Lorenzo Garcia

@Amarillo_Slim1 you need to get more educated on what the CPU narrative is about

This absolutely nukes the bullshit CPU narrative. uccl-project.github.io/posts/mkernel/

@jiayiy how is this different from skip softmax from nvidia?

🚀 BLASST just won Best Paper at #MLSys26! In this paper, we introduce a simple, training-free dynamic sparse attention mechanism that uses a single scalar threshold on online softmax statistics to skip negligible attention blocks. Unfortunately I won’t be there in person, but please say hi to my awesome coauthors! 🙌 Paper: arxiv.org/abs/2512.12087

@always_ff_rohan > FPGAs can hit 400 GFLOPs, basically replacing a 16-core server. 400 Gflops is nothing in the grand scheme of things. A single Graviton 4 core can do around 360 bf16 GFlops, so i don’t know which 16 cores are you planning to replace with this fpga

Baidu and a bunch of other companies have already been using FPGAs for AI. Meanwhile, Indian AI/DL startups are just burning money at GPUs. For instance on matmul, FPGAs can hit 400 GFLOPs, basically replacing a 16-core server. For DL workloads with small batch sizes, FPGAs can actually beat both GPUs and CPUs in terms of throughput.

@ezyang linear attention ops

Give me your O(num operators) PyTorch improvement ideas that you are interested in. Historical examples: making every kernel deterministic / support zero size. Not done: every kernel in your favorite DSL / batch invariant / masked / padded / device side size

@yacineMTB @difficultyang you’re delusional bro

@difficultyang i would simply just not use torch at all and write cuda directly

Let's say we wanted to rewrite PyTorch from scratch, because such a thing is topic du jour in the age of LLMs. What would the goals of such a rewrite be? What problems could a rewrite solve that incremental evolution from where the code is today not? 🧵

@bubbleboi taylor series approximation of exp is not an innovation. people have been doing this for ages

So Nvidia added the following innovations: -Taylor series like approximation of e^x -Horners method for soft max And we think GPUs aren’t just becoming LLM accelerators at this point?

@benjihyam aged like milk

ChatGPT is the next clubhouse. In 12 months, people will forget it existed.

@JennyTheDev @AnthropicAI you must be super dumb

@AnthropicAI 24,000 fake accounts and 16 million exchanges just to avoid training their own models. The "build vs steal" decision tree is getting wild.

We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.

@scaling01 @zephyr_z9 100x speedups are BS

This is what I'm most bullish on. Kernel optimization is just going absolutely vertical.

@gopinath9629 @mkurman88 vLLM does support CPU, if it didn’t work for you please raise and issue and we’ll fix it

@mkurman88 Llama.cpp doing well I guess but still its lack behind when its coming to multiple gpu and vllm don't support cpu i am currently expolaring exo and then distributed llama for my setup.

Is vllm broken? How can two identical requests produce different outputs? Temperature 0.0, seed 42 - how the hell can you ensure determinism?

@ivanfioravanti @awnihannun Everybody does that btw

Apple MLX team never stops. Here @awnihannun merging a PR on Saturday morning 💪🏻

@istoica05 17x speedup means that the baseline was either not written by experts or the op was not important enough to be optimised

Optimizing kernels for emerging hardware architectures remains a critical bottleneck today. This new blog post describes Autocomp, a new ADRS (AI-Drivern Research for Systems) framework that leverages LLM-driven search to automate this process and achieves up to a 17x speedup over expert-written code on AWS Trainium.

@suchenzang he doesn’t need that hairline yet he has it

@_fmla_ ilya would never need linkedin recruiters

be bold take risks aim to build super-intelligence before even reaching baseline intelligence

@noor_supernova7 .

For those who do not know, the occupation of Gaza has begun. If you’re scrolling, PLEASE leave a dot . it's just a dot.

@HasanEssam29636 .

I’m very hungry. My body is slowly falling apart from malnutrition, dizziness, and weight loss. If you’re scrolling, PLEASE leave a dot . it's just a dot

@chrisgpt Nobody wants to be garbage collected

can’t sleep, who wants to be added to a gc?

@GavinWax I lost brain cells reading this

Silicon Valley is a strategic national asset whose jobs should be legally reserved for U.S. citizens. It’s a matter of national security. This shouldn’t be controversial.

@SupaDupaChop That’s not what they’re saying in the video. They’re encouraging julani to strike tel aviv

@_jasonwei The transformer was invented to machine translation MT btw

There are traditionally two types of research: problem-driven research and method-driven research. As we’ve seen with large language models and now AlphaEvolve, it should be very clear now that total method-driven research is a huge opportunity. Problem-driven research is nice because you have a consistent and specific goal. The goal is usually virtuous, so it feels good to have a mission and identity. However, it just doesn’t work due to The Bitter Lesson. Basically everything in classical NLP (machine translation, summarization, chatbots) lost to simple scaling. ChatGPT is a prime example—it used nothing from chatbot research and certainly wasn’t the intended end goal of OpenAI’s 2022 research program, but was a huge hit because someone (John Schulman et al) figured out the right way to package large language models as a product. Method-driven research feels less stable because you’re constantly searching for problems and you have to be opportunistic. But I believe AI will allow method-driven research to dominate progress in most fields of science, one-by-one. The latest method (or “hammer”), as we’ve seen in AlphaEvolve, is ruthless search and optimization against a reward function (whether this requires RL or not is a separate discussion). Things that problem-driven researchers have been trying to solve for a long time like the kissing number problem will become nails hit by the hammer. Eventually the hammer will become bigger, stronger, and more general and will hit more and more nails. So a very important meta-skill for the next decade will be knowing how to create the right environments to use The Hammer. Ironically, the problem-driven researchers, who by definition are experts in a specific problem, are well-positioned to create these environments. If, that is, they can put down their egos and pick up the hammer.