Why Apple Will WIN The AI Race.. — Transcript

Full transcript

1. 0:00Yo, Greg. What's good, bro? How you
2. 0:01doing?
3. 0:02I'm [snorts] great, man. It's warm and
4. 0:04nice in California.
5. 0:05Nice, bro. I finally got some good
6. 0:07weather over here in New York, bro. We
7. 0:09had like our first 90° day a few days
8. 0:11ago. It was a brutal winter for us this
9. 0:13year. It's bad.
10. 0:14>> Yeah, it's really bad. Really bad, bro.
11. 0:16Gorgeous in in East Bay, San Francisco.
12. 0:20Mediterranean weather.
13. 0:22I wish I could like port that weather
14. 0:23here and have it 24/7. It'd be amazing.
15. 0:26Um
16. 0:26>> Just move. I love New York too much,
17. 0:29bro. I can't. I can't. Um but yeah,
18. 0:32great to have you on. I'm psyched
19. 0:34psyched about this conversation.
20. 0:36Um I think a lot of people have either
21. 0:39heard about Akash in the past or don't
22. 0:41know what you guys have been up to. I
23. 0:42know I recently when I when I checked a
24. 0:44lot of I was really really deep in the
25. 0:46Cosmos. So, I followed a lot of the
26. 0:48early Cosmos projects and you were got
27. 0:50you guys were one of the first Cosmos
28. 0:51projects that I checked when I was
29. 0:53looking. That was probably back in like
30. 0:542021, 2022. So, I know you guys have
31. 0:56been around for a while.
32. 0:57Um can you tell us a little bit about
33. 0:59your background and like what Akash
34. 1:01Akasha is, how you guys have grown over
35. 1:03the years?
36. 1:05So, my background I've been
37. 1:07open source
38. 1:08developer from most of my life. Um
39. 1:12and
40. 1:13focused heavily on distributed systems.
41. 1:16Mhm.
42. 1:17Working mostly for startup companies,
43. 1:19high growth companies here in Silicon
44. 1:21Valley.
45. 1:22Um
46. 1:23Uh before Akash, I founded a company
47. 1:25called AngelHack which defined the
48. 1:27modern day hackathon. At peak, we had
49. 1:29200,000 developers, 50 cities around the
50. 1:31world.
51. 1:32Uh it was largely responsible for the
52. 1:34web two acceleration in San Francisco.
53. 1:37Oh, wow.
54. 1:37>> It's funny because
55. 1:39we founded that company peak in 2012.
56. 1:42And what's happening now in San
57. 1:44Francisco reminds me of those golden
58. 1:46days of web two era.
59. 1:49But just
60. 1:5110x more intensity with AI. Mhm. So, uh
61. 1:55I would say like this is my third I
62. 1:57would say like second big San Francisco
63. 1:59boom
64. 2:01I I was part of.
65. 2:02Um
66. 2:04and the vision really you know been in
67. 2:07the cloud for a very long time in 2006 I
68. 2:10was
69. 2:11I designed Kaiser Permanente days.
70. 2:13Um
71. 2:14a cloud architecture before cloud was a
72. 2:16thing
73. 2:17that's when I discovered Amazon and I
74. 2:19fell in love with AWS um
75. 2:21being a big proponent being a big
76. 2:23advocate for uh
77. 2:25for for cloud only to realize all the
78. 2:27problems cloud has right so that led me
79. 2:30to embark on an open source initiative
80. 2:33to build a supercloud essentially
81. 2:36minimizing cloud providers um
82. 2:39to just to resource providers and sort
83. 2:41of bringing building a cloud of cloud
84. 2:43right so that eventually led to
85. 2:45uh what Akash is. Uh yeah.
86. 2:49So you see Akash is kind of like a a
87. 2:51decentralized
88. 2:53uh version of AWS or
89. 2:55like in in what way do you are they
90. 2:57comparable?
91. 2:59Akash is a first supercloud so
92. 3:02to go back a little bit history of
93. 3:04clouds right like
94. 3:05>> Mhm. uh you know when the internet began
95. 3:08every company was running their own data
96. 3:10center or you know colos or whatever
97. 3:14you own your own hardware.
98. 3:15And owning hardware is not the easiest
99. 3:17thing to do and got
100. 3:19you know
101. 3:20you can't just deploy hardware easily
102. 3:22takes a long time then cloud came and
103. 3:24gave us an amazing comfort solution that
104. 3:27you know where
105. 3:28you don't need to own the hardware but
106. 3:30just lease it so instead of owning a
107. 3:31house rent the house that's our model
108. 3:33phenomenal greatest scale they
109. 3:36really proved out that you can scale
110. 3:38very fast and web two is all about uh
111. 3:41just meeting demand in just in time so
112. 3:43scaling when you have scale kind of
113. 3:45thing it worked great for web two. Um
114. 3:48but the problem with with a cloud is
115. 3:50well um if you take the
116. 3:52the TCO the total
117. 3:54uh uh uh cost ownership cost model, it
118. 3:58actually is way more expensive than
119. 3:59running your own hardware.
120. 4:01Because they nickel and dime you for
121. 4:02every little thing. On top of that, the
122. 4:05world was going more towards an open
123. 4:06source
124. 4:07first model, whereas cloud was taking
125. 4:11open source software and closing them
126. 4:12down and reselling them all. There were
127. 4:13a lot of problems with cloud.
128. 4:15And we also noticed this is a massive
129. 4:19increase in in data consumption, right?
130. 4:21So, the rate at which it was increasing
131. 4:23was I think it was doubling or or
132. 4:26quadrupling every every year or
133. 4:27something.
134. 4:28And we we laid out a model and we
135. 4:31predicted that there's no way in hell in
136. 4:3310 years cloud be able to keep cloud
137. 4:35will be capable of actually handling the
138. 4:38demand. So, that then we were and
139. 4:43another big trend we noticed was there's
140. 4:45a lot of computer out there just
141. 4:46everywhere. It's not like in a central
142. 4:47place, right? So, that led us to on an
143. 4:51idea. It was not us, it was just whole
144. 4:53industry all together on an idea of a
145. 4:56super cloud in the sense. Can we
146. 4:57actually build a cloud provider that is
147. 4:59not a single provider, but it can
148. 5:01connect to all all cloud providers and
149. 5:03become like a a layer on top of it.
150. 5:05And that was proposed by Cornell in
151. 5:082015, I believe.
152. 5:11We we liked that idea.
153. 5:13And we implemented the first ever super
154. 5:16cloud. So, the Akash is not a cloud
155. 5:19provider, but a cloud of clouds, right?
156. 5:22So, Akash in Sanskrit means the sky. Sky
157. 5:25is where the the clouds live.
158. 5:27It's what we call a super cloud. That
159. 5:29means AWS, you know, could be a part of
160. 5:32Akash network as an offering.
161. 5:35What Akash gives you that the other
162. 5:37traditional cloud providers don't is
163. 5:39freedom. Freedom in terms of choosing
164. 5:41what provider you want
165. 5:43based on whatever metric you want to
166. 5:45you want to base on be it cost, be it
167. 5:47reliability, be it efficiency, be it
168. 5:48speed and whatnot, it gives you this
169. 5:50incredible choice of, you know, moving
170. 5:54between different cloud providers,
171. 5:56um and uh
172. 5:58or even distributing your workloads
173. 5:59across different providers. It could be
174. 6:01cloud provider, it may not be cloud
175. 6:02provider, could be compute. Now with
176. 6:04home node, it could be a home compute.
177. 6:06So, the idea is to bring all the compute
178. 6:08that's available and offer a single
179. 6:10place for you to for the user to to
180. 6:12shop.
181. 6:14Think of Akash like a mall, you know,
182. 6:15instead of going to a single store, you
183. 6:17go to a mall, you can get lot of
184. 6:18varieties.
185. 6:20Um that's sort of thing.
186. 6:21Interesting. That That makes a ton of
187. 6:23sense. That's That's a great great
188. 6:24explanation. Um yeah, I used to be a
189. 6:26software engineer. Um I actually went to
190. 6:28school for for comp sci and was a
191. 6:29software engineer after I graduated. So,
192. 6:31I know a little bit about about that. I
193. 6:33know So, if people were to go to
194. 6:36um I don't If you deploy application,
195. 6:38people would deploy their application
196. 6:40across different like AWS instances and
197. 6:42there's a UI for doing all of that. And
198. 6:44going back and forth, so people could do
199. 6:46that the same way just through Akash and
200. 6:49then you guys route them to wherever
201. 6:50else?
202. 6:52Yeah, so compute in Akash is not on AWS.
203. 6:54I mean, I said AWS could be. Uh and
204. 6:56there's an initiative to bring Actually,
205. 6:58there was an initiative uh to bring Aka-
206. 7:00AWS onto Akash. Okay. Uh and I think
207. 7:03it's going to continue, but right now
208. 7:05most We have about 70 odd providers and
209. 7:07most of them are independent data
210. 7:09centers or independent
211. 7:11uh companies uh that have this compute.
212. 7:14Um uh because, you know, you know, uh
213. 7:19I think the big advantage right now for
214. 7:21Akash is cost and access, right? So, if
215. 7:23you buy Amazon, you're not going to see
216. 7:24a cost difference. Although, there is a
217. 7:27uh initiative to bring Amazon reserve
218. 7:28compute. Reserve compute is
219. 7:31um
220. 7:32compute that's purchased by people ahead
221. 7:34of time for much cheaper than what you
222. 7:36would otherwise pay on demand. And most
223. 7:38of the time, this computer is just
224. 7:40sitting dormant. So, it's AWS compute,
225. 7:42but it's not being used. So, there's a
226. 7:44there's a case to be made to put that
227. 7:46computer on a cache and offer what exact
228. 7:49same Amazon computer for significantly
229. 7:50cheaper. So, arbitraging essentially.
