AlphaProof Code proves Signal's cryptographic code
Google DeepMind's AlphaProof Code framework mathematically proved key security properties of the Signal protocol's actual code, marking the deepest formal verification of a widely deployed cryptographic protocol's implementation so far. The agent-driven effort took four weeks; expert teams historically spent years on such work. — deepmind.google
Hardwired pause would end training-capable chip production
A hardwired-pause plan for frontier AI would halt production of chips that can train new models, allowing only chips that run existing ones. Such a pause could hold for at least a decade, and likely indefinitely, without trusting China or tanking the economy. — lesswrong.com
Minimax H3 upscaler halves size, wins on most scales
A new version of the Minimax H3 video-model upscaler uses half the parameters of its predecessor while beating it on most scales. The model enlarges video in compressed latent form before decoding, accelerating high-resolution generation without a costly pixel round-trip. — huggingface.co
Stopping dangerous AI outweighs regulatory caution
Stopping dangerous AI development soon matters far more than the risk of overregulating, and pausing would not prevent humanity from restarting later. Public involvement and alarm-raising are comparatively safe, high-value moves that build political will. — lesswrong.com
Heavy AI coding breeds fragile codebases
Heavy reliance on AI coding agents produces brittle codebases: shallow tests, over-engineered helpers, and shrinking human understanding of the code. The result is more customer-facing incidents, slower recoveries, and a degraded experience for users and developers. — evanxmerz.com