Anthropic’s preview-only Mythos impressed the UK’s AI Security Institute, completing a 32-step network takeover simulation other models couldn’t. The takeaway: outcomes scale with token budget, pushing security toward a compute-and-cash contest.
Anthropic’s Claude Code prompt-cache TTL reportedly dropped from one hour to five minutes, and developers say their quotas are draining far faster in long, high-context sessions. The Register has more on the pricing mechanics, possible bugs, and what Anthropic may change next.
A recent article by Bryan Cantrill revisits Larry Wall’s “laziness” as an engineering constraint that drives cleaner abstractions and simpler systems. He argues LLMs can amplify bloat without human taste and limits.
A recent post by the author breaks down “gates” in AI prompts—explicit conditions that must be met before an agent can move on. Unlike rules that can be hand-waved, gates force checkable steps (like holding URLs) and pair well with external “hooks.”
In a X thread, Chayenne Zhao argues that many agent frameworks waste tokens in ways that undercut key inference optimizations like prefix caching—hurting cost and throughput in long sessions. The takeaway: better agent–inference co-design may unlock big efficiency gains.