Bas Büller

Hey all, I will be at GTC next week talking about all the work my team and I did on large-scale MoE training in JAX on GPUs! We decided early on to have a fully dropless training stack to avoid token dropping. (1/7)

@__tinygrad__ Tinygrad will use thunderkittens as a backend?

PyTorch frontend Halide rangeify E-graph symbolic ILP memory planner (MODeL) ThunderKittens backend Pure Python drivers

I am hiring highly skilled performance engineers for my team! You will be working on optimising pretraining for models >100B params on O(1000s) of GPUs, and hardware-aligned architecture design. We are cooking a lot of very exciting projects and I can safely say you will have a lot of fun! Link in thread. <3

@BlackHC My thoughts

Isn't this the whole OpenAI Zürich office? 😂

@levelsio Get that 5090 if it’s available. More flops more ram

Nice, what about this set up?

@levelsio Vcs really are their customers

Landing page built for VCs not customers

Knuth shows us the way. Again:

(1/5) We’ve never enjoyed watching people chop Llamas into tiny pieces. So, we’re excited to be releasing our Low-Latency-Llama Megakernel! We run the whole forward pass in single kernel. Megakernels are faster & more humane. Here’s how to treat your Llamas ethically: (Joint with @jordanjuravsky, @stuart_sul, @OwenDugan, @dylan__lim, @realDanFu, @simran_s_arora, and @HazyResearch)

@dhh Same here

This has been happening to my wife's iPhone. Totally wild that nobody at Apple appears to own this. That it can just be a thing for years on end. Wonder how many flights or other critical appointments have been missed over this?

@JVal48103 Yes

I bought my husband a new wallet probably 5 months ago. He has no desire to switch over to it yet. The left is his current wallet. My dad has done the same thing. Men, is this a universal thing? Do you have a hard time giving up the old, faithful, falling apart wallet? 😅

@yacineMTB Human powered ones ;)

What is the safest wood cutting tool I could possibly use? I am extremely terrified of losing my fingies

@_rchaves_ Just Python, pretty sure it will be more convenient

I am coding the same agents in all LLM frameworks to compare them side-by-side: - DSPy - Langgraph - Google ADK - PydanticAI - InspectAI what else should I write it on?

@wordgrammer Hmm, I think it’s mostly that Python is easy to use as the frontend of your DSL. Python calling out to native code is really simple because of CPython and a simple and relatively unassuming runtime. For these reasons I’d love to see more LISP as it does better on those aspects.

Odd realization. 95% of compiler optimizations for machine learning occur just-in-time. Which is the main reason everyone uses Python. It’s not because Python is easy to learn; it’s because Python is easy to build JIT compilers for

@andersonbcdefg For some absurd reason 95% of programmers only go for frameworks and are scared of anything below. They somehow don't see that frameworks usually enshittify the whole thing!

@andersonbcdefg Call it shiny toy syndrome. This is also a general pattern I find with hf code. It looks simple at first glance, try do anything serious with it and the spaghetti code hits you.

wait. distributed training with pure pytorch is not that bad. why did we all collectively get gaslit into using accelerate...

@tsoding Tsoding lives both in the future and the past

Yes, I created a simple Stack-Based Programming Language that lets me describe how to cut and concatenate VODs and the interpreter automatically translates it to a sequence of ffmpeg commands. I've been using this system to edit my VODs for 4 years already. It's called Markut.

@_xjdr Hold my beer, seems like I’ve got new plans for tonight.

@ludwigABAP Synergies between model and cli will probably become the differentiating factor. Claude 3.7 supposedly is already better in Claude cli than cursor, will only get worse

every lab is gonna release the same typescript TUI with the same features that only integrates to their model and all of them will get washed by the one made by someone who knows how to make great TUIs and isn't tied to one specific model provider

@abacaj Although I agree Gemini is in a league of its own right now, the ease with which people switched to Gemini can also happen in reverse. But Google’s massive infra lead is what I’d be afraid of if I were one of its competitors, now that they understand how to build sota models.

OpenAI and Anthropic are cooked. Google finally made the best model - not just in benchmarks but real use. I don’t know anyone in my circle using Claude anymore, a move I did not see coming

I think AI is going to usher in a gold age of infra, not obviate it. It's just so clear that good CS fundamentals result in better AI built systems. Vibe coding works better with type safety, languages where syntax maps closely to semantics, referential transparency, tight scoping etc. These approaches have never been widely adopted in CS because they are hard for humans, and in particular novices. But they're not hard for AI. And they map so much better from natural language descriptions.