Natalie Wolchover checks back in on string theory more than a decade after the field entered a profound, public crisis. The occasion: a January preprint by Henriette Elvang and collaborators claims to derive the Veneziano amplitude — string theory's foundational scattering formula — starting from quantum field theory axioms alone, without ever invoking strings. If it holds, it reframes what string theory is. But the landscape problem hasn't moved: the theory still permits something like 10⁵⁰⁰ possible vacua, and no one can navigate to ours. This is the state-of-the-art, written for the rest of us.

A portrait of Astrid Eichhorn, who leads the asymptotic safety program — a competing approach to quantum gravity that requires no extra dimensions, no new particles, only a fractal-like structure at short scales where coupling constants cease running and lock in place. What makes it compelling isn't the elegance alone: in 2009, Shaposhnikov and Wetterich used asymptotic safety to predict the Higgs mass at roughly 126 GeV. We measured it at 125.09 GeV. That's one more empirical hit than string theory has landed in forty years. Read this alongside the Wolchover piece for a clear-eyed picture of where quantum gravity actually stands.

OpenAI acquired Astral — the company behind uv, ruff, and ty, the three tools that now underpin most serious Python projects. Willison notices the structural parallel: Anthropic acquired Bun (the JavaScript runtime) last December. Both moves insert an AI company directly into the tooling layer that developers rely on daily, before they ever open a chat interface. He raises the concern worth sitting with: OpenAI could plausibly use control over uv as leverage in its competition with Anthropic. He doesn't think they will. But now the option exists.

A survey of ~6,000 C-suite executives across the US, UK, Germany, and Australia finds 90% report no measurable AI impact on productivity or employment — and the paper prompts economists to exhume the Solow Paradox: "You can see the computer age everywhere but in the productivity statistics." The piece isn't another round of AI skepticism. It's a real economic question: the IT revolution did eventually show up in productivity data, but it took 15–20 years of organizational restructuring first. The bottleneck was never the technology. The question for AI is whether the lag is the same, shorter, or whether something is genuinely different this time.

Lambert's thesis for 2026: scale arrived in 2023, reasoning models arrived in 2025, and this year is shaping up to refine and integrate the workflow rather than introduce a new paradigm. He's genuinely uncertain — not dismissive — about whether the current period constitutes meaningful progress. The title captures something precise: some forms of improvement consume the capacity for further improvement. A good thing to sit with if you find yourself trying to track AI progress against some mental model of what "real" progress looks like.

Presolar grains — microscopic crystals inside meteorites that predate the formation of the Sun — are now being used to adjudicate a dispute between two origin stories for the solar system: a nearby supernova trigger versus stellar winds from a massive AGB star. The grains carry isotopic signatures that should, in principle, distinguish them. Short and genuinely lovely: geology as a time machine for cosmochemistry, with material older than the sun sitting in a lab at Boston University.