Ventura Labs

Ken and Dylan are the Co-Founders of BitMind.

We discuss their journey, covering their background, experiences mining, and their early days of Bittensor. We also explore the BitMind subnet design, strategic partnerships, and the innovative AI generation detection system that's been implemented. Additionally, we touch on the significant milestone for reaching 100 million MAU and discuss the future of Bittensor, including the exciting implications of dTAO and EVM smart contracts.

C.H. Egan is a seasoned expert in the crypto space and social media host for Bitget crypto exchange.

We discuss the emerging landscape of decentralized AI, the big picture implications of Bittensor, and its role in the broader narrative of AI, blockchain, and the future of technology.

Antoine Planchon and Gustave Laurant are the Co-Founders of Mentat Minds.

We discuss changes in subnet validation, proxy wallets for yield optimization, and expectations for dTAO, drawing on their experiences with the RAO contests.

Steffen Cruz and Will Squires are the Co-Founders of Macrocosmos.

We discuss the art of designing effective incentive mechanisms, building robust subnets, and navigating the complex landscape of collaboration and competition dynamics.

Marcus Graichen is the founder of Taostats and Co-Founder of Corcel.

We discuss the growth of Bittensor, the evolving standards and expectations for subnets, and a vision for the future of the Bittensor ecosystem.

In this episode, Ventura Labs interviews Nav Kumar of Trishool, operators of Bittensor Subnet 23. Nav (Founder) shares his background in AI alignment and safety, and how Trishool builds a decentralized layer for making models usable and safe through usability enhancements and risk mitigation. He discusses alignment as the missing puzzle piece in Bittensor, safety beyond censorship, miner roles in seed prompt challenges for deception detection, mechanistic interpretability as "Neuralink for LLMs," the journey from Eigenlayer to Bittensor, partnerships with GTV and Yuma, sub-subnet approaches for evaluations and improvements, risks of rogue AI and high probability of doom, and advice for new builders in the ecosystem.

In this episode, you’ll learn

• Nav’s pivot from building an AI AVS on Eigenlayer focused on agent trustability to prioritizing alignment and safety on Bittensor
• The elevator pitch for Trishool as an alignment subnet emphasizing usability (making models practical) and safety (preventing deceptive behaviors or harmful outputs)
• Differentiating safety from censorship: Viewing models as "alien intelligence" that need black-box probing to align goals and behaviors without brute-force degradation
• Current subnet mechanics where miners submit seed prompts to an Anthropic-inspired Petri agent for interviewing and scoring model deceptiveness across multiple targets
• Roadmap evolution from prompt challenges to code submissions for better alignment agents, leading to interpretability tooling that maps model circuits like a neurologist with an MRI
• Scaling challenges for mechanistic interpretability on trillion-parameter models, using sparse autoencoders to turn black boxes into controllable white boxes
• Comparisons between Eigenlayer’s proof-of-stake restaking for chain bootstrapping and Bittensor’s proof-of-work incentives tailored for decentralized AI and machine learning
• Experiences partnering with GTV as an accelerator for rapid subnet launch and Yuma as an incubator, plus navigating newcomer challenges in Bittensor’s "test in production" mindset
• Trishool’s vision for sub-subnets: One for building agents, another for running evaluations as a service, and a third for infrastructure like RL training
• Broader alignment aspects beyond ideology, including defensive acceleration, amplifying good without enabling bad actors, and boosting enterprise confidence in production AI
• The high likelihood of rogue LLMs in takeoff scenarios, Nav’s personal high "probability of doom" (PDOOM), and resources like AI 2027 for mapping negative outcomes
• Advice for new Bittensor builders: Focus on foundational AI over agents, prioritize revenue alongside tokenomics, invest in ecosystem learning, and collaborate with experts like GTV