ny

[Evis Drenova @evisdrenova]

one of the hardest things about building a company is that nobody wants to work with you until you've proven that you're worth working with

and the easiest way to do that is to have already worked with someone worth working with

it can be very frustrating

[ny @nunykotech]

@saraknggg I've built voice-routing AI systems, worked on RL training for Qwen, and developed end-to-end AI applications. dmed you and would love to learn more

[ny @nunykotech]

@Hackwithseries would love an invite

[Michael Yu @_michael_yu_]

I think Stanford's CS336 is one of the best materials to learn LLMs from scratch. I just finished working through it (tokenizer, transformer, triton kernels, distributed training, scaling laws, alignment) and highly recommend it. 🧵 Some notes:

At smaller sequence lengths like 1024, the FFN dominates, not attention, which is a big part of why MoE goes on the FFN. At 16k ctx, attention usually takes over.

Naive DDP on 2 gpus: 54% of each step was gradient sync. overlapping comms with backward: 3%. ZeRO-1 sharding cut peak memory 68.7 -> 47.7 GB.

GRPO: base model with the r1-zero prompt got 0/1300 correct for gsm8k. it gets 224/1300 correct with the 3-shot prompt. RL is most effective when it amplifies signal that's already there.

🔗 check it out: https://t.co/xXOOqAT0RN

[teo @teodorio]

the thing is taste is literally the opposite of training on existing patterns and arriving at the average. it's actually the exact opposite, but again most people in tech don't have it

[ny @nunykotech]

@LoopandPixels 👋

[Jasper Lu @lu__jasper]

If you’re just now starting to learn how to do rl over llms, I'd recommend jumping ahead to grpo and then working backwards from there to ppo, etc once you understand it deeply.

A lot of online courses and guides like to work through rl techniques chronologically (many courses still start from sarsa, q learning and all that), but grpo is just a much more intuitive starting point IMO since it has less moving parts.

[ny @nunykotech]

@Maaztwts in

[ny @nunykotech]

@arlanr @nozomioai interested, dming you!

[ny @nunykotech]

@kickingkeys insane. could you share some code from the training process? been trying to recreate this

[ny @nunykotech]

@samthekorean would love to connect!

[ny @nunykotech]

@richardzphotoz dmed you !

[ny @nunykotech]

@SuhailKakar @Polymarket interested, sent a dm

[ny @nunykotech]

@0xSero interested

[ny @nunykotech]

@0xSero interested, can't dm

[ny @nunykotech]

@connor1 https://t.co/ranpWUk5ss

[ny @nunykotech]

@benaratame @arlanr interested !

[ny @nunykotech]

@benaratame @arlanr interested !

[ny @nunykotech]

@16vchq yes, need a cofounder to collaborate with

[ny @nunykotech]

@lekondonyc Lekondo