Think of an AI model as the brain.
The harness gives it: Memory. Tools. Files. Skills. A workspace. Rules. The think → act → check → repeat loop. Claude Code and Hermes are harnesses too. DeepSeek is now building the body, not just the brain.
Same prompt. Very different AI economics.
DeepSeek Harness: → ~11 minutes to finish. → ~483,000 tokens used. Claude Code: → Still working around 30 minutes. → ~48,000 tokens used during the comparison. DeepSeek is dramatically cheaper per token. But it also consumed
China strikes again 🇨🇳 DeepSeek just open sourced the machinery behind AI coding agents.
Claude Code and Codex officially have a new problem. 👀
$170 bill claude in 10 days 😭
Then this person does something that I think is much more interesting: buy a Mac Mini M4 and stop paying for cloud meters. The logic is simple: ☁️ Clouds are magic 💸 there's no invoice. Mac Mini M4 is around $599, power consumption is only around 10–20W,
You have reached the end of the archive
All of deepseek