230. 7:53Got it. Got it. Greg, hello. I'm I'm I'm
231. 7:56Banks. We're having a little bit of
232. 7:57technical difficulties. I'm the um
233. 7:59I'm the resident [ __ ] here at Market
234. 8:01Bubble. So,
235. 8:03I represent the average layman. A lot of
236. 8:05people in my chat, a lot of people that
237. 8:06follow the show um
238. 8:08aren't well equipped to have to follow
239. 8:10the type of conversation that you guys
240. 8:11are having. To those types of people,
241. 8:14what would you say and how would you
242. 8:15describe your product and what you're
243. 8:16building? And how does it relate to
244. 8:18them? How does it affect them? Why
245. 8:19should they care? Yeah.
246. 8:22The simplest analogy which I absolutely
247. 8:24hate, but I'm going to do it anyway, is
248. 8:26Airbnb for compute.
249. 8:28Okay, so you can rent out your compute.
250. 8:30Everybody has compute, you can rent it
251. 8:31out.
252. 8:32Yes, you can if you have dormant
253. 8:34compute, you can rent it out. Why do you
254. 8:36hate that analogy? I think that's really
255. 8:37really strong. I like that. I hate
256. 8:39analogies in general.
257. 8:42>> [laughter]
258. 8:43>> But, this particular one is very
259. 8:44simplified and
260. 8:45it's been abused by so many other people
261. 8:48and
262. 8:48you know whatever. You've talked about
263. 8:51the the demand side a lot on how like
264. 8:52the demand has gone gone up a lot and
265. 8:55it's going to continue to go up a lot.
266. 8:57Um I guess where is a lot of the the
267. 8:59revenues that you're currently seeing
268. 9:01with Akash? Like what are the the use
269. 9:02cases that are that people are using
270. 9:04their um is using for? Like I know AI
271. 9:07inference has been a big one that people
272. 9:09have been their spare um GPUs for. Um
273. 9:13how do you guys feel feel like you fit
274. 9:15in with like the open source AI
275. 9:16community and open router and those
276. 9:18platforms? Oh, yeah. The biggest uh
277. 9:21use we are seeing I mean Akash we we
278. 9:23Akash is a decent trust computer
279. 9:25network. So, we don't exactly know what
280. 9:27people are running on because we have no
281. 9:29access into their computer. But, we only
282. 9:31know what people tell us, right?
283. 9:34Uh because use is inference any any day.
284. 9:36I think like nine over 95% is inference,
285. 9:40right? Um especially now when you have
286. 9:42so many of these open source models and
287. 9:45um you know, there are uh
288. 9:48there's a lot of effort on like using
289. 9:50smaller models, using mixture of experts
290. 9:52and whatnot. Mhm. Uh so, there are a lot
291. 9:54of lot of small, large, all kinds of
292. 9:57inference uh
293. 9:59uh models that are doing inference on
294. 10:00Akash. Um uh from a customer's
295. 10:03standpoint, uh like uh historically,
296. 10:05Venice has been Venice actually started
297. 10:07on Akash. Oh, wow. Yeah. Uh um they uh
298. 10:12went to a different model, but now now I
299. 10:13think they're coming back. I think now
300. 10:14they're using some some Akash, but I
301. 10:16think they diversified their their
302. 10:18supply quite a lot. Um we have like
303. 10:20Venice like people, a lot of them, that
304. 10:23use Akash quite a lot today. Uh there
305. 10:25are a lot of the inference players. Um
306. 10:26>> Mhm.
307. 10:27lot of uh
308. 10:29uh AI products actually. There's there's
309. 10:31you know, one of my favorite use cases
310. 10:32which does uh imaging. So, the scope
311. 10:35quite a lot of
312. 10:36uh
313. 10:37uh
314. 10:38uh image generation, video generation
315. 10:40stuff happening on Akash. Yeah.
316. 10:42I would say 87%
317. 10:45of usage on Akash comes from
318. 10:49non
319. 10:50wallet users. Akash has two ways to pay.
320. 10:52You can pay using a credit card that
321. 10:54becomes that converts into crypto in the
322. 10:57background or you can pay directly with
323. 10:59crypto. Um we are seeing 87 most usage
324. 11:02coming from non-crypto uh on or on on uh
325. 11:06non-crypto payment uh and that's mostly
326. 11:08AI.
327. 11:09So, we're uh retooling a lot of our
328. 11:11tools to be a lot more mainstream more
329. 11:13than crypto
330. 11:15friendly. That's impressive. That's
331. 11:17that's cool. Um and uh so, if for people
332. 11:20who do have some spare GPUs, um if they
333. 11:23wanted to rent them out to you guys,
334. 11:26like what do the economics on that look
335. 11:28like? And I guess
336. 11:29what does it look like to do like uh can
337. 11:31I also use when can I use the GPU? Like
338. 11:33how long do I have to give it to you
339. 11:34guys for? Like what is the Yeah, uh
340. 11:37great depends on the GPU obviously,
341. 11:38right? So, I mean if you're if you're
342. 11:40having H100s, mhm,
343. 11:43um which are really in high demand right
344. 11:45now. Yeah.
345. 11:47If you bought H100s about a year ago, uh
346. 11:50putting that on Akash right now will
347. 11:52amortize you within 9 months.
348. 11:54Oh, wow.
349. 11:55>> [laughter]
350. 11:56>> What?
351. 11:56>> The economics are not. That's crazy.
352. 12:00It's not because I I saw
353. 12:01>> That's insane.
354. 12:02Uh that was not the case all the time.
355. 12:04For some reason H100s are out. No one
356. 12:06has H100s right now. They're super high
357. 12:08demand.
358. 12:09And
359. 12:11it is crazy how these economics are like
360. 12:13turning out. I don't even know what to
361. 12:14make of it because, you know, you have
362. 12:16your depreciation cycles, right? Your
363. 12:17standard, you know, whatever
364. 12:18depreciation cycles, yep, you know,
365. 12:205-year depreciation, you you go to near
366. 12:22zero after 5 years.
367. 12:24Uh
368. 12:26H100s are reversed.
369. 12:29The value is actually going up. They're
370. 12:30like Rolex watches right now. How?
371. 12:32>> There's no
372. 12:33It's crazy, I know. Uh
373. 12:35I remember buying H100s uh about a year
374. 12:37ago for 210 210 a 8X node. Now, the
375. 12:42minimum used node is going for 275.
376. 12:45What [laughter] the [ __ ]
377. 12:46>> And
378. 12:48I saw quotes for like 300k. Like it's
379. 12:50just nuts. How is that
380. 12:53Why?
381. 12:55It's I have no idea. Uh I really I mean
382. 12:58I have some some
383. 13:00uh assumptions, but I have no idea why
384. 13:03there's a sudden [laughter] drop in H100
385. 13:06supply that's in the last
386. 13:08I would say
387. 13:102 weeks, maybe 3 weeks. Uh-huh. So, Um
388. 13:15Yeah.
389. 13:15So, I know that you spoke on how the
390. 13:18depreciation cycles of GPUs, they're
391. 13:19always building the next newest what
392. 13:22next best one, next more efficient one.
393. 13:24Um and I know there's like a a a
394. 13:26lack of ability availability to do that
395. 13:29like it's very difficult in how many
396. 13:30they can make.
397. 13:32Um
398. 13:33I guess with all the investment that's
399. 13:35happening there, do you think that the
400. 13:37depreciation cycle is going to be like
401. 13:38less
402. 13:40aggressive or
403. 13:41No, let me put some numbers to you,
404. 13:42okay? So
405. 13:44No, there are two ways to use AI. There
406. 13:46is vanilla, which is using ChatGPT. You
407. 13:48can ask questions, you get answers.
408. 13:50Uh-huh.
409. 13:50>> The second way is agent tech, right?
410. 13:52>> Right. So agents running cloud code, uh
411. 13:54mass uh open cloud, those are agent
412. 13:57tech. Mhm. Um
413. 13:59agent tech usage is exponentially
414. 14:02greater in terms of tokens usage uh on
415. 14:05on on um on charge on on the models,
416. 14:08right? Like exponentially. Mhm.
417. 14:10>> Order of magnitude of magnitude of in
418. 14:12order of magnitude, right? Um
419. 14:14I think I burned through
420. 14:16a
421. 14:17my entire So Claude gives you limits as
422. 14:19to how much you can you know there's
423. 14:20daily and weekly limits. I think I've
424. 14:22burned my weekly limits in 2 days. Cuz
425. 14:24my wife is visiting uh you know her mom.
426. 14:26>> [laughter]
427. 14:28>> That's how it would be, bro.
428. 14:29>> Amazing. My wife and
429. 14:30I'm I'm goblin mode right now, so I
430. 14:33by myself in the garage at home and just
431. 14:35coding a lot. That's [ __ ] funny. And
432. 14:39agents are just significantly different,
433. 14:41right? You can do so many things you
434. 14:42can't really do with regular I can't
435. 14:44imagine doing things without agents now.
436. 14:46And it turns out I mean AI users
437. 14:49globally, there are about a billion
438. 14:50users using AI including ChatGPT,
439. 14:53including Gemini,
440. 14:54including Google AI itself, which most
441. 14:56people don't know they're actually
442. 14:57using.
