Hubert Drazkowski

It should be a review not a rebattle.

A plain question about explanation even with LLMs at hand can create huge separation between a cohort students. Esentially, how well a person prompts. To what extent a persons is aware of the width and depth of explanation possibilities and consequences. What is this a proxy of?

It is not a rebbutal it is a rebattle.

@burkaru @yudapearl Really nice paper. I would love to see more microeconomics with causal reasoning flavor.

I took a UCLA class on causal inference taught by @yudapearl more than 10 years ago. Recently together with my co-author Egor we applied causal DAGs to persuasion. I believe it is worth a read: causality.pdf share.google/XFwcoxZMdlHmiE…

PAC-Bayes

Controversial names in science list

I try not to be affected by weather, not talk about weather, not chase the weather. It is just weather.

Believe me or not, but one day I have heard from someone with a Full Professor (working with data) say along the lines of „I don’t believe in karge sample studies, because in large sample you can find anything.”

Causal inference

Greeks + Vega

Start from the questions and go back to the beginning with the slides challange.

Polish universities should also do a semester of research exchange abroad, say top students, scholarship. Connections.

Does an IV on a snapshot of autoregressive process ever exist? …exclusion restriction what?

The light is always greener on the other side of the crosswalk.

When you do a phd in 6 years you should have 2 papers, when you do a phd in 3 years you should have 10.

Accent within a sentence - good, within paragraph even better, within a story, invaluable. Accent within a word? Are you insane or what?

“How much” is a generalization of “how many”.

Is being old like being young, but sleep deprived, just all the time?

Te EEG proof of no free will is no proof of anything. Consciousness is not free will. I am waiting for the EEG predicting the answer/decision to a question before it is being asked.

To be fair we should evaluate CVs from a debiased/masked (potentially personalized) generative model on the same level set multiple times ~ not the factual ones. The big problem is the model < human still in many aspects: intangible information, reality modelling etc. At that point the machine could do it end to end. Any in-between solutions?

Often people jump and use specialized language with every word loaded with context, definitions and properties behind it, assume the convention of “standard”. Looks professional, it is compressed and efficient locally, useful. Globally that is not the point.

To have fencing of what it is not and what it could be in the paper/presentation; to have all of the assumptions and their alternatives. Hard, but crucial.