Connito launches decentralized training network: splits model into experts, competitive updates improve whole
markjeffrey · x · 2026-08-05
Connito announces a decentralized training network built on Bittensor, with a different premise than systems like IOTA and Pluralis. Instead of splitting one training run across workers, it splits the model itself into individual experts.
Core mechanism:
- Workers bring their own expertise and techniques and compete to improve specific experts.
- Winning updates get integrated back into the shared model.
- Signal shows that training a particular selection of experts as a partial model can improve the performance of the whole system.
The Unsupervised team says they've obsessed over decentralized training for the past year and consider Connito one of the most interesting new approaches, having invested in it.
Related event: Connito Launches Decentralized MoE Training Network(2 posts)→
More from Venture
- Pika Launches API Club: Aggregates GenMedia Models at Up to 88% Lower Cost — dr_cintas · 2026-08-05
- Bain Capital VC Partner: AI-Era Software Firms Should Embrace 20% Growth and Cut Costs — aaref · 2026-08-05
- Gwern Retires from Writing to Launch Personal AI Models Startup Guardian Angel Inc. — NathanpmYoung · 2026-08-05
- Convex raises $57M Series B: AI agents are reshaping backend architecture choices — TAbrodi · 2026-08-05
- Airtable Deal Lesson: $500M ARR with 20% Growth Only Yields 2.5x Valuation — jonchu · 2026-08-05
- Monetizing AI Agents: Is Putting Workflows Behind a Paywall Viable? — j4ys0nj · 2026-08-05