443. 14:58Out of that, only 2.4 million people use
444. 15:01agents.
445. 15:03I don't know, do you guys use agents to
446. 15:04do anything?
447. 15:04>> course. Of course. Yeah, we uh
448. 15:07We use it to do everything. We do we we
449. 15:09have like this show prep. I'll I'll
450. 15:10share my screen with you. There you go.
451. 15:13Yeah.
452. 15:14That looks like Claude code. Yeah, we do
453. 15:15a whole
454. 15:16Yeah, we do a whole
455. 15:18run a show like it's We have We have
456. 15:20imaginations and dreams of eventually
457. 15:22building like an AI co-host that joins
458. 15:25us on the show and kind of real-time
459. 15:27sites Right. examples and content. We
460. 15:30can ask it in real time like, you know,
461. 15:32what markets are doing and
462. 15:34have it build a personality and Are you
463. 15:36familiar with Joe Rogan show, obviously?
464. 15:38Of course, yeah. Yeah.
465. 15:39>> He has like an off-screen like co-host
466. 15:41and his name is Jamie. You never see
467. 15:42him.
468. 15:43>> Sure. And occasionally you'll hear him
469. 15:44and he'll come chime in when it's
470. 15:45appropriate.
471. 15:46>> Jamie agent. Yeah, Jamie agent. Jamie
472. 15:48agent.
473. 15:49>> Perfect. Bubbles. We're going to make
474. 15:50our ours like a hot girl, I think. But
475. 15:53>> [laughter]
476. 15:53>> Um eventually we want to ultimately
477. 15:55build it to the point where the entire
478. 15:57show is produced by these agents.
479. 16:00So you can understand the benefits of
480. 16:01agents. It's incredible, right? It's my
481. 16:03thing. It's unreal. 2.4 million people
482. 16:06are using agents and that is causing the
483. 16:08crunch. I had to pay a four times more
484. 16:10for my memory. The same memory I paid
485. 16:12like about last year.
486. 16:13>> No, I'm getting [ __ ] whacked, too.
487. 16:15Claude flipped it and I don't know. It
488. 16:17was
489. 16:18It was super affordable at first and now
490. 16:19I'm spending [ __ ] three, four
491. 16:20thousand dollars on on AI a month. But
492. 16:23it's well worth it for for me and just
493. 16:25just even the sake of tuition, learning
494. 16:27it. Exactly. You mentioned all the
495. 16:30options. Obviously, there's a
496. 16:31endless amount of options and what it
497. 16:33feels like for the average consumer and
498. 16:35user of AI, user of agents is
499. 16:37there's no really discernible
500. 16:38difference. Like for the average person
501. 16:40like
502. 16:41Z was talking about how his mom uses AI
503. 16:43and like the average Joe who's looking
504. 16:45up [ __ ] recipes for banana bread and
505. 16:46how to get to his favorite coffee shop
506. 16:48and what the UV is outside and what the
507. 16:50price of Solana is. You're not going to
508. 16:51notice a real difference, but obviously
509. 16:53for somebody like you who's building in
510. 16:54it and as you said, I can't even imagine
511. 16:56not using it. Which one do you prefer?
512. 16:58Which which is your favorite?
513. 17:00Uh depends on what I'm doing. Uh really
514. 17:02depends. If I'm thinking
515. 17:03>> Gun to your gun to your [ __ ] head,
516. 17:06Anthropic, OpenAI. Which one? Uh
517. 17:09Anthropic Claude, 4.7. I mean, Anthropic
518. 17:11OpenAI. With with the quickness, by the
519. 17:13way.
520. 17:14>> Yeah, that was fast. It's
521. 17:16It's Anthropic on this side, baby. You
522. 17:17already [laughter] know.
523. 17:19My Codex is amazing. Codex is really
524. 17:21good too for coding. So, really depends
525. 17:23what I'm doing, right? There's no one
526. 17:24model that fits all, right? I'm not
527. 17:26going to use
528. 17:27Like, for example, for thinking, when
529. 17:29I'm like brainstorming and coming up
530. 17:30with ideas, I use, you know, um
531. 17:33Opus 4.7. When I'm actually coding, I
532. 17:36use a different model because writing
533. 17:38Once you plan out and actually executing
534. 17:40the plan to to code, you don't need
535. 17:42Opus. Opus is like extremely intelligent
536. 17:46overkill
537. 17:47for just writing coding tasks, right? I
538. 17:49use a cheaper model. I actually use
539. 17:50Akash ML for like actual coding, Oh,
540. 17:53cool. um our own like, you know, service
541. 17:56because it's significantly cheaper than
542. 17:57using Opus, right? So,
543. 18:00but overall, I think general-purpose
544. 18:01Opus 4.7 would be my Is it Hermes or
545. 18:04Hermes? Do you use that over Open Claw?
546. 18:06Like, what's the deal?
547. 18:07>> Hermes all the way, baby. Well, I mean,
548. 18:09I love Nous. I mean, I mean, disclosure,
549. 18:11I'm an early investor in in Nous as
550. 18:13well.
551. 18:14And Nous is a big user of Akash, at
552. 18:16least in the early days. Nous Research
553. 18:17built Hermes.
554. 18:19Um so, a lot of Hermes training was done
555. 18:21on Akash, too, early days. Um so, I love
556. 18:24Hermes. It's just night and day for me.
557. 18:27And they ship really well. The team is
558. 18:28incredible. Um you know, yeah. So, It's
559. 18:32interesting. Obviously, you're a power
560. 18:33user of AI again. Um I have a more
561. 18:35normie friend. Some of the audience may
562. 18:37or may not know him. It's It's honestly
563. 18:39um
564. 18:39It's irrelevant to the topic, but he
565. 18:42exists in like traditional like music.
566. 18:44He's a music exec, and he manages
567. 18:48artists, and that's the field that he's
568. 18:49in. And I originally put him on to Open
569. 18:52Claw. I'm like, "Yo, you can automate so
570. 18:53much shit." And just even just like
571. 18:56the way [clears throat] that I use it in
572. 18:57meetings and and and just sharing
573. 18:59transcripts and documents and emails and
574. 19:01all that [ __ ] Like, the bare-bones
575. 19:02stuff. Put him on early earlier, and we
576. 19:05went for a walk 2 days ago and kind of
577. 19:07caught up on things, and he was pilling
578. 19:08me hard on um
579. 19:10Hermes? Yeah, Hermes. Hermes. Yeah,
580. 19:13yeah, yeah. Hermes got great memory. Uh
581. 19:14I also use Hermes with my Obsidian
582. 19:17walls. Obsidian's [ __ ] Obsidian's
583. 19:19goated.
584. 19:20It is. It is. I mean, I'm like random I
585. 19:22take my
586. 19:23my phone to my shower now because random
587. 19:25thoughts and just like talking
588. 19:27[laughter] to
589. 19:28That's sick, bro. You're locked in, bro.
590. 19:30You're locked in. I'm just like Yeah, we
591. 19:32know. We we know what kind of random
592. 19:33thoughts are going through your head in
593. 19:34the shower now.
594. 19:35>> [laughter]
595. 19:36>> Hey hey Grok, suck me, baby.
596. 19:40Cuz showers are my best time, you know,
597. 19:41the best time to think, right? But you
598. 19:42want to take it down. I'm working on
599. 19:44like five, six projects at the same
600. 19:45time, right? Like I'm always like I'm
601. 19:47always talking, always talking.
602. 19:49I'm a talking so much now
603. 19:51like I don't type anymore, right?
604. 19:52Because, you know, typing is like Yeah,
605. 19:55slow. four four K minute at this point.
606. 19:57Frank T. God hit me with a [ __ ] bar
607. 19:59when he was filling me on and onboarding
608. 20:01me into like Open Claw and all this [ __ ]
609. 20:03and he said um
610. 20:05English is the new coding language. I
611. 20:07maybe I botched that, but it's like I
612. 20:10don't know, it's a bar.
613. 20:11>> Yeah, it's it's it's
614. 20:12>> Yeah.
615. 20:13I have
616. 20:15I have a hierarchy system for my agents
617. 20:17like
618. 20:18you know, because it can get extremely
619. 20:20chaotic. Imagine all this system, so I
620. 20:22develop my own system. I think
621. 20:24everybody's developing their own system.
622. 20:25I think there's a new like abstraction
623. 20:27stack that's being built on top of
624. 20:28agents that is kind of forming. We're
625. 20:30not seeing that yet as products, but
626. 20:33but I can totally see where that's
627. 20:34going. Um
628. 20:36uh it's incredible. So anyway, coming
629. 20:37back to my 2.4 million. Only 2.4 million
630. 20:40have experienced agents. Imagine if that
631. 20:43number is 10x. What's that going to
632. 20:44matter? You're going to see. Yeah. 2.4
633. 20:46million is leading to the current crisis
634. 20:48where we have no H100s available at all.
635. 20:51Right. Uh what is 24 million going to
636. 20:53look like? I don't know. Like I think
637. 20:55we're in uncharted territory. Um
638. 20:59uh you know, if you look at the supply
639. 20:59chain, you have obviously have the
640. 21:01compute, you have inference, you have
641. 21:02the chips, and then you have the energy,
642. 21:04which is where I'm focused quite a lot.
