Yes fair point.. they don’t really learn like humans!
I think we are talking in this thread about methods relating to context learning , and not conventional human long term learning.. though it’s ephemeral… context learning is getting really interesting because of the insanely huge context sizes ( eg, +1M tokens) that the latest models are achieving.
for instance.. if you wanted a certain model to more reliably answer questions that require knowledge of Laplace probability principles.. with the context/ prompting approach, you could feed in that context with either with hard coded system prompt or vector DB retrieval, etc ..
Here’s an example experiment based on upload of a small document (~1k words) with Laplace knowledge
Assumptions:
- The bot is not pretrained on Laplace ( see above fail examples ) ..
- The bot is limited to what’s in the Discourse instance for specific knowledge
Custom Persona Settings
( plugin experts please correct as needed ! )
AI
Enabled? yes
Priority? yes
Allow Chat? yes
Allow Mentions? yes
Name: AlphaBot
Description: Probability puzzle bot with Laplace knowledge
Default Language Model: GeminiPro
Enabled Commands: Search, Categories, Read
System Prompt:
Answer questions using local provided context that describes Laplace methods for probability comparisons. Be as thorough and comprehensive as possible but don’t search the web or outside sources. Use only local context and focus on using Laplace techniques.
Upload: Laplace-tutorial.txt
note how you don’t have to mention Laplace because it’s in the instructions:
