Hasty Briefsbeta

The LLM Lobotomy

5 hours ago
https://learn.microsoft.com/en-us/answers/questions/5561465/the-llm-lobotomy
Copy Link
  • #LLM degradation
  • #Azure criticism
  • #model accuracy
  • The author is working on a product using Azure for LLMs and Audio Models, testing conversational flows with system prompts.
  • Over six months, the same LLM model has shown decreasing accuracy in responses despite using identical messages and prompts.
  • The author notes that newer models like gpt-5-mini and nano are slower and less accurate compared to previous versions like gpt-4o-mini.
  • Microsoft is suspected of intentionally degrading older models to push users towards newer, albeit inferior, versions.
  • The author criticizes this strategy, emphasizing the importance of accuracy and consistency, and considers moving away from Azure due to instability.
  • The author has documented proof of these issues and urges Microsoft to either improve products or maintain stable, backward-compatible services.
AboutLogin