Ryan Wesslen

Economics might seem -- from the outside -- like it's about competition. But really it's about creating the miracle of cooperation, where folks from all around the world enrich your day in a million tiny ways. It's that beauty that I'm worried about losing.

behold the CONNECTED SPLATTERPIE

kneel before my works, ye mighty, and despair

“Responsible AI” is a bad word at NIST now.

Editorial: A president just disrespected America in the Oval Office. It wasn’t Zelensky It’s time to say it plainly. America’s leadership has switched sides in the war. The American people have not, and they should speak up. In the past several weeks, the U.S. leadership has demonstrate...

NEW: A searing editorial in the The Kyiv Independent.

“It’s time to say it plainly. America’s leadership has switched sides in the war. The American people have not, and they should speak up.

“A president just disrespected America in the Oval Office. It wasn’t Zelensky.”

@kyivindependent.com

New paper <3
Interested in inference-time scaling? In-context Learning? Mech Interp?
LMs can solve novel in-context tasks, with sufficient examples (longer contexts). Why? Bc they dynamically form *in-context representations*!
1/N

New paper in Political Analysis on synthetic text data for training classifiers. Main idea: generate training examples with LLMs, then fit classifiers on synthetic (+real) text. Paper has validations and guidance.
Blog: andrewhalterman.com/post/synthet...
Paper: www.cambridge.org/core/journal...

What are some of the things you've learned about how LLMs (and LLM-powered systems like ChatGPT) work that were non-obvious but most helped you build a more effective mental model of how to use them?

Zero-Cost Custom Feeds on Bluesky A simple stack for generating custom feeds for Bluesky programmatically without a backend server

Wrote down the process to build your own custom feeds for Bluesky programmatically in Python and run it 100% free

Uses @skyfeed.app + @github.com actions to do periodic filtering and re-ranking and @cloudflare.social static pages to provide data to @bsky.app

There’s something fundamentally wacky going on here. That a) there’s an arrow of time in language is cool but b) that the magnitude varies a lot between languages? Wacky.

What's new with ML in production What's different about LLMs versus traditional ML

New post: Have been meaning to write something around what has fundamentally changed around the process of putting ML into prod now that we have LLMs. TL;DR: It's still just compression, we just don't control as much anymore.

vickiboykis.com/2024/01/15/w...

Prodigy in 2023: LLMs, task routers, QA and plugins · Explosion We have made a ton of new updates in Prodigy this year with v1.12, v1.13, and v1.14 releases. So we decided to write a post about them.

In 2023, we rolled out Prodigy v1.12-1.14 packed with new features like spacy-llm integration, prompt engineering, QA support like IAA metrics, task routing, and new plugins such as PDF and Hugging Face 🤗.

We highlight the many updates in our new blog post 🎉

explosion.ai/blog/prodigy...

The PyData NYC video from "Half hour of labeling power: Can we beat GPT?" by @ryanwesslen.bsky.social & me is now live!

We show how to use LLMs to speed up annotation, collect 1.2k examples & beat our baseline.

📺 Video: www.youtube.com/watch?v=Ta45...
📝 Slides: speakerdeck.com/inesmontani/...

Deck the halls with NLP! @explosion-ai.bsky.social's💀 t-shirts, tot bags, or mugs are sure to bring a touch of skeleton swagger this holiday season

Hello NYC ✨ Looking forward to seeing everyone at PyData NYC next week!

We'll have an @explosion-ai.bsky.social booth again with brand new swag, and @ryanwesslen.bsky.social & I will start the conf with our tutorial "Half hour of labeling power: Can we beat GPT?": nyc2023.pydata.org/cfp/talk/WQY...

Announcing ✨Prodigy-HF ✨

It's a new plugin that allows you to train @huggingface.bsky.social NER models directly on annotated data in Prodigy. It also provides a recipe to upload annotations to Hugging Face HUB!

Interesting attempt to categorize LLM hallucinations (from an international team including Stanford and Amazon) arxiv.org/abs/2310.04988

We've recently introduced Prodigy Plugins which extend the features of Prodigy by adding direct support for 3rd party integrations. One of these plugins is P... Prodigy-PDF for PDF annotation and OCR - Prodigy Shorts

I added support for PDFs for Prodi.gy in the past few weeks. So I figured I'd record a small demo.

www.youtube.com/watch?v=rwyz...

If folks are interested in working on detection models for PDFs -> let me know if there's recipes missing!

I wonder if people can guess what plugin I'll add next to prodi.gy.

We recently released ✨ Prodigy-ANN ✨ that allows you to use contextual search to find relevant subsets of data to annotate first.

To help explain this new feature, @koaning.bsky.social made a small demo to highlight the new feature 👀

youtu.be/jyu2nbjwfXw

Prodigy interface with PDF-Prodigy OCR extension example

The new OCR feature uses Pytesseract under the hood to attach parsed text to segments that you've annotated with the `pdf.image.manual` recipe.

If you want to learn more, the docs have plenty of extra info: 👀

prodi.gy/docs/plugins...

An example output of Prodigy's inter-annotator agreement command with annotation and agreement statistics.

Curious if your annotators are on the same page? Prodigy has just released v1.14.3 with built-in inter-annotator agreement (IAA) metrics to track and measure their agreement. In this 🧵, we'll review Prodigy's document-level IAA metrics. prodi.gy/docs/metrics