A team at Novelty Search, led by @trishoolai, has announced the development of a 0.8B guard model that is approaching state-of-the-art (SOTA) performance and is significantly smaller than Alibaba’s leading model. This advancement is part of their efforts in productizing Constitutional AI, making them one of the few teams outside of Anthropic in this niche. The model development benefits from competitive jailbreaks generated by Bittensor miners, which play a crucial role in identifying and mitigating vulnerabilities through adversarial prompts, enhancing the security of their AI systems.

SN23: SN23, also known as Trishool, is a Bittensor subnet focused on developing state-of-the-art AI guard models through decentralized red-teaming. It uses competitive inputs from miners to identify vulnerabilities and harden models like Halo. The team positions itself as the primary effort productizing Constitutional AI outside of Anthropic.
Anthropic: Anthropic is the AI research company known for developing Constitutional AI methods to align and safeguard large language models. The news highlights SN23 as the only other team actively productizing similar Constitutional AI techniques for guard models. This underscores emerging competition in decentralized AI safety approaches.
Bittensor: Bittensor is a decentralized network that incentivizes the production of AI-related digital commodities through a system of subnets, miners, and validators. Subnets like SN23 leverage miner competition for tasks such as generating jailbreaks to strengthen AI guard models. Recent community discussions emphasize its role in advancing decentralized AI red-teaming.
Trishoolai: Trishoolai is the X account and team behind Trishool (SN23), a Bittensor subnet dedicated to AI alignment and guard model development. Operated by Astroware AI with founder involvement, it focuses on using miner-driven adversarial testing to improve model robustness. The account recently discussed these efforts on the Novelty Search episode.
Const Reborn: Const Reborn, also known as Jake Steeves, is the host of the Novelty Search podcast and a core community member associated with the Opentensor Foundation. The account focuses on building neurons and advancing Bittensor ecosystem discussions. It recently hosted the episode featuring Trishoolai and SN23’s Constitutional AI updates.
Novelty Search: Novelty Search is the weekly Bittensor community podcast and open call hosted in the Bittensor Discord, organized by the Opentensor Foundation. It features in-depth discussions with subnet teams on their technical developments and ecosystem contributions. The latest episode covered SN23’s work on Constitutional AI guard models.

AI Red-Teaming: Decentralized subnets in networks like Bittensor are actively using miner competition to generate adversarial prompts that identify and mitigate vulnerabilities in AI systems.
Guard Model Development: Teams are advancing open guard models designed to act as gateways for securing AI applications against unsafe outputs and prompt-based attacks.