OpenAI boasts about success of reinforcement learning: is AGI now close?

OpenAI CEO Sam Altman raised the eyebrows of many when he said that artificial general intelligence (AGI) is just a matter of scaling up. Surely it’s not that simple, is it? Well, stunning new model achievements suggest that progress is happening very quickly.

First, the company reached a remarkable milestone in the field of AI development with its o3 model, a general-purpose AI system developed using reinforcement learning.

The model won a gold medal at the International Olympiad in Informatics (IOI), surpassing human benchmarks and outperforming specialized handcrafted models. Essentially, the model’s performance exceeded those of elite human competitors.

“Overall, these results indicate that scaling general-purpose reinforcement learning,

rather than relying on domain-specific techniques, offers a robust path toward state-of-the-art AI in reasoning domains, such as competitive programming,” OpenAI said.

In its new paper “Competitive Programming with Large Reasoning Models,” the company almost modestly concentrates on competitive programming.

However, the reason the o3’s achievement is actually a big deal is the fact that the model was developed using reinforcement learning.

This training methodology enables AI systems to improve iteratively by rewarding correct outputs and penalizing errors. It means that the model learns and improves through feedback, much like humans – but without actual human input.

“AlphaGo became the best Go player in the world without human guidance. It just kept playing itself until it mastered the game. Now, OpenAI is applying the same principle to coding – and soon to all STEM fields,” Matthew Berman, an AI enthusiast and entrepreneur, wrote on X.

OpenAI just dropped a paper that reveals the blueprint for creating the best AI coder in the world.

But here’s the kicker: this strategy isn’t just for coding—it’s the clearest path to AGI and beyond.

Let’s break it down 🧵👇 pic.twitter.com/NJwbq4kxRs

undefined MatthewBerman (@MatthewBerman) February 16, 2025

According to him, it now looks like every domain with verifiable rewards – maths, coding, science – can be mastered by AI just by allowing the model to play against itself.

Berman thinks the o3 model, a large reasoning model designed to generalize across a broad spectrum of tasks, is proving that AI can adapt, evolve, and even surpass human benchmarks.

“The IOI 2024 findings confirm that large-scale RL training alone can achieve state-of-the-art

coding and reasoning performance. In addition to its significantly improved problem-solving capabilities, we observe that o3 demonstrates more insightful and deliberate chains of thought,” OpenAI adds in the paper.

To many, it sounds like AI – or at least models built by OpenAI – is removing human limitations.

“That’s how we get to AGI,” says Berman. “AI just needs more compute – not more human intervention. We’re witnessing the birth of AI superintelligence in real time.”

“It won’t stop at coding. The same techniques will make AI the best mathematician, scientist, and engineer in history. The race to AGI is on.”

AGI is a concept for a machine that can learn and understand any intellectual task a human can – essentially, a type of AI that aims to mimic the human brain’s cognitive abilities.

In late January, Altman attempted to shut down the hype after rumors that OpenAI had already achieved AGI began to surface online. But in that same blog post, Altman claimed that his company knows how to build AGI.

What’s worrying is that OpenAI said in December that a successful transition to a world with “superintelligence” was “far from guaranteed” and openly admitted: “We don’t have a solution for steering or controlling a potentially superintelligent AI, and preventing it from going rogue.”

Even more concerning is that despite concluding that “humans won’t be able to reliably supervise AI systems much smarter than us,” OpenAI then disbanded teams focused on AI safety.