4th Grade - Dr. Julius Sumner Miller - Demonstrations in Physics!
#physics #televison #PBS

www.youtube.com/playlist?lis...

Hey Danny, there are at least two of us in Dallas now. Want to start a Roo Meetup?

"forces" is doing a lot of work in that booklet

If that piece of the prompt falls out of context, for whatever reason, you lose tool_use specificity

This is why all of major providers have a "tool=[]" array as one of the parameters for the API call. Prompt injection tool_use only gets you so far

In fact, we think that a new metric in the SWE Bench should be Instruction Following,

You do not suck at prompting, some LLMs just like to ignore us occasionally.

Even if we constantly remind LLMs to do X is gives us a big - NOPE. :-)

One of the biggest offenders is Gemini - if we tell it to do 1 simple binary thing every prompt - "Do not write comments for code you write – EVER" it still will do it 7 times out of 10. But... still writes pretty good code

Nope, same experience on the OSS agentic programming plugin - Roo Code, we regularly swap in / out new models as they roll out daily and find the same thing.

Less hallucination (since we have full control of context) but forgetting instructions is a very big deal.