If you're interested in semantic data processing, you can also check out related systems like DocETL from Shreya Shankar et al., LOTUS from Liana Patel et al., and Palimpzest from Chunwei Liu et al. (4/4)

GitHub - anishathalye/semlib: Build data processing and data analysis pipelines that leverage the power of LLMs 🧠 Build data processing and data analysis pipelines that leverage the power of LLMs 🧠 - anishathalye/semlib

I've just open sourced Semlib: github.com/anishathalye.... (3/)

Semlib: LLM-powered Data Processing Semlib is a Python library for building data processing and data analysis pipelines that leverage the power of large language models (LLMs).

These operators are implemented in Semlib, a new library I built to help solve a class of semantic data processing problems that is underserved by current tools such as agents and conversational chatbots.

More on the story and use cases here: anishathalye.com/semlib/. (2/)

I replaced Python's built-in functions with semantic operators. (1/)

Current first-world problem: avoiding em-dashes so my writing doesn't look "AI-generated." Been using 'em since 2014, and yep—my blog posts are in Common Crawl. I guess I did this to myself?

Linearizability testing S2 with deterministic simulation How we validate strong consistency

Always fun to see when a grad school side project grows to the point where it's powering real systems! Nice blog post from the S2 developers on how they do linearizability testing with deterministic simulation (using, among other tools, Porcupine): s2.dev/blog/lineari...

Coding agents follow "it's always your fault" a bit too well.

When they hit a compiler bug, they'll rearrange methods, tweak syntax, and go in circles rather than consider the tooling might be wrong.

Had to manually identify and minimize this mypy bug after Claude Code couldn't.

If you have suggestions for topics to cover in the next iteration of the course, please share them in this thread!

Missing Semester Classes teach you all about advanced topics within CS, from operating systems to machine learning, but there’s one critical subject that’s rarely covered, and is instead left to students to figure out...

Lecture videos: www.youtube.com/@MissingSeme..., Notes: missing.csail.mit.edu

Missing Semester has grown past 100K subscribers on YouTube. Appreciate all the engagement and support!

We plan to teach another iteration of the course in January 2026, revising the curriculum and covering new topics like AI IDEs and vibe coding.

GitHub - anishathalye/neural-style: Neural style in TensorFlow! 🎨 Neural style in TensorFlow! 🎨. Contribute to anishathalye/neural-style development by creating an account on GitHub.

Incidentally, this is how I first got interested in ML. github.com/anishathalye...

My favorite way to measure progress in AI: finding papers obsoleted by ChatGPT prompts

GitHub - anishathalye/lumen: Magic auto brightness based on screen contents 💡 Magic auto brightness based on screen contents 💡. Contribute to anishathalye/lumen development by creating an account on GitHub.

Code/binary here: github.com/anishathalye...

Ever get blinded when writing code late at night and you alt-tab from your dark-mode terminal to your browser? Made this little macOS utility to solve this little problem, just updated for the latest macOS.

No thanks to AI for hallucinating BrightnessKit.framework.

GitHub - cleanlab/aiuc-workshop: AI User Conference 2025 - Developer Day workshop AI User Conference 2025 - Developer Day workshop - GitHub - cleanlab/aiuc-workshop: AI User Conference 2025 - Developer Day workshop

We did a workshop at AIUC that: (1) implements a RAG app on top of Cursor's docs, (2) reproduces the widely-publicized failure from last week, and (3) shows how to automatically catch and reproduce this failure. All slides/code are open-sourced here: github.com/cleanlab/aiu... (5/5)

What’s the solution? I believe that one ingredient will be intelligent systems that evaluate the output of these LLMs in real-time and keep them in check, building on and combining techniques like LLM-as-a-judge, using per-token logprobs, and statistical methods. (4/5)

Why do such failures occur? These next-token-prediction models are nondeterministic and can be fragile. And they’re not getting consistently better over time—OpenAI’s latest models like o3 and o4-mini show higher hallucination rates compared to previous versions. (3/5)

It’s been over a year since the well-publicized failures of Air Canada’s support bot and NYC’s MyCity bot. And these AI’s are still failing spectacularly in production, with the most recent debacle being Cursor’s AI going rogue and triggering a wave of cancellations. (2/5)

We reproduced (and fixed!) Cursor’s rogue customer support AI. (1/5)

I wonder if there's anything special in the Cursor Tab completion model or system prompt that induces this behavior.

Coincidence, or genius growth hack? Cursor self-propagating through developer set-up instructions.

Real-Time Evaluation Models for RAG: Who Detects Hallucinations Best? A comprehensive benchmark of evaluation models to automatically catch incorrect responses across five RAG applications.

2/2
It works surprisingly well in practice.

cleanlab.ai/blog/rag-eva...

Hoping to see more of these real-time reference-free evaluations to give end users more confidence in the outputs of AI applications.

Is AI any good at evaluating AI? Is it turtles all the way down? We benchmarked evaluation models like LLM-as-a-judge, HHEM, Prometheus across 6 RAG applications. 1/2

And some repos are even organically suggested by ChatGPT. (3/3)

Some of this might be through web search / tool use, but for at least some, knowledge about the projects is actually part of LLM model weights. (2/3)

A substantial portion of traffic for some of my open-source projects comes from ChatGPT these days. Sometimes even a majority, beating traffic from Google. Time to prioritize LLM optimization over search engine optimization. (1/3)