#thecompaniesapi improving results from phi3 is a tedious task, but if it pays off we get rid of a ton of problems

#thecompaniesapi continue work on phi, lots of iterations, kinda exhausted lately

#thecompaniesapi few iterations on our ai flows, trying phi3

#thecompaniesapi continue working on phi3 ; first steps of subdomain support & association ; rewrite most of our AI impl

#thecompaniesapi continue on phi3; hell of a refactor

#thecompaniesapi work on phi3 again :) continue my search for the best prompt

#thecompaniesapi continue r&d on phi3, trying idea refinement from structured text content, then pipe to inference flow, unknown territories 🤣

#thecompaniesapi run my phi3-128k flow using llama3.1 and the results are mind blowing, it's insane how good llama is at conserving context and original purpose even when supplied with thousands of tokens; also shipped multiple hotfixes in robot UI; about to merge a month of work and then hop on fine tuning

#thecompaniesapi tests are over; phi3 is what we need; dropping lots of python code in favor of a single inference process to govern them all

#thecompaniesapi quantize phi3.5 to 4bit to use it in our inference server; same model size but 128k context length instead of 4k, I can now process huge chunks of texts without relying on batching

#thecompaniesapi port back my ai wrapper improvements on our existing llama stack and prepare fine tuning for phi3

#thecompaniesapi port back my ai wrapper improvements on our existing llama stack and prepare fine tuning for phi3

#thecompaniesapi watch the data being extracted for our customer ; switch back to phi3 branch again

trying fine-tuning #supabox

change fine tuning strategy and data structure and play with new fine tuned model #reva

#thecompaniesapi lunch with @LeCoupa and try new prompt techniques with phi

playing around with speeding up tools to run faster. I think anything over 200ms is going to make me rethink any/all usage

Work on 3rd prototype. #jovial

Work on 3rd prototype. #jovial

#thecompaniesapi improve the dx of our AI flows so we have faster feedback loop in development, noticed it was making me lazy that i had to wait for app rebuild to rerun everything