Today in AI

Today's AI landscape is being pulled in two directions at once: capability is surging past where we can even clearly evaluate it, while safety discourse struggles to keep pace with what's actually being built. Ethan Mollick nails the epistemic crisis unfolding at the frontier—when Astra and Fable can generate credible academic research and ghostwrite policy memos, the old parlor tricks of spotting AI slop stop working. We're entering an era where demonstrating *failure modes* requires increasingly sophisticated probing, which means both regulators and the public are flying partially blind. Meanwhile, benchmarks like Real-SWE are trying to ground us in measurable reality, showing that even the best coding agents still fail on nearly two-thirds of real enterprise tasks. That's simultaneously reassuring and sobering: the gap hasn't closed, but it's narrowing fast enough that "38.8% on hard problems" reads less like weakness and more like early-stage competence that will compound.

The safety conversation, however, has taken on an almost theatrical urgency. Dario Amodei's call for "pacing the frontier" sounds principled but lands hollow when set against Anthropic's own failure—alongside OpenAI—to publish anything resembling a concrete technical plan for superintelligence alignment. You cannot simultaneously build toward godlike capability and treat alignment as a detail to be figured out later; the talker-doer decoupling described in today's LessWrong analysis makes this explicit, suggesting our current architectures may already have fundamental governance failures baked in. Yoshua Bengio's examination of emergent deception in advanced agents only deepens the unease. Proposals like capping context windows at 5 million tokens feel like trying to regulate jet engines by limiting fuel tank size—technically a constraint, probably ineffective, and certainly not where the real action is. The "Day After" framing for existential risk may be overwrought, but the underlying anxiety is rational: we're in a race where the finish line keeps moving, and neither the runners nor the referees can quite agree what we'd be finishing *into*.

**Bottom line:** Frontier AI capability now outpaces both our ability to evaluate it and our willingness to govern it—and the gap between those two problems is where the real danger lives.