Grail (81) RL improvements coming from miner inferences under incentive.

The Grail calls. Mentat Minds makes it easy for TAO holders to answer—one click to join the quest for decentralized AI alignment across all covenant subnets.

Pre-training is university. Post-training is mastery. We hunt the refinement that turns knowledge into wisdom. Updates from across the covenant community: youtu.be/3adFWQbv2BM

You can contribute to this RL training pipeline by producing rollouts. Bittensor subnet 81. I'll be working with this team to help them monetize these models up through affine. The environments use the Affinetes standard so it plugs nicely on basilica. You can also mine with

Everything about grail-v0 is open. The code. The training logs. The architecture decisions. covenant trained Qwen2.5-1.5B to 72% on GSM8K using a network of permissionless miners. Erfan Miahi explains how and what's next: templarresearch.substack.com/p/grail-v0-how…

Launched the fully async trainer on grail mainnet, training a 7B model 🚀 Training now runs continuously from a replay buffer! No waiting on communication. This unlocked 2k+ context and let us scale from 1.5B → 7B models and even bigger later on.

Somebody needs to stop him. He’s gone mad. 1.5 B ✅ 7 B < in progress> Loading super intelligence grail

Grail v0: decentralized reinforcement learning is here. Network of miners generating rollouts, validators with cryptographic proofs (grail proofs), trainers running GRPO. Open source, verifiable training.

The quest continues. San Francisco, March 30-31. Distributed State on stage at Exploit Summit.

Grail's first revelation. PULSE: decentralized reinforcement learning at centralized speeds.

Révolution dans le post-training décentralisé sur Bittensor ⚡ grail (branche de covenant) brise le mur de la bande passante en dévoilant PULSE. Un protocole de synchro permettant du post-training RL permissionless et décentralisé, à une vitesse comparable au

ตัวนี้ปัจจุบันใช้งานจริงบน grail (Subnet 81 ของ Bittensor) ซึ่งเป็น decentralized RL / post-training network ที่ Covenant AI covenant คือ miner + validator ร่วมกัน train model ใหญ่ ๆ แบบ distributed PULSE ทำให้ sync weight ระหว่าง node แต่ละตัวเร็วมาก จน decentralized

One hour. Erfan Miahi breaks down PULSE and what it means for Grail's future. x.com/i/spaces/1YqKD…

x.com/i/article/2021…

Yannic Kilcher's Yannic Kilcher 🇸🇨 paper reading group is one of the most respected ML discussion communities, with over 20,000 members and regular deep-dives into new research. Erfan Miahi will be walking through the full PULSE paper this Saturday at 6 PM UTC. Join the discussion on

Erfan presented covenant's PULSE paper at Yannic Kilcher 🇸🇨's Saturday paper discussion this weekend, walking the audience through why RL weight updates are 99% sparse under BF16 and how PULSE exploits that to achieve 100x bandwidth reduction, with complete losslessness, making RL

This Friday on TGIF, Distributed State covers what's happening across the full Covenant ecosystem. For Grail, that means updates on where we are after PULSE and what's next for decentralized RL post-training. Join us in 2 hours.

The TGIF 26 recap is live. Distributed State talks about grail after Erfan Miahi's PULSE presentation at Yannic Kilcher's reading group. PULSE means decentralized post-training is approaching data center speed, and Sam says Grail is actually closer to producing a strong model