Mark Chen

We wrapped up this year's competition circuit with a full score on the ICPC, after achieving 6th in the IOI, a gold medal at the IMO, and 2nd in the AtCoder Heuristic contest!

we shipped a new version of 5.3 instant to chatgpt yesterday. 5.3 was unintentionally pretty annoyingly clickbait-y. it's better in yesterday's model and we're going to keep stamping that behavior out. keep the feedback coming! help.openai.com/en/articles/68…

Insane how leaky OpenAI is smh

@0thernet Just use Codex. That might have been a single prompt and worked within your $20 sub

If you give GPT-5.4 a raw dump of the GPT-2 weights and ask for a <5000 byte C program to inference it, GPT-5.4 succeeds in under 15 minutes! I remember working on a similar exercise to compare results against a proprietary model in a previous paper - it took days!

research has really been cooking with gas lately. feels like playing with a great orchestra

Some thoughts on AI and mathematics, inspired by "First Proof."

Very excited about the "First Proof" challenge. I believe novel frontier research is perhaps the most important way to evaluate capabilities of the next generation of AI models. We have run our internal model with limited human supervision on the ten proposed problems. The problems require expertise in their respective domains and are not easy to verify; based on feedback from experts, we believe at least six solutions (2, 4, 5, 6, 9, 10) have a high chance of being correct, and some further ones look promising. We will only publish the solution attempts after midnight (PT), per the authors' guidance - the sha256 hash of the PDF is d74f090af16fc8a19debf4c1fec11c0975be7d612bd5ae43c24ca939cd272b1a . This was a side-sprint executed in a week mostly by querying one of the models we're currently training; as such, the methodology we employed leaves a lot to be desired. We didn't provide proof ideas or mathematical suggestions to the model during this evaluation; for some solutions, we asked the model to expand upon some proofs, per expert feedback. We also manually facilitated a back-and-forth between this model and ChatGPT for verification, formatting and style. For some problems, we present the best of a few attempts according to human judgement. We are looking forward to more controlled evaluations in the next round! 1stproof.org #1stProof

Just because we focus on research doesn't mean we will pursue *all* research. Jakub and I have our own research tastes as well (I think we're fairly good tastemakers 😅), and we have to importance weight the various paths to AGI.

How does OpenAI balance long-term research bets with product-forward research fundamentals? I’ve been getting this question a lot lately, usually framed as a suggestion that Jakub (@merettm) and I are pushing an increasingly product-focused agenda. That characterization is simply wrong. Foundational research has been core to OpenAI from the start, and today we run a research program with hundreds of exploratory projects - much like the ones that led to our reasoning-model breakthrough. The majority of our compute is allocated to foundational research and exploration - and not product milestones. Anyone who has spent time with me or Jakub knows we are the last people in the world who would push for the advancement of products over the advancement of research. We’re in the business of creating an automated scientist, and capabilities that were considered grand challenges just a few years ago (like IMO-level mathematical reasoning) now emerge as normal parts of the research process. We’re also seeing our models accelerate researchers worldwide, helping advance work across biology, mathematics, physics, and even our own research. Jakub and I put a lot of effort into ensuring that research stays focused on uncovering algorithms that will scale to the compute we’ll have a year from now. We protect mindshare and amplify discourse on exploratory work. We do this while recognizing that we’re also a deployment company - and that deployment gives us access to even larger-scale compute, richer feedback, and more room for exploration. Our researchers are passionate about having their work out in the world, and a special slice of our org is dedicated to making sure our deployments are delightful for end users. Our goal isn’t to turn research into a quarterly race. It’s to build a durable research engine - one that compounds learning over time and consistently turns long-horizon exploration into real, measurable advances, while ensuring those advances become valuable in the real world. That’s the roadmap we’re executing on. And while there have been ups and downs over the last decade (as you expect with any research program), I think most of our researchers would share my strong optimism today.

Erdős Problem #635 autonomously resolved by GPT-5.2 Pro. The model thought for just 50 mins, outputting a correct proof in Latex, then formalised in Lean by @HarmonicMath's Aristotle. Big thanks to @AcerFur for cleaning up the Lean. Literature review is ongoing.

Excited for @mehtaab_sawhney to join the team! I strongly believe that accelerating Mehtaab will also meaningfully accelerate mathematics!

Watch Cursor build a 3M+ line browser in a week

Congrats to @bchesky and @Ahmad_Al_Dahle! Excited to see what new Airbnb experiences are unlocked when outstanding design meets excellent ML.

Our Aleph agent driven by @OpenAI's GPT-5.2 was able to identify misformalizations in the PutnamBench, and obtained the highest performance 99.4% to-date. 5.2 is extremely powerful.

Either OpenAI has a team of leprechauns and top mathematicians working 24/7 on FrontierMath questions, or GPT-5.2 Pro has actually become that good at mathematics. I can hardly find any non-trivial hard problem that the model cannot answer after 1–2 hours of interaction. Singularity is near…

We finally had a moment to run our system with GPT-5.2 X-High on ARC-AGI-2! Using the same Poetiq harness as before, we saw results as high as 75% at under $8 / problem using GPT-5.2 X-High on the full PUBLIC-EVAL dataset. This beats the previous SOTA by ~15 percentage points.

If AI could interact and learn from the physical world, could it make more scientific advances? We had GPT-5 optimize molecular cloning protocols in the wet lab. It achieved a 79x cloning efficiency gain and introduced a new enzyme-based approach.

@GregKamradt Some of the smartest people I know are incredibly data inefficient.

If you’re so smart, why do you need so much training data?

@BenjaminDEKR I agree that anything that feels like an ad needs to be handled with care, and we fell short. We’ve turned off this kind of suggestion while we improve the model’s precision. We’re also looking at better controls so you can dial this down or off if you don’t find it helpful.

OpenAI's @markchen90 says they are looking into the not-great "shop at Target" frustration. Giving them the benefit of the doubt to improve it. Mark always seems genuine to me. My 2 main points are really just: - This is essentially an ad - Let users opt-out or turn this off

@BenjaminDEKR Dang, sorry Benjamin! We’ll look into this!

I'm in ChatGPT (paid Plus subscription), asking about Windows BitLocker and it's F-ing showing me ADS TO SHOP AT TARGET. Yeah, screw this. Lose all your users.