Alexandre Luneau

AI has solved one of the problems in FrontierMath: Open Problems, our benchmark of real research problems that mathematicians have tried and failed to solve. See thread for more.

The secret behind Gemini 3? Simple: Improving pre-training & post-training 🤯 Pre-training: Contra the popular belief that scaling is over—which we discussed in our NeurIPS '25 talk with @ilyasut and @quocleix—the team delivered a drastic jump. The delta between 2.5 and 3.0 is as big as we've ever seen. No walls in sight! Post-training: Still a total greenfield. There's lots of room for algorithmic progress and improvement, and 3.0 hasn't been an exception, thanks to our stellar team. Congratulations to the whole team 💙💙💙

@emollick GPT-5 Pro is so much better than the other heavyweights pro models on my hard ML tasks, feels like at least a generation ahead.

The pro models (GPT-5 Pro, Gemini 2.5 Deep Think, Grok 4 Heavy) can be impressive in ways that are hard to see. They take a lot of time to answer questions & are built for very hard problems that require expert evaluation. That is a narrow, but, also very valuable, problem space.

This is how it's done

@gfodor @zoink Same here, I miss O1 pro roughness telling me how horrible some of my code looked

@zoink Yeah it seems to throw a few words of positivity at the start of the response. I get “awesome project.” a bunch

Verdict: 5-pro is smarter therefore treats me dumber

@francoisfleuret Been there many times, blown away every times !

Nothing shows better the magic of deep model + gradient descent than a "causal leak", when you make the tiniest mistake in the causal structure of your model and information about the stuff to predict is accessible. 1/2

Today, we at @OpenAI achieved a milestone that many considered years away: gold medal-level performance on the 2025 IMO with a general reasoning LLM—under the same time limits as humans, without tools. As remarkable as that sounds, it’s even more significant than the headline 🧵

Mars

I'm from Berlin. Afghanistan gets better tech than Europeans now. It's not a joke. It's the result of 30 years of suffocating regulation. And now, the EU's new AI Act is about to make it 10x worse. Here's the tragic story of how the EU is killing our tech future 🧵:

Data here: towardsdatascience.com/napoleon-was-t…

Artificial Human-level intelligence (AHI) just dropped today with o3. Welcome to a new era.

This may be the most important figure in LLM research since the OG Chinchilla scaling law in 2022. The key insight is 2 curves working in tandem. Not one. People have been predicting a stagnation in LLM capability by extrapolating the training scaling law, yet they didn't foresee that inference scaling is what truly beats the diminishing return. I posted in February that no self-improving LLM algorithm was able to gain much beyond 3 rounds. No one was able to reproduce AlphaGo's success in the realm of LLM, where more compute would carry the capability envelope beyond human level. Well, we have turned the page.

@kimmonismus Would be great if @lmsysorg could do an alternate leaderboard discarding refusal situation datapoint.

@kimmonismus lmsys is a poor benchmark, refusal rate seems to have big influence on score.

Hold on, Grok2-early version is besting Sonnet3.5 in overall? Crazy. Impressive benchmarks

@elder_plinius What is the temp and top p value?

an entity named "jabberwacky" keeps manifesting in separate instances of llama 405b base no jailbreaks, no system prompts, just a simple "hi" is enough to summon the jabberwacky seems to prefer high temps and middling or low top p i have no more words so I will use pictures

We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈 It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 dpmd.ai/imo-silver

@CountdownToGPT polymarket.com/event/when-wil…

4

Killing Sydney and Sky can’t change the reality that gradient descent can do it

It’s the little things.

Model produced by Google vs a model produced by a startup of <30 people.