🚀 AutoGPT Roadmap - Knowledge acquisition 🎓

[Pwuts]

Objective 🎯

AutoGPT should be able to find, collect, and aggregate information more efficiently than a skilled human.

SWOT 🧭

Strengths 💪

• Access to cheap LLMs → inexpensive bulk processing
• LLM often has prior knowledge it can use to choose an effective angle of inquiry

Weaknesses 🚧

• Reliance on Selenium for viewing/scraping pages: Selenium+Chrome devours RAM
• LLM Hallucinations

Opportunities 🚀

• Parallel exploration/aggregation
• Vector search/filtering/grouping
• Open-source Knowledge Gathering Engine has market potential

Threats ⚠️

• Outcome of LLM operations can change significantly with the same prompt when switching models
• Many websites have anti-bot systems or patterns that can hinder robotic accessing, especially parallelized access
• Reliance on rate limited / CloudFlared APIs can decrease parallelized capacity & reliability
• Processing many/long web pages → $$$

Related 👪

• 🚀 AutoGPT Roadmap - Web interaction 🦾🖥️ #6967

Actionables

• Integrate OSit
• Integrate Perplexity
• Implement parallel exploration on inquiry
  • Scrape and process multiple web pages in parallel
  • Bonus idea: Use a (temporary) embedding repo to select the most relevant pages/sections
  • Bonus: Select most relevant links using semantic sort/search and scrape those as well
• PoC knowledge memorization
  • Once objective knowledge has been acquired, this could be cached in a local memory store
  • Care must be taken not to re-use cached knowledge (e.g. “current stock price of stock X” will go out of date very quickly)
• Expand memorization capabilities to user scope
  • Rather than starting fresh with every task, we could allow sharing/re-using knowledge between tasks