Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

Training is not the same as chatting: ChatGPT and other LLMs don’t remember everything you say

May 29, 2024 - simonwillison.net
The article discusses misconceptions about the training process of Language Learning Models (LLMs) like ChatGPT. Contrary to popular belief, these models do not learn or memorize everything said to them. They function as stateless calls, only considering the current conversation and the user's new prompt. Starting a new chat effectively wipes the model's short-term memory. The training process involves pre-training on large amounts of text data and reinforcement learning to improve conversation quality. Once trained, the model remains static and unchanged.

However, the article also highlights concerns about LLM providers potentially using user data to improve their models, as their terms and conditions often allow this. While some providers offer opt-out mechanisms, these are often opted-in by default and can be confusing. The article also mentions the introduction of "memory" features in some LLM tools, which allow them to remember small details for use in future conversations. Misunderstandings about how these models work can lead to misguided policy decisions and legislation.

Key takeaways:

  • ChatGPT and similar tools do not directly learn from and memorize everything that users say to them. They function as stateless function calls, starting each conversation from scratch.
  • There is a common misconception that 'training' these models involves them learning from each conversation, which is not the case. Training involves a lengthy and complex process of identifying patterns in large amounts of text data.
  • While these models do not learn from individual conversations, many LLM providers do have terms and conditions that allow them to improve their models based on usage, which can lead to privacy concerns.
  • There is a risk of policy decisions being made based on misconceptions about how these models work, potentially leading to measures that address invented rather than genuine risks.
View Full Article

Comments (0)

Be the first to comment!