This week in the Cosmos: subnet dashboards complete, a new SDK launches, and SN9 goes to NeurIPS

Covering the latest developments at Macrocosmos.

Welcome to our roundup, where you can discover all about our recent updates, and learn more about what Macrocosmos has been up to.

Subnet 37 gets a dashboard

SN37, our fine-tuning subnet, has launched its dashboard. This gives a real-time view of our ongoing competition, allowing people to see where we’re situated in the LLM landscape. This is the fifth dashboard we’ve released, following both SN13 and SN1, released last month. Users can see miner model improvements over time, MMLU and MMLU Pro benchmarking results, and a live miner leaderboard.

This dashboard means every one of our subnets now has supporting visuals. Each release has confirmed the importance of having live and interactive data to serve our subnets’ users, and the ongoing effort to add transparency, legitimacy, and accessibility to the Bittensor ecosystem. Our SN37 dashboard exemplifies this by adding context to the community’s work and framing our current and previous efforts against benchmarks and timelines.

Atom, our SDK, has been released

We’ve released our own SDK, Atom, designed to enable rapid development within the ecosystem, with tools for miners, validators, and subnet builders. Currently, Atom can help to form a generic miner and validator backbone, as well as provide organic scoring tools (plus other chain-related tools). However, in the future, its utilities will expand.

This is being used by Macrocosmos, with all our subnets being compatible with it, but the SDK’s suitable for all subnets, even those that don’t belong to us. Atom is not a product, but more of a public good toolkit. Having run so many subnets, all with distinct use cases, architecture, and incentive mechanisms, we’ve learned a lot about what common challenges arise, and so we built this SDK to help alleviate them, and lower the barrier of entry to participate.

Subnet 9 team present at NeurIPS

SN9’s machine learning lead, Alan Aboudib, and Machine Learning Engineer, Rodrigo Lopez Portillo Alcocer presented at NeurIPS 2024 in Vancouver. They discussed federated learning, data-mixing, and the relationship between our pre-training subnet and our four others.

They broke down how SN9 overlaps with SN37, how our data-scraping subnet (SN13) can be used to improve models as they get trained, and the interconnectivity SN37 has with SN1 (Apex). Rather than their presentation simply being a showcase of pre-training on Bittensor, it was a display of how to grow a decentralized ecosystem and toolkit.