Back in 2018, Globo had what many companies would consider a nice problem to have: we had about 100 million users. Unfortunately, advertisers weren't particularly interested in buying impressions for "100 million humans of unknown characteristics"
https://t.co/hh3yzZ20ww
Interesting observation: GLM 5.2 for planning + GLM 5.1 for coding works surprisingly well on my personal projects. On my work projects, though, the quality drops a lot, so I end up using Opus for planning and Sonnet for coding instead.
Strange example in The Book of Why: using the remaining unvaccinated 1% as the counterfactual for vaccination. They're not comparable, they differ in healthcare access, income, and other confounders. That's selection bias, not causal inference.
I've built many versions of ML models over the years, chasing small performance gains, usually by adding new features. But one thing I almost never did was run a McNemar test to check if those gains were actually significant. How much of those "improvements" were actually real?
Reading Karpathy's autoresearch got me wondering: could it be used to improve tabular models? My guess is that it would shine more with MLP-based models, where it can explore architecture changes. For XGBoost, Optuna already does a great job searching the hyperparameter space.
Things that annoy me in the databricks notebooks: It now show what are the cells executing in the table of contents, it no show in what section i'm in on the table of contents, dont have the ctrl+a, ctrl+b and ctrl+m... i miss jupyter lab
@databricks notebooks are cool, but I really miss esc+a, esc+b and esc+m that I have in local jupyter lab. Sometimes when I want to be fast, I do the things locally with sample data, then I go to the databricks cloud to final execution
@Decolar O atendimento diz que a compra na passagem é na verdade uma compra de direito pra participar de uma corrida de checkin online. Só ganha o direito de voar quem faz checkin primeiro, os demais perdem o dinheiro e a viagem
Comprei passagem na @Decolar apesar das recomendações contrárias. Me ferrei. Deu ovebooking e vou perder minha viagem e eles não resolvem nada... Nunca mais...
For most production workloads on Kubernetes with HPA, a moderate target (around 60–75%) with realistic requests and no/loose CPU limits is usually the best default. A “90% target + high limits” strategy causes worse latency, slower reaction, and more noisy-neighbor risk.
No magic. No AI. Just me, a spreadsheet, and the growing realization that perhaps the robots weren’t replacing us that quickly. Then I ran the company’s shiny AI classifier and calculated precision and recall. The results? Recall: a painful 42%
https://t.co/AnXJKfGYIp
@gabrielskyz This is the 4th year I'm using it, and the first year that I did not have to change the code due to changing in the csv exported by B3. Maybe its a good idea to opensource it now