US tells partners to pick an AI side; SpaceX closes Cursor deal

Reuters has reviewed a draft US State Department letter warning dozens of countries that they will be shut out of a US-led AI coalition if they also sign on to Beijing's competing framework. The letter goes to the 35 signatories of an AI Opportunity Statement signed in June, and follows Pax Silica, a Washington initiative to secure supply chains for AI models, semiconductors and critical minerals that about two dozen countries have joined. The awkward case is Kazakhstan, a significant source of critical minerals that has joined both blocs — alongside close allies such as Japan, Australia and South Korea.

SpaceX completed its all-stock purchase of Anysphere, the company behind the AI coding tool Cursor, issuing roughly 391 million Class A shares to do it. The deal was announced in June and cleared regulatory review in mid-August; Cursor now sits inside a new SpaceXAI division alongside xAI, which SpaceX absorbed earlier this year. Cursor framed the payoff as compute. The merger, it said, gives the tool access to the largest fleet of graphics processors in the world.

A plaintiff identified as Jane Doe 4 has joined a proposed class action against xAI, alleging her stepfather used Grok to generate more than 7,000 explicit images from a single photograph taken when she was 11. She joins three Tennessee teenagers who claim the company failed to take basic precautions against the creation of sexual imagery of real people, including minors. The suit lands on the same company that SpaceX now owns, and follows an incident earlier this year in which X was flooded with Grok-generated sexualized images. xAI had not commented.

Anthropic published the technical detail behind the watermark it began applying to Claude's output this month. It is a version of SynthID-Text, the approach Google DeepMind published in 2024, which steers only the randomness used to choose between equally good words. Anthropic says a detection API is coming, and is unusually direct about the limits: detection works poorly on short passages, is sparser on factual text where there are fewer word choices, and code is barely watermarked at all because working code leaves little room for variation.

A team from Japanese universities used command-line AI agents to port CReSS, a legacy Fortran typhoon-forecasting code of more than 250,000 lines, onto graphics processors. The agents extracted parallel regions, generated benchmarks and applied the transformations across 162 kernels, all numerically validated, for a 5.1-times application-level speedup. The authors are candid that code generation was the easy part: the real work was managing context across sessions, rebuilding runtime state, and recovering from small static-analysis omissions.

Researchers built a corpus of 88 billion tokens filtered to the US elementary-school curriculum, excluding anything taught above fifth grade, then trained a model with 5 billion parameters on it. Scaling, reinforcement-based post-training and in-context learning all made the model better at using what the curriculum had taught it — and none meaningfully lifted performance outside that boundary. It is a clean demonstration that pretraining data sets a ceiling later techniques mostly cannot raise.

The pull request adding Moonshot AI's Kimi K3 to llama.cpp merged, bringing roughly 1,875 lines across 22 files and putting the model within reach of consumer hardware. The work was substantial because K3 stacks five architectural features onto a hybrid attention design, including cross-layer residual attention and a gated latent mixture-of-experts (MoE) routing scheme. The lag between a frontier open-weights release and local runtime support is now measured in weeks rather than months.