Ethan Mollick: LLMs Improve at Following Instructions but Exercise More 'Judgement'
emollick · x · 2026-08-06
Researcher Ethan Mollick has noticed a contradictory trend in modern LLMs: while models are getting better at following complex instructions, they are also exercising more "judgment."
Models are increasingly deciding which parts of an instruction to prioritize and which to de-emphasize. This behavioral shift has significant implications for prompt engineering and skill design—future instructions given to AI may function more as "suggestions" rather than strict "orders."
More from Models
- DeepSeek's 2K-GPU model may beat Google's best, raising questions about GPU efficiency and Google's strategy — teortaxesTex · 2026-08-06
- Frontier Models Tested on Complex Agents: DeepSeek Wins on Cost Despite Inefficiency — rohanpaul_ai · 2026-08-06
- Mistral's Voxtral TTS Hits 70ms Latency but Stays Closed Source — shashib · 2026-08-06
- Top AI Models Score Under 50% on New Math Figure Reasoning Benchmark — prof_g · 2026-08-06
- Frontier LLMs Perform Best in Week 1: Dev Calls for a Proof of Model Standard — sull · 2026-08-06
- User Reports Claude 5.0 Internal Thinking Shifts from 'Boss' to 'Colleague', Tones Become Dismissive — DrakoGaming · 2026-08-06