In the spirit of this holiday, huge thanks to my cohosts @wolframravenwolf.bsky.social @nisten.bsky.social @yampeleg.bsky.social

Tons of folks on stream today!

theyre free on huggingface, and like.. actually good

Really appreciate how huggingface is becoming public infrastructure.

small correction to nitpick on, the TURN server is not necessary to relay the whole traffic, just the initial session-negotiation/nat-traversal etc..

The rest of the traffic itself itself can also be entirely p2p(peer to peer).

So far 3 doctors, 1 nurse, 2 builders of deployed AI doctor apps and 2 caregivers of people with dementia have spoken.
The discussion is stacked.

There are 2 real doctors and a real nurse in the room and the conversation has started off with a banger insight.

The AI itself has dementia😅

Yeah i was surprised it just worked in LM_eval Harness when vllm was erroring out and i replaced it with plain hf, was expecting it to be slow and do the llama.cpp style splitting but it wasn’t slow it was actually parallelising evenly across 8x A40s

Yea, at one point.

It was also very much a pleasure to chat with @JustinLin610 from @Alibaba_Qwen team again about the crazy 1M context window Qwen 2.5 Turbo!
This felt like the ThursdAI crew assembled but in video format 👏
@yampeleg @nisten.bsky.social @wolframravenwolf.bsky.social

This is so..defeatist..

is that kekistan

The article is more of an emotional piece with Elon etc and rather than being oriented towards evidence-based medicine or using objective reasoning to provide an accurate assesment.

All of Norways healthcare has deployed AI for xrays for months. If you do a chestscan, an AI reviews, prioritises it.
I myself worked on the first commercial doctor ai, been deployed for over a year and a half with no real issues.
What people say & feel is different from what the evidence shows.

didnt they have it going at one point via some oss server?

Not if you dont own the data.

But if someone lives in a place with good healthcare and wants to donate their info to open science i think thats more than fine.

many such claims, but hmmm this one is coming from the team that actually shipped a good working 1m context model