Claude AI is not intelligent | Stochastic Architecture
summary
LLMs act as probabilistic generators not reasoners, so Claude and Claude Code produce non‑deterministic outputs, stochastic parrot behavior, and hallucinations that are expected. Practical experience shows thousands of generated lines, tests written to /tmp and discarded, and one vendor version that deleted hundreds of source files. Open‑source projects built—Domo Actors (Java/zoomX), TypeScript and Python ports, and Domo Tactical—accelerate work but require human craft; using generated Javadoc heuristics raises a correctness signal from ~21% to ~67%, proving speed is abundant craft is scarce and hallucinations are by design.
AI-written summary · about
chapters (6)
- 0:00Introduction to talk
- 2:17What is stochastic architecture
- 15:43LMs strengths and weaknesses
- 27:53Speed vs craft in software
- 36:37Play small for context window
- 44:18Final thoughts and Q&A
related talks
30mUsing AI to Create Better Requirements and Designs
30mWhen AI Takes Over: DevOps Nightmare Unleashed
45mAI's Bitter Lesson for SOCs: Let Machines Be Machines
01h 06mHow Claude Code Works
18mMastering Claude Code for Data Professionals
59mIt AI n't What You Think!
21mHow Anthropic Uses Claude
01h 06mHow I Tamed Claude