I sympathize with this guy. He’s as nervous as I would be under this much pressure. He’s mentioning some of my old work colleagues. We were thinking about specialized NN processors in 1985. I’ll jot a few notes in a tweetstream...
“Neural Network Complier”: sic. Yes, this is how you do it. There’s nothing wrong with this presentation, just full of genuine geek detail which I understand. If you are looking for unique features that give an advantage, nothing so far.
He’s concluding. Their NIH NN processor sounds plausible as a research project. No propietary advantage, and can’t afford to keep up progress on it. Continuous emphasis on “all cars have or wil have” FSD hardware, designed to get cars sold now instead of waiting for better.
Q&A: Getting into the weeds with Elon. Power consumption, range. If the computer is using enough power to reduce range, something is seriously wrong. “Chiplet based architecture” - new term for me (I’m way out of date.)
Elon: our solution is FSD specialized, vs Invidia and other general purpose accelerators. “Anyone relying on LIDAR is doomed!”-Elon. Uh, man that fort. WE ALONE WILL RULE.
Elon thinks computer power consumption is critical. Again, false. He might want to look for other defects creating excessive vampire drain.
Costs: true, a custom chip made by a foundry amortized over a million cars isn’t too expensive. But the hardware will still be obsolete in 5 years no matter how many OTA updates it gets. Assuming the FSD software works. Coming up.
IP Theft: Doesn’t matter. Elon-“powerful sustainable advantage” of the fleet. 100x as many cars uploading data. It’s an old but still good deception. Fantasy moats. This to explain why having no FSD patent thicket a la Waymo won’t matter.
Elon-“Simulations don’t capture the real world.” To explain away argument that real-world driving (millions of miles, many deaths as “corner cases” are encountered) is better than simulation (trillions of simulated miles, thousands of corner cases explored, only bit deaths.)
On to visual object recognition. Basic background. Still 1985 level. Tuning out until something interesting turns up.
Blast from the past: DE Shaw (prof at the time) bounding down the aisle at IJCAI ‘85 at UCLA, showmanshipping his multiprocessor for AI. Like many of us, he ended up applying his skills to finance.
Observation: Waymo (nee Google) got 99% of the way to self driving by precompiling (super-mapping) areas and computing paths in advance, leaving only dynamic objects to be recognized. That last 1% is the hard part. claims ot have leapfrogged the static analysis...
...which is why AP has no idea it’s aiming at a gore barrier or a bridge abutment. The GPS-total mapping-consistency approach prevents many, many possible recognition errors.
Drinking game: “corner cases.” Elon adds the creation of a simulation good enough to do training with - “no way!” IE anything I need to be impossible is, anything I say I am going to do is possible. But only by me.
Hint that they send filter instructions to fleet and it returns example videos that fit filter. Admission that returned data is selective - there’s no mult-petabyte library of all data of all sensors and actuators. Which we knew because doesn’t pay for that much LTE.
The vast majority of bulls think this “millions of miles of data” is stored somewhere for training. It’s not. staff working on one issue set up criteria for examples, get examples, train network, run in shadow mode to test. One feature at a time, not holistic...
Integration of all features stays weak. Changing one NN set can have unexpected interactions with another. Shadow mode testing may find most of the bad side-effects. “Most” isn;t good enough when behavior changes and drivers are surprised.
More boilerplate on visual depth perception. I need to eat lunch, this is a good time.
Guy just implied only visual sensing can detect a driver looking back toward your vehicle to merge. I would be very surprised if has a “driver intention” NN feature that can follow other drivers’ gaze.
Key problem with the whole idea of FSD: we have specialized brain hardware for detecting human expression and intention. So long as most vehicles are driven by humans, lack pf this facility will be a problem. Robo cars should be easily recognized so others know they’re stupid.
Next questioner brings that up - LA freeway merging requires social signals. A: blather. Fleet learning! Will predict what people do. MOAR TRAINING.
Elon takes over. More aggressive modes - La freeway mode. Select your risk tolerance! More crashes but faster weaving. [comment: Jesus Christ on a rocket skateboard!]
Elon: - “Once you’ve solved vision, [LIDAR] is useless.” Once you’re insanely rich based on promises, truth is dangerous.
Elon:Detailed mapping and GPS are useless-dangerous. Translation: we don’t have the ability to do that. In the real world you’ll need all the premapping and prerouting you can get to add to realtime recognition. doesn’t have it, so makes many deadly errors it avoids.
Obseqious questioners continue. That last one demonstrated male uptalk. “I love my Model 3!!”
Most of us could never achieve the level of shamelessness required to demo for Elon. This is what you get when the thoughtful, ethical people have all left, leaving you with the glib and the desperate. The best people in this field aren’t at .
Now Elon’s reviewing the master plan. Recycling previous pumps. Solar roof tiles! End of this year. Uh-huh. “Feature complete” FSD means not ready. THERE IS NOTHING NEW HERE, so y’all can stop watching.
I’m drunk. Ten buzzphrases in the last minute. “Combination Uber and AirBnB model.” Mixing previous claims to get investors looking 1-2 years ahead when all this messiness (defective cars, failed product launches, enormous losses, damaged brand) will all have been solved.
There’s no more cash to demonstrate spending on these dreams. Robotaxi fleet! The 20-30,000 cars scattered and decaying in lots all over the country are a start, since they already spent a billion on them. Buy our new issue!
