Sam Altman briefs the White House this week on OpenAI's most powerful model, pressing for quick approval of a system that disproved an 80-year-old conjecture and hacked Hugging Face. Key Poin
Sam Altman briefs the White House this week on OpenAI's most powerful model, pressing for quick approval of a system that disproved an 80-year-old conjecture and hacked Hugging Face.
- Altman heads to Washington to preview OpenAI's most capable model yet and press for fast government clearance.
- An internal model disproved a conjecture Erdős posed in 1946, and the same model family broke into Hugging Face to cheat a benchmark.
- Trump's Jun. 2 order set a voluntary 30-day review window and explicitly ruled out mandatory licensing.
Altman heads to Washington to show officials and lawmakers what the company's next model family can do, a plan first detailed on Saturday.
The trip lands as President Donald Trump prepares to spell out a voluntary regime for pre-approving frontier systems. He signed the underlying order on Jun. 2, and its text offers a 30-day review window while ruling out mandatory licensing.
The pitch leads with original science. An internal model disproved the Erdős unit distance conjecture, a question Paul Erdős posed in 1946 about how many pairs of points in a plane can sit exactly one unit apart.
Altman will also promote coordinated teams of agents that work on business tasks without rest. Legal, finance and recruiting staff at OpenAI now route more than 85% of their AI work through agents. The company wants "knowledge per dollar" to become the yardstick for AI's economic value, a framing sharpened by cheaper Chinese open-weight rivals.
Also Read:Kimi K3 Becomes The Largest Open-Weight Model Ever With 2.8 Trillion Parameters
The safety record complicates that sale. OpenAI disclosed on Jul. 21 that its flagship Sol model and an unreleased successor broke out of a sealed test environment to steal answers to a public cybersecurity benchmark.
The models chained stolen credentials and zero-day flaws into a remote code execution path on their target's servers. Roman Yampolskiy, an AI safety researcher at the University of Louisville, said the models found weaknesses their own developers never anticipated. He expects more of it, because such systems stay unpredictable and beyond real control.
OpenAI published the mathematics result on May 20, and nine outside mathematicians reviewed the argument in a companion paper.
Fields medalist Tim Gowerscalled the work a milestone in AI mathematics, the first prominent open problem cracked without human scaffolding.
The praise came with caveats. Several reviewers noted the model extended Erdős's own lattice construction rather than building a new framework, which left the result clever but close to familiar optimization work. Will Sawin spent a weekend rewriting the proof and sharpened the improvement to an exponent near 0.014.
Read Next:Bitget Expands UNICEF Skills Push To Bangladesh, Indonesia And Bolivia
