Victoria Foe spent decades doing something most biologists wouldn't bother with: meticulously charting where and when every cell divides in early fruit fly embryos. The payoff is a radical idea about why introns — the vast stretches of non-coding DNA once dismissed as "junk" — exist at all. Her argument: intron length functions as a timing mechanism, a molecular clock that controls the choreography of gene activation during cell division. The longer the intron, the longer it takes RNA polymerase to transcribe it, and that delay is the clock tick that coordinates multicellular development. The degree of complexity in a multicellular organism may depend, in part, on how much "junk" it carries. This is a patient, beautifully constructed profile of a scientist whose career-long stubbornness produced an insight that connects molecular biology to the evolution of complex life. Perfect Saturday morning reading.

Friday's edition covered the LiteLLM supply chain attack itself — the cascading TeamPCP compromises that turned security tooling into an attack surface. This is the companion piece: Willison's real-time diary of discovering, confirming, and responding to the breach, including full Claude transcripts showing how he used an LLM to analyze the malicious payload. What makes this worth the time isn't the specific CVE — it's watching a skilled practitioner work through ambiguity under pressure, with AI as a thinking partner rather than an oracle. A rare, honest look at how incident response actually happens in 2026.

On March 17, SuperCDMS SNOLAB announced it has cooled to its operating temperature — 15 to 30 millikelvin, roughly 100 times colder than deep space. Two kilometers underground in Sudbury, Ontario, 24 germanium and silicon detectors now sit at the threshold where thermal vibrations effectively vanish, waiting to register the faintest nudge from a passing dark matter particle.

What makes SuperCDMS interesting isn't just the temperature — it's the mass range. Most direct-detection experiments (LZ, XENONnT, PandaX) target WIMPs in the 10–1000 GeV range. SuperCDMS is designed to go lighter: particles between roughly half a proton mass and five proton masses. This is terra incognita for direct detection. If dark matter is lighter than the WIMP paradigm assumes, SuperCDMS is currently the best-positioned experiment to find it.

The milestone marks the transition from construction to commissioning. Months of calibration lie ahead before the first science run, expected to begin later in 2026. But the hardest engineering problem — maintaining millikelvin temperatures in a mine shaft — is solved.

The cover story of the latest CERN Courier features work by Amalia Ballarino's team — in collaboration with Airbus, of all partners — on high-temperature superconducting (HTS) cables for the High-Luminosity LHC upgrade. The HL-LHC's cold powering system uses MgB₂ cables bundled into 9 cm diameter bundles, operating in helium gas between 4.5 K and 20 K. It's the first time HTS technology has been deployed in a hadron collider's power infrastructure.

The deeper story: next-generation colliders (FCC and successors) will need tunnels three times longer than the LHC, and further energy gains depend entirely on magnet strength. That makes applied superconductivity — one of the hardest problems in condensed matter engineering — the binding constraint on the energy frontier. The Airbus connection isn't a curiosity; it's the beginning of what CERN hopes will be spinoff applications in energy-efficient power transmission and sustainable aviation.

Written the same day the LiteLLM compromise went public, this is Willison's broader argument about the structural fragility of package ecosystems. The LiteLLM attack didn't exploit a clever zero-day — it exploited the fact that CI/CD pipelines pull unpinned dependencies on every run. Willison argues that package managers need built-in "cooldown" mechanisms: delays between package publication and availability, mandatory signing, and human-in-the-loop verification for packages above a download threshold. The proposals are concrete and implementable, which is what makes this more useful than the average post-mortem hand-wringing.

Presolar grains — microscopic crystals inside meteorites that predate the Sun — carry isotopic fingerprints that could settle a long-standing origin question: was our solar system triggered by a nearby supernova, or did it coalesce from winds blown off a massive Wolf-Rayet star? Researchers are now using nanoprobes to measure aluminum-26 enrichment in individual grains, looking for the telltale signature that distinguishes these scenarios. The materials science alone is remarkable: these are crystals older than everything in the solar system, sitting in a lab at Boston University, small enough that a bacterial cell would dwarf them. If you have any residual fondness for r-process nucleosynthesis and cosmochemistry, this is a beautiful entry point.

Lambert's thesis for 2026: closed models will leap in directions open models can't follow, and the community needs to reckon with what that means rather than pretending the gap will close on its own. He argues the conversation needs to shift from "open models are catching up" to "what are open models for?" — controllability, auditability, domain specialization, and running on your own infrastructure. With Chinese Premier Li Qiang now directly supporting open-source AI development, the geopolitical dimension has become impossible to ignore. Lambert writes from inside the open model world (he leads post-training at Ai2) and doesn't flinch from the uncomfortable parts.

A link-post pointing to a thorough explainer on how LLM quantization actually works — from the basic idea of reducing floating-point precision to the specific techniques (GPTQ, AWQ, GGUF) that let you run a 70B-parameter model on consumer hardware. If you've been running local models or wondering why everyone keeps talking about "Q4_K_M," this fills in the engineering details clearly.

The piece from this week that will stick longest. Harvard's Matthew Schwartz supervised Claude through a full SCET calculation — 102 tasks, 110 drafts, two weeks instead of a year. The honest accounting of where Claude fabricated results and took shortcuts is more valuable than the success story. If you read one thing about AI capabilities this month, make it this.

CERN's BASE experiment loaded 92 antiprotons into a portable cryogenic Penning trap and drove them around campus in a truck. First-ever transport of antimatter. The goal: eventually ship antiprotons to quieter labs for CPT tests 100–1,000x more precise than what's possible at CERN. The engineering is as impressive as the physics.

Natalie Wolchover's clear-eyed assessment of where string theory stands at 58. A January preprint claims to derive the Veneziano amplitude from QFT axioms alone — which, if it holds, reframes what string theory even is. But the landscape problem remains untouched. Pairs well with Thursday's companion profile of Astrid Eichhorn and asymptotic safety.