newsfilter
← Interviews
Y Combinator35 min

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

Jared, Harj, Sam Altman, Diana Alvear, Mark Mandelmann, Francesc Campoy Flores, Melanie Warrick, Kendrickson Jr., Harshi, Jake Heller

Episode, Podcast
  • AI models are projected to surpass human capability in chip design and circuit layout, addressing NP-complete routing challenges for industry leaders like NVIDIA and Intel while eliminating current development bottlenecks.
  • Weekly improvements are anticipated from a current "worst point," with scaling laws and a target of four additional orders of magnitude in capability driving toward an estimated 10^13 dollars in spending.
  • General Artificial Intelligence (AGI) and Artificial Superintelligence (ASI) are predicted to emerge within a timeframe of four to fifteen years (thousands of days), potentially solving physical laws to enable space colonies, climate solutions, and abundant energy.
  • Specific advancements like O1 reasoning capabilities, supported by a new chain-of-thought reasoning dataset and pre-cached diagrams to manage processing speed, are identified as the key missing link for accelerating scientific discovery across all fields.
  • Future model iterations (O2, O3) are expected to further unlock capabilities, while new development patterns integrate Retrieval-Augmented Generation (RAG) with reasoning to automate complex engineering tasks such as component selection and CAD simulation.
  • Industry value creation is shifting toward proprietary vertical data, high-accuracy benchmarks requiring the final 10% precision for industrial applications, and robust UI layers with high switching costs rather than mere model access.
  • Operational deployments have demonstrated an 85% accuracy rate in automating complex customer support edge cases, with a strategic shift anticipated from merely viewing model outputs to editing and directing their reasoning steps.
  • Significant competitive advantages are expected for startups in mechanical, electrical, chemical, and bio-engineering, alongside a potential decline in standalone AI coding agents as general reasoning models outperform dedicated coding infrastructure.
  • Societal narratives are forecast to transition from fear to an era of abundance, contingent upon technologists successfully deploying these technologies within the specified four-to-fifteen-year window to solve high-value complex problems.