Jannes Gladrow

Surely TACO still applies? Anyone?

Same for housing.

All the angst in genAI orgs or even Nvidia because of R1 is overblown. If you want to roll this model out to millions of users, you'd still need a ton of compute. Hyperscalers are safe.

And, from a safety perspective, the apparent ease with which deepseek caught up to OAI would only leave (massive?) restrictions on who gets to use compute as a viable strategy? 🍿 potentially ahead.

Also, r1 re-establishes the brutal ML principle that simple techniques win. GRPO is almost PPO: no MCTS or similar needed. Just guess until the end, then verify.

how ironic it would be if the first people that are actually AIed away are those from fields where truth can be established in an automated fashion... #r1

So I set up some recurring GPTtask using the new feature. But .. how can I stop it again?

First positive thing to note is an impeccable 5G coverage quite unlike London #vancouver

On my way to #vancouver. First time. What should I see, hear or eat?

Can't believe spiking was NOT considered as an offense until now.

SimpleNN with a poetic touch. Is chatgpt trying to tell me sth?

OpenAI shipping just in time for #neurips. Whatever it is (O1?) it might again dominate the conference like 2 years ago.

Looking at you SF and London

Something something about supply and demand... #yimby

The official Llama3 repo is not using flash attention? Probably just a minimal code skeleton instead of full open source.
github.com/meta-llama/l...

Living in London means dreading the province but simultaneously getting to terms with living in a dark mouldy shoebox forever

Same for secret-tax-trick-that-will-save-you-millions content

Is there a word for the fear of accidentally clicking on a WWII YouTube video provoking the algorithm to from then on only ever show you super-secret-nazi-plan-in-WWII content?

Advent of code seems like the right kind of activity for a day like this one

Love my ultra I but I did switch out the pickups for vintage ones... Not sure what the differences are to the ultra II

Anyone seen any projections for what O1-style models will mean in terms of cost/compute?

Train back from Cambridge on Wednesday got delayed indefinitely. Train to cambs this morning got cancelled, train back now is delayed.

Morning commute in icy England starts to feel like Christmas

Maybe now?

Who's on here in #ml #ai to follow?

I guess it's a sign of general number inflation (aka Moore's law) that a quadrillion is now a normal number in machine learning