Noam Brown
Showing 1–3 of 3 transcripts.
- Sequoia Capital30 min
OpenAI’s IMO Team on Why Models Are Finally Solving Elite-Level Math
Alex Wei, Sheryl Hsu, Noam Brown, Sonya Huang
Alex Wei, Cheryl Hsu, and Noam Brown led a focused sprint that enabled AI models to achieve gold medal performance at the International Math Olympiad, marking the first instance of a system reaching this elite human benchmark. This breakthrough relied on scaling test-time compute to over 100 minutes of reasoning and deploying self-verification protocols that allowed the model to correctly identify unsolvable problems rather than hallucinating solutions. While the team utilized general-purpose techniques to advance broader reasoning capabilities for future scientific research, they explicitly noted that current constraints prevent solving complex combinatorics or Millennium Prize problems within the required timeframe.
- Sequoia Capital45 min
OpenAI's Noam Brown, Ilge Akkaya and Hunter Lightman on o1 and Teaching LLMs to Reason Better
Noam Brown, Ilge Akkaya, Hunter Lightman, Sonya Huang, Pat Grady
OpenAI's O1 model, internally codenamed Project Strawberry, introduces a paradigm shift by employing "inference time compute" to enable systems to engage in extended, self-correcting reasoning processes akin to human System 2 thinking. This architecture has delivered unprecedented capabilities in STEM domains, allowing the AI to solve complex Olympiad-level programming problems, pass research engineer interviews, and assist in scientific discovery by bridging the gap between difficulty in generation versus verification. While the project faces limitations in speed and creative tasks compared to predecessors like GPT-4, its demonstrated ability to scale performance through increased thinking time marks a significant advancement toward the operational goal of Artificial General Intelligence.
- Sequoia Capital51 min
Founder Eric Steinberger on Magic’s Counterintuitive Approach to Pursuing AGI
Eric Steinberger, Sonya Huang, Noam Brown, Sonia
Former DeepMind collaborator Eric Steinberger founded Magic to develop vertically integrated AI software engineers capable of achieving general-domain, long-horizon reliability through increased inference-time compute. Challenging the industry's reliance on standard benchmarks, the company recently open-sourced a "hashless eval" methodology that forces models to process entire context windows rather than exploiting retrieval heuristics. Steinberger's strategy prioritizes a lean, high-velocity research team focused on proprietary model training to build "colleague-tier" agents that automate complex software tasks with over 99% reliability.