Table of Contents
Inference Labs has agreed to hand over Bittensor Subnet 2 to former team member Dan, who will lead its development as DaniLabs. The transition shifts SN2’s focus from DSperse’s cryptographic verification work toward running AI models for developers and applications.
DaniLabs confirmed the takeover on Oct. 16, saying the subnet would focus on fast, affordable AI inference. Its service is in development, with no data retention among its priorities.
Dan, known online as @opendansor, is returning to a project he helped build. Inference Labs said he joined from the Opentensor Foundation more than a year ago and that his applied research helped scale DSperse.
From Verifying Computation to Serving Models
Under Inference Labs, SN2 worked on zero-knowledge machine learning, or zkML, using cryptographic proofs to verify AI computations. Inference Labs said the subnet’s developers and community produced more than 6 billion cryptographic proofs and developed ways to improve zkML performance.
The DSperse research paper, which lists Dan Ivanov among its authors, explains an approach to reducing the cost of that verification. Rather than converting an entire model into a cryptographic circuit, DSperse divides computation into segments, or "slices," and applies proofs to selected parts. Audits, replicated computation, or economic incentives help maintain consistency across the wider system. The framework allows verification to cover part or all of an inference pipeline, with proof boundaries aligned to the model’s structure.
DaniLabs’ announced direction changes the product emphasis to delivering inference itself, the process of running a trained model to generate an answer. Its stated goals for developers concern the time, cost, and data handling in each request.
No Data Retention, Speed, and Cost
DaniLabs, short for Distributed Accelerated Neural Intelligence, plans to process prompts and responses without storing them after a request. The service under development also aims to reduce the time to a useful response and lower inference costs for applications.
The planned software spans both the subnet and the service built around it. The team’s GitHub organization outlines four components: the Bittensor protocol with validators and miners, hosted backend services and APIs, a frontend platform, and ManagedVNetIR, a downloadable Rust-based miner kit. The team will add repository links as those projects become public.