643. 21:05I testified before Congress on energy
644. 21:08about a year ago. I called called it
645. 21:09out. I said energy is going to be a big
646. 21:10challenge and you know, turns out energy
647. 21:13is a big challenge. People like Elon or
648. 21:15not their first principle thinking,
649. 21:17they're not they're going to space to
650. 21:19solve the energy crisis. I think there's
651. 21:21a lot more we can do on Earth too to
652. 21:23solve the crisis. I've made some bets
653. 21:26early bets on distributed training,
654. 21:28distributed inference systems that can
655. 21:30leverage a distributed grid to show
656. 21:33energy
657. 21:34home node for example we we launched
658. 21:36with intention of actually tapping into
659. 21:38a distributed uh
660. 21:40a grid, right? So it's just going to get
661. 21:43worse. I mean now even if you look at
662. 21:44the energy, right? Like most energy most
663. 21:47new data centers are using LNG which is
664. 21:50liquid natural gas because you can't
665. 21:52really have any other way to power these
666. 21:54systems.
667. 21:55So that led to another crisis with
668. 21:58turbines and transformers. Now we have a
669. 22:00four-year lead time to get a transformer
670. 22:02and a 14-year lead time to get a
671. 22:04turbine. To get a turbine
672. 22:07It's nuts. I can't like any if you look
673. 22:09at the entire supply chain
674. 22:11every
675. 22:12part of the supply chain is constrained
676. 22:13right now.
677. 22:15I've never seen anything like this
678. 22:16before. Yeah, no you talked a lot we've
679. 22:18we've been look talking a lot about
680. 22:20energy and how US is so like slow on
681. 22:22building it. Do you think that China I
682. 22:25don't know if you have a good answer to
683. 22:26this. Do you think that China is in like
684. 22:27a better place energy-wise as far as
685. 22:29building out their capacity? And
686. 22:31I know Jensen I don't know if you if you
687. 22:33watched the Jensen Huang
688. 22:34interview did you watch saw that?
689. 22:36Yeah, well he he was talking about how
690. 22:38Nvidia they should be selling to China.
691. 22:41And it seems like there might be some
692. 22:43kind of way for I guess it seems like we
693. 22:46have the chips and they have like the
694. 22:48energy capacity. Like do you do you
695. 22:50think there's anything that could happen
696. 22:51there or what do you think about Well, I
697. 22:53mean
698. 22:54theoretically yeah, but I think it's
699. 22:56it's mostly geopolitical and
700. 22:59Trump did approve us selling H200s and
701. 23:0359 days, but China didn't want them
702. 23:05because they want their own stack now.
703. 23:06Oh, wow. It was actually under the Chips
704. 23:09Act and the sorry, the under Biden
705. 23:10administration we started export
706. 23:12controls on chips. Now that led to a
707. 23:14local
708. 23:15chip industry by Huawei. Huawei I think
709. 23:17is doing 7 nanometer nanometer chips and
710. 23:20they're going to catch catch up to 3
711. 23:21nanometer very soon. Yeah. But China is
712. 23:24forcing
713. 23:25China is like aggressively building
714. 23:27local stack. And that's that's a risk
715. 23:30and that's a challenge because now
716. 23:31you're going to have a China stack and
717. 23:33you're going to have a western stack.
718. 23:34China makes four times more energy than
719. 23:37than than US. They have eight terawatt.
720. 23:40>> Eight terawatt installed capacity. US
721. 23:41has about two terawatt installed
722. 23:42capacity. Four times. Damn. China
723. 23:45produces China puts
724. 23:47adds about 36
725. 23:50sorry, adds about a gigawatt of solar
726. 23:53capacity every 36 hours.
727. 23:55A gigawatt. Gigawatt is [laughter] what
728. 23:57a
729. 23:58nuclear uh power plant produces.
730. 24:01And you know, so a gigawatt can house
731. 24:04Every 36 hours?
732. 24:05Every 36 hours. Oh, [laughter] [ __ ]
733. 24:07They have the and it's they have the
734. 24:09energy and we're talking about energy
735. 24:10infrastructure they have the battery
736. 24:11infrastructure, they have the rare
737. 24:12earth. They have Yeah. the entire supply
738. 24:15chain to build batteries, to build solar
739. 24:17panels and we don't have We lost control
740. 24:20of the supply chain a long time ago. So
741. 24:22I don't know if that's ever going to
742. 24:23come back, right? Like
743. 24:24>> Yeah. We can't produce batteries in
744. 24:26America anymore and you need batteries
745. 24:28for for solar.
746. 24:29I think we're starting to onshore
747. 24:31the
748. 24:33the panels now, but you need rare earth
749. 24:35minerals. Rare earth minerals are
750. 24:37controlled by China, so there's there's
751. 24:38another you know, bottleneck there. I
752. 24:40don't know, it's it's kind of a crazy
753. 24:42place right now. Uh
754. 24:43There's a massive re-indust-
755. 24:45industrialization effort by the current
756. 24:47current administration to bring back the
757. 24:48rare earth minerals. The reason why we
758. 24:50want to
759. 24:51uh what do you call
760. 24:53not invade, but I guess like well, let's
761. 24:56call it invasion invade Greenland is for
762. 24:58rare minerals. Right, yeah. There's a
763. 25:00new deal with the Ukraine for rare
764. 25:01minerals.
765. 25:03All this is classified information. We
766. 25:05don't exactly know how much quantities
767. 25:07these countries have but but yes,
768. 25:09there's a big push. You'll see any
769. 25:11geopolitical issue and you it really
770. 25:13comes down to rare minerals
771. 25:15for for America. Yeah, we we talked
772. 25:17about a little bit MP like MP materials.
773. 25:19I don't know if you know that company.
774. 25:21Like what MP they they're one of the
775. 25:23companies in the US that is trying to
776. 25:25fix that problem like
777. 25:27the rare earth metals that we do have
778. 25:29here. It's the I think that it's the
779. 25:30Mountain Pass mine in California but
780. 25:32Trump invested in them because of that
781. 25:34that situation that you're talking
782. 25:35about.
783. 25:36But it it seems like the biggest risk
784. 25:38for like tech and the US stock market is
785. 25:42that dynamic that we have. China has an
786. 25:44advantage in energy
787. 25:46and an advantage also. Don't you think
788. 25:48they're ahead on like the open source
789. 25:49side or do you not agree with that?
790. 25:51>> They are. They are leading
791. 25:52I mean all the all the open source
792. 25:53models are from China. Yeah, so leading
793. 25:56in so many different things.
794. 25:57Where is the US leading then?
795. 25:59>> Well, our our I think our tech companies
796. 26:01and the close source models are much
797. 26:03better.
798. 26:04Well, I'll have to ask you this. What do
799. 26:05you think that how do you think the gap
800. 26:07exists right now between like the close
801. 26:09source models the US tech companies like
802. 26:12Anthropic Open AI and then the open
803. 26:14source models which is I think it's
804. 26:16Alibaba who's releasing Qwen.
805. 26:19Qwen there is Minimax 2.6
806. 26:22there's there's a bunch of these. Yeah.
807. 26:24Yeah. So
808. 26:26the Qwen 2.6 apologies
809. 26:29so
810. 26:30how far behind? They're not very far
811. 26:32behind. Like Qwen 2.6 is better it's an
812. 26:35open source model
813. 26:36from China is better at coding than Opus
814. 26:41uh
815. 26:424.6.
816. 26:44Wow.
817. 26:44>> So the previous generation
818. 26:46of Opus. 4.7 is still number one for
819. 26:49agentic use cases. I think it can run
820. 26:53I think 3,000
821. 26:55parallels sessions I can't remember I
822. 26:57can't remember the stats but anyway so
823. 26:59but the point is for me
824. 27:01software coding model which is what most
825. 27:04agent take agents want they they just
826. 27:08one maybe one more one version behind
827. 27:11because like you can have like Entropic
828. 27:14can come the close models can
829. 27:16launch and the open models are just
830. 27:18going to steal from the close models.
831. 27:20Right.
832. 27:20>> There's nothing you can stop
833. 27:21distillation, right?
834. 27:23>> once they're out it's easier to make the
835. 27:24the open source version of it, right?
836. 27:26>> Yeah, once you have a really good close
837. 27:28model you distill from that close model
838. 27:30to an open one.
839. 27:32Yeah. So
840. 27:34open source is going to win, right? Like
841. 27:36there's no way in hell
842. 27:38you I mean there's like lots of
843. 27:39conversations about banning open source
844. 27:41now like during
845. 27:44Biden administration that was a thing
846. 27:46then Trump said he's not going to do it.
847. 27:48Now there's another conversation in the
848. 27:49White House to to regulate open source
849. 27:52AI quite a lot because of this problem.
850. 27:55I want to go back to that I want to go
851. 27:56back to that 2.4 million figure cuz it
852. 27:59seems obviously quite low for what this
853. 28:01tech is offering in terms of just like
854. 28:04helping people be productive and
855. 28:06optimizing their lives and and the way
856. 28:08that they do things day-to-day. There's
857. 28:10a couple interesting like angles and
858. 28:12pieces to this for me and kind of in
859. 28:14line with what you were just saying. The
860. 28:16first one being
861. 28:17why is that number so low and like
862. 28:20I don't know I guess I guess Open Claw
863. 28:22and Hermes feel like how I would imagine
864. 28:25computers and and internet felt 70s 80s.
865. 28:28Uh-huh. Flow the barrier to entry is
866. 28:29quite thick.
