I actually think Perturb’s approach makes sense, but not for the usual “decentralized AI” reasons people throw around. AI security is becoming a much bigger problem as models get more capable and agents start operating with less human oversight. And this is where @perturbaix has a structural advantage. Most red teaming is still done in-house or with static open-source tools. Perturb turns adversarial attack discovery into a continuous, incentivized competition, then uses those findings for actual model training. The problem is that finding attacks and turning them into something companies will pay for are two different things. Even if miners keep producing strong adversarial examples, the real question is whether enterprises will treat the resulting scores, datasets, and certificates as credible enough to act on. Moving into a real product with continuous scanning + hardening could give them more control over that distribution layer. We’re also seeing the subnet mature (scanning + training live, product phase starting). But I wouldn’t confuse subnet activity with proof that the commercial model works yet. Leaving pure research mode means Perturb now has to prove it can generate demand and revenue without just relying on emissions. My view is that continuous, external adversarial pressure could become just as important as model performance over the next few years. If Perturb can productize that without the output remaining mostly internal to the subnet, the direction is justified. Otherwise, it risks staying an interesting research network rather than infrastructure. SN26 at these levels is interesting either way. #Bittensor $TAO #Perturb #AI