The AI That Scored 100% on the World’s Hardest Math Exam
Let me start with the one that actually made me sit up straighter. On July 23, two Chinese tech companies, Huawei and Xiaohongshu, announced their AI models each achieved a perfect score on this year’s International Mathematical Olympiad.
For context: 666 contestants from around the world took part. Only seven humans achieved full marks. And now AI has joined that elite club.
“We are delighted with this result, because achieving a perfect score at the IMO is extremely challenging,” Xiaohongshu said. The models, which included Huawei’s “Celia” and Xiaohongshu’s “dots-note-3.0”, solved all six problems within a specified time limit with no human intervention.
What’s striking is how fast this happened. At the 2024 IMO, Google achieved a silver-medal score, solving four of six problems over two to three days. Two years later, AI is solving all six perfectly.
One venture partner at Menlo Ventures tested four leading AI models—including OpenAI, Anthropic, and China’s Kimi K3—on this year’s IMO questions. All four scored 42 out of 42. His conclusion? “The frontier of AI has officially moved well past IMO math”.
The AI That Never Hallucinates—Even in the Dark
While AI was acing math tests, researchers at KAIST in South Korea were solving a different problem: hallucinations.
Multimodal AI models—which process text, images, and audio together—often misinterpret sensory information. They mistake bright areas in thermal images for light reflection. They claim to hear sounds that don’t exist because an object appears in a video.
The KAIST team developed two technologies that fix this. The first, DNA optimization, helps AI accurately understand special camera sensors like thermal and X-ray. The second, MAD, prevents confusion between visual and auditory information at the source.
Here’s the clever part: both technologies work without costly model retraining. They can be applied to autonomous vehicles operating at night or in bad weather, robots in smoke-filled environments, and medical image analysis.
The lead researcher put it simply: “It will serve as a foundation for building multimodal AI that can be trusted in real-life and industrial settings”.
The AI That Plans Like a Human Scientist
On July 15, researchers at Zhejiang University unveiled the Qiushi Engine—an AI scientific discovery system that can independently organize and continuously advance research.
Unlike conventional AI tools that execute individual instructions, the Qiushi Engine continuously evaluates progress and adjusts its strategy. It decides what to do next, generates follow-up tasks, and can sustain reasoning across thousands of steps.
Here’s how it works in practice: researchers provide an open-ended research objective. The system carries out literature review, theoretical analysis, experimental design, programming, data analysis, and result evaluation—adjusting its plans based on experimental feedback.
In one demonstration, given only an open objective, the Qiushi Engine ran a 206-step research process lasting about 21.5 hours and proposed four candidate directions. In another task, it completed in about six hours what would typically require researchers several weeks of work.
The team emphasized that AI is becoming a research collaborator, not a replacement. Humans set the objectives and validate conclusions. AI does the sustained exploration.
The AI That’s 10,000 Times More Efficient
And then there’s the one that genuinely surprised me.
Northwestern University engineers built a brain-like electronic device inspired by the cerebellum—the part of your brain that handles reflex reactions without you even thinking about it.
Your cerebellum doesn’t waste energy analyzing every moment. It constantly monitors the world for the unexpected and springs into action only when something changes. Northwestern’s device does the same thing.
In proof-of-concept experiments, it identified abnormal heart rhythms within one-fifth of a heartbeat with more than 98% accuracy. And it required roughly 10,000 times fewer computer operations than conventional AI approaches.
The lead researcher, Mark Hersam, put it perfectly: “The cerebellum is excellent at ignoring the expected and reserving its resources for reacting to the unexpected. That approach ultimately translates into lower energy consumption, and that is where we achieve orders of magnitude improvement”.
This could enable a new generation of low-power, always-on AI for wearable health monitors, self-driving cars, and autonomous robots.
The frontier of AI has officially moved past math. It’s now moving into science, safety, and efficiency.
And that’s worth paying attention to.