Simon Loewe

Economist and Data Scientist at the Burning Glass Institute | Labor economist and applied econometrician | PhD in Economics from @PSEinfo | Ex-physicist

[Simon Loewe @LoeweSimon]

@MaximeRivest If I go through 3rd party services like groq, there are DSPy native options for making inference as fast as possible right?

[Simon Loewe @LoeweSimon]

@MaximeRivest Tbh rn, it is bc I am only starting on these big pipelines, and I haven't gotten around to it. Which is a terrible answer :D And tbf there are some tasks where I felt that o3/o4-mini was necessary, but that is probably bc I haven't tried enough optimization with things like DSPy

[Simon Loewe @LoeweSimon]

@MaximeRivest Not to mention the speed-up as well. At that number of requests BatchAPI is much much faster than synchronous prompts.

[Simon Loewe @LoeweSimon]

@MaximeRivest Makes total sense. For prompt optimization, I don't use batch, but for full runs I need to. I often want to run prompts on millions of examples and there a 50% price difference is the difference between affordable and unaffordable!

[Simon Loewe @LoeweSimon]

@MaximeRivest Thanks a lot, these are super useful. With all of that, I am still baffled though why BatchAPI is not supported as I assume that this would be extremely easy to do and cost savings of 50% is a lot. Rn I would like to use DSPy for a project, but without the cost savings, I can't.

[Simon Loewe @LoeweSimon]

@MaximeRivest That's cool! Only starting to look at DSPy but I am assuming that simply extracting the optimal prompt from the optimization process and sticking into the OAI API will not cut it because DSPy does a whole lot of behind the scenes structuring?

[Wharton Human-AI Research @WhartonHumanAI]

That’s a wrap on the #Wharton AI and the Future of Work Conference!

Over two days, we explored how #AI is reshaping the workplace...from org design and labor markets to innovation, policy, and generative AI.

Follow us for more at the intersection of AI & business. #WhartonAI

[Simon Loewe @LoeweSimon]

@antoine_chaffin @LightOnIO This looks super cool! Has somebody tried to see how it does in the SBERT benchmarks (https://t.co/KCVdpuZQOR)?

[Simon Loewe @LoeweSimon]

@SamuelMullr This is very cool, congrats, but the 10K limitation is definitely restrictive... All the data I have worked with is way bigger than that.

[Simon Loewe @LoeweSimon]

@economeager @ben_golub Out of curiosity, can I ask why? For instance compared to Casella and Berger or other options?

[Sahil Chinoy @sahilchinoy]

i'm on the job market with a paper about political sorting in the labor market! it's perhaps harder to choose your coworkers than your friends or who you follow on twitter -- could the workplace be less segregated than other places where democrats and republicans tend not to mix?

[Simon Loewe @LoeweSimon]

@Kaju_Nut Though I guess less the RG stuff tbf

[Simon Loewe @LoeweSimon]

@Kaju_Nut Daniel Harlow's New lecture notes go in that direction I would say.

[Alex Imas @alexolegimas]

Not a comment on study (which isn’t linked) but I used to work the night shift (8pm-3am), entering data of manufacturing firms from hard copies.

I thought I’d be more productive if I listed to industrial & trance music super loud on my headphones. And I was! Doubled my entry rate and it was fun.

My boss came by three days later. I had a big smile on my face thinking he was there to compliment me.

He let out an big sigh. “Alex, I don’t know what you’re doing back here but your error rate is like 80%”.

I was fired by the end of the week.

Lesson for me: things that may seem like they make me more productive might…not.

[Simon Loewe @LoeweSimon]

@arieda_muco Thanks for the interesting post, was looking into this right at the moment. I'm not entirely convinced though, because as far as I can tell (maybe I'm wrong), now I don't really own the remote. eg if I want to make the full repo public, it seems a bit weird. Might be wrong! thx

[Shengwu Li @ShengwuLi]

DNC: No one can take apart a car and reassemble it into two copies of itself.

Wapo fact check: This is false. By carefully slicing the car into non-measurable sets and applying the axiom of choice...

[Stanford AI Lab @StanfordAILab]

arXiv -> alphaXiv

Students at Stanford have built alphaXiv, an open discussion forum for arXiv papers. @askalphaxiv

You can post questions and comments directly on top of any arXiv paper by changing arXiv to alphaXiv in any URL!

[Americans For Tax Fairness @4TaxFairness]

BREAKING: The IRS has recovered $1 BILLION in unpaid taxes from millionaire tax cheats.

This is what happens when you fund the IRS.

"Now that we are resourced, we can do the job of ensuring that they pay."

Anyone trying to cut IRS funding just wants to protect rich tax cheats.

[Vincent Pons @VinPons]

Les candidats arrivés en 3e position qui, par intérêt personnel et calcul médiocre, refusent de se désister, sont individuellement responsables d'une possible victoire RN.
Electeurs, il faudra être stratégiques et voter pour le candidat non-RN le mieux placé.
#legislatives2024