🔥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

Reka AI has launched Reka Flash 3, a new open source 21B parameter reasoning model - the highest scoring model of its size Key details: ➤ Artificial Analysis Intelligence Index of 47, beating almost all non-reasoning models with just 21B total parameters ➤ Stronger than all

what simple tricks are there to improve vanilla LoRA training for LLMs (especially in DPO)? LoRA+ by Soufiane Hayou is easy to try- scale up the B gradients. What about simple ways to better initialize the matrices? PiSSA looks easy to try (bad name..)

My go-to test for any ‘real-time voice AI’ demo: try to interrupt me while I’m talking. So far, none can.

git checkout -b matthen/qwen3

false summit (a 3D fractal created from Weierstrass functions + rendered on GPU)

very cool application of RL tuning- fun to see the reward hacking and how it was solved