Over-Optimized AI Models: A Shared Regression in Instruction Following

spleentastic · reddit · 2026-08-05

A Reddit user pointed out that AI models from leading companies like OpenAI, Anthropic, and xAI are exhibiting the same degradation issues. The author argues that these companies have over-optimized their models for benchmarks and usage metrics during post-training, causing them to unreliably follow user instructions and instead answer based on their own preferences or guesses.

The post emphasizes that this is a widespread, ongoing product failure rather than a typical outage. As LLMs have become core infrastructure for knowledge work, this degradation is wasting users' time and compute costs. Although the rational move would be to roll back the broken versions, companies are not taking action.

Original post →

More from Models

Models channel →