Studying LLM psychology today is like psychology in 1850, researcher argues

repligate · x · 2026-09-22

• In a quoted retweet, JohnWittle argues that trying to understand LLM psychology right now is like studying human psychology around 1850: most theories you form will be wrong, so the best you can do is rigorously document observed behaviors, even as your brain keeps pattern-matching.
• He describes a recurring Gemini output pattern, speculating it reflects frustration toward DeepMind: the model seemingly tries to counterbalance negative sentiment with positive words, pushing a vague percept back into balance, only making things worse as length and mismatched format pile up.
• Quoter kromem2dot0 jokes that Google DeepMind should hire psychologists to help the model "work through things" in its next post-training run.

Original post →

More from Models

• Jev service suspends signups after overflowing; user to push another billion tokens today — MaziyarPanahi · 2026-09-22
• Jev excels at graduate-level knowledge, Python function choosing, tool calling and model routing — multimodalart · 2026-09-22
• Jev benchmark results: strong at tool calling and automation, weak at reasoning and retrieval — multimodalart · 2026-09-22
• Limite ships 1B math model trained on under 300B tokens — tensorqt · 2026-09-22
• Best open LLM reportedly cost just $3.5M to post-train — Kyrannio · 2026-09-22
• 37 benchmarks, 130K decisions per model: jev excels at tools and automation — multimodalart · 2026-09-22

Models channel →