867. 28:31Um
868. 28:32I don't know what's your what's your
869. 28:33opinion on that? Who do you think has
870. 28:35the best chance to win in that regard in
871. 28:36terms of like mass distribution mass
872. 28:39adoption really onboarding the average
873. 28:41person to kind of you know whatever the
874. 28:43iPhone equivalent is what Apple did for
875. 28:45compute computers traditional computers,
876. 28:48who is, you know, who is
877. 28:50kind of in the leading position to do
878. 28:52that, put AI truly in the hands of
879. 28:54everybody. Um, that's the first piece. I
880. 28:56want to talk about like the
881. 28:58general general sentiment around AI
882. 28:59after that and how negative people,
883. 29:01specifically Americans, are around that.
884. 29:03How do How do you think that'll affect
885. 29:04legislation moving forward and how
886. 29:06that's a really [ __ ] bad thing,
887. 29:07obviously, for
888. 29:08the US and and this race to AGI.
889. 29:12Mhm. All that [ __ ] There's just a lot
890. 29:14to unpack there. Yeah.
891. 29:15Yeah.
892. 29:16It Now,
893. 29:18this is a
894. 29:20very fast-moving uh this space moves
895. 29:23very fast, right? Who's in the best
896. 29:24position to succeed? I'm going to take a
897. 29:27contrarian take here. I think it's
898. 29:29Apple. I think it's Apple, too. I said
899. 29:31that Yes! And you're clearly a [ __ ]
900. 29:33genius. Did you guys hear that?
901. 29:35>> [laughter]
902. 29:35>> Look at that one. I think it's Apple,
903. 29:36also. Why? Cuz they're already
904. 29:38>> Why? Because
905. 29:39their
906. 29:41UI is exceptional. They're They know how
907. 29:44to
908. 29:45>> right? They're already in everyone's
909. 29:46phone. Yes, dude. Yes. But along with
910. 29:48the distribution
911. 29:49>> too.
912. 29:50Put it next to this one. I'm a [ __ ]
913. 29:51genius. No, I'm just kidding. I [ __ ]
914. 29:53>> No, I don't know if you use the Apple
915. 29:54intelligence. It's horrible
916. 29:56horrible models, but the experience
917. 29:57[laughter] is phenomenal, right? Like,
918. 30:00if you use Siri, terrible in terms of
919. 30:03quality, but the experience is great. Uh
920. 30:06the the intelligence where you can
921. 30:07proofread and whatnot, I use that quite
922. 30:08a lot on my phone. Um experience is
923. 30:12great, terrible model, right? But if you
924. 30:14can replace a model, Yeah. uh you're
925. 30:17going to get an excellent And there's
926. 30:18another Why thing? Yeah. I feel like if
927. 30:21Apple tomorrow released Siri AI and it
928. 30:24was even 20% as capable as Opus 4.7 or
929. 30:28Chat GPT, what's the what's the most
930. 30:30up-to-date, 5.5 or whatever?
931. 30:31>> Yeah. I think if it was even 20% as
932. 30:33capable, your grandma could do her
933. 30:34birthday card and it'll tell you what
934. 30:36the price of Bitcoin is and all that
935. 30:38good [ __ ] Um
936. 30:40and they just pinned it to the top of
937. 30:41your iMessage and you can talk to Siri
938. 30:43the way that you talk to your friends
939. 30:44and type with that way and you can give
940. 30:46it access to your phone and it can help
941. 30:48you organize that. It can reply to
942. 30:49people for you and it just became an
943. 30:51extension of iOS and iMessage. I think
944. 30:53they just won on the spot though, like
945. 30:56at least US. And the ecosystem power
946. 30:59they have, they have every Apple TV.
947. 31:01Literally.
948. 31:02>> All the You wake up your phone updates
949. 31:04one your phone updates one day and their
950. 31:06chat GPT or their Claude is just pinned
951. 31:08right there in your messages. Yeah,
952. 31:09yeah. I mean, it's already here. It's
953. 31:11just like terrible more I don't know
954. 31:12what they did whatever. I think they
955. 31:13fired the guy who did it now they have a
956. 31:15new CEO and all that, right? So, but
957. 31:18they have the best chance to succeed and
958. 31:20they're also pretty good. They've been
959. 31:21doing AI for a lot. The MLX thing they
960. 31:24have the the machine learning you know,
961. 31:26libraries they have in Apple's really
962. 31:27good. The Mac Studios the the M5s come
963. 31:32with 141
964. 31:35unified memory gigabytes of unified
965. 31:37memory which is actually really powerful
966. 31:39to run local models. The Mac Mini the
967. 31:41Mac Mini meme meme. That's where I have
968. 31:43all my [ __ ]
969. 31:43>> the mini the the studio
970. 31:45the the bigger machine
971. 31:47the M 12. Yeah. Uh
972. 31:49>> Well, that's some [ __ ] that you're on.
973. 31:50You're [ __ ] taking over the world
974. 31:51with AI. I'm talking to it about [ __ ]
975. 31:54So, so so minis are small, right? The
976. 31:56the studio is the bigger one and they're
977. 31:59very powerful. You can run local models
978. 32:01really good. Imagine
979. 32:03clustering the studios. A lot of people
980. 32:04are buying studios because I have a
981. 32:06studio, too. Because in my home when I'm
982. 32:09using AI try to use local AI more than
983. 32:12cloud AI because of privacy. I do a lot
984. 32:14with with my AI. I mean, my AI is
985. 32:16connected to my camera feeds, my
986. 32:18microphone feeds. It has access to my
987. 32:20entire life. And I don't don't want that
988. 32:23on the cloud. No no way in hell, right?
989. 32:24So, I use a lot of local models and
990. 32:27imagine all this local everybody having
991. 32:28Mac minis, Mac Studios can connect those
992. 32:31local local models or connect those
993. 32:33those computers using a cache or
994. 32:35something where you can like, you know,
995. 32:37sell. Imagine the the power Apple has to
996. 32:40create an incredible network of these of
997. 32:43these like local machines and use that
998. 32:45for AI. I mean, there's a lot of
999. 32:47opportunity Apple has.
1000. 32:48If they don't [ __ ] it up, I think
1001. 32:49they're going to be massively
1002. 32:50successful.
1003. 32:51You You said you think that open source
1004. 32:53is going to win. I think that's a a
1005. 32:54really
1006. 32:55That's an interesting take. If If you do
1007. 32:57think that open source is going to win
1008. 33:00the race in AI, um I guess what areas do
1009. 33:05you think are currently underexplored
1010. 33:07that are going to become more popular in
1011. 33:09the future? I know you talked a little
1012. 33:10bit about distributed training and
1013. 33:11distributed inference.
1014. 33:13Um and then also, how do you um
1015. 33:16how do you feel that like crypto is
1016. 33:18going to fit into that realm of like
1017. 33:20open source AI? Cuz it seems like crypto
1018. 33:21should benefit from open source AI a
1019. 33:23lot.
1020. 33:24>> yeah, I have a thesis on that. So, the
1021. 33:26biggest unex I mean, under invested and
1022. 33:28underexplored area is distributed
1023. 33:30training. I think it's such a big such a
1024. 33:32big opportunity and there's a lot of
1025. 33:33companies that I'm really excited about.
1026. 33:35Pluralis is one. Uh they haven't
1027. 33:36launched yet, but I'm really excited
1028. 33:37about Pluralis. Yep. Uh Zeus was working
1029. 33:40on something. I don't I don't know how
1030. 33:41far uh I haven't caught up with them on
1031. 33:43the
1032. 33:43on the distributed training piece. And
1033. 33:45then Gensyn, which is a another I think
1034. 33:47it's a crypto It's crypto token device.
1035. 33:49They're very very good. They're They've
1036. 33:50been working on distributed training as
1037. 33:52well. No one has quite figured out the
1038. 33:54incentive structure, which I think is
1039. 33:55going to be uh
1040. 33:57a big unlock once we figure out the
1041. 33:59technology. Technology is very hard.
1042. 34:01Um like Pluralis uh is able to achieve
1043. 34:05something called heterogeneality in
1044. 34:07training. So, what that means is
1045. 34:09uh I don't want to go too technical, but
1046. 34:11right now training uh the big limitation
1047. 34:13for training is you got to have the same
1048. 34:14chips, no matter what. So, if you're
1049. 34:16training 8100s, you got to have 8100s.
1050. 34:18Yeah.
1051. 34:19So, you can't mix and match them, and
1052. 34:20that's a big big challenge for a lot of
1053. 34:22companies because
1054. 34:23um
1055. 34:25uh
1056. 34:27uh
1057. 34:28Sorry.
1058. 34:29Um so, a big challenge for a lot of
1059. 34:31companies because, you know, you're
1060. 34:32going to have different chips, but if
1061. 34:33you can do heterogeneality, you can
1062. 34:35unlock
1063. 34:36uh distributed training from homes where
1064. 34:38you have different types of chips,
1065. 34:39right? Like I might have 4090, someone
1066. 34:41have 5090, someone have H100. You can
1067. 34:44combine all that to actually train. The
1068. 34:46efficiency is only efficiency difference
1069. 34:48is only 1.6x
1070. 34:51uh compared to a centralized model,
1071. 34:52which is not a bad thing, which is not
1072. 34:55as great as a centralized train model,
1073. 34:56but it's really good good
1074. 34:59for as a model. So, I think that's one
1075. 35:01area. Um uh and second area, I mean,
1076. 35:04it's it's being talked about from like
1077. 35:06big labs like Anthropic talks about it
1078. 35:09quite a lot. The you know, Anthropic
1079. 35:10co-founder, I I I've talked to the head
1080. 35:12of research for Anthropic who's been,
1081. 35:14you know, at the company since the
1082. 35:15beginning.
