
barbiercha
48 posts



















Sharing a short update on where we are right now. $TAO At this stage, the project has moved into a release-preparation phase. The main focus over the past period has been on stability, cleanup, and making sure everything is production-ready. We’ve been tidying up internal data, migrating almost all required components to the production environment, and handling a lot of the necessary but less visible work that sits behind a solid launch. Alongside this, we’ve also made a few focused improvements: - Refreshed parts of the landing page. - Added a new feature for downloading promotional materials for miners. - Continued improving and optimizing the codebase. Most of the work right now is intentionally routine and detail-oriented – final checks, refinements, and alignment across the system – but it’s exactly what allows the platform to feel stable and cohesive as a whole.

An announcement will be made next week $TAO 🤝




Just came across an update on Quasar (SN24) on Discord and thought it was worth discussing. @QuasarModels Apparently, the Quasar SN24 mining system is basically finished. They’re in the final stretch now, cleaning up a few bugs and running longer stability tests before launch. What caught my attention is how they’re structuring this subnet. It’s set up as a collaborative but very competitive environment, focused entirely on pushing ultra-long-context inference forward. Stage 1: is all about attention performance. Right now, Quasar uses a token-by-token (non-parallel) attention mechanism, which obviously doesn’t scale well once you start pushing 100k+ or even million-token contexts. "So the challenge they’re putting out is pretty straightforward: Can miners build faster, chunk-parallel attention implementations that actually hold up at extreme sequence lengths"..? Miners are expected to: • Write custom CUDA kernels • Implement chunk-parallel attention • Optimize tokens/sec for very long contexts Validators then compile and benchmark everything, measuring throughput across different context sizes (100k, 1M tokens, etc.). What’s interesting is the reward structure. It’s pure performance-based: >>1st place takes 60% >>2nd–4th split the remaining 40% No subjective judging, just whoever is fastest across sequence lengths. Feels like one of those subnets where low-level optimization really matters, not hype or narratives. Curious what others think, does this kind of performance-only competition actually move long-context inference forward, or does it just favor a few highly specialized CUDA devs?





