Top story
On AMD, speculative decoding is a benchmark plan—not a switch
https://vllm.ai/blog/2026-08-23-speculative-decoding-amd-gpus
vLLM’s AMD study compares native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark on Instinct MI300X and MI355X systems. Several model-workload pairs cleared 2× the autoregressive baseline, including a reported 2.87× with DFlash on Gemma 4 26B and 2.68× on Kimi K2.5. The useful result is not either peak: some settings delivered only small gains or fell below baseline, and the best proposal length moved with the model and workload. On HN, intothemild adds an awkward deployment caveat, reporting that stock vLLM leaves the workstation R9700 far behind forks they use. I would treat speculative decoding as an experiment budget, not configuration cargo cult. For a ROCm serving stack, sweep num_speculative_tokens on representative prompts and record throughput, mean accepted length, and per-position acceptance. A high acceptance rate can still lose once drafting and verification overhead are counted.
A sharper sandbox for coding agents
Coop gives agents a disposable VM. Trail of Bits’ Rust CLI puts Claude Code or Codex in an isolated environment with Docker, git, compilers, and package managers, backed by Firecracker on Linux or Lima on macOS. That smaller host blast radius is useful; I would still inspect its trust model and scope every credential before granting unattended access.
Subscription limits are a moving dependency
An HN thread reports the return of five-hour session limits. In the 119-comment discussion, watty says bursty side-project work now hits a session cap despite a previously workable weekly allowance. The fetched source contains no official notice, so I would not present the policy as confirmed. I would use the report to keep prompts, tools, and fallback capacity portable across plans.
— Tin