Someone Else's Words in Your Answer: A Design Review of Cloudflare's GLM-5.3 KV Cache ExposureHow a refused prompt crossed between two machines and came back with other customers' words, what Cloudflare and SGLang changed, and what's still open.27 min read·anandms101.hashnode.dev2ASNov 7
Your AI Remembered. Did It Learn the Right Lesson?An agent can preserve your words but lose their limits. Following one travel request shows how records become lessons and reach the next task.9 min read·anandms101.hashnode.dev11GOct 27
An Attention Window Is Not a Checkpoint: Inside vLLM's 128-Token ReplayA coding agent returns with another test result. Following that request reveals why vLLM saves less state between turns, then rebuilds a recent window without reproducing it exactly.10 min read·anandms101.hashnode.dev12KAOct 11
The Model Is Untrusted Code: A Design Review of Meta's MuseHow Muse contains a prompt-injected agent, replayed against one malicious email, plus the questions I'd push on and the parts worth copying.17 min read·anandms101.hashnode.dev1MSep 29
Toon vs JSON: The Battle for Token Efficiency in LLM ApplicationsDiscover how Toon format reduces LLM token usage by 20% compared to JSON. Real data, cost analysis, and practical examples for OpenAI and Anthropic APIs. Sa6 min read·anandms101.hashnode.devNov 8, 2025