Erfan Miahi

Covenant-72B is the largest LLM ever pre-trained over the open internet. In the weeks after launch, Bittensor's TAO token nearly doubled, from about $180 to over $330, adding roughly $1.5 billion in market cap, with market coverage citing Covenant-72B as a key catalyst. The launch post drew 1.8 million views on X. On the All-In Podcast, Chamath Palihapitiya called the run "a pretty crazy technical accomplishment," and NVIDIA CEO Jensen Huang likened it to a modern Folding@home.

I was one of the main contributors on both the research and engineering teams behind it. The model has 72 billion parameters and was trained on about 1.1 trillion tokens by 70+ independent peers, each running 8 NVIDIA B200s, with no data center and no central cluster. It beats Meta's LLaMA-2-70B on MMLU (67.1 vs. 65.6) and ARC-Easy. Its chat version posts the best IFEval (64.7%) and MATH (26.3%) scores among the chat baselines we compared, and the run maintained about 94.5% compute utilization over commodity internet links.

As DeepR Analytics' founding ML research engineer, I helped build its automated trading pipeline from the ground up and turned reinforcement learning research into daily trading profit. I built a PyTorch pipeline with simulated market environments that trained over 1 million RL trading agents, adding more than $35 of daily PnL per asset per share.

I implemented state-of-the-art papers and designed my own RL algorithms across 10+ research projects, adding over $20 more of daily PnL per asset per share, and automated hyperparameter search (Optuna on Ray Tune). On the engineering side, I cut epoch time 9x and EC2 cost 4x by optimizing the data pipeline, simulators, evaluation, and async I/O, and moved training and backtesting to AWS ECS and Fargate so the stack scaled on demand across 100+ assets. I also raised test coverage to 85%, cutting production bugs by 45% across the ML stack and the C++ trading engine.

I led Stoa at Pluralis Research, the world's first RL post-training run powered by consumer Macs. Fourteen everyday Macs in Paris, Zurich, Dublin, and Toronto generated all the training experience for an 8B mixture-of-experts model (LFM2.5-8B-A1B), while a single NVIDIA B200 did the learning. The Macs had no cluster or cables between them and talked only through cloud storage, yet 14 of them were enough to keep the B200 fully busy.

The model learned an agentic research task: answering biomedical questions by searching 16 million paper abstracts over multiple rounds of tool calls. I designed the system so Macs could join, crash, or leave mid-run without breaking training: workers ran MLX inference and pulled PULSE weight updates about 110x smaller than full checkpoints. At the same time, DPPO token gating and a staleness cap kept off-policy training stable. I also built an inference engine from scratch, with a scheduler included, specialized for pipeline parallelism over wide-area networks (WAN-PP); it runs over 10x faster than existing engines on both time-to-first-token and tokens-per-second. Finally, I wrote the public launch post on the Pluralis blog.

PULSE is my method for making distributed RL post-training over 100x cheaper to network, and Hugging Face now ships it in TRL. After every RL step, trainers must push fresh weights to inference workers: for a 7B model in BF16, that is 14 GB per sync and about 20 Gbit/s of bandwidth to keep GPUs busy. I showed that about 99% of those weights don't change bit-for-bit between steps because of how BF16 precision interacts with RL learning rates, and I designed PULSE to send only the values that changed.

The result: 14 GB down to about 108 MB per sync, bandwidth needs cut from 20 Gbit/s to 0.2 Gbit/s, and bit-identical training. Hugging Face implemented the approach in TRL, and Fireworks cited the paper in its write-up on the RL infrastructure behind Cursor's Composer 2. I took PULSE from idea to paper to production, running it in both Grail and Stoa.

At Covenant AI, I led the development of Grail, one of the first open, decentralized networks for post-training RL on LLMs, which runs live on Bittensor. Anyone can contribute GPUs to generate rollouts; validators check the work, and a trainer learns only from verified data. To prevent cheating, we fingerprint hidden states at 4 bytes per token and check token probability traces, so contributors can't swap models or tamper with outputs.

I also led R&D on off-policy RLVR, maintaining stable training with 16 steps of policy staleness and fully asynchronous inference and training, and integrating PULSE for weight sync over the public internet. On a 1.5B Qwen model, Grail lifted MATH accuracy from 12.7% to 47.6% (nearly 4x), GSM8K from 57.9% to 72.2%, and AMC 2023 from 7.5% to 25%, closely matching an on-policy TRL baseline.