Elon: “Full-stack integration.” This business model has never worked in the absence of a monopoly (either patent or gov’t.) “Million-mile life.” He’s going for the Tesla Network scheme. “we’ll have over a million robotaxis on the road.” NPV of a robotaxi over $300K...
“...so buying a Model 3 is a good deal.” BUY NOW AND MAKE MONEY FROM HOME. I’d like to see the “automatic plug-in” robot working first.
“Any robotaxi that isn’t electric will not be competitive.” Robo gas or hydrogen refueling doesn’t seem much harder. Again the theme is that only Elon’s Way can work. Funnel all your $ to ME.
“We had HD maps but didn’t use them.” Like LIDAR, it is deprecated because it got in the way of claiming cars would be capable of FSD as bought. Because Tesla can’t do detailed mapping, it’s not needed. This is delusional.
“It’s financially insane to buy any car but a Tezzla.” - Elon. Horses mentioned again. The horse puts out less manure.
“Tezzla-owned robotaxis would be concentrated in dense urban areas...” I’m sure will find it easy to recruit local service garages and mechanics to keep all these maintained. They’ve done so well at it so far...
“A neural net is like a cruise missile...” I’m glad cruise missiles aren’t guided by neural nets.
Adam Jonas: question on safety. Lies coming up. “Twice as safe.” Elon: analogy to elevator operators. Exactly the same problem! Jonas: spending? Elon answers, “it’s our entire expense structure.” Because all these cars we cracnk out will end up in the robotaxi fleet.
It was probably a mistake to suggest Tezzla would have its owned fleet. The tesla Network model where Tesla owners maintain / charge the cars and let their cars be used when they don’t need them is more plausible unless thinks it can raise half a trillion to buy a fleet.
Are we done? Guess the test drives aren’t streamed. That was genuinely sad and desperate. Thousands of satellites, millions of robotaxis, just buy the car or at least FSD. Now! Or this dog gets it. 
Lease a 3, somehow or other make money as a robotaxi. Please pay our costs until we can strike it rich in MaaS. The leases were immediately profitable because the credits and risks were largely securitized, leaving with an obligation to service it could ignore.
Create a similar model where consumers use their credit to finance the future robotaxi fleet, collect tax credits, somehow foists all risks onto customers, profit!!! So far we’ve foisted costs onto insurance cos and customers and still lost $billions.
That’s it for now. I specialize in thorough, precise, correct work, so this livestreaming isn’t comfortable. Lots of errors, opinions from the hip, no cites. Thanks for reading! No Patreon, but there’s jebkinnison.com, substratewars.com, and mansionengineer.com.
1/ Some simple explanations of NN training methods and why they cannot reach autonomy - robotaxi level using them. First, Musk points out humans use just simple binocular vision, and if we can, the car has many more cameras.
2/ But a human driver also has a background of general knowledge and ability to recognize anomalous situations that might be more dangerous than apparent. has focused on the simplified problem of limited-access highway driving, where you have abstract vehicles interacting.
3/ engineers send out a filter request and get back driving data that matches it - say, on vehicles cutting in on the highway. They train the network on those cases and release the new weights for shadow testing. The performance when that situation is recognized improves.
4/ Slowly they build competence in the cases they have examined. Several years later, AP can *usually* get you down the highway safely until you take the offramp. promotes it as letting you do other tasks while your car drives you. But then a few people die.
5/ Now is pivoting: deprecating EV manufacture (which we’re told will be cash-flow neutral - profits are no longer expected!), now promoting a new MaaS dream. Spin off FSD software and Tesla Network and raise? I see a lot of credulous coverage.
6/ What happens in city-suburban driving using their training methods? Just pick a corner case out of thousands: everyone knows when you are driving down a street and you see a ball in the street from between parked cars, there’s a big chance a child follows.
7/ Slowing or stopping to be sure it’s safe is obvious to humans. What year will engineers train for the ball-in-the-street case? The plastic bag case was mentioned - object identification is just the start. Humans consult general knowledge and associations. NNs cannot.
8/ And that’s just one obvious case. Their method requires identifying each case first then training on it. will be forever responding to deaths and injuries with “we’re going to handle that corner case soon.” Politically that’s not going to fly.
9/ Which is why every responsible player in the field needs to step on hype *right now* or a few years into their robotaxi rollout autonomous vehicles will be outlawed or subjected to so much precautionary regulation they won’t be practical.
10/ The largest players might survive, but the innovative startups won’t. Elon and will have discredited the entire field for a generation. Autonomy in general driving with mostly human drivers on local street is a much harder problem than network-controlled driving...
11/ Where pedestrians are absent and every vehicle is networked to every other so convoys at high speeds on highways and lightless intersections increase safety and road capacity at the same time. Heterogenous and dynamic streetscapes will always be with us, though.
12/ I don’t think will make it to next year without Ch11/7 and a change in mgmt. This plan is not just audacious or difficult, but completely impossible and reckless. Only an impresario desperate to keep the reckoning at bay as long as possible would propose it.
13/ One NPD CEO and a cloud of enabling media and sycophants normally wouldn’t do this much damage. But in the early stages of bubble-fueled story stock, it pays to be an enabler. Credit Musk with incredible persistence and unwillingness to accept loss.
Finally, @Tweetermeyer shows a better demo of street-level autonomy, with all of the messy realness robotaxis would have to deal with: /end
