OpenAI just shipped GPT-6 Astra and had to gate part of it, the model got good enough at finding zero-days on its own that it crossed their "critical" cybersecurity line. If you're on a security team, that vetted defender access is worth requesting. The 99.9% ARC-AGI-3 number going around is from a favorable harness though. Standard harness: 62.7%.

100% Agreed. A model that learns from your own production traffic is the first form of AI that behaves like an asset. It appreciates with use, while everything rented sits on the books as an expense. The software analogy is the strongest part of this: everyone writes their own app layer, nobody writes their own OS. Frontier models are turning into the OS!

Okay this one's genuinely exciting! Every self-improving agent until now turns ONE knob, edit the scaffold or train the weights. SIA turns both, in a single loop. And it rips: LawBench: 50% → 70.1% once weights kick in CUDA kernels: 14x speedup RNA denoising: +502% over baseline And the tell is on denoising—this is the part that's genuinely awesome. The scaffold-only loop, across every single iteration, never found a fix. Then the first weight update just... figured it out: a two-line clip that rounds imputed counts to non-negative integers. A basic biological constraint that no prompt, no scaffold, no amount of clever engineering ever surfaced. The model learned it on its own. That's the whole thesis in one moment! The harness changes how an agent searches. The weights change what it knows. And you really, really want both.

Hot take: 90% of "bad AI outputs" are a data freshness problem, not a model problem. You're blaming GPT-4 for confidently answering with info that was accurate 3 months ago. That's not a model failure. That's infrastructure debt. Read more on how to solve this here! stateai.substack.com/p/you-d…

This is awesome! You can get Replit Core FREE for 1 month AND gift someone Replit FREE for 1 month. Great win for builders!