I haven't used ChatGPT for a month now and haven't missed it. Today felt like a good day to cancel my subscription.

We are hiring Members of Technical Staff (Research Engineers)! Current LLM agents lack reliability, creating a gap between demos and production. We solve this by automating the complex workflow of debugging, evaluation, and iteration required to make agents robust. 👇

- Iran is in a humanitarian crisis. - Thousands are reported dead in 72 hours. - We are past the point of solidarity. Empty words do not stop bullets. Action does. - The world must intervene now.

For years, Iran’s masked plainclothes regime thugs have abducted and murdered citizens with absolute impunity for wanting prosperity and refusing to fear them. Officials call it law enforcement and smear protesters as paid agents of the state’s enemies. This must end. Iranian people must prevail!

Found myself repeating this to several students at NeurIPS: When you’re choosing an internship or a job, what you work on and who you work with matter way more than the logo. Don’t optimize for brands. Become the brand!

Hiring researchers & engineers to work on –building reliable software on top of unreliable LLM primitives –statistical evaluation of real-world deployments of LLM-based systems I’m speaking about this on two NeurIPS workshop panels: 🗓️Saturday – Reliable ML Workshop 🗓️Sunday – LLM Evaluation Workshop

Woke up to this email this morning - Wow, I won a NeurIPS award?! - …runner-up, but I’ll take it. - Wait, I didn’t submit a paper. - Ah, I’m chairing the session and I’m supposed to give the award. Huge congratulations to the actual winners and runners-up!

If you're at @neuripsconf.bsky.social on Dec 6, don’t miss our panel session at @unireps.bsky.social with Ahmad Beirami, Sara Hooker and more to be announced! 🚀

Will be at NeurIPS Thu Dec 4 to Sun Dec 7, excited to reconnect with old friends and make new ones. If you are excited about AI engineering (orchestration, evals, and optimizing scaffolds), we are hiring! On Saturday I’ll be on panels at the Reliable ML & UniReps workshops.

Once you see a math concept geometrically, it becomes much easier to think about, and it’s hard to go back to any other way of seeing it.

I am sorry for what many of my excellent former colleagues are going through. Layoffs can be emotionally challenging for everyone, whether you are directly affected or not.

The math that LLMs can do today is novel enough to be considered publishable, but it's not the kind of math that would be consequential.

My thoughts on the broken state of AI conference reviewing. www.linkedin.com/feed/update/...

Let's regress from here to AGI!

This is the conclusion slide of a talk I gave more than a year ago on RL/Alignment! It still holds true today.

This also applies to telling your story (e.g., in a CV, bio, interview, etc). Focus on what you have accomplished and what you are excited about doing next; not just where you did it!

The actual unpopular opinion is that the notion of senior and junior authors should be abolished. It has completely diluted the notion of scientific authorship and created this entire industry of free-riding, head-in-the-clouds, incompetent PIs/managers. List down exact contributions instead. [+]

I occasionally get messages asking how to follow my path and get into Meta, DeepMind, or similar places. That is the wrong question. Do not focus on the brand! Focus on what you want to work on, then find the opportunity that fits your goals best.

Related to this if a paper turns out to have a major error in it, you’re supposed to throw yourself under the bus not your students.

Corollary: If you lack bandwidth or expertise to act as the verifier, then you shouldn't sign up to be the senior author of a paper!

Unpopular opinion: When a paper has a senior mentor and a junior mentee, the senior author must make sure the claims are correct and well supported. They must check every claim and gate the submission until it meets that bar.

This is the recipe for many provable claims: Make enough assumptions and narrow down the claim, then prove a narrow result with caveats. Present it as broad, hide the caveats, and declare “XYZ is provable!”

Today, Industry research is focused on short term (3-6months) bets. Academics have an opportunity to balance their portfolio with medium term (1-2 years) and long term (5-10 years) bets. Putting all academic efforts in short-term basket is suboptimal!

When I worked in corporate, I was often first in the office because that routine worked for me. It was a personal preference, not a benchmark for anyone else. We should not judge commitment by hours, especially in research. We should look for thoughtful work and steady progress.

Common mistake in LLM prompting projects: jumping into full-scale pipelines (datasets and inference) without testing feasibility. Iterating at scale is expensive and time-consuming. Start with ONE example to validate the hypothesis, verify context, debug the design, then scale.

Thoughts that are explained clearly are more respected. Clarity is a scarce skill. Many of us (me included) leave out key context and make people work too hard to understand us. AI models should get better at this: not just reasoning, but communicating with the right amount of context.