1083. 35:16Um they are also looking very deeply
1084. 35:19into distributed training. They don't
1085. 35:20talk about publicly, but I know talking
1086. 35:22to labs and talk talking to people on
1087. 35:24the frontier, uh they're developing
1088. 35:25these frontier models, distributed
1089. 35:27training is going to be a big deal. And
1090. 35:29they're already seeing challenges to
1091. 35:30compute, right? They have to. Um I think
1092. 35:33second area that is uh
1093. 35:36underexplored is provenance. Like as you
1094. 35:39get a lot of
1095. 35:43um
1096. 35:44fake
1097. 35:46AI generated content AI slop, how do you
1098. 35:49the the there's enormous need for
1099. 35:51provenance like to to verify that uh you
1100. 35:54know
1101. 35:55the goods that are created are legit,
1102. 35:57not created by AI, right? Because the
1103. 35:59value will be players, I believe uh you
1104. 36:01know, just like, you know, if you
1105. 36:03compare like if you have you two
1106. 36:04paintings, they look exactly the same,
1107. 36:06one is AI generated, one is human
1108. 36:07generated, if you try to sell them,
1109. 36:10which do you think is going to sell?
1110. 36:11Human generated, right? Obviously.
1111. 36:13There's value in human generated stuff.
1112. 36:15Uh it's not the quality in the outcome
1113. 36:17of the product, but it's the effort that
1114. 36:19goes into producing something that
1115. 36:20people value. Like I value things that
1116. 36:22are handmade versus things that are
1117. 36:24machine made, right? Because there's
1118. 36:25effort and there's uh talent that goes
1119. 36:27into it. I think that's another area
1120. 36:29providence area is under explored. Um
1121. 36:32and uh
1122. 36:33and privacy is under explored, too. You
1123. 36:36know, in in we're looking at TEEs and
1124. 36:38all that stuff, but
1125. 36:40the a lot of
1126. 36:41like challenges with trusted execution
1127. 36:43environments, right? Like side channel
1128. 36:45attacks and whatnot. I'd love to see ZK.
1129. 36:47I'd love to see
1130. 36:49FHE the the you know, federated the
1131. 36:52uh
1132. 36:53homomorphic encryption like in in AI. I
1133. 36:55think that's severely under explored
1134. 36:57under explored. Um
1135. 37:00Yeah, I don't know. There are a lot of
1136. 37:01deep technical problems that that are
1137. 37:03that are that I can think of. But these
1138. 37:05three are on my top top of my head.
1139. 37:07Awesome. Uh that's interesting. That's
1140. 37:09that's really helpful context. That's
1141. 37:11cool. I've heard of Pluralis. Kel
1142. 37:12actually, I don't know if you know Kel
1143. 37:14on Twitter Kel XYZ, he talks
1144. 37:15>> Okay, yeah, yeah. Yeah, yeah, he's smart
1145. 37:17super smart dude. Pluralis is amazing.
1146. 37:19Really good team. They've been silent. I
1147. 37:21mean, they're like a research team,
1148. 37:22right? They're they've been silent, but
1149. 37:23I know they're coming up with an amazing
1150. 37:26thing in a few weeks. Mhm. Uh and
1151. 37:28they're going to be using a lot of
1152. 37:28compute from Akash, so I'm excited.
1153. 37:30Gotcha.
1154. 37:31>> For Pluralis. Um
1155. 37:32>> Yeah.
1156. 37:33Yes, I guess I'll ask questions
1157. 37:34specifically about Akash. I know you
1158. 37:36guys partner with Venice. Um Venice
1159. 37:38seems like they've gotten a lot of the
1160. 37:40attention over the past um few weeks
1161. 37:42over there
1162. 37:43in the chat. Um it seemed like they have
1163. 37:45a really cool integration of like their
1164. 37:47token and the platform. So, how do you
1165. 37:50think about the You know, you have also
1166. 37:52have a token with your platform. How do
1167. 37:53you think about the integration between
1168. 37:54the two and like the advantages of being
1169. 37:56a a crypto platform?
1170. 37:59Right now with AI, I wouldn't call it an
1171. 38:01advantage. Or disadvantage. I was going
1172. 38:03to say or disadvantage.
1173. 38:05Also,
1174. 38:05the token was serious.
1175. 38:09The good things and bad things. The good
1176. 38:10things is tokens give us a lot of
1177. 38:12leverage in terms of
1178. 38:14uh
1179. 38:15subsidies in terms of behavioral like
1180. 38:19engineering or whatever you want to call
1181. 38:21it. There are incredible ways you can
1182. 38:23use tokens to bootstrap networks and
1183. 38:24there are a lot of good things. But the
1184. 38:26bad thing is the user experience is
1185. 38:28terrible, right? I mean, you like
1186. 38:29initially Akash was just token. Like if
1187. 38:32you go to console, you have to have AKT
1188. 38:34in order to
1189. 38:35>> Yeah. uh use compute and that uh
1190. 38:38severely limit
1191. 38:39>> I mean, crypto and AI, it sounds like
1192. 38:42you're jumping you're jumping through
1193. 38:43[ __ ] 25 hoops to get there, right?
1194. 38:46Like for the It's so challenging. I
1195. 38:47mean, we would have like our conversion
1196. 38:49rates would be like less than like 1%
1197. 38:51like from from the people from people
1198. 38:53that want to use Akash to like people
1199. 38:54that actually end up using Akash. It was
1200. 38:56terrible. Uh and we had we lost so many
1201. 38:58big deals. Uh we lost a big deal with
1202. 39:00Nvidia because nobody wants to have
1203. 39:02tokens because they their finance teams
1204. 39:05Huh.
1205. 39:06Nvidia was using Akash in the through an
1206. 39:08acquisition was using Akash in the
1207. 39:09beginning. Um
1208. 39:11you know, and then uh we had their CF
1209. 39:14their finance teams just basically blew
1210. 39:16up saying that they don't want to use
1211. 39:18anything with crypto. Cuz they can't
1212. 39:20hold AKT
1213. 39:21on a balance sheet. Yeah.
1214. 39:23>> Just as far as like general sentiment to
1215. 39:25the average person, the average layman,
1216. 39:26you're fighting two battles, right?
1217. 39:27Obviously, the overwhelming kind of
1218. 39:29feedback from the average person anti-
1219. 39:31anti-AI. And I feel like while most
1220. 39:33people are kind of just unbothered,
1221. 39:35unconcerned, uninterested in crypto,
1222. 39:37crypto is kind of
1223. 39:39for the average person, at least where I
1224. 39:41come from, my corner of the internet,
1225. 39:42crypto equals scam. Like the two words
1226. 39:44are like synonymous.
1227. 39:45>> It is. And it's sadly, yeah. unfortunate
1228. 39:47cuz you're fighting cuz as you said, the
1229. 39:49tech stuff, you know, you see it.
1230. 39:51Obviously, there's a ton of [ __ ]
1231. 39:52benefit there, but
1232. 39:53Yeah. if nobody's coming through the
1233. 39:54front door, it's like [ __ ] Yeah. It's
1234. 39:57night and day. So, we have another
1235. 39:58product called Akash ML where we removed
1236. 39:59crypto completely. Still got the brand
1237. 40:01name and that is hitting all-time high
1238. 40:04usage. I mean, even today we hit
1239. 40:05all-time high on tokens there. So, when
1240. 40:07you remove crypto
1241. 40:09Akash ML is the inference uh as a
1242. 40:11service from Akash. Uh-huh. Uh it is for
1243. 40:13uh basically, you can get all the all
1244. 40:14the good models. Uh you can use Akash ML
1245. 40:17with your cloud code or your open code.
1246. 40:19You can switch like your back-end model
1247. 40:21to use Akash ML. I use Akash ML quite a
1248. 40:23lot for my coding where I use Opus 4.6
1249. 40:274.7 for thinking and writing a plan on
1250. 40:28what I want to do. Once I have a solid
1251. 40:30plan, I switch to Akash ML. I save a lot
1252. 40:32of money like a lot of tokens. So, it's
1253. 40:35a very very very good. Um
1254. 40:37Uh and so it's an inference as a service
1255. 40:40and we get a lot of usage there
1256. 40:41especially if you don't want to go
1257. 40:42through all the pain of setting up a
1258. 40:43model. Running a model is very
1259. 40:45expensive, right? And if you're running
1260. 40:46like a big model, it costs you like I
1261. 40:47don't know like thousands of dollars a
1262. 40:49month just to run it for yourself. But
1263. 40:52if you use the shared model like Akash
1264. 40:54ML where you you can use a model that's
1265. 40:56shared with other people, it's
1266. 40:57significantly cheaper. So, Akash ML is
1267. 41:01designed for AI devs and non-crypto
1268. 41:03folks, right?
1269. 41:04And that is our biggest growing product.
1270. 41:06Um And that directly translates to to to
1271. 41:11to you know, that hosts on Akash and it
1272. 41:13drives usage to Akash. From a token
1273. 41:15model standpoint, right?
1274. 41:17I think the advantage is
1275. 41:20the way Akash token is tied to the usage
1276. 41:22is burned. Like every time
1277. 41:24every time you know, you you use
1278. 41:26computer on Akash, a portion gets burned
1279. 41:29using a BME model. So, there's a there
1280. 41:31is a usage-based economic model and we
1281. 41:35have more more burns than than mints now
1282. 41:37because of usage that's going up and up,
1283. 41:39right? So, if this continues, I think
1284. 41:42it'll be incredible thing from economic
1285. 41:44standpoint. Um but it also gives us
1286. 41:47an ability to create incentives, right?
