Google DeepMind published AlphaEvolve on May 14: an evolutionary coding agent that pairs Gemini models with automated evaluators, letting it rewrite entire codebases and score candidates against a hard metric rather than a chat grade. It has been deployed across Google's computing ecosystem for over a year — a Borg data-center scheduling heuristic that continuously recovers an average 0.7% of Google's worldwide compute, a 32.5% speedup on a FlashAttention kernel, and a functional simplification of a highly optimized arithmetic circuit for matrix multiplication.

On mathematics it found a search procedure that multiplies two 4x4 complex-valued matrices in 48 scalar multiplications, the first improvement over Strassen's 1969 result in that setting, and established a new lower bound for the kissing number in 11 dimensions with a configuration of 593 outer spheres. DeepMind's own summary is that it rediscovered the best-known solution in roughly 75% of cases and beat it in about 20%. The loop is not exotic: an LLM proposes code, an evaluator judges it, evolution keeps the winners — and the output is human-readable code that can be debugged and deployed.

The Bureau of Industry and Security announced on May 13 that it is rescinding the Biden administration's AI Diffusion Rule, issued January 15, 2025, whose compliance requirements were due to start May 15. BIS instructed enforcement officials not to enforce the rule, said it will publish a regulation formalizing the rescission and a replacement rule later, and framed the original as burdening exporters and downgrading dozens of countries to second-tier status. For four months that framework — including a new control on AI model weights under ECCN 4E091 — had been the planning assumption for where frontier training and inference could legally sit.

In the same release BIS moved the opposite way on China: new General Prohibition 10 guidance states that using PRC advanced-computing ICs — naming Huawei's Ascend 910B, 910C and 910D — risks violating US export controls anywhere in the world, carrying the prospect of criminal and administrative penalties, and warns of consequences for allowing US chips to train or serve Chinese models. China's commerce ministry rejected the guidance the following Thursday. Net effect: blanket country tiers out, government-to-government deals and per-chip enforcement in.

Nvidia and HUMAIN, the Public Investment Fund's new AI subsidiary, announced on May 13 a partnership for AI factories in Saudi Arabia with a projected capacity of up to 500 megawatts powered by several hundred thousand Nvidia GPUs over five years, first phase an 18,000-GPU GB300 Grace Blackwell supercomputer with InfiniBand. Saudi Data & AI Authority gets up to 5,000 Blackwell GPUs for a sovereign AI factory, Aramco Digital builds alongside, and HUMAIN takes the country's first Omniverse Cloud. AMD said the same day it would supply chips for 500MW and that HUMAIN had committed $10 billion; Nvidia stock closed up more than 5%.

Three days later the US said it had agreed a path for the UAE to buy some of the most advanced American AI chips, alongside an Abu Dhabi campus announced May 15 — G42-built, 5 gigawatts, 10 square miles, billed as the largest such project outside the United States — with reporting putting the annual figure near 500,000 Nvidia H100s and roughly 100,000 of those going directly to G42. The Gulf deals are the first practical output of the rescinded diffusion rule, negotiated by the same policy officials while the export framework was being dismantled.

OpenAI released HealthBench on May 12: 5,000 multi-turn conversations between a model and a user or clinician, each graded against a physician-written rubric, 48,562 unique rubric criteria in total, created by 262 physicians who have practiced across 60 countries. The paper went up on arXiv on May 13 and the benchmark code ships in OpenAI's open simple-evals repository. Scores run from GPT-3.5 Turbo at 16% through GPT-4o at 32% to o3 at 60%, with GPT-4.1 nano outperforming GPT-4o at 25 times lower cost; the deliberately hard subset tops out at 32%.

Two design decisions are worth stealing: criteria carry signed weights, so wrong or unsafe behavior is penalized rather than merely unrewarded, and grading is done by a model whose agreement with physician graders OpenAI measured and reported. Physicians quoted by STAT flagged exactly that arrangement — a lab grading its own models with a model-based grader can hide errors the grader shares. The benchmark is deliberately unsaturated, which is what makes it usable for tracking a moving frontier rather than measuring saturation.