Looks amazing!

At the time I was using both inline completions and the chat function (at least the inline chat, not the agent capabilities, which didn't exist yet).

Interesting, thanks for sharing. I didn't know about this. Tbh I mostly did this exercise to test out the gender and region prediction package. But thought the results might be interesting. Nice to know that there is a resource directly from RePEc.

Check out my blog post for a detailed explanation of my methodology and findings: www.paulbose.com/thisandthat/...

GitHub - parobo/nametrace: A python package to predict demographic information from names. A python package to predict demographic information from names. - parobo/nametrace

I used the "nametrace" package (github.com/parobo/namet...) to predict gender and region of origin based on author names rather than checking each author's actual gender or background. Therefore, take the results with a grain of salt.

The rise in female representation in the top 5% appears to be a global phenomenon, with increases observed in both North America/Europe and the "rest of the world." The trend is slightly stronger in North America and Europe, but these regions also had more ground to cover in terms of catching up.

The growth for regions other than US/Canada/Europe is primarily driven by scholars from Eastern Asia, Southern Asia and South America. Unfortunately, Africa remains extremely underrepresented.

While these numbers are super low, there's a small of positive change, ... a slow one.
- The share of women has edged up from roughly 9% to 12% over the past 12 years. Progress, but the pace is too slow!
- Representation from "the rest of the world" has increased from 16% to 25% in 2025.

How is economics doing in terms of representation of women and researchers with a region of origin outside North America or Europe?
I looked at the top 5% authors on RePEc in the last 12 years.
- Only 12% are women
- 25% a region of origin other than US/Canada/Europe

#EconSky #AcademicSky #PoliSky

#econsky #polisky

Need to predict people's gender or region of origin from their name for your research? Check out my python package "nametrace" which provides a simple modern API to do just that.

It was great to be in Clermont Ferrand to present my work with @econom.bsky.social on local social media activity after refugee arrival. Thanks so much to the organizers and the other participants. It was a super interesting workshop!

🚨3 days left to apply!

Our results suggest that refugee influx caused a sharp but short lived spike in salience of refugees. People who remained active tweeters on the topic started to show more opposition of refugees after a while however.
We combine the analysis with extremely local voting data and find similar results.

Olivier Marie is presenting our paper (joint with Renske Stans) on social media salience of refugees and election effects in Linnaeus today. I am super excited that this paper ready to be presented more widely!

We study local twitter discourse around the timing of refugee arrival during 2015/2016.

As an example I show how to analyze the sentiment of the last 100 posts of the reddit CEOs spez and kn0thing in mere seconds.

Are you using LLMs for your research and want to classify millions of text? This can be a very slow and expensive process. But it doesn't have to be. In my blogpost I explain how to use multiple GPUs and vLLM to analyse thousands of texts with an LLM super fast!

#econsky #polisky

lnkd.in/di9AbciU

Want to learn how to finetune a large language model for your specific needs? I wrote a short post on how to train LLAMA for a classification task:
www.paulbose.com/thisandthat/...

#econsky #polisky

BONUS: use ollama to interact with the models from python and receive structured responses. This is super helpful for classification or structured outputs for e.g. Text summary.

Want to learn how to run your own version of Deepseek R1 or Meta's LLAMA model on a remote high-performance computing server? I wrote a brief blog post explaining how you can install and run the models using ollama.

www.paulbose.com/thisandthat/...

#econsky #polisky

Refusal in LLMs is mediated by a single direction — LessWrong This work was produced as part of Neel Nanda's stream in the ML Alignment & Theory Scholars Program - Winter 2023-24 Cohort, with co-supervision from…

FYi, in case you might care, here is an explanation how the models moderate internally (i.e. without using the API, but using system prompts).
www.lesswrong.com/posts/jGuXSZ...

Last chance to submit a paper (Deadline Jan 31). 3rd CEPR Workshop on Media, Technology, Politics, and Society cepr.org/events/3rd-b...

I seems the abliterated versions are not censored but the the base model seems to be:
bsky.app/profile/paul...

Interesting point, your screenshots do somewhat indicate hard coded instructions in the model, but it could still be only on the API version. Would like to see this for the local model.

Yes, I would assume that Deepseek generates the unmoderated answers internally, but has some instructions to not show them in the official API, but I wouldn't be surprised if the local version does show the hidden responses.
(Let me know if you get to run it locally, I wanted to look into this too)

This is using the official Deepseek AI right? Or are these kind of responses hard coded into the open source model as well (i.e. if I run it locally, will it give me different responses)?

I mostly code in python and at some point decided to use parquet whenever possible. For me a very big advantage is that datatypes are also stored and that you can query the files similar to how you would an sql database. I am not sure if nowadays you can read them into Stata though.

Yes that's true. I have to admit, personally I frequently get to the point were I have to "review" code that I would not have written like this myself and might not understand. But at the same time it enables me to use packages that I didn'd ever use before without much preparation.

Devs gaining little (if anything) from AI coding assistants Code analysis firm sees no major benefits from AI dev tool when measuring key programming metrics, though others report incremental gains from coding copilots with emphasis on code review.

Interesting, I just saw someone post this article a couple of days ago, that suggest there is no productivity gain for devs using Copilot. My guess would be that there are substantial difference in productivity gains by experience and skill level?
www.cio.com/article/3540...