Emerson1

useful ideas on how to think about improving AI, from Deep Mind researcher. Focuses on the duality of rapid advances yet persistent uncertainties in the field, and what needs more focus.

FSD working better for youtubers in OK!

For they are the aristocrats of our time. The rest, mere peasants and shareholders can drive themselves,and buy the stock. And be happy those prolific video makers, and musk driving up the stock price with perpetual FSD excitement.

The reality is that Waymo and Cruise use huumans in the loop, to make those businesses run, no one has magic AI - and musk will have to do the same when he launches the robotic taxi service (24/7 huumans remotely overseeing/troubleshooting the fleet).

If it works out, Tesla will eat into Uber's massive revenue - much larger than Tesla's. This will go a long way to grow Tesla's top line while EV sales are slow.

Shareholders don't need actual self driving AI cars, or honest marketing, they just need revenue growth and the stratospheric stock price - that's why our favorite car company employee is apparently getting a $44B bonus (may come in handy with any FTC,SEC,DOJ legal fees).

There are good and smart people working on AI elsewhere, tesla seems just fine as is.

(as an aside: Tesla dipping into Uber's revenue, with a tesla taxi service- might be safter bet than the simmilar strategy of Starlink dipping into Telecoms' selling of local internet access).

how could you go about testing this on say llama3 70b with two 4090s - vllm supports tensor parallelism, would the expectation be that inference would be faster with p2p? How would you update the nvidia driver tho? Thanks any thought appreciated