1287. 41:50Now now there's a big challenge for
1288. 41:52onboarding providers using token
1289. 41:54economics, we can actually create
1290. 41:56attractive incentives for providers to
1291. 41:58get to be competitive and whatnot. So, a
1292. 42:00lot of things you can do really well
1293. 42:01with tokens. And especially if you're
1294. 42:02doing distributed training, I'm very
1295. 42:04excited for token. Yeah.
1296. 42:06Uh you can see a world where let's say
1297. 42:08Pluralis, okay? Good example. Pluralis
1298. 42:10where
1299. 42:11if I contribute my computer on model,
1300. 42:14right? If there's a model somebody wants
1301. 42:15to try, say you you know Ansem wants to
1302. 42:17train a model. You don't have computer
1303. 42:19but you have an idea. You have the code,
1304. 42:20you have the data, but you need you know
1305. 42:22millions of dollars to compute, right?
1306. 42:23You can go to Pluralis and be like, "Hey
1307. 42:25look, I'm willing to pay I need a
1308. 42:27million dollars worth of compute. I
1309. 42:29don't have a computer right now, but I'm
1310. 42:30willing to share my future profits with
1311. 42:32with the people that provide compute."
1312. 42:34Oh. So, Pluralis running nodes or
1313. 42:36whatnot, I'm a node runner. I'm you know
1314. 42:38I have nodes. I have 590s at home. I can
1315. 42:41then participate in this model training,
1316. 42:43get some token that represents my
1317. 42:45ownership or my contribution, and when
1318. 42:47the model goes to inference where you're
1319. 42:48making money, hosting on Akash Amel or
1320. 42:50whatever,
1321. 42:51you can get a portion of that
1322. 42:54portion of that revenues directly for
1323. 42:57participation.
1324. 42:57>> Wow. Now that I see as a It's really
1325. 42:59cool. extremely disruptive model. No one
1326. 43:01has done this so far and think the
1327. 43:03people are starting to look into it. We
1328. 43:06call this op closed weights open source
1329. 43:08model where the model itself is open
1330. 43:10source, but the weights are closed only
1331. 43:12accessible for
1332. 43:13for people that are doing inference.
1333. 43:14There's a lot of opportunity for this
1334. 43:16incredible tokenism token mechanism
1335. 43:18design that are going to that are going
1336. 43:20to come with like
1337. 43:22with with distributed training. I'm
1338. 43:24really excited about and that's going to
1339. 43:26be very disruptive. In second uh
1340. 43:29VVV has done a phenomenal job Venice
1341. 43:31with with their DM model, right? They
1342. 43:33they really understood that you can give
1343. 43:35out free like free inference
1344. 43:38and you can get really good inference
1345. 43:39when you can actually use Opus directly
1346. 43:41using Venice. Not too many people know
1347. 43:43this, but using your DM you can actually
1348. 43:46get free
1349. 43:47free
1350. 43:48um free compute free inference that you
1351. 43:51are otherwise have to pay money for for
1352. 43:52a cloud. So, I sometimes use Venice, you
1353. 43:55know, when when I want to when I want to
1354. 43:56use
1355. 43:58when I have I have some VVV that got air
1356. 44:00dropped to me super early that is worth
1357. 44:03a lot of money [laughter] now. Yeah.
1358. 44:05So, and that I use that as staking and I
1359. 44:07get my DMs and I do all kinds of cool
1360. 44:09things. Very very good product if you
1361. 44:11haven't used it. I was a big user.
1362. 44:14I use Venice
1363. 44:15quite a lot. Um
1364. 44:17Uh I use I mean I used to talk about
1365. 44:19Venice about few years ago quite a lot
1366. 44:21before before uh all this hype. Uh for
1367. 44:24anything medical related I used Venice
1368. 44:26quite a lot, you know. Uh
1369. 44:27>> [snorts]
1370. 44:28>> anything that I want privacy I use
1371. 44:29Venice. I don't really trust
1372. 44:31uh Claude or Oh, by the way, if you
1373. 44:33Claude will use your data to train. So,
1374. 44:35you just be very careful what you give
1375. 44:36it, right? So, don't give anything you
1376. 44:37want to you want to be become a a part
1377. 44:40of a training set, right? Um
1378. 44:43So,
1379. 44:44and they'll report things that they find
1380. 44:46suspicious. It doesn't matter. You can't
1381. 44:48It doesn't have to be suspicious for
1382. 44:49you, but if they think it's like I don't
1383. 44:51know. I mean, well, that's concerning.
1384. 44:53I've I've literally sent Claude pictures
1385. 44:54of my my penis.
1386. 44:56>> [laughter]
1387. 44:57>> Yeah, that's part of your training For
1388. 44:59medical medical reasons. I'm like it's
1389. 45:01like, you know, it's like my It's like
1390. 45:03my best friend. He's like my accountant,
1391. 45:04my lawyer, my doctor, all of it. Be very
1392. 45:07careful. It's going to snitch on you.
1393. 45:08Yeah, it's going to leak. Uh so, so
1394. 45:10whatever you do, Claude.
1395. 45:12But Venice is going to be great because
1396. 45:13Venice doesn't retain data. Privacy I
1397. 45:14think privacy is is a big uh
1398. 45:17big challenge with AI and love Venice.
1399. 45:19It's It's a great tool. So,
1400. 45:21that that It seems like to me like if
1401. 45:22you believe two things that open-source
1402. 45:25AI is going to dominate in the future,
1403. 45:26continue to be more popular, and also
1404. 45:28that AI demand and the demand for
1405. 45:30agentic
1406. 45:31um AI is going to continue to go up, it
1407. 45:34seems like there's a lane for really
1408. 45:36really cool token mechanism design,
1409. 45:38which is what you were talking about. Um
1410. 45:40and that like a lot of the issues that
1411. 45:42crypto projects have had in the past is
1412. 45:43there was like there wasn't a lot of
1413. 45:45revenues tied to these tokens. Like it
1414. 45:47was a future growth story that you were
1415. 45:48betting on something happening in the
1416. 45:50future. But now it's like the demand is
1417. 45:51real,
1418. 45:53um and it seems like you can tie tokens
1419. 45:54to that demand in a really
1420. 45:57um clear way. So, that's that's awesome
1421. 45:58to hear. I think that's that's really
1422. 45:59dope. Uh you said this token token
1423. 46:02economics models are going to be a
1424. 46:03winners, right? Because you can like
1425. 46:04helium direct from the job B and me
1426. 46:07with the cash we kind of adopt a similar
1427. 46:09models
1428. 46:10with with the demand that's going crazy
1429. 46:12now. I think time that back to the token
1430. 46:14economics is going to be beneficial.
1431. 46:16Yeah, I'm actually very excited. Now
1432. 46:18that things are slowing down a little
1433. 46:19bit. There's a lot of capitulation a lot
1434. 46:21of like market correction and protocols
1435. 46:24dying what not. I think the real
1436. 46:25economic no longer you're going to have
1437. 46:26just speculation just like hey I have a
1438. 46:28token for governance. I think that's not
1439. 46:30going to fly much. It's not it's not
1440. 46:32going to work anymore. You got to tie
1441. 46:33down. You got to have some metric that
1442. 46:35you're going to optimize for and and pay
1443. 46:36attention to.
1444. 46:38I'm I'm very excited for this new new
1445. 46:40wave of token economic design. I think
1446. 46:42it's all like
1447. 46:44We see this quite a lot, right? Like
1448. 46:45every time there's a boom and a bust,
1449. 46:47you know, you see the fundamentals go
1450. 46:48up. I only hope that this time
1451. 46:51it looks like market is recovering quite
1452. 46:52a bit now. At least the the fundamental
1453. 46:55projects are gaining some some you know,
1454. 46:59some attention. Yeah. I really hope that
1455. 47:01we don't diverge into pure gambling like
1456. 47:03we always do after this first and then
1457. 47:05meme coins. I have no idea what what
1458. 47:07this is going to be. I really really
1459. 47:09pray that the gambling won't return.
1460. 47:12>> Yeah.
1461. 47:13You know.
1462. 47:14Yeah, I also I also would prefer the
1463. 47:17real project. I mean, it's interesting
1464. 47:18for me cuz I as a trader I kind of
1465. 47:20identify with the areas I think are
1466. 47:22going to get a lot more attention.
1467. 47:24And at one point that was NFTs at one
1468. 47:26point that was meme coins. But like I
1469. 47:27think it's really really important is
1470. 47:29like why I really enjoy having you on
1471. 47:31and having other people on the pod is
1472. 47:32like to talk about the projects that are
1473. 47:34really doing doing well and have like
1474. 47:36real fundamentals behind them.
1475. 47:38So yeah, I do think the market's
1476. 47:39turning. I mean, stocks are [ __ ] at
1477. 47:41all-time highs, bro. Everything else is
1478. 47:42ripping. Crypto is down and everything
1479. 47:44else is ripping and we're just now
1480. 47:46starting to see the few really solid
1481. 47:48crypto protocols start to do well like
1482. 47:50hype all-time highs today. Zcash has
1483. 47:52been doing well. So I do I do think that
1484. 47:54there's a like there's a lot of projects
1485. 47:57that weren't solid and now we're seeing
1486. 47:58like the quality ones do well, which I
1487. 48:00think is going to continue for sure. So
1488. 48:02>> Yeah, and I think another point about
1489. 48:04how crypto and AI are going to win
1490. 48:07together. Like the crypto stuff is very
1491. 48:09hard to use, right? Like you know Akash
1492. 48:10stuff. If you remove the credit card
1493. 48:12stuff, it's very hard to use in general
1494. 48:13speaking. There's a lot of things that
1495. 48:14you don't get
1496. 48:16that you get with with traditional
1497. 48:18application you don't get with crypto.
