🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

we're hiring for our engineering team Prime Intellect right now. primarily looking for builders in rust, python, typescript and more

⢀⣀⣀⣀⣀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣀⡀⠀⠀ ⠘⢿⣿⡿⠿⢿⣦⡐⠈⡀⢂⡖⠿⠻⢽⣿⣿⠃ ⠀⠈⣿⣿⣮⠂⡝⡻⣆⡰⣻⡣⠜⡧⢾⣞⠁⠀ ⠀⠀⠙⠛⠿⢶⡤⢉⢮⠝⠢⢠⡵⠾⠛⠁⠀⠀ ⠀⠀⠀⣠⣾⣷⡰⠥⣪⡞⡬⡔⡸⣳⡄⠀⠀⠀ ⠀⠀⠀⠈⢿⣿⣾⢷⡋⠙⣶⣷⣟⣯⠃⠀⠀⠀ ⠀⠀⠀⠀⠈⠙⠛⠋⠀⠀⠉⠛⠋⠁

Awesome work by Mario Sieg to accelerate quantization of pseudo-gradients in decentralized training settings like DiLoCo - already integrated in pccl (prime collective communication library)

prime intellect mentioned

Reppin the homies companies at the gym

I endorse this message 🫡

Average Prime Intellect meeting

will brown cooking up some amazing async RL things! dev branch reveals async off-policy steps, which can easily give you 2-3x because of higher GPU utilization👀

i'm teaming up with Kyle Corbitt from openpipe to teach a class about agents + RL :) we'll be teaching the class on Maven 🏛 starting june 16. as far as we know, this is the first course of its kind anywhere to bridge RL + LLM agents, and we’re really excited to share some of our

fun fact: @primeintellect has around 20 employees total everyone is exceptional at what they do. you have a lot of autonomy, and that comes with a lot of responsibility. we're hiring, but not rapidly. we want someone really, really good for this role. sound like fun?

co-led a session with the awesome avwtr at Edge Esmeralda yesterday on what we claim to be the ultimate scientific bottleneck- our own miscalibrated conceptions of what science even is an example this is how we've reduced words/concepts to v narrow meanings. people can

rocm-smi coming to a heterogeneous synthetic data generation run near you soon

another easy to generate verifiable task: ask an LLM to generate a JSON that adheres to a very complex (LLM-generated) pydantic model Coming soon to the most diverse RL dataset out there, along with many other tasks that are not math and coding

prime intellect masterplan

A 256x256 matrix memory at 2048 tokens of sequence length can achieve 98.6% retrieval accuracy. With random embeddings, accuracy would saturate at 67% and with optimized embeddigs at 80%. Combined with a 4-layer LSTM decoder and a 1-layer store gate, the accuracy increases to

Seems like no one saw this either, scraping arxiv manually seems to be the way. Pretty cool paper on rl for creative writing on Qwen3 32B base, and most interestingly it's one author from the Star Writing Team (haven't heard of them). They seem to have access to the 32B base tho

ai tpot is missing lots of banger RL arxiv releases every week but Grad isn't. we gotta post the weekly docs he creates for our paper reading group somewhere

async RL is faster that synchronous counterpart. this might be the first time in ML history where an algorithm is naturally async at scale. we realized two things at prime 6 months ago: * RL will be as much compute intensive as pretraining, pushing frontier capability * for