#thecompaniesapi early LLM results are bonkers but the qop/s are relatively slow; setting up results storage to train smaller models; one step at a time

#thecompaniesapi mind blown by early LLM pipeline results ; being a noob in such subject is so fascinating

#thecompaniesapi start a new customer dataset extraction, also tried to run our new LLM on a way smaller GPU with success :)

#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 lots of improvements on our robot, now track per-job metrics, also track input/output tokens for all our AI queries, started collecting dataset on sample domains for our fine-tuning

#thecompaniesapi lots of fixes on the queue ; trying to get the llm stable ; refactor of sources

evaluate an LLM experiment to clean up my blog #aiplay #blog2

study Stanford’s Alpaca instruction model finetuning for LLaMA source data #life

Fix fast LLM model failing with Error 400 and add a fallback to other model in case it errors #jobician

nlp pipeline refactoring work 6 hours #consulting

#thecompaniesapi add storage pipeline for our new extraction feature; add endpoints to our storage server to automatically build fine tuning datasets from our production extraction

Read [Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters](arxiv.org/pdf/2408.03314)

improve #mirage SSE API so that we can stream more than LLM output chunks and provide JSON data for new secret #crisp product

listing pipeline #alertbnb

Upload pending datasets to S3 (Data Science course) #zg

got actually working dataset for clv #ml4all

Ran some local LLM tests 🤖

Chunk dataset big files and import to DB (Data Science project) #zg

cache recs for longer to save on model inference costs #pango

#thecompaniesapi working on llama3.1 with large context window, trying to improve some parts of our AI flow before merging it; now properly respects industries & naics constraints for current sample