FSD Release Notes Leaked, Cybertruck Report, Twitter Crossover — Transcript

Full transcript

1. 0:00hey everybody rob Mauer here and today
2. 0:02we're going to be talking about new FSD
3. 0:04Beta release notes 10.69.3 has started
4. 0:07to roll out to employees so we'll go
5. 0:08through those also got news on Tesla in
6. 0:10China a report on the Cyber truck
7. 0:12production timeline so I'm crossover
8. 0:14news between Tesla and Twitter and a few
9. 0:16other items as well another
10. 0:17outperformance for Tesla today starting
11. 0:19to add up a little bit this week Tesla
12. 0:21up a tenth of a percent to close at
13. 0:23227.82 while the NASDAQ was down nine
14. 0:26tenths of a percent but again probably
15. 0:28doesn't matter too much as we head into
16. 0:29the fomc meeting tomorrow that'll
17. 0:31probably be the major determining factor
18. 0:33for where the market goes from here
19. 0:34alright today we're starting off with
20. 0:36some FSD beta news of course yesterday
21. 0:38we talked about how Tesla had started to
22. 0:40push out
23. 0:4110.69.2.4 so no different release notes
24. 0:43on that probably just bug fixes but at
25. 0:45the same time it looks like Tesla
26. 0:46started to push out 10.69.3 a new
27. 0:49version out to employees we've been
28. 0:52waiting for this for a couple of weeks
29. 0:53and this version does contain
30. 0:54significant changes which are captured
31. 0:56in the release notes which have leaked
32. 0:57as shown here on Twitter by homars where
33. 0:59you've got about 15 different bullet
34. 1:01points here and as usual it's fairly
35. 1:03technical in terms of the language so I
36. 1:05think it's very helpful to break these
37. 1:06down and put them into broad categories
38. 1:08of course there is going to be some
39. 1:09overlap between these categories but
40. 1:11chunking them up I think makes it a
41. 1:13little bit easier for us to dig through
42. 1:14and understand what we're really seeing
43. 1:16in this update for me I'm really seeing
44. 1:18four broad categories performance
45. 1:19related changes really improving latency
46. 1:21and things like that control so
47. 1:23adjusting things in the planer and then
48. 1:25lanes and also vulnerable Road users or
49. 1:28vrus so we'll go through those
50. 1:29categories but first just three kind of
51. 1:31General bullet points that don't
52. 1:32necessarily fit specifically in those
53. 1:34categories Tesla has upgraded the object
54. 1:36detection Network to Photon count video
55. 1:38streams and retrained all parameters
56. 1:40with the latest Auto label data sets
57. 1:42with a special emphasis on low
58. 1:43visibility scenarios so this is a big
59. 1:45deal elon's talked a lot about this over
60. 1:47the last year this is Tesla bringing in
61. 1:49the most unfiltered possible data set
62. 1:50that they're getting from their cameras
63. 1:52not doing any post-processing on it and
64. 1:54using that as the input for their object
65. 1:55detection neural network and presumably
66. 1:58that gives them the most robust data set
67. 2:00to make the best decisions with and also
68. 2:02save some processing where that
69. 2:03filtering step might apply otherwise
70. 2:05last year Elon described this on Twitter
71. 2:07by saying that one of the improvements
72. 2:08to FSD Vision involves training with
73. 2:10actual Photon counts so removing the
74. 2:12filters used to make pictures pretty to
75. 2:14the human eye a common example of that
76. 2:16might be an infrared filter which we
77. 2:17wouldn't want to see in our pictures or
78. 2:19videos but could be helpful to the
79. 2:20neural network in figuring out what's
80. 2:22going on around the same time last year
81. 2:24Elon on the Lex Friedman podcast said
82. 2:26that this will allow Tesla to see much
83. 2:28better in the dark than you could
84. 2:29possibly imagine due to small
85. 2:30differences in photon counts so it's a
86. 2:33big deal and Tesla chose to put this
87. 2:35bullet point first it's also exciting
88. 2:37that Tesla is using Auto labeling on
89. 2:39these raw Photon count video stream data
90. 2:41sets that's probably a key factor in
91. 2:43even being able to make this change in
92. 2:45the first place all right the other
93. 2:46couple of points in the general category
94. 2:48Tesla says that they have improved the
95. 2:49architecture for better accuracy and
96. 2:51latency higher recall of far away
97. 2:53Vehicles lower velocity error for
98. 2:55crossing Vehicles by 20 and improved
99. 2:57vulnerable Road user Precision by 20
100. 2:59percent think those ones are pretty
101. 3:01self-explanatory and then finally in the
102. 3:02general category Tesla upgraded the
103. 3:04occupancy Network to align with Road
