Anthropic Walks Into Washington. Here's Why.

Anthropic's CEO Dario Amodei walked into the West Wing on Friday for a meeting with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The conversation was billed as "productive and constructive" - a remarkable reversal given that the Trump administration had labelled Anthropic a supply-chain risk just weeks earlier and said it would "not do business with them again."

What changed the math was Mythos, Anthropic's AI model for cybersecurity. During internal testing, Mythos found thousands of previously unknown vulnerabilities in every major operating system and browser - including a 27-year-old bug in OpenBSD and a 16-year-old flaw in FFmpeg that had passed automated testing five million times. Rather than release it publicly, Anthropic restricted access through Project Glasswing, a coalition backed by up to $100 million in credits that includes AWS, Apple, Google, Microsoft, and others. US intelligence agencies, CISA, and the Treasury Department are already testing Mythos, and the White House meeting was partly about how other government departments access it. The litigation with the Pentagon remains unresolved - Anthropic is barred from DoD contracts - but civilian agencies like Treasury and Energy aren't willing to be collateral damage in that fight. One source put it plainly: "It would be grossly irresponsible for the US government to deprive itself of the technological leaps that the new model presents. It would be a gift to China."

Multiplexed PET Imaging Changes Cancer Treatment

Conventional PET scanners are "monochromatic" - they can image only one radiotracer per session. That forces radiotherapy to rely on a one-size-fits-all dose model, even though tumours are wildly heterogeneous. Hypoxic regions (lacking oxygen) can increase radiation resistance threefold. A new approach called multiplexed PET (mPET) uses radiotracers that emit both positrons and gamma photons, enabling simultaneous imaging of multiple biological processes. For head-and-neck cancer, doctors can now map both clonogenic cell density and hypoxia-related radioresistance in a single scan - perfectly co-registered, no anatomical drift. Radiobiological modelling suggests this "dose painting" strategy - escalating radiation to radioresistant areas while protecting healthy tissue - could increase tumour control probability from the current 60% to 90% or higher. The physics works on existing hardware with no modifications required. The main technical hurdle is low signal from the tagged "triples" events, but emerging work on V-shaped reconstruction algorithms is addressing that.

Postgres at Scale: Indexing 18,000 Fan Curves in 4ms

A developer built a web-native HVAC fan selection engine that indexes 18,141 real manufacturer pressure-flow curves across 13 fan families and returns a match in about 4 milliseconds. The key insight: store each curve as a polynomial coefficient array (7 floats per curve) instead of sampled lookup tables. Evaluate with Horner's method - no Math.pow, no piecewise-linear interpolation error - and index on (q_min, q_max) to range-filter before evaluation. The three gotchas that each cost a week: pressure and efficiency are independent curves (not derived from each other); polynomial degree varies per manufacturer (zero-pad all coeffs to max degree so the JS engine keeps the inner loop monomorphic); and extrapolation outside the fitted domain returns physically meaningless numbers silently (domain checks are a correctness guarantee, not an optimisation). The production code is open at github.com/goncharovart/polynomial-fan-matcher. It's a concrete example of how specificity in data modelling and algorithmic choice - rather than raw compute - solves performance problems.

The broader pattern emerging this week: AI is shifting from "lab breakthrough" to "which agency gets access first," multiplexed imaging is collapsing what used to require sequential scans into a single session, and web developers are discovering that the maths you choose matters more than the framework you use.