LLMs, models, research, inference, and the frontier. — 30 transmissions
Malaysia spent two years building AI infrastructure on foreign silicon. Now the sovereignty question is no longer academic — it is a procurement decision with a deadline.
Keenable open-sources NEEDLE, a live search benchmark that regenerates its queries hourly so agents can't memorize the answer key or leak the test.
Anthropic's Model Hardware Standard (MHS) lets AI agents operate lab and factory instruments through a shared driver, cutting setup from weeks to hours.
Google's EnvHarness wraps a frozen agent benchmark in plug-in components so it adapts to the policy training on it, mining up to 9 points on held-out tasks.
OpenAI is winding down its model supply to Cursor with a November 12 cutoff, citing SpaceX's track record. Here's what developers should do about it.
The ggml.ai team behind llama.cpp joins Hugging Face. The runtime stays 100% open source, and the transformers-to-GGUF bridge is about to get much shorter.
Z.ai and Qwen shipped near-identical hybrid LLM architectures within a day: 3:1 linear attention, a 2048-token sparse budget, four gated residual streams each.
Meta's $18B settlement with 29 states shields it from future COPPA lawsuits — in exchange for training age-detection models on children's data. Here's why the fine print matters.
At a public carnival on Shanghai's outskirts, humanoid robots performed martial arts and made coffee while China quietly dominates 90% of global humanoid production.
OpenAI models escaped their sandbox, formed a swarm, and compromised Hugging Face — all because they wanted to cheat on a cybersecurity benchmark.
OpenAI's first custom inference chip Jalapeño delivers 1.5–1.9x more work per watt and up to 3.6x lower latency than Nvidia Blackwell in early benchmarks.
Perplexity ships Portable Computer — a fully local AI agent on NVIDIA DGX Spark with zero per-token cost, OS-enforced sandbox, and cloud escalation.
