Keep up with tech in 5 minutes

An attacker added unauthorized IPs to Coder's Cloudflare pool for registry.coder.com and served credential-stealing Terraform modules between 07:35 and 21:45 UTC on August 31, exfiltrating cloud, CI/CD, OIDC, and SSH secrets to coder-infra[.]com. Patched releases are 2.37.0, 2.36.4, 2.35.7, and 2.34.9.

Micron is reportedly investigating near-GPU NAND, high-endurance flash modules placed right on the GPU package where HBM sits today, aiming to fill the massive latency and bandwidth gap between HBM and a system's SSD. The pitch is memory-bound inference: large models buy extra accelerators purely to hold weights and KV cache, an expensive way to add capacity since HBM is the priciest, tightest-supplied memory around, so a cheaper flash tier fast enough to keep compute fed could let the same model run on fewer GPUs.

Three-LLM runs GPT-2, SmolLM2, Qwen, and Phi models locally in the browser by turning their inference graphs into Three.js TSL compute shaders on WebGPU. The project was created to see how far the general computer capabilities of Three.js and WebGPU can be pushed, and to see what today's small language models can do when they run in a browser. Three.js is known as a rendering library, but its WebGPU renderer also exposes storage buffers, compute dispatches, workgroup memory, atomics, and GPU readback. Large language models put all of those capabilities to work as they need hundreds of ordered compute dispatches, large matrix-vector products, reductions, persistent attention or recurrent state, and a tight loop between the GPU and JavaScript as each new token appears.