AI vendors cut you off, and open models are the escape hatch

A hacker reverse engineered a Samsung TV and found an on-board AI listening. His OpenAI account was then reportedly suspended. Commenters said that if you can't use an LLM for reverse engineering without risking your account, its value drops sharply. One compared it to a utility cutting you off for not streaming enough telemetry. Several pointed out that Chinese open models like GLM can already do a lot of this work. Moonshot shipped Kimi K3.1 Pro the same day, and the title debate was whether "We Must Pace the Frontier" is a mission or a slogan.

The related Anthropic thread ties in. Anthropic stopped letting Claude Code subscriptions power OpenClaw, citing "outsized strain on our systems." Commenters read that as "when you fully use the capacity you paid for, that's too much." Others wondered whether it hurts startups. In the local-model thread, one person said Qwen 3.6 27B dense performs about like Claude Haiku 4.5. Others said tokens per second on local hardware is still too slow.

The pattern here is that closed labs set the terms, and builders keep an exit ready. Those terms cover which tools you can run, how hard you can use your plan, and what counts as acceptable work. Open-weight models are the hedge, even though they trail the frontier on speed and polish.