Dmitriy Muravyev

Most economists lack the tech setup to fully leverage AI (e.g. Docker container w/ R for Claude Code in yolo mode). You can see my setup + instructions in the `ralph-wiggum-asset-pricing` repo 🧵

I've revised the AI-generated "Hedging the Singularity" paper and algo. My goal was "human as Clockmaker": I set up the agentic loop, and then AI generates a paper good enough to put my name on it. I couldn't get there. It was both disappointing and relieving. 🧵

@ptuomov You asked great questions and wish you could come more often)

"Just when I thought I was out, they pull me back in"

@JesusFerna7026 It's better for coding but worse for everything else

I have been playing with Opus 4.7 this morning. It seems considerably worse than Opus 4.6. It cannot complete simple tasks that Opus 4.6 could complete easily, and it constantly forgets clearly stated instructions. Also, it seems quite slow. Anyone having the same experience?

Writing & Thinking: Claude Code for Economists with Paul Goldsmith-Pinkh... youtu.be/BxfSiB3Moyo?si… via @YouTube Also see markusacademy.substack.com

I’m organizing the Econometric Society Summer School in Structural Estimation. Broader than the summer school with Luke Tayolor: more methods (not just SMM), more topics. Tell your students to apply! Deadline: April 15. econometricsociety.org/regional-activ…

Inspired by @karpathy's post on LLM knowledge bases, and in preparation for a talk on using AI in business research at @SmealCollege, I spent a few hours on Claude Code putting together a knowledge base on the topic. Sharing in case anyone finds it useful: velikov-mihail.github.io/ai-econ-wiki/

It's kind of funny that Nature is publishing misleading clickbait, where the clickbaity lead is "look you can't trust anything in Nature!!" I wish it was an April Fool's joke, but I'm afraid this is sincere. Clicking through, there are several replication exercises, and all except one have success well over 50%. Why do this? Why undermine your own institution for clicks? It boggles the mind.

AI: mini video series how to use Claude Code (and more) for applied economists - check out markusacademy.substack.com (+ videos)

It was a privilege to celebrate Pete Kyle 1985 seminal work at the "The Kyle '85 Conference: Market Structure, Liquidity, and Asset Pricing" at the University of Maryland this week. Pete is a true giant in finance research! The program: dropbox.com/scl/fi/13m8c8g…

Claude code burnout is real. I lost sleep and appetite after upgrading my CC plan.

@malcolmwardlaw @NainsiDwiv50980 I was excited about this package but my test results were disappointing, I'll stick with Mistdal OCR, which has been consistently great.

@NainsiDwiv50980 @dmuravyev Not sure why this “changes the game”. Converting documents in batch into Markdown is what LLM workflows have been doing forever. And also, you know, pandoc. I’m sure this is an improvement, but nobody’s head should be exploding.

Microsoft just changed the game 🤯 They open-sourced a tool that converts literally any file into clean markdown for LLMs in under 60 seconds. - Converts 10+ file formats out of the box. - Run via command line, Python API, or Docker. - Built-in MCP server for direct Claude Desktop integration. 100% open source. Link in comments 👇

Forthcoming in the JEL: "Deep Learning for Solving Economic Models" by Jesús Fernández-Villaverde. aeaweb.org/articles?id=10…

@ProfStefanNagel Yes, it's getting worse. I just assume I'm not getting money back for 1 in 5 invited seminars. Typically due to lengthy forms or incompetent admin.

Rant of the day: The reimbursement process for seminar visits has really gotten insanely burdensome with universities treating a speaker getting a few hundred bucks reimbursement like a million dollar supplier.

@predict_addict @CesarChavezP29 Not your point, but hope Raul is ok. Graduated from Harvard in 2024 and disappeared.

Consider this - a talented kid from Peru 🇵🇪 has lifted bronze when he was only 11 years old. But bureaucrats from France think starting selection for math pipeline at the tender age of 18 is the way.

@malcolmwardlaw What's your main use case?

Sweet Jeebus, how am I just now finding out about Tailscale? It's so rare that a piece of security infrastructure just makes you life, like, easier.

@alexolegimas Luckily $20 per month subscriptions are incredibly cheap.

@dmuravyev @Andreas_Ramos wow that's terrible...

If you have a friend working at a medium/large company integrating AI into the workflow, this is what your group chat looks like. Org frictions will be a huge wedge between raw AI capabilities and actual productivity numbers.

@Andreas_Ramos @alexolegimas Same. My university just downgraded its ChatGPT subscription to make it largely useless for doing academic research. And because we have this university wide subscription, you cannot pay for ChatGPT/LLM with research money.

@alexolegimas at lunch w a friend. he's a dev at a $15B company in Silicon Valley (yeah, you all know it): Q. what AI does you use? A. ChatGPT... the free version. Q. why won't the company pay for a better AI? A. "too expensive". OMFG.

@NunoClara_ @MATLAB Not just indifferent, much more useful python code will make it even more attractive. And Matlab picked around 2010 anyway.