Enjoyed speaking with @DeltaInstitutes about going from information theory to ML, recent safety alignment/RL work, and lessons on RL for LLMs that stuck! Check out the podcast episode here: lnkd.in/eb6dWHDv

In 2009, a prominent signal processing professor said the market was tough and h-index ≥6 was needed just to get a faculty interview. We now seem to be drifting toward the same bar for PhD program entrance.

Congratulations to the Google team on the release of the newest Gemini Image generation model! 🍌🍌 I am super impressed with what the model did here (no other model gets even close -- including Google's previous model). This is truly bananas!

What are the founders going to own? 🤔

Cannot believe this needs to be said loud: Do not ride a motorized bike 30+ mph on a mixed-use trail. It’s beyond reckless and puts walkers and cyclists at grave risk! P.S. “Electric” doesn’t make it safe at 30+ mph.

My thoughts on acceptance rate artificially kept low: If we focus on merit based acceptance with -claims are substantiated with theoretical/empirical evidence -claims push the science envelope by epsilon Then we still end up with ~25% acceptance rate and won't need to artificially reject any papers

also works for teaching. when I taught information retrieval, lecture 1: The Information Access Problem, lectures 1-2: evaluation. the rest of the semester presents algorithms and methods leaning on those first lectures to analyze and compare. should work for many courses

Being yelled at by both sides of an argument often only means either of two things: - You have a rational stance that balances all the nuances that either side is missing. - You have gone completely crazy.

When we have a good verifier for a capability, we already know how to distill that capability into the model through RL. In most cases, building the verifier is the bottleneck!

Concretely, a research project proceeds in this order: 1. Problem definition 2. Evaluation metrics and success criteria 3. Solution Too often, steps 1 and/or 2 are skipped or treated as afterthoughts.

Too often, researchers propose solutions before they’ve defined the problem first.

A research project needs test beds of different granularity. Iterating at the large scale is expensive and hard to debug Validating on smaller models is helpful in moving fast by ruling out ideas that are unlikely to work My alignment research is driven by thinking about a ternary language model

Most “robustness” work (adversarial, shift, etc.) is just training on reweighted samples (augmented, model-generated, or mined). OOD generalization then comes from: (1) inductive bias (2) similarity to train data (3) luck The 3rd one is the most important of the three.

Impactful papers that *shift* the paradigm are generally met with resistance. A *public* review of Shannon's paper publication: "The discussion is suggestive throughout, rather than mathematical, and it is not always clear that the author’s mathematical intentions are honorable."

None of our impactful papers have had an easy path through traditional venues. Most cited paper? Rejected four times. Most impactful paper? Poster at a conference. But none of it matters because arxiv makes everything work

pass@k is an effective decoding paradigm to improve models at math and also for an attacker to jailbreak the models. Now that we know we are decoding from the model with pass@k or an adversary is jailbreaking the model with pass@k, how should we think about RL? A short 🧵

This is how offline calibration + PPO compares against GRPO on helpfulness BT rewards. Would be curious to see how this might help your usecases.

We are wired to see our own best and others' worst. Maturity is flipping the lens: Be a generous fan of others and a tough editor of yourself! Self-awareness is underrated!

The best AI researchers zoom at three abstraction levels: - High: paper-level ideas & math - Mid: code-level implementation - Low: GPU/TPU reality (kernels/memory) Low exposes bottlenecks. High accelerates exploration. Mid makes it real. The job is to translate between them!

Early 2020 at Meta, we were building an assistant agent using SOTA 2B LLMs with tool APIs (messaging/calling/timer/etc). We were still addressing coreference resolution, named entity recognition, robustness to paraphrases, etc. Amazing how much progress has been made since!

I had an interaction yesterday where someone basically called me stupid. While I didn't appreciate it, I neither blocked nor reported this person, and simply moved on! Today, I realized the person is "suspended". Now I wonder whether this was the reason, and if yes I think this is not healthy!

I’m still receiving several inquiries each week from folks looking for full-time roles and internships on *my team at Google*!

Post-training research was fueled by the KL-regularized RL mathematical foundation. That led to a lot of algorithmic research and a ton of progress over a few years. This helped us learn how to "distill" metrics back into models. But today we are optimizing workflows/agents.

Leaderboards lack granularity! When we say model A is better than B at text, it is too coarse. Even when we say A > B in instruction following, it is still too coarse. I am apparently a tail user as I find my personal ranking drastically differs from leaderboards. We need more holistic evals!