Claude's Max tier goes off-topic during an ML system design review
A user relied on Claude for designing a system. Starting with the medium tier, the model gave fewer suggestions but they were all correct.
After finishing the design, the user switched to the Max tier to get an overall review. At that point the Max model lost track of the context and kept bringing up chatbot-app technologies like SSE (Server-Sent Events) and , which have nothing to do with ML , feeding a string of misleading information until the user corrected it.
Key points
- Medium tier gave fewer but accurate suggestions for ML
- User switched to Max tier for an overall design review
- Max tier went off-context, bringing up unrelated SSE/ topics
- Misleading information continued until the user manually corrected it