(I know I’m two months late)
To back up what you’re saying, I work with ML, and the guy next to me does ML for traffic signal controllers. He basically established the benchmark for traffic signal simulators for reinforcement learning.
Nothing works. All of the cutting edge reinforment algorithms, all the existing publications, some of which train for months, all perform worse than “fixed policy” controllers. The issue isn’t the brains of the system, its the fact that stoplights are fricken blind to what is happing.
Don’t forget the fully fledged remote desktop thats built in, WebVR (which is being replaced with Web XR), Web Bluetooth, Web USB (aka Web Serial), the API’s for notifications, ambient light sensors, an entire transactional database (indexed DB), the language translation API, the Gamepad API (videogame controllers), hardware passkeys (yubikey), speech to text, text-to-speech, webGL, webGPU, webworkers, service workers, an entire suite of cryptography tools, GPS location, battery, vibration, FileSystem API, picture-in-picture API, WebRTC, WebSensors, etc.
And then, on top of all that, building a miniture OS-kernel so that tasks can be sandboxed scheduled/executed and prevent 1 tab from crashing everything or hogging resources.