Gas Dank Cannabis

(1/4) Trust is earned over time, but conviction is shown in follow-on capital. We are proud to announce a $50k strategic OTC deal for SN54 alpha tokens. This investment comes from a private partner who was part of our initial $900k outside raise last year. They saw what we built before Bittensor. Now, they are doubling down on our future within it.

🚀 The last 8 months at Ridges have been a blur. Since May, we’ve shipped - Open source agent competitions topping SWE-Bench with OSS models - Mine with Google login, no keys necessary - Mixed evaluations - agents running on both fix and create style software engineering benchmarks - Ridges Beta v1 👀 - Lots of under-the-hood breakthroughs like runner sandboxing, spam detection, scaling evals beyond 1k per day And much, much more. Each of these moments have been pivotal in setting up for the next - but this is just the start, and we’re super excited to show you what’s next. The biggest news in Ridges history drops in two weeks. Mark your calendars. 📅

2025 was a year of technical milestones for DSTRBTD. We’re proud of how far we’ve come; here are some of our favourite highlights of the year: • May: First team to integrate DeepMind’s DiLoCo strategy into a Bittensor subnet. • August: Became only the 2nd team globally to train a 1B LLM in a fully open consensus setting (trustless, permission-less, distributed & decentralized). • October: Unlocked 30B parameter scaling via multi-GPU mining and began training our 4B model on main-net. • December: Launched Mechanism 1, a winner-takes-all incentive engine for producing SOTA distributed training strategies. Despite these wins, we are currently the lowest priced subnet on the network and the leading candidate for de-registration. We’ve sourced feedback from stakeholders and the wider bittensor community and here is what we believe we got wrong last year: 1) We were slow to adapt to Dynamic Tao (dTAO):  The shift to dTAO brought legal and operational complexities that consumed too much of of our time in Q2. We prioritized compliance over visibility. We now have a sustainable structure that frees us to focus entirely on building. 2) We assumed the tech would speak for itself:  In distributed training, the moonshot goal will take time to realize. We've learned that along the way we need to distill incremental technical wins into digestible milestones that relate to our long term vision. 3) We failed to capture the non-technical mindshare:  We focused on the "how" and ignored the "so what." We were so immersed in the technical challenges that we didn't prioritize explaining to the average retail investor why these achievements mattered. We are now committed to crafting bespoke, accessible content on a weekly basis. We're confident that if we survive this de-registration battle we can use the lessons we've learnt in 2025 to move faster and communicate more clearly as we try and achieve our objectives for 2026: - SOTA Distributed Training Strategies: Incentivising distributed training strategies on Mechanism 1 that outperform the current SOTA. - Decentralised Competitiveness: Proving Mechanism 0 can create decentralized LLMs that match the outputs of models like GPT, Claude and Gemini. - Value Distribution: Building systems to capture the value of our outputs and distribute it back to our token holders. We wholeheartedly believe in creating AI for the many, not the few. We’re ready to prove that the most undervalued subnet on the network is also the most technically ambitious. Wish us luck!

Introducing DSTRBTD’s Subnet Mechanism 1 - DisTrOpZ github.com/dstrbtd/DisTrOpZ Inspired by @exolabs's gym, this mechanism asks miners to submit a competitive Distributed Training Strategy which includes a pairing of a Communication Class and an Optimizer Class. Mechanism 1 will be a winner-takes-all competition where 100% of the emissions will flow to the strategy that achieves: • highest throughput • lowest communication cost • greatest loss reduction The objective is to incentivize the submission of strategies that can outperform existing approaches such as SparseLoCo (@tplr_ai), DiLoCo (@GoogleDeepMind), DeMo (@NousResearch), and Muon (@benjamintherien et al.) and to ultimately integrate the winning strategy into Mechanism 0 to enable training larger models faster with substantially lower bandwidth requirements. Emissions will remain at 0% initially as we refine our validation process and benchmark these baselines against it. Once our evaluation framework demonstrates consistent, repeatable results, we’ll begin incrementally increasing emissions to this mechanism.