Long-read asks: is Anthropic training the weirdness into Claude?
MaziyarPanahi · x · 2026-09-13
A multi-part article questioning why Claude models score great on evals yet still feel strange to use: constant babysitting, careful prompting, output checking.
Key points:
- Cites a spicy moment (23:25) in a Dwarkesh episode where speakers argue Sonnet/Opus are worse than GLM and Kimi despite claimed distillation of logits from Mythos — clearly labeled as their take, not fact.
- Their explanation: access to a teacher doesn't cover how it handles all messy real-world requests; you still need many diverse, realistic prompts. Router and proxy data might help Chinese labs there, though the author stresses their training data is unseen and he won't treat it as fact.
- The author then goes on his own investigation.
Related event: Developers question why Claude benchmarks well but feels quirky in use(2 posts)→
More from Models
- ARC Prize 2026 closes in 3 days as Yi-Chia Chen tops ARC-AGI-3 at 75.56% — GregKamradt · 2026-10-31
- Fable 5.5 Max stuns with gorgeous SVG of a pelican riding a bicycle, headed to Claude Code — ChrisGPT · 2026-10-31
- OpenAI reportedly to train on users' Daybreak sessions with no opt-out — dyn___ · 2026-10-31
- NVIDIA Nemotron 3.5 scales LLM self-verification, +9-10 points on agentic benchmarks — sarahcat21 · 2026-10-31
- Mistral Large 4 scores 94% on 1,273 USMLE-style questions but flips 44 at temp 0 — MaziyarPanahi · 2026-10-31
- Researchers find 5 personality structures in LLM weight space, Big Five-like — luke_drago_ · 2026-10-31