Programming has quietly turned into a practice of making micro-wishes to a genie. The art is making the wishes in the right order in just the right way to eventually get what you want.

Continuing the journey of optimal LLM-assisted coding experience. In particular, I find that instead of narrowing in on a perfect one thing my usage is increasingly diversifying across a few workflows that I "stitch up" the pros/cons of: Personally the bread & butter (~75%?) of

Introducing the Environments Hub RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI

We’re releasing new Codex features to make it a more effective coding collaborator: - A new IDE extension - Easily move tasks between the cloud and your local environment - Code reviews in GitHub - Revamped Codex CLI Powered by GPT-5 and available through your ChatGPT plan.

Calling all #NLProc student researchers! 📢 The EACL Student Research Workshop 2026 is on the horizon. Get ready for key updates and announcements. We'll be posting them right here. Stay tuned! eaclmeeting #EACL2026

Stay tuned!

i think we reached some sort of threshold where i can use these language models to do large sweeping refactors like, before they could only add code but now they can remove code. that's new

it's kind of hard for me to stress this i'm one shotting -500 +600 pull requests like this is a intermediate engineer muscle replacement

"LLMs aren't good for programming" "yeah, I guess", i say, as I merge the third +98 -380 refactor of the day at 8:12 am

By popular request: you can now branch conversations in ChatGPT, letting you more easily explore different directions without losing your original thread. Available now to logged-in users on web.

I think congrats again to OpenAI for cooking with GPT-5 Pro. This is the third time I've struggled on something complex/gnarly for an hour on and off with CC, then 5 Pro goes off for 10 minutes and comes back with code that works out of the box. I had CC read the 5 Pro version

Many of you do not know the trauma of having to write out Java on paper for the comp sci AP exam and it shows

|￣￣￣￣￣￣￣￣￣￣￣￣￣￣￣￣￣| |Don't Push To Production On Friday| |＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿| \ (•◡•) / \ / —— | | |_ |_

If you're not watching EuroBasket today, you don't like basketball.

> be Google in 2017 > small team drops “Attention Is All You Need” on arXiv > execs nod politely, go back to selling ads for socks > let Transformer gather dust for 5 yrs like a vintage Beanie Baby > be Noam Shazeer, OG wizard > quits, builds AI-boyfriend app

I finally got around to making a tool to compare completions from SFT vs. RLHF trained models. This is a mini site for the RLHF book that I've wanted for a while. rlhfbook dot com slash library It's always been hard to say what RLHF does to a model within a more complex

Training on messy instructions makes LLMs more robust and often better on clean prompts too. Llama 70B trained with 100% noisy instructions leads BBH under noisy tests, and also reaches the top on MMLU. The problem is that small wording changes can flip answers even when

Cool new work by Ahmed Alajrami (w/ Xingwei Tan)! 👉🏼 Instruction-tuning on noisy instructions doesn't harm, and in some cases, improves LLM performance 👉🏼 LLMs don't really care about the instruction so much, they can infer what they should be doing from the input examples?

Just published in TMLR! 🚀 ElChat offers a more robust way to adapt chat LLMs to new languages using only unlabeled data. No base model needed. Better performance and chat abilities. 📄 openreview.net/forum?id=6IdoI… Work done with もりし@YANS浜松 Aline Villavicencio Nikos Aletras #NLProc #TMLR

New paper on efficiently adapting chat LMs on raw (non-chat, IT) target language data