Thanks for sharing. Elon has assembled some amazing talents in that team and, in aggregate the AI/ML, software and hardware development capability is incredible. Blew my mind.

He's very explicit that they are developing this capability in house and will consider licencing it to other vehicle manufacturers. But does that mark the end of Data Labelling As A Service more generally? And if so when? I agree that in the long term, it is bad new news for APX. But it will be hard for other player to emulate these capabilities.

If the APX State of AI report is to be believed, then in the short-to-medium term the number of clients are growing and the size of budgets is expanding.

I am very interested to see the next APX results and the narrative around them.  

Overall, I agree with your proposition. It is bad for APX, but over what timescale, I am less than clear.

Yea it is bad for Appen especially when the entire first section was dedicated to Tesla's past attempts at labelling. Autolabelling made manual labelling redundant. Although, there is the counter argument that since Automakers are terrible at software they may pay for data labelling anyway (GM cruise, Waymo). 

My problem with Tesla is why are they spending soo much resources mapping out the environment? That ain't AI. I guess they are doing it to enhance the simulation for training purposes but there are way too many edge cases and the environment always change. It's still supervised machine learning and I guess they are developing it "responsibilly". The tradeoff is that it will take a long time for the segnet to go through the march of 9's. 

The part about predicting behaviours from agents is AI. Especially at parking spots and watching others before responding. They are using reinforcement neural networks and I like the montecarlo tree of doing things. Very similar to how Deepmind made Alpha Go with various policy networks. To get real FSD they need to get to Alpha Zero which unfortunately is unsupervised deep learning. Alpha Zero code is much shorter than Alpha Go so I don't get why they need to make supercomputer like Dojo. They are admitting that the neural networks will get larger so they need faster computing. I don't see that trend being sustainable. 

At this stage I much rather buy a Tesla and plug in a Comma 3 for autonomous driving (Comma AI cost less and they are atleast putting resources to solve self driving cars instead of world building :D) Tesla mentioned cones too many times during the presentation but they are not Waymo haha. Also the humanoid robot will be atleast 4 years away from actually working. Autonomus humans are much harder than self driving cars.

Bear in mind the purpose of the presentation is to attract talent with ambitious goals no one has solved.  

I am sceptical that the equation is that simple. With massive growth in AI not all data sets will be amenable to auto labelling and even to train the autolabeling algorithm you will require human annotated data. My understanding remains that its that last few percentage points of accuracy that really challenges any computer only trained dataset. 

To me the situation is sort of similar to when a competitor pops up in a new market at first its concerning before you realise it is because there is a real profitable opportunity there. There is a real need for these labled data sets.

Of course I could be wrong but willing to see what the company says and the numbers it reports rather than accept that all computers will be teaching themselves in five years. 

Just look at the financials. They do not match up with the exponential growth in AI training data demand.  

I said it before and will say it again. THESIS IS BROKEN. IMO to hold on is hoping future innovation will slow...........and it wont. 

Cathie Wood says AI training data costs are dropping 68% pa

Costs are dropping because machines are training machines.....