104. 3:06surface instead of ego which is the
105. 3:08vehicle or improved detection stability
106. 3:10and improve recall at Hillcrest so that
107. 3:13one's a little bit trickier just to make
108. 3:14my best guess at that I would imagine
109. 3:16that that means that Tesla is kind of
110. 3:17anchoring the position of things that
111. 3:19are appearing in the occupancy Network
112. 3:21to the road surface rather than to the
113. 3:23vehicle's position which then as I say
114. 3:25there seems to improve the stability of
115. 3:26the positioning and the recall at
116. 3:29Hillcrest when things may be out of
117. 3:31vision for a period of time alright and
118. 3:34more specific categories here we've got
119. 3:35three bullet points on performance
120. 3:37related things so things that are
121. 3:38reducing latency allowing the system to
122. 3:40make quicker decisions the first one
123. 3:42here is an 80 reduction in the run time
124. 3:44of candidate trajectory Generations I
125. 3:47believe these will be the generations of
126. 3:48the possible paths the vehicle may take
127. 3:50so to achieve this Improvement Tesla
128. 3:52says that they have converted an
129. 3:54expensive trajectory optimization
130. 3:55procedure into a lightweight planar
131. 3:57neural network I think this is exciting
132. 3:59because Tesla has talked about using C
133. 4:01plus plus for a lot of the planning and
134. 4:03control work so where Tesla can convert
135. 4:05some of those functions to neural
136. 4:06networks there's potential for both
137. 4:08better efficiency and better
138. 4:09Effectiveness so it sounds like that's
139. 4:11what Tesla is achieving here with the
140. 4:13trajectory generation
141. 4:14next performance Point Tesla has
142. 4:16converted the non-vulnerable road user
143. 4:18attributes Network to a two-stage
144. 4:19Network which they say has reduced the
145. 4:21latency reduced incorrect Lane
146. 4:23assignments of Crossing Vehicles by 45
147. 4:24and reduced incorrect Park predictions
148. 4:27by 15 finally then Tesla says that they
149. 4:30have reduced the best case input to
150. 4:32Output or object Photon to control
151. 4:34system latency by 26 through the
152. 4:36combination of adaptive planar
153. 4:38scheduling restructuring of trajectory
154. 4:40selection and parallelizing perception
155. 4:42compute I think that's mostly a summary
156. 4:44of some of these other changes so main
157. 4:46point for this category being that Tesla
158. 4:48has changed a number of things that will
159. 4:49allow Tesla to as they say here make
160. 4:52quicker decisions and improve reaction
161. 4:53time next a few bullet points here on
162. 4:56control Tesla's improved control for
163. 4:57vehicles cutting out of their desired
164. 4:59path by better modeling their turning
165. 5:01and lateral Maneuvers thus avoiding
166. 5:03unnatural slowdowns anytime I see
167. 5:05improvements on unnatural slowdowns that
168. 5:06makes me happy that's probably as I've
169. 5:08said before my most common intervention
170. 5:10Tesla's also improved longitudinal
171. 5:12control while offsetting around static
172. 5:14obstacles by searching for feasible
173. 5:16vehicle motion profiles so that would be
174. 5:18better control of vehicle speed while
175. 5:19going around things that are stationary
176. 5:21and then Tesla also notes improved
177. 5:23longitudinal control smoothness for
178. 5:25in-lane vehicles during High relative
179. 5:27velocity scenarios by also considering
180. 5:29relative acceleration in the trajectory
181. 5:31optimization now I'm not certain but I
182. 5:34hope I really really hope that this
183. 5:36means that if there's a vehicle in front
184. 5:37of you and it speeds up that your
185. 5:38vehicle will speed up and follow it a
186. 5:40little bit more closely by factoring in
187. 5:42the time that it takes to get back up to
188. 5:44that relatively High Velocity if anyone
189. 5:46has a different interpretation of that
190. 5:47please let me know that would be amazing
191. 5:48if that is what that means finally on
192. 5:51control I'm not sure if this really fits
193. 5:52into control or maybe I'm just
194. 5:53misinterpreting something here but Tesla
195. 5:55has added control for more precise
196. 5:57object geometry as detected by General
197. 5:59occupancy Network so I guess just
198. 6:01improving the controls as Tesla has now
199. 6:03moved towards the general occupancy
200. 6:05Network to be more adaptive to the more
201. 6:08precise information that Tesla is now
202. 6:09getting all right last couple of
203. 6:11categories here probably a little bit
204. 6:12more self-explanatory first being Lanes
205. 6:15Tesla has reformulated the auto
206. 6:16aggressive Vector Lanes grammar which is
