Stephen Fernandes

@Sid_Arora_18 @LTIatCMU @Meta @shinjiw_at_cmu Congratulations Siddhant 🎊

Finished my PhD at @LTIatCMU . Excited to start at @Meta as Research Scientist, where I’ll be working on speech processing and conversational AI for wearables like smart glasses. Grateful to my advisor @shinjiw_at_cmu, collaborators, friends, and family for all the support.

The CUDA Moat has been broken ✨️

@ylecun @julianboolean_ he was just trying to get your attention, to just do reach maxxing

@julianboolean_ You simply did not understand the argument.

It's interesting to think about how LeCun got this so wrong In a sense, he was perfectly correct. LLMs almost always get answers "wrong" - if by "wrong" you mean that somewhere in the reasoning trace there was a misstep But we don't care about the reasoning trace and its numerous misfires; we only care about the final answer. "So "the probability that any produced token takes us outside the set of correct answers" is meaningless - we can't define correctness until the last token. There is no exponential divergence.

@maharshii whats the training regime ? one wrist curl for every kernel launch ?

what being a nerd does to you

the polarization in mindset that AI brings

@0xSero @huggingface Open source must win 🏆

Best models to run on your hardware level I'll be doing this every week, I hope you guys enjoy. ---- 8 GB ---- Autocomplete for coding (like Cursor Tab) - huggingface.co/NexVeridian/ze… - huggingface.co/bartowski/zed-… Tool calling, assistant style - huggingface.co/nvidia/NVIDIA-… ---- 16 Gb ---- Here things get better: Multimodal - huggingface.co/Qwen/Qwen3.5-9B - huggingface.co/Tesslate/OmniC… - huggingface.co/unsloth/Qwen3.… ---- 24 GB ---- - The best model you can get (thanks Qwen) huggingface.co/Qwen/Qwen3.5-2… - Great model (strong agents) huggingface.co/nvidia/Nemotro… - Mine hehe huggingface.co/0xSero/Qwen-3.… I'm doing a weekly series

the original TurboQuant paper tested on A100 with models up to 8B. 6 days later, a bunch of strangers on the internet had it built and running on: - Apple Silicon M1 through M5 - NVIDIA 3080 Ti through DGX Spark Blackwell - AMD RX 6800 XT and 9070 - a 10-year-old Tesla P40 - an 8GB MacBook Air - models from 3.8B to 70B across 6 architecture families - 30+ independent testers along the way we found new optimizations the paper didn't cover and failure modes it didn't test. the fact that a loose group of people across the world can read a paper, build implementations from scratch, stress-test across hardware none of us could individually afford, and push the research further in under a week is genuinely one of the best things about this era. the tools and the community make it possible. open source is something else.

@jedmaczan This is dope ✨️✨️✨️ awaiting ...

I built a tiny-vllm in C++ and CUDA - paged attention - continuous batching - educational - 100% human-written™ And now I writing a course where you will build your own vLLM yourself. Still work in progress, I'll finish by the end of April. All for free ofc, just a GitHub repo

I've been seriously telling everyone around even these non tech novices who barely even use a laptop ... just get a macbook and experience prestine design in software and hardware getting an iPhone is subjective and optional but getting a mac isn't.

Its officially an epidemic right now. Every single day, another tech-illiterate grifter aka wanna be Steve Jobs starts yapping about their billion-dollar "unicorn" idea like they've just cracked the matrix. Like bro first just learn to center a div

The VC-funded subsidiary's $20/month plans are gradually coming to an end. You can barely make ends meet on the $20 Claude Code/Cursor subscription anymore. These multi-trillion parameter models were never economical to run at $20/month. You were just subsidized by VC money.

@asmah2107 isn't RadixAttention in SgLang more superior in terms of inference performance than PagedAttention in vLLM ?

Best KV cache out there to Speed up inference ? Anything better than vLLM (PagedAttention) ?

@etash_guha Congrats Etash ✨️✨️✨️

Career Update: I’m joining Anthropic on the pretraining team! Excited to learn from all the brilliant and creative people there. Let’s go train some models!

with the release of TurboQuant google just gave everyone a free performance upgrade on all of their existing hardware. and with this all of a sudden GPUs became marginally more economical to run inference on

for around $1200 a year you could argubly run a few mini inference projects

If this even hypothetically turns out to be true, we could: - one shot prompt and cure cancer - one shot prompt to rewrite every software on earth in direct machine code

@brahma_4u @YouTubeIndia @YouTube YouTube premium is the only way ! i pay youtube HAFTA just so that i dont see his face

Dear @YouTubeIndia @YouTube Kindly do something about this guy, he ruins every video.

non-tech founders right now

lets assume we theoretically scaled VJPEA architecture to 1 trillion parameters and fed it with the whole youtube corpus what possibly could JEPA understand about the world? would it understand concepts like gravity, wind, texture, viscosity etc .. all coalescence together ?

PEAK FLEX !!! 🫣🤯