making Claude 3.5 Sonnet to come up with a pytorch code which finds a model which computes the ahrefs DR impact of a backlink. With manual exploration I came to a correlation of 66%. 5 minutes of Claude -> pytorch already pushed it up to 72%

Claude 3.5 just blows my mind sometimes. It's incredibly intelligent naturally. I am going through final tests on a new App I'm building to generate, clean, and enrich datasets. I was testing the enrichment portion last night and it just knows things. Like it understands when cleaning data and I'm asking it to add thousand separators that some of the data may not be a number in the traditional sense. I had a dataset that has population, median_income, and zip codes in it, but it only modified the population and median_income data without the need to specifically prompt for it. This might seem trivial, but its ability to reason without guidance amongst large sets of data is quite remarkable. #datamachine

trying to find a better data science model for backlink analysis - spent around 16h for a promising new version - which turns out is worse than before 🙁

reverse engineering ahrefs DR log scale - my manual approach was MUCH better than 2h of claude loops. Interesting: data science is another thing which works better by hand (or with external help) than with AI

#topbubbles rafactor model updates

predictive analysis #consulting

Tried Claude instead of ChatGPT, much better at code but still not perfect #blogrecorder

continued on my ahrefs DR reverse engineering. Around 10h investment, old model: correlation: 89%, new model: correlation 69% 😭

Full refacto of AI predictions and credit management #melies

fine tuned data cleaning to handle the limits on Claude 3.5 Sonnet Context Window.. now continues processing where it left off and we've added the ability to fact check and complete, incomplete datasets. I really want to add a diff view instead of just clobbering existing data, but we're getting into scope creep. Can come after launch. #datamachine

#thecompaniesapi work on datasets filtering & classification UI ; we can now review outputs from Claude/GPT and directly grab the dataset to send for fine-tuning

beta model running #s2c

keep struggling with the attribution model #reva

👨‍💻 re-create the "is the car damaged?" from kaggle | train the model on 64 picture and accuracy of the model is 92% 🤯

ask chatgpt to do regression analysis but unsure about its results because what do I know? #yelmair

try new noise model #photoai

continue experimenting with different models #ai

eval model

play with GPT-4 to compare against fine-tunned model and cry because it does it better but is expensive as hell #reva

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