iiinova — where your digital self keeps evolving

Yoav Artzi @yoavartzi.com · 05/11/2025

It's a Thursday though ....

Yoav Artzi @yoavartzi.com · 03/11/2025

Pushed a big update to LM-class (v2025.2) -- this second version makes a much more mature resource Many refinements of lecture slides + significant improvements to the assignments Many thanks to @ch272h.bsky.social, Yilun Hua, and Shankar Padmanabhan for their work on the assignments

140

Yoav Artzi @yoavartzi.com · 28/10/2025

Cornell is recruiting for multiple postdoctoral positions in AI as part of two programs: Empire AI Fellows and Foundational AI Fellows. Positions are available in NYC and Ithaca. Deadline for full consideration is Nov 20, 2025! academicjobsonline.org/ajo/jobs/30971

031

Yoav Artzi @yoavartzi.com · 22/09/2025

Joey Ramone wrote a song dedicated to his stock investor My contribution to your ICLR deadline soundtrack (or if you are looking for investment advice ¯\_(ツ)_/¯ ) : www.youtube.com/watch?v=vbJx...

110

Yoav Artzi @yoavartzi.com · 08/07/2025

@colmweb.org decisions are out, and so are we The strength of submissions this year amazed us! Many many hard decisions 😩 + Aditi, @eunsol.bsky.social , @ranjaykrishna.bsky.social 😴😴😴

0221

Yoav Artzi @yoavartzi.com · 04/07/2025

This particularly has problematic side effects

070

Yoav Artzi @yoavartzi.com · 03/07/2025

Tokenization is still an open problem

080

Yoav Artzi @yoavartzi.com · 23/06/2025

Sounds like RAG all the way...

010

Yoav Artzi @yoavartzi.com · 18/06/2025

ChatGPT gives some numbers, but who knows how reliable they are

000

Yoav Artzi @yoavartzi.com · 12/06/2025

@colmweb.org discussion period ended, and now we entered the meta reviewing period. It's not simple getting reviewer engagement nowadays, but slow and steady we made descent progress, even if shy of the ideal of early and wholly. Progress bars -> 🧵

110

Yoav Artzi @yoavartzi.com · 31/05/2025

I would worry about the rate of false positives. The emergency system is already overloaded, and in general not efficient. My activity tracking is not exactly what you want to measure, but it shows everything is last minute

210

Yoav Artzi @yoavartzi.com · 31/05/2025

I am not convinced this is the best way to compute this, but here goes. Definitely n per reviewer is too small to get anything meaningful

100

Yoav Artzi @yoavartzi.com · 29/05/2025

There are many reviews in the middle. These are per review

110

Yoav Artzi @yoavartzi.com · 29/05/2025

Starting to track daily participation for @colmweb.org

071

Yoav Artzi @yoavartzi.com · 25/03/2025

Help me understand why it's different than "my multi-step decision process model is not perfect (i.e., biased), so it introduces bad/biased decisions into the output"... because I don't know if I want to start using "sabotage" instead of "error" in my papers

1101

Yoav Artzi @yoavartzi.com · 05/03/2025

It's now public! My postdoc call is for the inaugural postdoc as part of this $10.5M gift for a new AI fellows program at Cornell. There's a lot more in this program, so more exciting things to happen here real soon! news.cornell.edu/stories/2025... Application: forms.gle/tiydAChgV1wL...

0112

Yoav Artzi @yoavartzi.com · 28/02/2025

Why choose when you can have both?

020

Yoav Artzi @yoavartzi.com · 19/02/2025

No surveys here, so I ran this on X. So this brings about the next question: why are we still doing two columns in some venues?! Or tiny fonts in others? We need templates that fit how people read: fonts, columns, margins, etc

4140

Yoav Artzi @yoavartzi.com · 18/02/2025

Oh, and, yea, there are lovely scaling behaviors as well, both on performance and stability frontier! All credit to @giomonea.bsky.social In collaboration with @xkianteb.bsky.social and @abosselut.bsky.social

000

Yoav Artzi @yoavartzi.com · 18/02/2025

A lot of the heavy lifting of in-context learning (ICL) is thanks to the semantics of labels (i.e., "positive" for sentiment is much more meaningful than "label1"). Can you ICRL without any semantics? Yes, you can! -> solid lines, or even better with a few exemplars -> dashed lines

100

Yoav Artzi @yoavartzi.com · 18/02/2025

First, we got ICRL to work out of the box without any specialized inference methods. It is actually just a matter of sampling and prompt construction. So, (good) LLMs can RL in-context! Purple💜 curves ...

