Saatvik

bioinformatics & local AI
17
all views reflect my personal opinions

[Saatvik @saatvikv_]

@NolmerPortion What do you mean? Does the CMP 90 potentially have more GDDR6 on the actual board?

[Saatvik @saatvikv_]

@PeasantSmith how do they ablation methods work?

[Saatvik @saatvikv_]

@Morrowmake @borgerntendies drop the seller

[Saatvik @saatvikv_]

@MuhsinunC i think that’s interesting. something I ponder is an LLM being to tune its own weights in a sense and “prune” certain associations/things it doesn’t like how the human brain is able to prune its own memories or neurons. more of a thought than something that could happen though

[Saatvik @saatvikv_]

@needmorevram @ItsmeAjayKV don’t think so 💔; only to trusted partners

[Saatvik @saatvikv_]

@axitek @unsloth 512 token decode for the measurement? will do rn on my 3xv100s (16) + 3060

[Saatvik @saatvikv_]

Thanks Peasant! Super happy to have found the new optimization trick for GLM :)

[Peasant Smith @PeasantSmith]

It's been a while since I released twice in a day, but if I do hold them back there is just too much stuff to go into q single release which creep up the amount of testing required

So...

Maya v1.1.52 is out 🚀

⚡ GLM decodes about 8-9% faster on NVIDIA GPUs, with the same answers
🔴 AMD on Windows: setup finds your card on its own, failed updates tell you exactly how to fix them, and one engine error no longer takes the server down
🧠 DeepSeek: a wrong reasoning_effort value in the config now gets a clear message in seconds, instead of failing after a full model load
🤝 Community additions: experimental AMD MI50 support, faster DeepSeek on RDNA 3, and more

https://t.co/lTy1AXl8gY

[Saatvik @saatvikv_]

@PeasantSmith benchmarks of the the 73 gb quant?

[Saatvik @saatvikv_]

@ItsmeAjayKV i assume so or else why would you get a 5090 when u can get a pro6k blackwell

[Saatvik @saatvikv_]

node 1 is online 👀

[Rohan Arun @RohanArun]

The Auto-research harness now supports Antigravity thanks to @saatvikv_ ! Shockingly, Google models are very good at auto-research/RSI. 🤯

I'm working on a cool benchmark that rewinds frontier research repos like Integer Mult, QRH, and these muse kernels to test all the models and harnesses and find the fastest and cheapest combinations.

We tested Gemini 3.8 flash to hillclimb very fast and it even beat Opus 5.5 and Astra to some matched goals.

I suspect Google has trained on code for RSI, and this is a spiky feature their models are very good at. Argon wouldn't allow it and kept stopping for security.

Gemini 3.8 flash just climbed 26 ranks on the Open Kernel Challenge overall leaderboard to #70.

I'm working on benchmarking the cost/breakthrough and will publish soon.

I also published more CUDA optimization best practices as baked in prompts, and multiple people have used it to get on the leaderboard for the Muse Open Kernel Challenge now:

https://t.co/JZepY0J1p9

[Saatvik @saatvikv_]

@RohanArun Wow thanks for the feature!! Have been a day 1 believe of gemini, any advice since i’m running swarms of 3.8 flash agents!

[shaun @Biggest]

Distillation is a fine thing. Can we stop going through a new cycle of "omg they distilled!!" every time a Chinese model is released?

It's why Qwen4 took longer than expected, and based purely on my own tests, it has been well worth it. The new checkpoints are real improvements.

[Saatvik @saatvikv_]

@PeasantSmith thanks so much peasant!!!

[Lior Pachter @lpachter]

edgePython v0.5.0 released at https://t.co/NOMbVAEfzw (`pip install edgepython`)

Matches the forthcoming edgeR 5.0 with several extra features as well as some edgeR bugs avoided.

Major speedups (4-6x faster, camera >100x), I/O improvements, and proper support for kallisto.

[Saatvik @saatvikv_]

@lpachter Thanks for accepting the PR and letting me fix bugs Dr. Pachter!

[Saatvik @saatvikv_]

@PeasantSmith Definitely the >25 TPS so it’s usable for defent work and not just overnight

[Serio_ai @Multi_Serio_Ai]

아니 이런것까지 다 오픈소스로 풀면 무엇이 남나요?

[Saatvik @saatvikv_]

@unbug Niko will take care of it

[unbug @unbug]

We just hit 300 members today—woke up to 100+ new people joining overnight. Come join these top DIY hackers using extreme optimization to make old hardware run Qwen3.8-Flash and GLM5.3-Flash. 🚀https://t.co/5W0ilKD1QY