/compact cannot be trusted on its own.
using /compact with a prompt e.g "/compact focus on auth, saving state, forget the debugging we did." is the way
I have led a toothless life, he thought. A toothless life. I have never bitten into anything. I was waiting. I was reserving myself for later on—and I have just noticed that my teeth have gone.
Jean-Paul Sartre
Workflow engines like @temporalio are the real agents
Being forced to find the boundary between deterministic and non-deterministic tasks, minimizes hallucinations and improves self-consistency
cc: "The Log is the Agent" @yoheinakajima
While trying to take a decision, I ended up forgetting that the events that resulted in the decision were not independent. I realized that while attending an introductory about how LLMs are bad at causal inference, but good at making the building blocks (DAGs) that aid it.
using LLMs to code gets me 80% of the way
Bigger the change, lower the percentage
Doing multiple validation passes gets me close to the ideal output, 80% -> 96% (80+20*0.8) etc
Nothing beats a set of equivalence class tests
@DSPyOSS implementation provides some instant wiins
1. DSPy Refine - intelligent retries, policy enforcement
2. MultiChainComparison - self-consistency checks via voting mechanisms
Adding FewShot examples should also be a quick win
Still experimenting with GEPA, need 30 examples
Wrote up some flashcards and practice problems to help myself retain what @reinerpope taught.
Hope it's helpful to you too!
Suggest more below and I'll add them.
https://t.co/2gQcO24uIW