120

Yoav Artzi @yoavartzi.com · 18/02/2025

We recently pushed an update to our in-context RL paper. Usually, updates don't justify a post, but this one is exceptionally contentful -> 🧵 tl;dr: all the findings are stronger, and the behaviors are super cool! arxiv.org/abs/2410.05362

1184

Yoav Artzi @yoavartzi.com · 10/02/2025

(roughly) Copying @docmilanfar.bsky.social 's post from X to here, because this is an exceptional talk by Michael Jordan: www.youtube.com/live/W0QLq4q...

180

Yoav Artzi @yoavartzi.com · 09/02/2025

I don't have the time to do research along different segments, but quickly found this index for construction costs. Line items like building depreciation should account for this increase to be realistic. www.mortenson.com/cost-index?u...

200

Yoav Artzi @yoavartzi.com · 06/02/2025

Nothing makes you more hopeful about the future than helping COLM! Please ack positively to your AC/reviewer invites! Naomi already did ❤️

090

Yoav Artzi @yoavartzi.com · 30/01/2025

The world has gone mental. Jiayi Pan does a (very!) cool experiment showing R1-like pipeline in a very restricted domain (this is amazing, and the right way to conduct such experiments). Then it escalates pretty quickly to a $30 reproduction 🤦‍♂️ I guess the VCs already signed over the billions

1183

Yoav Artzi @yoavartzi.com · 10/01/2025

In case this didn't work, ChatGPT preemptively generated a nonsense alternative as well ¯\_(ツ)_/¯

010

Yoav Artzi @yoavartzi.com · 10/01/2025

Can someone create pbcopy-file? Don't disappoint ChatGPT Seriously though, is there a way to copy a file to the clipboard so I can paste it somewhere in finder? The file itself, not its content or its path.

320

Yoav Artzi @yoavartzi.com · 05/01/2025

London started in 2003. Asking a 29 year old about the difference…. 🙄

180

Yoav Artzi @yoavartzi.com · 18/12/2024

Again, this is by (the infinite) far our best call for workshops!

110

Yoav Artzi @yoavartzi.com · 17/12/2024

This is so far our best call for papers! We completely re-thought what a CFP is, and.... mostly copied last year's while surgically increasing some numbers

3152

Yoav Artzi @yoavartzi.com · 06/12/2024

I don't know if odd, but definitely opaque Training from scratch 👇

020

Yoav Artzi @yoavartzi.com · 06/12/2024

Is Llama3.3-70B trained from scratch? It sounds like it huggingface.co/meta-llama/L... So what's the cause of the delta from 3.2? Is online preference optimization just RLHF with an on-policy algorithm (~PPO)?

370

Yoav Artzi @yoavartzi.com · 22/11/2024

I am very excited about this work, and how strong the results are. Relying on natural and implicit interactions signals is challenging, but a serious game changer Part of the excitement is because this continues a research thread that started with my **first** paper from 2011 👴

1192

Yoav Artzi @yoavartzi.com · 22/11/2024

No. Instructions become less diverse (and maybe more templat-ish), but they remain well understood by humans (actually better understood, because performance goes up) This is the older papers where we saw this first: arxiv.org/abs/2108.04812

030

Yoav Artzi @yoavartzi.com · 15/11/2024

So this happened while I was already on my early flight back home :) Congratulations Omer!

0220

Yoav Artzi @yoavartzi.com · 13/11/2024

40 faculty lines there, 50 there ….

270

Yoav Artzi @yoavartzi.com · 13/11/2024

An absurd way to chill at EMNLP is the nearby dog park: maps.app.goo.gl/QsKuEy7wYSjA... Bonus: it’s full of tiny lizards running around 🦎 there was a lot of fun and excitement

180

Yoav Artzi @yoavartzi.com · 08/11/2024

Pretty interesting AI outline in the new GScholar extension. They really tried to make it a navigation aid, rather generating what would appear to be replacement summaries

130