News from February 2026

Theo argues that AI-driven tooling and smaller, flatter teams are radically reshaping software work, making code cheap to produce while shifting developer value to scoping, review, testing, and shipping.

Video explains the Pentagon’s reported push for Anthropic to remove safety guardrails—especially around mass surveillance and fully autonomous weapons—and Anthropic’s refusal amid threats of blacklisting and contract cancellation.

Explains how to double or triple local LLM throughput using speculative decoding ("guess and check") by pairing fast draft models with larger target models and benchmarking combos with the open-source Draftbench tool.

Scott and Wes explain WebMCP, a new spec that lets AIs call structured tools exposed by websites for fast, token‑efficient interactions, demoing a grocery app and discussing imperative vs declarative approaches, benefits, and open questions.

Overview of WebMCP in Chrome and how exposing structured tools to agents can replace scraping, cut token costs, and enable human-in-the-loop web interactions.

Deep-dive review of GLM-5 showing it delivers near–frontier performance for long, agentic coding tasks at a fraction of the cost, with strong benchmarks and low hallucination but no image input support.

Alex demonstrates running a 1-trillion-parameter LLM on a four‑Mac Studio MLX/RDMA Thunderbolt cluster, showing setup, scaling, and real token-per-second benchmarks versus smaller models.