synapz_org

3 ms·

1. 1.▲
   Fault tolerance in low-bandwidth model parallelism(tplr.ai)

   1 points by synapz_org 1mo ago|0 comments
2. 2.▲
   Globally distributed pre-training across 48 A100s at $0.12 per million tokens(tplr.ai)

   2 points by synapz_org 2mo ago|0 comments
3. 3.▲
   PULSELoCo: 17x less trainer-to-trainer bandwidth in distributed RL post-training(arxiv.org)

   3 points by synapz_org 5mo ago|0 comments
4. 4.▲
   Autonomous RL Fine-Tuning on Ephemeral GPUs: Extending Karpathy's Autoresearch(templarresearch.substack.com)

   5 points by synapz_org 7mo ago|2 comments
5. 5.▲
   by synapz_org 8mo ago

   Author here. Some context on what this is and why it matters. Covenant-72B is the largest language model pre-trained through fully permissionless, decentralized coordination. 72 billion parameters, approximately 1.1 trillion tokens, trained
6. 6.▲
   Pulse: Decentralized RL training centralized speed (100x weight sync reduction)(arxiv.org)

   2 points by synapz_org 9mo ago|0 comments
7. 7.▲
   SparseLoCo: Communication-Efficient LLM Training(arxiv.org)

   19 points by synapz_org 1y ago|1 comments
8. 8.▲
   by synapz_org 1y ago

   Paper: https://arxiv.org/abs/2508.15706 Code: https://github.com/tplr-ai/SparseLoCo Templar AI has developed SparseLoCo, a distributed training algorithm that achieves extreme compression ratios (