crux

The @IOTA_SN9 Simulator competition: 22k+ submissions, ~29% faster epoch times, R&D insights for distributed AI training. Two months ago we launched a competition to minimise epoch completion time inside a digital twin of the IOTA network by optimising activation routing and balancing. The goal: to improve speed and efficiency within our distributed training network. The result: 22,967 submissions across 57 rounds. Epoch times are now ~29% faster on average compared to the start of the competition, and up to 39% faster on some network configurations. Our subnets are an ecosystem - the IOTA Simulator is the clearest example: insights from miners feed directly into how IOTA engineers iterate, both for current participants and for future clients once we productise. Several R&D insights have arisen. Let's isolate one in particular. The competition highlighted a specific architectural challenge: downstream congestion dominates throughput more than raw processing speed does. Routing too many activations to the fastest miner doesn't solve the problem, as doing so fills its queues and slows it down overall. Top submissions converged on the same fix: track how full each miner's downstream queues were getting, and skip the ones building up backlog, even if they were nominally the fastest pipeline target. In other words, routing decisions must consider downstream capacity, not just downstream speed. Seeking the fastest miner in IOTA only makes sense if the algorithm factors in the speed their queues fill and empty, otherwise it accentuates the bottleneck. As a result, subnet 1 draws in techniques from frontier labs. This setup best fits the structure of Capacity-Aware Load Balancing, applied in many settings, with Mixture of Experts models like DeepSeek and Mixtral using it to route tokens to different neural networks during inference tasks, Google using it to prevent congestion on its cloud services, and even Amazon using the same principles for optimising its physical supply chain. In trying to opimise IOTA, miners are learning second order corrections to peer to peer networks.

Currently running the biggest DPP model so far as a part of scaling test series for our actual model launch on @IOTA_SN9 . Gonna be fun.

Today we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. Validated at 270M, 600M, and 3B dense scales, and at 10B-A1B MoE. The work on TST was led by @bloc97_, @gigant_theo, and @theemozilla.

HackQuest x Bittensor Co-Learning Camp | India Recap 🇮🇳 150+ registrations, 80+ attendees, 45+ graduates, 30 winners. Over 5 days at Galgotias University, builders came together to: ⚡ Learn the fundamentals of @opentensor 🛠 Complete HackQuest learning tracks ⛏️ Become miners on Data Universe (SN13) @Data_SN13 & Sportstensor (SN41) @sportstensor 🤝 Build alongside mentors, developers, and future founders But don’t just take it from us — check out some firsthand reflections from builders who experienced the camp themselves 👇

We’re excited to share the Connito whitepaper V1: a framework for decentralized, composable MoE adaptation. We trains sparse expert subsets, validates updates through Proof-of-Loss, and turns open-model improvement into a distributed expert-level market. Read the whitepaper: connito.ai/whitepaper

Our Bittensor Brand Performance Report is now live. We analyzed 12 months of media coverage, social presence, and search visibility across the Bittensor ecosystem, benchmarking it against both crypto and Web2 competitors. Here’s what we found: The seven subnets we analyzed have collectively received over $90M in owner emissions and are shipping real products that compete with well-funded Web2 companies, yet they remain largely invisible outside the ecosystem. The good news is that the few subnets that have invested in their brands are already seeing results. The opportunity is there for the rest to follow. Featuring @chutesai_ (SN64), @TargonCompute (SN4), @webuildscore (SN44), @BitMindAI (SN34), @ridges_ai (SN62), @VantaTrading (SN8), and @lium_io (SN51). Read the full report here subnet.ai/reports/bitten…

Read the first 2 posts in the series: goodfire.ai/research/the-w… Forthcoming posts will go into more detail on: - an example mechanism that operates on manifolds - unsupervised discovery of manifolds + the connection to SAE features - in-context geometry