That reads kinda meh

Robo Taxis are mostly bullshit, because we can’t even handle scenes for robots in a controlled environment.
What will this robo taxis do, if snow or dirt disrupts their view?
What will they do, if they can’t scan a speed limit sign at a road construction site?
That stuff doesn’t work for most current automobiles that show you the speed limit, which they think is right, how should an automatic system adapt?
Even if you preprogram them for the cities, you want to use them in, there are still ever changing things like constructions.

And the AI stuff reads like the last years.

Somehow we seem to not have anything really new and exciting

As I said, the architectural changes are quite cool

As far as I’ve understood it mostly comes down to splitting it up into multiple expert systems, so you don’t need to activate the complete system with every request

But I’ve only scratched the surface…

Also, open source… The weights are made publicly available.
None of the training data or systems

Edit: regarding “open source”:
Also Meta’s Llama is on huggingface, just like deepseek. I still wouldn’t talk about transparency here

Distilling OpenAI and Llama models probably also helped quite a bit

Although I must admit, that the architectural changes are pretty cool

but I have to add, that I’ve just started reading into the topic a few weeks ago and don’t really have any real practical experience, besides checking out some huggingface docs I got linked yesterday and stupid me hasn’t thought about looking there…
So everything I say is probably bullshit o⁠:⁠-⁠)