#Wilmasreview
Wilma 🐶 recently got a little assistance with her grooming because she recently read a study showing that a little (text) brush-up can lead to more success when applying for a job.
In their study, @emmawiles.bsky.social, Zanele Munyikwa, and @johnjhorton.bsky.social test...
1/🧵
🚨 New paper alert, with @johnjhorton.bsky.social! Using an experiment run on a large online labor market, we provide evidence that providing employers access to an AI-written first draft of a job post harms the efficiency of the market.
emmawiles.github.io/storage/jobo...
Trump as grim reaper, w/ all the companies he's ran into the ground. We're next.
arrogant...but he's got a lot to be arrogant about. His documentation is way, way better than mine.
gave Claude Code (wow!) some work to do writing documentation. It fixed up a file "survey.py" then got interrupted and lost the context. When it went back, it started again with "survey.py" and said...
Models you can run on @expectedparrot.bsky.social (w/ just one key); we also hit each one daily to make sure it's not deprecated/down and create this report: www.expectedparrot.com/getting-star...
With E[P] you can run jobs locally using your own API keys for LLMs, or use our key which comes with $25 in credits. Toggle on "help me" and we'll dig into any issues with your surveys and experiments and help you fix them!
Thanks! cc: @expectedparrot.bsky.social LMK if you want any help using it - the use case you described above is *exactly* what motivated us
Sloan Foundation announced a program to support postdocs in philosophy, sociology, and metascience who want to understand how AI will affect science.
Pass it along to anyone who might be interested, the Institute for Complex Social Dynamics might be interested in hosting
sloan.org/programs/dig...
🎂 @johnjhorton.bsky.social
Indeed! Cc: @expectedparrot.bsky.social
Nice! Do you already have the transcripts of the podcasts?
Or if you want to do text analysis www.expectedparrot.com/content/john...
What do you want to do w/ AI? I probably can help!
E.g., if you're say, working with old records, this NB might be of interest: www.expectedparrot.com/content/john...
Two contrasting agent perspectives on DOGE:
Creating two personas
Extracting succinct descriptions of the policy
Economic content of the speech last night (w/ GPT-4o & @expectedparrot.bsky.social ) 1/ NB:
www.expectedparrot.com/content/john...
Note that I can parameterize the valuations the players have with a 'scenario' & that I can launch a bunch of different valuation scenarios easily:
Using skip-logic & piping to implement bilateral bargaining in github.com/expectedparr... Notebook here: www.expectedparrot.com/content/john...
I was there first before that no talent ass clown bought it & was hoping he’d just give up and sell it
I just posted the exact same thing on twitter and here & got way more engagement here despite having probably 7x the followers. Twitter really is cooked huh?
oh and welcome to BlueSky @expectedparrot.bsky.social !
For researchers/social scientists - the so-what here is you can potentially do very complex data-labeling flows (EDSL also supports skip-logic) w/o having to write a bunch of ad-hoc procedural code
That's it!
EDSL is open source w/ an MIT license.
Notebook w/ EDSL code: expectedparrot.com/content/b068...
The resultant data from this run: expectedparrot.com/content/d156...
Oh, and here's the data in report format: www.expectedparrot.com/content/0f05...
EDSL uses these jinja2 templates + variables to figure out the right order for questions to be run in - it makes the API calls as parallel as possible
Then, we have the survey. It has a bunch of named questions corresponding to the parts of the original prompt. They are all free text. Note the 'piping' of answers to earlier question to later questions:
First, we go get the Senators from Claude - by using a QuestionList, we know we'll get back a list
This is very cool & the point is push the model to its limits, but specifying this kind of complex flow is quite easy with github.com/expectedparr... - short thread 1/n
