it's fine but I'm spoiled by largely flying JetBlue's Airbus lines, which are ABSOLUTELY nicer

on the way back I get an A319 instead of the E175 which I think makes for a crappy plane hat trick

of course these days I don't need tires, my wheels are 711mm of solid steel

I absolutely relate to "forgetting the license plate number you've had for 7 years" though

I'm just really baffled by the Google Photos mention here. Did you forget that you sometimes leave your home? Did the idea that "weather happens" really require a computer vision algorithm?

I still think that's a better introduction than your sophomore year track times, though

Massachusetts state agencies, personally. The MBTA and DOT in particular only post alerts on Twitter!

...I can see the vision of "with enough layers the model will figure out the optimal preprocessing itself!"

That said, I don't usually have that many layers, and if you already know preprocessing that more or less works then it's WAY more efficient to just do it!

I will say, back when I was doing more ML work we often found that we could skip a lot of preprocessing steps and only have relatively small performance impacts. A long way downstream from raw sensor data, and it rarely saved enough cycles to be worth the hit. But still...

1/

separately, I also remain an "AI-ready" skeptic -- I don't think it's reasonable to expect producers to maintain use case specific (rather than domain-specific) metadata and I think users have a responsibility to build their own cleaning and ingest pipes

At AGU a speaker referred to ARD as being somehow less rigorous than "AI-ready data" and that's just simply incorrect! We've still got so many competing specifications for "AI-ready" data, while ARD means making your data rigorously interoperable!

yeah sorry that's what I was trying to point at -- but I just woke up and did NOT make that very clear 😅