1498. 48:20Agents don't have the problem.
1499. 48:21>> [clears throat]
1500. 48:21>> Agents can
1501. 48:22>> [laughter]
1502. 48:23>> Exactly. very easily. Yeah. I was
1503. 48:25surprised how well agents were using
1504. 48:27Akash. It's like remarkable different.
1505. 48:30Um and I use Akash with agents now, only
1506. 48:32with agents now. I don't I don't deal
1507. 48:33with the command line directly. But
1508. 48:35we're able to build so much now, so so
1509. 48:38fast. It's unbelievable speed. Before
1510. 48:42like I'm talking about 6 months ago Mhm.
1511. 48:44or even 3 months ago, our mode was
1512. 48:46focus, focus, focus. Yeah, even though
1513. 48:48we have a million ideas, really good
1514. 48:49ideas, right? Uh you have to stay
1515. 48:52focused on single doing one thing, one
1516. 48:54thing only. Yeah.
1517. 48:55>> But that has flipped now.
1518. 48:57Um because the cost the failure cost is
1519. 48:59low. Right? You can put you have an
1520. 49:01idea, you can prototype that quickly, go
1521. 49:03to market very quickly,
1522. 49:05fail or succeed, you you'll know very
1523. 49:07quickly. Yeah. Yeah, and the failure
1524. 49:10cost has gone down, but the opportunity
1525. 49:11cost has gone up. Cuz if you don't do
1526. 49:13it, someone else is going to do it.
1527. 49:15>> Somebody else is, yeah. Right? Because
1528. 49:16the AI cuz you can produce this thing so
1529. 49:19you know, the world has flipped now.
1530. 49:21What is that going to look like in terms
1531. 49:22of businesses and how they're exploring
1532. 49:24opportunities? For us, we want to
1533. 49:27uh
1534. 49:29if you have an idea, we want to go to
1535. 49:30market as quick as possible.
1536. 49:32Um you know, and then
1537. 49:34uh you know, whether we fail or succeed,
1538. 49:36I'll proceed our way instead of
1539. 49:38rejecting ideas. So we're now a lot more
1540. 49:41open-minded to a lot more and incredible
1541. 49:43things are happening across the team. Um
1542. 49:45all engineers now use our token maxing
1543. 49:48basically. Mhm. And we measure their uh
1544. 49:52uh their their performance based on how
1545. 49:54much tokens they use.
1546. 49:55>> Really?
1547. 49:56I mean, you see the outcomes, but
1548. 49:57everybody's producing outcome. I mean,
1549. 49:58there's a lot of outcomes that's that's
1550. 49:59very impressive. But also, like if
1551. 50:01you're not using tokens efficiently,
1552. 50:04uh you no longer have a place in the
1553. 50:05company. Like we are very very hardline
1554. 50:08on it. Um uh and if you ask me a
1555. 50:11question that you didn't ask Claude or
1556. 50:13AI before, that's another red flag on
1557. 50:15your uh
1558. 50:17on your thing. That's hardcore.
1559. 50:19Yeah, we I mean, because I don't know
1560. 50:21how much you use it. You can connect
1561. 50:22using MCPs, you can connect every every
1562. 50:24data source. Like we use linear for our
1563. 50:27management, we connect that, we use
1564. 50:28HubSpot for whatever CRM, we use
1565. 50:31uh Google Analytics, we use uh Postgres
1566. 50:33database, Metabase. We have so many
1567. 50:35tools that we use. You can all of them
1568. 50:37to Claude. You can just ask it
1569. 50:38questions. Oh, what is our conversion
1570. 50:40rate? How is our conversion rate
1571. 50:41affecting our revenues? What is our All
1572. 50:44these questions that business questions
1573. 50:45that you normally have to go and ask
1574. 50:47your data guy to go build up your uh you
1575. 50:49know, build build up your dashboards or
1576. 50:51whatever. That's no longer the case. You
1577. 50:53literally go and ask. If you're a
1578. 50:55marketing person, you have a question
1579. 50:56about data, you go to uh you go to
1580. 50:58Claude and you ask Claude uh before you
1581. 51:01approach an engineer to to to pull the
1582. 51:03report for you. So, the amount of
1583. 51:06friction is so low in terms of getting
1584. 51:09data, the amount of asymmetry is so
1585. 51:11little in terms of information. There's
1586. 51:13no excuse for you
1587. 51:15uh
1588. 51:16uh to not get the information that you
1589. 51:18need if you haven't tried. So, our bar
1590. 51:20has increased a lot now for anybody that
1591. 51:22comes uh and works for us, you have to
1592. 51:24learn how to use the systems, and you
1593. 51:27have to um now we're already we're also
1594. 51:30like evolving our systems in a way
1595. 51:32It's going to sound a little technical,
1596. 51:34but I think it's important. Um
1597. 51:36How you know,
1598. 51:38Brian uh uh
1599. 51:41Coinbase uh allegedly
1600. 51:43uh is uh
1601. 51:45you know, having non-developers develop
1602. 51:47production code. Um you did you see that
1603. 51:50thing? They fired a bunch of people. And
1604. 51:52that's funny and a little dangerous
1605. 51:53because you you don't really have I
1606. 51:55think AI is not yet there in terms of uh
1607. 51:59uh having a non-developer just push
1608. 52:01production code. I mean, non-developers
1609. 52:03can build apps easily now with AI, but
1610. 52:06to build a production quality app, it
1611. 52:08takes a lot of effort. But, I started
1612. 52:10asking the question, uh why does it
1613. 52:12matter? Does it matter? Do you have good
1614. 52:15good good good quality?
1615. 52:18So,
1616. 52:18uh turns out it does matter if your
1617. 52:22systems are built in a way that are
1618. 52:24brittle. But, what if you rebuild this
1619. 52:26>> on them. Yeah. No, so what if you build
1620. 52:28your internal architectures to be more
1621. 52:31modular, more fault tolerant in the
1622. 52:33sense like oh, you know, how how to
1623. 52:36build a gigantic like ERP system we're
1624. 52:37building, right? Uh to
1625. 52:40to connect all your diverse all your
1626. 52:42data sources and like sort of visualize
1627. 52:43and all that stuff. If we made
1628. 52:45individual modules independent so that
1629. 52:48even if they fail, they can fail
1630. 52:50independently without taking down the
1631. 52:52whole system. Yeah. So, if you
1632. 52:53modularize and if you box your
1633. 52:55applications and you make them small
1634. 52:57applications, you can actually have
1635. 52:59non-engineers build your applications.
1636. 53:01Like, you can. Yeah. But, you just need
1637. 53:03to understand the the the the the
1638. 53:06failure surface. If you reduce the
1639. 53:08failure surface because AI is really
1640. 53:10good for like starting new projects and
1641. 53:12shipping them quickly, but it's not so
1642. 53:13good at retrofitting into old projects.
1643. 53:16Um especially if an old project has more
1644. 53:18than 2,000 lines of code, AI goes really
1645. 53:19down. The bigger the context, right? So,
1646. 53:22what if you can what if you can create a
1647. 53:25system where you have very small
1648. 53:26context, smaller code bases, uh small
1649. 53:30failure surface,
1650. 53:31uh very strong validation uh criteria,
1651. 53:34and very strong hyper modularity. I
1652. 53:36think there is a new uh this is what I'm
1653. 53:39talking about, new architectures for
1654. 53:40software engineering are emerging with
1655. 53:42this agents as the first users or first
1656. 53:45builders.
1657. 53:46Agent first building and we developing
1658. 53:48all these things are I'm just enjoying
1659. 53:50because all these
1660. 53:52knowledge you gain over time like 25
1661. 53:54years of writing software, right? All
1662. 53:56this knowledge you gain over time are
1663. 53:58coming back now all this modularity, all
1664. 54:01these like services oriented
1665. 54:03architectures, all these architectures
1666. 54:04that lost long time are coming back with
1667. 54:07these agents it's it's a lot of fun to
1668. 54:09be in the market. services architecture
1669. 54:11>> [laughter]
1670. 54:12>> Microservices remember that stuff
1671. 54:13that thing is gone but That's what I
1672. 54:15used to do.
1673. 54:17Right, right. So there's a big hype
1674. 54:18about it but no they went to monetary
1675. 54:20policy where it's too hard to maintain
1676. 54:22microservices but you need microservices
1677. 54:24now because
1678. 54:25you want an individual service to fail
1679. 54:28instead of the whole whole whole system.
1680. 54:30That's actually cool.
1681. 54:31>> So
1682. 54:32exactly. So it's it's kind of fun to go
1683. 54:35back to the old concepts.
1684. 54:37Well yeah, we we talked about a lot man.
1685. 54:39I me and my we kind of align with you
1686. 54:41like how you said it's it's
1687. 54:43non-developers can now build things. We
1688. 54:45talk about it all the time like the the
1689. 54:46way that you change your your work
1690. 54:48professionally.
1691. 54:50We got we got Maine waiting in the back.
1692. 54:53He's going to hop on with us in a
1693. 54:54second. I appreciate you coming on. This
1694. 54:56has been great and appreciate it. This
1695. 54:58has been great. Yeah, it's been awesome
1696. 54:59talking to you. We got to bring you back
1697. 55:01on in like a couple months see how much
1698. 55:03stuff has changed. Let's do it.
1699. 55:05>> Cool bro.
1700. 55:07Yo later. Love Greg.