On episode 996 of "Is AI coming for your job too?" We present, Agent Bain vs. Agent McKinsey. A New text-to-SQL benchmark for the business domain (CORGI): https://t.co/Cr7L40rJRk Joint work among Cornell Bowers CIS, Cornell Johnson School of Management, & Gena AI (@YueeeLi_

If for some Finger Lakey reason you happen to be in Ithaca, come to my talk at the Cornell Johnson School of Management. Sage Hall noon 21st Oct. https://t.co/8HSSisgP9Q Data Access for Everyone with AI For everyday people, databases are a foreign concept. There is a

read our paper: https://t.co/lMdUnKHpjF contribute to our open source benchmark and data: https://t.co/bbK7uS7zg1 evaluate and submit online: https://t.co/Cr7L40rJRk send questions to: rt529@cornell.edu or unsojo@cornell.edu

A few interesting insights**: 1) AI is relatively bad at identifying latent patterns in data like seasonal sales trends 2) AI is worse at giving data-informed recommendations (future) than data-informed explanations (past). 3) AI relatively bad at giving advice on planning 4)

Our new CORGI benchmark features 10 hand-curated databases representing modern enterprises spanning retail e-commerce such as Lululemon, DTC enablement like Shopify, and c2c like Airbnb and TheRealReal. Our databases are jacked 😳 with more tables and relations than seen in

People, it's the data.

This is my lecture from 2 months ago at @Cornell “How do I increase my output?” One natural answer is "I will just work a few more hours." Working longer can help, but eventually you hit a physical limit. A better question is, “How do I increase my output without increasing

Our work serves as a case study of how applying domain expertise can reduce LLM cost -- Using generic LLM methods bluntly results in unnecessary expenses. Leveraging knowledge of SQL and databases can level the playing field for efficient applications.

Preprint is available here: https://t.co/F0kmxLIo4U Open source package available here: https://t.co/umRgA9CdmX [soon to be integrated to PyPI txt2sql]

With 14B Qwen + N-rep you can get better performance than o3-mini. Go from 46 cents per query to 3.9 cents for comparable performance.

Using multiple schema representations, you can reduce the number of calls required in self-consistency that are sometimes 100+ LLM calls!

Is your text-to-SQL AI worth the cost? We introduce "N-Rep," a new faster and 10x cheaper approach for text-to-SQL without chain-of-thought reasoning or expensive fine-tuning. Joint work with @genadotco @andreawwenyi @dmimno 🧵

Adding @_KarenHao @sheonhan @restofworld as this might be of interest!

Special thanks to @Xianbao_QIAN for helping brainstorm and @tsmullaney for inspirational book!

China is a nation with over a hundred minority languages and many ethnic groups. What does this say about China’s 21st century AI policy?

This suggests a break from China’s past stance of using inclusive language policy as a way to build a multiethnic nation. We see no evidence of socio-political pressure or carrots for Chinese AI groups to dedicate resources for linguistic inclusivity.

In fact, many LLMs from China fail to even recognize some lower resource Chinese languages such as Uyghur.

LLMs from China are highly correlated with Western LLMs in multilingual performance (0.93 - 9.99) on tasks such as reading comprehension.

Do Chinese AI Models Speak Chinese Languages? Not really. Chinese LLMs like DeepSeek are better at French than Cantonese. [ https://t.co/22ZGZZI1Us] Joint work with @andreawwenyi @dmimno 🧵

Feels so odd to be teaching a history of AI when it's very actively being contested in 2025, though time will tell what gets written!