Chidhambararajan R (a.k.a Chidha)

@aidenybai Bro it should be the opposite

introducing cli-to-server turn any CLI into an API server you can call npm i cli-to-server

@MatternJustus Do you have any hiring page? I am interested

Currently in BLR to help build our model training team! With our recent momentum, it's clear to me that we have a shot at building a frontier research org in India. For this, we need more research talent If you want to work on post-training 100B+ param models, I'd love to meet

@adxtyahq Bro that's the worst c curve ball I have seen

just found out that arr[i] and i[arr] both compile in C++ and return the same result

@xieenze_jr @yitongli165665 My bad just now saw the prefix.. amazing work!

@xieenze_jr @yitongli165665 How did you fix the precision mismatch problem in rl rollouts

🚀 Excited to share Sol-RL (Speed-of-Light RL) — a new high-efficiency preference alignment method for Diffusion RL, primarily developed by first author Yitong Li (@yitongli165665 )! It uses a smart two-stage design: FP4 for ultra-fast massive rollouts and quick filtering of high-contrast samples, followed by BF16 high-precision optimization on the selected data only. Achieves up to 4.64× faster convergence while delivering better alignment results on SANA, FLUX.1 & SD3.5-L. 📄 Paper: arxiv.org/abs/2604.06916 Let’s push Diffusion RL forward together! 🔥

@thsottiaux Make a 50 bucks one too!

We did it, say hi to the $100 plan! It should be the sweet spot for a ton of you. It comes with a ton of codex usage. And yes we are resetting the limits again too as I mentioned yesterday. Let’s keep building!

@interaction >says no signup >Opens the page has a sign up Wt

Starting today, personal superintelligence is just one tap away. No download, no signup. Text Poke for free now: Poke.com 🌴 — 0:00 – What's Poke? 0:50 – Introducing Poke Recipes 1:25 –  Create a Recipe in 10 seconds 1:43 – Earn on Poke 2:44 – Build with npx poke 12:58 – Recap 13:36 – Parisian Love

There are a couple of issues with your proposal I remember attempting qkv attention pruning in my in home experiments a while ago. I did PCA on the qkv vectors with calibration data, this often shows high compressibility. (Similar to your approach) The results where something like: Top 4 dims: 95% energy Top 16 dim: 98% energy Top 32 dim: 99% energy And this goes quite exponentially Just to be void of biases from the calibration data, did some SVD analysis on the qkv weight matricies and it showcased way higher dim requirements for the same vectors for maintaining similar energy Then decided to drop the idea because: Attention qkv reduction is analogous to vector db retrieval. (In case of vector db the compressed vector's recall with the original should be atleast 99.9% for it to be acceptable). Bringing the same idealogy to qkv matricies the energy retained should atleast be 99% or 99.9% for it be close to lossless And this requirement meant higher number of dims preserved post the compression logic which can potentially reduce the wins You did mention in your paper that you calibrated on wiki test and the perplexity drop was minimal. (But perplexity often doesn't paint a good picture of performance drops) Google's turboquant having close to lossless perf (in 3-3.5bits) makes cause many papers have mentioned these models can't store more than 3bits of information per float A bit more emphasis on benchmarks again I made this comment in LinkedIn too, just putting it again here incase you are not active there

@hrshc7 Bro need not put ChatGPTed summaries.. the community would love simple explanations in your own terms

Everyone’s busy arguing about “which model is best”… meanwhile a tiny 4B model is out here quietly doing the job 👀 Just saw Gemma-4-E4B casually identify sea animals from images in a single session, no drama, no insane setup, just working. And that’s the part people are missing. We’ve been conditioned to think: bigger = smarter cloud = necessary expensive = better But this flips all of that. A small model, running locally, handling vision tasks end to end… without begging for APIs or burning money per request. Not perfect, not magical. But good enough to be useful and that’s way more important. Because once something is: fast private and basically free …it stops being a “tool” and starts becoming part of your daily workflow. The shift isn’t loud. It’s practical, and already happening. But sure… keep debating benchmarks while this runs on someone’s laptop 🚀

@xai Bro cutout of this slop and focus on coding agents bro

Introducing Quality mode on Grok Imagine – powered by our most advanced image generation model. Quality mode gives you enhanced details, stronger text rendering, and higher levels of creative control. Now available on web and mobile. Try it at grok.com/imagine

@lateinteraction @prithivida Yeah that was the db company I was talking about

@prithivida @TheSeriousProg there's for example Mixedbread dot com

overwhelming evidence for late interaction / multi-vector models yet again :-) > even after finetuning, single-vector models lag far behind multi-vector embeddings, which achieve significant performance gains and exhibit greater robustness to catastrophic forgetting.

@lateinteraction Untill recent progress from a particular db company wasn't the ram demand and storage demand for multivector embeddings pretty high? Yeah I agree on the transformer vs Convnet example though

@TheSeriousProg It’s kind of like saying in 2026 that “transformers like BERT are too expensive so we use a ConvNet”. It’s an incorrect/lazy excuse.

@ChShersh Well you don't have any problem uploading it to claude endpoints though..

Wait, am I the only one who thought all this time that Claude Code was already open-sourced? Why would you feed all your private data into a closed-source project?

@DirhousssiAmine I mean grpo is opd on policy distillation

@TheSeriousProg This is GRPO not distillation :)

Been going down a massive rabbit hole with numerical stability in RL training lately.🕵️‍♂️🕵️ Take a look at these two GRPO sanity runs. Exact same model, identical task. One climbs perfectly, the other completely flatlines. The only difference? The dead run is in bf16, the successful one is fp32. What do you think the problem is with these runs? Drop your best guesses below !

@DirhousssiAmine Because sampling model policy and distillation model policy are closely aligned in similar precisions. Which is not the case with mismatched quantizations.

@TheSeriousProg haha yes, but why ?

@elonmusk Musk bro make a coding LLM way more money over there

Try Grok Imagine

@PriyanshuP1405 I don't even find my 16gb gpu sufficient

If you're a Machine learning engineer or a quant researcher and find 16 gb to be sufficient stay away from me

@Yuchenj_UW Imo in the future presumable software engineers will be valued more for number of average runs per line of code which they right This is just going way tok much in the hype loop

If you had two software engineering offers: > One pays you $500k/year salary, but covers zero LLM tokens. > One pays you $400k/year salary, but gives you $500/day free LLM tokens. Which one are you taking?