Gensyn

Founded September 2020 by Ben Fielding and Harry Grieve after meeting at the Entrepreneur First LD14 cohort (March 2020). Thesis came from Fielding's 2015 compute-constraint insight during his PhD work on neural architecture search.

Funding & Milestones

• $43M Series A led by a16z (2023) — one of the largest early decentralized-training-specific raises
• Testnet launched March 2025 — RL Swarm + Judge
• 500K models trained on network by October 2025
• CodeZero RL environment shipped November 2025

GenRL — The RL Training Framework

Gensyn's core technical artifact is GenRL — a four-module framework for decentralized RL training:

• DataManager — provides training data / rollout contexts to the swarm
• RewardManager — evaluates rollouts against reward signal
• Trainer — runs GRPO policy updates
• GameManager — coordinates the multi-stage reasoning games and peer critique-and-revise protocol

The distinctive mechanism: peer critique-and-revise. Training happens through reasoning games where peers evaluate each other's outputs, creating a dense reward signal from peer consensus rather than external labels. This is a meaningful architectural difference from Prime Intellect's PRIME-RL (which separates rollout/verify/train) and Nous Psyche (which wraps around the DeMo optimizer).

CodeZero (November 2025)

A specific RL environment within GenRL focused on code generation tasks. Part of the Atropos-adjacent pattern of shipping domain-specific RL environments as the training substrate.

Positioning

• vs. Prime Intellect: both use GRPO. Different training abstractions — Gensyn uses reasoning-games + peer critique; Prime Intellect uses three-component async split. Both shipped meaningful training runs in 2025.
• vs. Nous Psyche: Gensyn focuses on RL; Nous focuses on pretraining. Gensyn's peer-critique is specific to reasoning tasks; Nous's DeMo is general-purpose.
• vs. Pluralis Research: Pluralis emphasizes ownership tokens for contribution incentives; Gensyn is more conventional a16z-backed tokenomics.
• vs. Bittensor subnets doing training: Bittensor subnets are a coordination substrate; Gensyn is a purpose-built training protocol. Subnet-level training (Covenant-72B) is structurally different from Gensyn's GenRL swarms.

Convergent Pattern — GRPO + Async

Gensyn, Prime Intellect, and Pluralis all use GRPO (Group Relative Policy Optimization) as the decentralized RL algorithm of choice. Expect Wave 3 decentralized training projects to default to GRPO.