Chains vs Agents
Discover more from 🕵️♂️ ABC's for AI Agents
Notes from the webinar:
What is an agent?
While loop - while you’re iterating - use the language model you check to see what step to take and if its done or not
ReACT type of stuff - the action is a finish action is when you’re done.
Grey area - using a router in the middle
Plan an execute agent ->
Use the language model to determine what actions to take
agents = planning + memory + tool use
Autonomy - limited form of intelligence - trust this thing on its own and do things on our behalf.
What is a chain?
- pre-determined set of steps or sequence of events
Swyx and smol developer:
Link to slides: https://docs.google.com/presentation/d/1L_CHsg26sDxPmKj285Ob5T2xsAUejBlfiGQSnsSHTk0/edit#slide=id.p
The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT
Memory at its most superficial level is a tool. These become so specialized that they deserve a special place in the stack
AutoGPT is very well known but not practical
recommend: Tech’s two philosophies - ben thompson (Stratechery)
TRICK: babyAGI - create diagrams with your code base
5 different levels of automated driving → Human driver vs AI Driver
Engineers being able to wield AI more effectively makes them more productive
Biggest challenge: whole program coherence - shared_dependences.md and name variables
Markdown is great for prompts
code injections
structure
Planner is too primitive but read the voyager edition from nvidia - https://arxiv.org/abs/2305.16291
2020 - 2022: Open AI Playground - completions
End of 2022 - Chat
2023 - chat with Bing ← chat + tools/skills
March: Chat with plugins
April/May: Autogpt/babyAgi
Chat with plugins/Bing
plugin usage that isn’t single iteration
There’s some kind of planning (for-loop) with a task queue and task usage
Short term gains from doing feature engineering
trade off - capability vs efficiency
just pay for open ai
What prompting technique is best for having agents use tools?
Few shots prompting