207. 6:19a fun reminder back to AI Day 2 when
208. 6:20Tesla talked about building their lane
209. 6:22model off of language models anyway
210. 6:24they've improved the Precision of lanes
211. 6:26by nine percent recall of lanes by 19
212. 6:28recalled Forks by 51 and they say this
213. 6:31includes a full Network update where all
214. 6:33components were retrained with 3.8 times
215. 6:35the amount of data Tesla has also added
216. 6:38a new road markings module to the vector
217. 6:40Lanes neural network which improves Lane
218. 6:41topology error at intersections by 39
219. 6:44presumably that means that Tesla is now
220. 6:46looking at more things on the road that
221. 6:48can be then Auto labeled and used to
222. 6:51improve the perception of the lane
223. 6:52Network finally on Lane's Tesla says
224. 6:55they've improved decision making for
225. 6:56short deadline Lane changes around Gores
226. 6:58which is the triangle area when one road
227. 7:00splits into two so you've got the
228. 7:02physical area where those roads split
229. 7:04but then also the markings on the road
230. 7:05indicating that that split is coming
231. 7:07Tesla says they now have richer modeling
232. 7:09of the trade-off between going off Route
233. 7:11so going the wrong way when those splits
234. 7:13happen or the trade directory required
235. 7:15to drive through the gore region so
236. 7:17obviously by that they're talking about
237. 7:18the road markings of the gore region but
238. 7:21kind of an interesting one here as I
239. 7:22wouldn't think Tesla would be too eager
240. 7:24to drive through those very often even
241. 7:26if it seems like it could be
242. 7:27accomplished anyway final category here
243. 7:29is vru's vulnerable Road users Tesla
244. 7:32says that they have converted the VRU
245. 7:33velocity Network to a two-stage Network
246. 7:35which reduced latency and improved
247. 7:37Crossing pedestrian velocity error by
248. 7:39six percent and Tesla's reduced to false
249. 7:41slowdowns for pedestrians near
250. 7:42crosswalks by using a better model for
251. 7:44the kinematics of The Pedestrian so I
252. 7:46think both of those pretty
253. 7:47self-explanatory but again anything
254. 7:49reducing false slowdowns in a safe way I
255. 7:52think makes the product a lot better so
256. 7:54that's everything in 10.69.3 definitely
257. 7:56seems like a significant update
258. 7:57hopefully employee testing goes well and
259. 8:00some people outside of Tesla can start
260. 8:01to get this version over the next week
261. 8:03or so alright next up we've got a bit of
262. 8:05an update on Tesla in China Moneyball on
263. 8:07Twitter is reporting that the China
264. 8:09insured units for the last week of
265. 8:10October here have come in so October
266. 8:1224th through the 30th at 8 1079 Vehicles
267. 8:15if we combine that with previous insured
268. 8:17vehicle reports we can see that October
269. 8:193rd through the 30th is about 17 300.
270. 8:22that's not going to tie perfectly with
271. 8:24the retail sales or domestic sales
272. 8:26figure that the cpca reports in a week
273. 8:28or so but should be pretty close then
274. 8:30you also have to factor in October 1st
275. 8:31and 2nd and October 31st which is not
276. 8:33covered here but it gives us a ballpark
277. 8:35of somewhere around 20 000 vehicles for
278. 8:38domestic sales during the month of
279. 8:39October now this makes it pretty clear
280. 8:41that Tesla is still doing a little bit
281. 8:42of that delivery wave type of stuff but
282. 8:44this would be if it is 20 000 the
283. 8:47highest first month of the quarter for
284. 8:49domestic sales in China so far still the
285. 8:51numbers should be expected to pick up
286. 8:53significantly for November and December
287. 8:55especially following the price cuts that
288. 8:57we saw on October 24th which like we
289. 9:00talked about likely chosen by Tesla
290. 9:01because they're ready to shift a little
291. 9:02bit more of their production back within
292. 9:04China also Shanghai related we've got a
293. 9:06report from Bloomberg today that Tesla
294. 9:08is sending over a couple hundred workers
295. 9:10from gigashanghai to Fremont to help
296. 9:13improve production at Fremont I think
297. 9:15this is actually pretty exciting because
298. 9:16we just saw Giga Shanghai go through
299. 9:18their production upgrade very smooth
300. 9:20huge increase in productivity in a very
301. 9:22short period of time those people have
302. 9:23seen that take place or rather have made
303. 9:26that happen and hopefully can bring in
304. 9:27some insights and improve productivity
305. 9:29at Fremont as well all right next we've
306. 9:31got an update from Reuters on Cyber
307. 9:33truck production today I don't think
