The Silicon Takeover: OpenAI’s New AI Outsmarts 99.8% of Human Coders
OpenAI—the company behind the launch of ChatGPT—has just released a new paper to demonstrate how its AI will outsmart 99.8 percent of coders in the world. Its AI beating the best of human coders in the world will open a new frontier in how we interact and use AI to augment human intelligence.
The paper titled Competitive Programming with Large Reasoning Models reveals a strategy that is not just specific to coding but to other spheres, as well – science, maths, research, etc. A path to building first, Artificial General Intelligence and then ultimately Artificial Superintelligence -the Holy Grail of AI development.
It echoes what Sam Altman -the company’s CEO – has said in the past that his company understands what it will take to build Artificial General Intelligence.
“We are now confident we know how to build AGI as we have traditionally understood it, we believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.”
For Altman, it is no longer a question of “if”, it is now a question of “when”. And he believes that at least AGI will be in 2025.
Inside OpenAI’s Master Plan for Achieving AGI
According to the publication from OpenAI, reinforcement learning (with verifiable rewards) applied to large language models (LLMs) significantly boosts performance on complex coding and reasoning tasks.
Reinforcement learning with verifiable reward is simply a way for AI to train itself, by trying a bunch of different things to see which one is rewarded and then keep improving on itself. You don’t need humans to do the training.
OpenAI o3 in December was the 175th the best competitive programmer in the world. That value, according to Sam Altman is now around 50 and may hit the number 1 spot in 2025. The Benchmark for achieving Gold in the Olympiad in Informatics (IOI) is 50. By that standard, o3 is already better than most coders.
It achieved those results without relying on hand-crafted inference heuristics. It seems to be that removing humans from the loop is the key to achieving massive improvement. o3 trained itself to be at par with elite human competitors.
“Our findings highlight that increasing reinforcement learning training compute, coupled with enhanced test-time compute, consistently boosts model performance to nearly match the best humans in the world.
Given these results, we believe o-series large reasoning models will unlock many new use cases for AI in science, coding, math, and many other fields.”
The paper demonstrates that you can boost model performance to nearly match the best humans in the world (and even surpass them) by increasing reinforcement learning training compute, coupled with enhanced test-time compute.
You can read the full paper here
See : Elon Musks: “Grok 3 Is Outperforming Anything that has been released”
