:30:50,496
perspective on what's happening

357
00:30:50,496 --> 00:30:52,596
with Bitcoin on Capitol Hill

358
00:30:52,596 --> 00:30:54,336
that'll be on the TFTC feed this feed

359
00:30:54,336 --> 00:30:56,336
if you're listening to this on the podcast on Wednesday

360
00:30:56,336 --> 00:30:57,816
so go check that out

361
00:30:57,816 --> 00:30:59,836
and we'll be back next Monday