Why The Next AI Breakthroughs Will Be In Reasoning, Not Scaling

The hosts connect Sam Altman’s long-standing belief in AI-assisted science to o1’s focus on reasoning. They argue that models need to think through problems to help with scientific and engineering work, and point to chip design as an early example.

YC startup Diode demonstrates a workflow that turns a high-level wearable heart-rate monitor request into component choices, a circuit diagram, and a PCB layout. The team says GPT-4o could not reliably handle system design, while o1 could select components using datasheets. They describe combining 4o mini for extracting information from PDFs with o1 for component selection and reasoning.

Camper lets users create CAD designs through natural language. In the demo, it runs several simulations to design airfoils for specified flight conditions, a task the hosts say would usually require engineering expertise and solving complex equations.

The hosts argue that breaking work into steps and evaluating each step remains important, even if models handle more of the reasoning themselves. For vertical software, they suggest building large, domain-specific eval sets from work inside customer workflows, where useful data may not be publicly available. They also point to integrations, user experience, and distribution as parts of a product’s advantage.

YC company Giga applies AI to customer support, where complex edge cases have limited adoption of automated agents. The hosts report that, during a hackathon, Giga combined o1 with careful evaluation and reduced its error rate from about 70% to 5%; they also describe improved handling of complex cases. They see potential for reasoning models in engineering and other applications that affect the physical world.