Amit Prakash

@FieroTy @OpenAI congrats ! Hoping for more rigorous evals in the healthcare field

Hi twitter friends, I’m joining @OpenAI to work on applied evals! A few years ago my day job was studying microbes Somehow this feels like a natural progression 😄

@noah_vandal modal since i get both cpu and gpu sandboxes. heard daytona is good, but i havent tried that

what is the best sandbox provider right now?

@ashebytes @Yuchenj_UW even if he uses 8xH100 for 30 days it's going to be around 8,500$ another 8xH100 for inference for any open weights model (minimax models) so approximately < 20,000$ probably ? under assumption: gpu's are running 24/7 for 30 days and local inference.

@Yuchenj_UW how much do you think running your autoresearch would cost?

OpenAI just dropped a training challenge: Train a <16MB language model in 10 minutes on 8×H100s and minimize held-out loss on a fixed FineWeb dataset. Basically NanoGPT Speedrun. They’re sponsoring $1M in compute. I can summon my autoresearch army to win it… if I have time.

@WorkshopLabs congrats on the release! time to read the blog 😋

Letting a provider see all your data is the price of admission for AI. We're changing that. Introducing Silo, the first private post-training and inference stack for frontier models, with hardware-level guarantees that we can’t see your data. Privacy without compromises. 🧵

@thkostolansky @PrimeIntellect

@PrimeIntellect were the talks recorded by any chance??

Prime Intellect Day Saturday — March 14 (Pi Day) 4-11pm — The Melody of SF Talks on frontier model training, agentic RL, computer use, coding agents, diffusion LLMs, automating AI & science, robotics — and why the singularity will not be singular.

@jino_rohit his gpu mode lecture on CuTe is a goldmine! youtube.com/watch?v=vzUhbD…

Cris Cecka from NVIDIA, the creator of CuTe, liked my blog on CUDA layouts. You should read it too! 1. primer to understanding layouts - jino-rohit.github.io/blogs/02_layou… 2. layout algebra in CuTe - jino-rohit.github.io/blogs/03_layou…

@vincentweisser @ndrewpignanelli Hopefully prime-code soon 👀 actually i had a chance to speak with one of the claude code MTS. They mentioned their approach is currently different but they said they might have plans to roll out officially if it works well :) [ conversation on Feb 15th ]

@ndrewpignanelli we might outaccelerate them 🫡

there's just no way anthropic doesnt drop RLM or continuous improvement soon with all the chatter im seeing on the topic

@_xjdr amazing research drop🔥!

Read the blog by @WorkshopLabs and worked on this while reading. Hope it helps out someone who plans to read the blog :) Link : #json=FY0SyC8j8hqCwCA-46U2C,LYa0t4a1eclgg_mZqvb0XA" target="_blank" rel="nofollow noopener">excalidraw.com/#json=FY0SyC8j… Any suggestions or improvements lmk :)

📌 OpenAI Codex Meetup

another banger added to reading list for the weekend 💯

this is good advice btw, if you wanna deeply learn kernel optimization on nvidia GPUs just naturally follow the arch progression

some interesting blogs I have been reading recently : > Random Labs Slate : randomlabs.ai/blog/slate > Unsloth + Nemotron Blog : unsloth.ai/blog/rl-enviro… > cutlass : kapilsharma.dev/posts/learn-cu… by @kapil_sh_ > cute dsl and layouts

5pm: “Claude usage limit reached. Your limit will reset at 7pm..” Me from 5pm to 6:59pm:

@ellen_in_sf @Apple @mem0ai @taranjeetio @openclaw if possible, try using helios by @snwy_me or hermes agent by Nous Research.

just got a new launched @Apple macbook m5 max on my new role at @mem0ai Thanks @taranjeetio should i gave full control to run @openclaw or run 20 claude code terminal to claudemax my work

@karpathy this is exactly how I imagine a "bigger IDE", everything I need at my disposal with minimal time spent switching between things High hopes for vision glasses to solve this

@amit05prakash I wanted to buy a bigger monitor and discovered that others had the same idea

Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming.

@eliebakouch genuinely your tweets have taught me so much about model training! Enjoy your break and I really hope you keep up the amazing work! 🥳

today is my last day at hugging face feeling really grateful to have worked with such an amazing team and learned so much along the way. i’m proud of what we accomplished together, especially the smollm series. building that project from scratch, putting so much into it, and getting to iterate on a model and training recipe that pushed the frontier for its size was really rewarding i hope i was able to play a part in making model training more accessible and in pushing the open model ecosystem forward. i’m also very thankful to hf for giving me the chance to share my passion for llm research, especially here, and to connect with so many awesome people things can get quite intense in this field, but i’m still very excited about the next challenges and about the good this technology can do but first, taking a few weeks break :)

@victormustar this + karpathy's auto research packaged together can probably serve as a great automated research gui

llama.cpp UI now has MCP support 🔥 Working super well, make sure you are up to date: ``` brew install llama.cpp ``` then ``` llama-server --webui-mcp-proxy ```