⚠️ There is a problem hiding inside every AI model deployed in the world right now. $TAO's Perturb SN26 just went live on mainnet. Making it one of the most important subnets on the network. Your AI model scores 94% accuracy in testing. Looks great. Ships to production. Then someone changes fewer than 0.3% of the pixels in an image a change invisible to any human eye and that same model misclassifies completely. Confidently. Catastrophically. This is called an adversarial attack. Medical imaging models used for cancer detection can be fooled this way. Autonomous vehicle systems can be made to ignore a stop sign. Fraud detection systems can be made to approve fraudulent transactions. Facial recognition controlling physical access to secure buildings can be bypassed. The scariest part? Most organizations shipping AI products have never tested for this. Not because they don't care. Because the tooling to do it at scale simply didn't exist. Until now. Think of it like a security testing firm except instead of one team of experts working on your model for 6 weeks and charging $200,000, you have a global network of miners competing around the clock to find every possible vulnerability. Validators present a challenge: here's an image of a tabby cat. Fool the AI model into thinking it's something else but do it with the smallest possible change to the image. A change so small no human would notice. Miners compete to find that attack. The one who fools the model with the least visible change, and does it fastest, wins. The winning technique earns $TAO emissions. Every single attack is verified by an LLM semantic checker so a valid attack genuinely fools the model on a meaningful label, not just a technical integer class mismatch. The verification is tight. What makes this extraordinary is the compounding property. Every day miners compete, the network gets better at finding attacks. The dataset of adversarial examples grows. The techniques improve. No centralized security firm has this. They charge you $200K, write a report, and leave. Perturb gets stronger every single day permanently. The market is real and it's accelerating fast. The EU AI Act classifies AI systems in healthcare, autonomous vehicles, critical infrastructure, and hiring as high-risk requiring mandatory robustness testing before deployment. Non-compliance fines go up to 30 million euros or 6% of global revenue. That's not a soft guideline. That's a compliance requirement with teeth. Enterprise procurement teams are already demanding adversarial robustness certificates before signing AI vendor contracts. A Perturb certificate can unblock deals worth orders of magnitude more than the subscription cost. This is a $1.43 billion market growing at 26.1% annually projected to hit $11.6 billion by 2033. And there are zero competing subnets in this category on Bittensor. Zero. The roadmap is real and well-staged. • Phase 1 launches with EfficientNet-B5 a production-grade image classifier that's hard enough to attack that basic methods fail, but accessible enough that any serious miner can participate with an RTX 3080. • Phase 2 adds architecture diversity CNNs, Transformers, hybrid models. • Phase 3 expands into LLM text attacks. Phase 5 targets vision models with over 1 billion parameters. Perturb’s real prize isn’t testing. It wants the winning attack systems to become on-chain robustness certificates cryptographically immutable, regulator-auditable, and impossible to forge. Something no vendor’s self-reported compliance document can ever match. Take a problem the centralized world handles poorly expensive, static adversarial testing run it as an open competition, pay the best solutions, and let the network compound. The team is building in public. Sigurd is active in Discord and they’re looking for design partners right now to help build the dashboard. If you want to shape this, the door is open. perturbai.io $TAO DYOR

Introducing Aurora, a new optimizer for training frontier-scale models. We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks. Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs. By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity. What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.