AI chiefs flip: from "slow down" to "floor it"
Good Morning from San Francisco 🌉 Silicon Valley, who once begged for AI guardrails, now wants them gone. Their new battle
Research
AI Learns Complex Reasoning Without Human Training Data
Researchers have created an AI that teaches itself complex skills without any human examples. Unlike current systems that learn from human data, this AI generates and solves its own challenges. Early tests show it outperforming traditional models - but questions remain about its limitations.
1 min read
Researchers from Tsinghua University, Beijing Institute for General Artificial Intelligence, and Pennsylvania State University have developed an AI system that learns to solve complex problems without using any human training examples. The breakthrough opens a path to AI systems that can teach themselves increasingly sophisticated skills.
The new method, called Absolute Zero, lets AI propose and solve its own tasks, learning through trial and error. Published on arXiv in May 2025, the research shows the system matches or outperforms AI models trained on thousands of human-created examples.
"We wanted to remove the bottleneck of needing human-curated data," says lead author Andrew Zhao from Tsinghua University. "Our system creates its own practice problems in coding and mathematics, then improves by solving them."
The researchers tested their system on coding challenges and mathematical reasoning tasks. The AI not only learned to handle these problems effectively but also showed an ability to transfer skills between different types of tasks. When tested on coding problems, it performed better than specialized systems trained specifically for that purpose.
The system works by generating its own programming exercises and mathematical problems. It then attempts to solve them, learning from both successes and failures. This self-directed learning approach produced better results than traditional methods requiring human-created training examples.
The technology proved effective across different AI model sizes and types. Larger models showed greater improvements, suggesting the method becomes more powerful as computing capability increases.
Various leading AI labs have highlighted the growing challenge of finding enough high-quality training data as AI systems become more advanced. This research offers a potential solution by letting AI systems generate their own learning materials.
Why this matters:
- AI systems might soon learn complex skills without relying on human examples, removing a key limitation in AI development
- The success of self-directed learning in both coding and math suggests this approach could work for other types of reasoning tasks
Read on, my dear:
Read next
The Office Politics of AI: Use It But Don't Tell
AI helped them work faster and better. Then their colleagues found out. New research reveals an unexpected tension in today's workplace: using AI tools can boost performance but damage professional reputation. Data from 4,400 workers shows why many now hide their AI use
AI System Cracks Soccer's Secret Language
AI's new soccer superpowers aren't just tracking the ball—they're reading the whole game. How a Shanghai research team built a system that spots what even veteran refs miss, outperforms GPT-4, and could revolutionize how we watch the world's most popular sport.
AI Tool Turbocharges Lab Analysis, Ends Scientists' Band-Spotting Blues
Imagine spending hours squinting at lab gel images, counting bands by hand. That's what scientists have done for 40 years—until now. A new AI system can spot every band in seconds, even in messy gels that trip up existing tools. The secret lies in its unique approach.
AI Gets Smarter with Just One Example, Researchers Find
What if everything we thought about AI training was wrong? Researchers just discovered that artificial intelligence can master complex tasks by studying a single example—matching the performance of models trained on thousands. The secret lies in finding that one perfect example.