Thread by @Tweetermeyer

I'll start my #AIDay reactions thread on a positive note: taking the public inside the latest technology and not being afraid to go over a few heads (basically everything before the dancing "robot") is something the AV sector needs more of. Good for Tesla for leading on this.

The "real" part of the presentation was interesting and impressive... but impressive in terms of what the team has been able to do with an extremely limited hardware platform (sensor suite). It was not particularly impressive relative to what major L4 AV developers are doing.

Musk keeps talking about "real world AI" but basically everything substantive they discussed was foundational work. "Real world AI" development, in the AV context, starts with what Musk calls "the march of nines" and we saw very little about that kind of work. Where the 9s at?

The upshot of all of the "real presentations" for me was: quite impressive, cool stuff, but no real compelling evidence that Tesla is even close to mastering safe SAE Level 5 driving or going driverless. Pragmatically, that seems like more than enough to keep the teams/$$$ busy!

SAE Level 5 means full autonomy in all domains. As an AI problem that means making statistical inferences about ~all of reality~, at something close to 100% accuracy. The major AV developers are developing L4 to limit the complexity of that task, just to have a fighting chance.

If you look at AVs and robotics more generally, the only viable products have been super focused: automate a very specific task and design it to fit the task's needs: Roomba, is one example. Welding cells in manufacturing. Specialized, not generalized.

General tasks require general intelligence. We know how hard that is. We also know that humanoid robots have been tried. Development on Pepper the Softbank robot started in 2004, and it keeps getting fired from customer service jobs because it annoys more than it helps.

Boston Dynamics is the canonical example: it's been working for years, and its major output is... viral videos. If Musk just wants Tesla to make viral robot videos, and the human dancing in a robot suit suggests that is one goal, then sure. Have at it!

See, Musk already tried an ambitious project in specialized robotics: "The Alien Dreadnought" that was supposed to build cars with zero humans at blistering speed. That was such a disaster, Tesla now makes cars largely by hand in a tent. Key quote: "humans are underrated."

If building robots to automate the very specific workflow processes of a car plant was hard, at least it makes sense as a robotics project. You know what it needs to do, so you automate from there. A "general helper bot" is infinitely harder to make useful because you don't.

Limiting the domain and scope is important because variables/complexity scales non-linearly. Even if this "helper bot" were extremely limited in tasks it performs, usefully and consistently performing even one task in an infinitely-variable domain is INSANELY HARD.

Basically, if you find Tesla's ambitions here compelling, I beg of you: please, please read @lauramajor's book "What To Expect When You're Expecting Robots." She covers all the dynamics/factors that push robotics toward specialization (and much more) in a super approachable way.

Ultimately, AI Day's main accomplishment was to widen the divide between people who take Musk on blind faith and those who think it's worth assessing things for themselves. If you're in the latter camp, the evidence overwhelmingly points to this being a recruiting gimmick at best

For five years now Tesla has taken millions of dollars from the public for "Full Self-Driving" and Tesla doesn't appear to be anywhere close to real driverless operation. Heck, we're coming up on the 9th anniversary of the "100% solar Supercharger" promise. How's that going?

Musk basically knows that he can say anything at this point, and tons of people will lap it up. He clearly knows he doesn't have to convince the actual experts: their informed objections will be shouted down as jealousy, lack of vision, or competitive pique.

My main point on all this is: just listen to the experts. Ignore me, I'm a parrot on this topic... just repeating/applying what I've learned from experts. There are many brilliant people in robotics, with very real accomplishments, and we should hear what they have to say.