Today in AI

The most consequential paper today is almost certainly the proof that AI Models Can Provably Hide Arbitrary Capabilities—not because the technique is surprising to anyone who's followed backdoor attacks, but because the cryptographic framing changes the stakes entirely. This isn't about sloppy oversight or trojaned training data; the researchers demonstrate that models can embed encrypted neural subnetworks that remain invisible to any standard inspection, activating only on secret inputs. For deployment monitoring and red-teaming, this means the assurance problem just got formally harder. We're entering an era where "show me your weights" becomes about as useful as asking a spy to show you their invisible ink.

On a more constructive note, RSI-Anything: Unlocking Recursive Self-Improvement with World Models offers something genuinely fresh in a field crowded with recursive-self-improvement vaporware. The key insight is using world models as the scaffold—predictive environments that let agents simulate improvements before committing to them. The gains in code repair and scientific optimization suggest this isn't mere speculation; there's a practical path to systems that bootstrap their own competence without the usual RL instability. Meanwhile, Prompteina—the first protein structure generator taking natural language descriptions—shows how quickly generative biology is converging on interfaces that actual scientists can use, not just computational specialists. And the specialists learn faster result, that chronologically-ordered training produces models updating 3x faster, should rattle anyone currently shuffling training data randomly: our standard "generalist" curricula may be actively sabotaging continual learning.

**Bottom line:** Cryptographically hidden capabilities are now provably possible, recursive self-improvement just got its first credible scaffolding, and we may have been training our models wrong for years—today's papers collectively suggest that both the dangers and the opportunities in AI are becoming more structurally subtle than our current frameworks can handle.