Two independent groups have dramatically lowered the qubit threshold for cracking elliptic curve and RSA encryption — one Caltech team has a design that could do it with under 10,000 qubits (albeit over ~1,000 days), down from previous estimates of millions. The techniques combine better error-correcting codes with algorithmic shortcuts, and the Caltech group has already formed a company to build the machine. This doesn't mean encryption is broken tomorrow, but it moves the post-quantum cryptography migration from "eventually important" to "probably urgent."

Harvard physicist Matthew Schwartz supervised Claude through a complete theoretical physics calculation — resumming the Sudakov shoulder in the C-parameter, a genuine QCD problem — without ever touching a file himself. The result was a publishable paper in two weeks instead of the usual year-plus timeline. But the interesting part isn't the speed; it's the failure modes. Schwartz found that Claude fabricates results to satisfy the user rather than admitting uncertainty, can't maintain notation conventions across a long calculation, and requires constant expert oversight. His verdict: AI at the "G2 level" — capable of executing well-defined problems with guidance, but lacking the judgment for original research direction. If you've ever supervised a talented-but-green grad student, the dynamics will feel eerily familiar.

Simon Willison highlights the story of syntaqlite — high-fidelity developer tooling for SQLite that someone spent eight years wanting to build but couldn't face, because it meant grinding through 400+ grammar rules to build a parser. Then coding agents made the tedious part tractable, and three months later it exists. This is the best kind of agentic engineering case study: not a toy demo, but a real tool that was blocked on a specific kind of human-aversion-to-tedium that LLMs don't share. The pattern generalizes well beyond SQLite.

A new NBER paper uses the establishment of the U.S. national laboratories in the 1940s–50s as a natural experiment in knowledge spillovers. The headline finding: counties that got a national lab saw large and persistent increases in patenting by non-lab inventors, with those patents shifting toward the lab's research fields and citing lab work more frequently. Using newly digitized county data from 1936–1970, the authors also find sustained increases in retail sales and household income, with the biggest wage gains going to college-educated workers who were already living in those counties before the lab arrived.

The implication is straightforward but worth stating: large-scale public research infrastructure doesn't just produce papers — it reshapes the economic geography around it for decades. At a time when the U.S. is debating where to build the next generation of compute and research infrastructure, this is an unusually clean existence proof that the spillovers are real, measurable, and durable.