308. 9:35there's anything really new here they're
309. 9:36reporting that Tesla is aiming to start
310. 9:38mass production of the Cyber truck at
311. 9:40the end of 2023 according to two people
312. 9:42with knowledge of the plants how they
313. 9:45say mass production here so I don't
314. 9:46think that's really any shift in the
315. 9:47timeline oh indications have been the
316. 9:49middle of 2023 for the start of
317. 9:51production so of course it's going to
318. 9:53take some time to get that to mass
319. 9:54production which is somewhat of an
320. 9:56arbitrary thing anyway usually for Tesla
321. 9:58we're talking about that being a 5 000
322. 10:00per week number that would actually be
323. 10:01incredible if Tesla is targeting that
324. 10:03for the Cyber truck for the end of 2023
325. 10:05I would be surprised if they could hit a
326. 10:07number like that but this report may not
327. 10:09be talking about it in that way anyway
328. 10:11all right the next Tesla item that's
329. 10:13been getting a lot of attention over the
330. 10:14last day or so I've gotten a few
331. 10:15questions on this CNBC has reported that
332. 10:18Elon has enlisted some help from about
333. 10:2050 Tesla employees to help out with some
334. 10:22of the changes that are going on with
335. 10:23Twitter right now of course we've seen
336. 10:25on Twitter Elon talking a lot about this
337. 10:27today I think regardless of how you feel
338. 10:29about it it is exciting to see how
339. 10:31quickly Elon moves regarding Tesla
340. 10:33employees helping out for this period of
341. 10:35time though I don't care about it at all
342. 10:36I trust Elon very much to make the
343. 10:38proper organizational decisions that
344. 10:40need to be made for both organizations
345. 10:41and do so in a way that is not
346. 10:43detrimental to Tesla we've already seen
347. 10:45the competition for resources if you
348. 10:47want to frame it that way between Tesla
349. 10:49and SpaceX there's often crossover work
350. 10:51between employees for those two
351. 10:52companies and they manage just fine
352. 10:54Tesla can do the same here with Twitter
353. 10:56for a little bit and eventually that
354. 10:58will probably Circle back around and
355. 10:59Tesla will be able to benefit from
356. 11:01Twitter employees Elon at the top makes
357. 11:03that easy to facilitate and again I
358. 11:05trust Elon to make those proper
359. 11:06decisions without harming Tesla which by
360. 11:08the way Elon still the biggest
361. 11:09shareholder of Tesla still the majority
362. 11:10of his net worth in test so it's not
363. 11:12something that I'm worried about alright
364. 11:14last few items here Panasonic has
365. 11:16announced yesterday that they are going
366. 11:17to start building their new battery
367. 11:18plant that we've talked about in Kansas
368. 11:20this month they say that they are aiming
369. 11:22to begin mass production by March of
370. 11:252025. next Toyota has reported their
371. 11:28fiscal second quarter so it'll be
372. 11:29confusing but their end of September
373. 11:31quarter numbers and bottom line their
374. 11:34net income which was 461 303 million yen
375. 11:38which converts over to about 3.1 billion
376. 11:40US Dollars was below Tesla's third
377. 11:43quarter Gap net income of about 3.3
378. 11:46billion dollars so pretty crazy
379. 11:48Milestone especially to be achieved with
380. 11:50such a disparity in scale at the moment
381. 11:51but fun to see after all the criticisms
382. 11:54of Tesla being the most valuable
383. 11:55automaker over the years well now people
384. 11:58are starting to see why that is anyway
385. 12:00net income was down 28 percent
386. 12:02year-over-year for Toyota and earnings
387. 12:04per share was down 30 percent looks like
388. 12:06the revenue was up 22 so not exactly
389. 12:09sure what's driving such bad
390. 12:10profitability and I know they've got
391. 12:12foreign exchange headwinds on their cost
392. 12:14structure but without spending more time
393. 12:16analyzing it kind of tough to sort all
394. 12:18those things out finally then today
395. 12:20congratulations to SpaceX for the
396. 12:22successful flight and landing of the
397. 12:24Falcon heavy for the first time since
398. 12:272019. always incredible to see those two
399. 12:30boosters come back and land
400. 12:31simultaneously so congrats and that's
401. 12:33where we'll wrap it up for today so as
402. 12:35always thank you for listening make sure
403. 12:37you're subscribed and signed up for
404. 12:38notifications you can also find me on
405. 12:39Twitter at teslapodcast and we'll see
406. 12:41you tomorrow for the Wednesday November
407. 12:432nd episode of Tesla daily thank you
408. 12:47[Music]