newsfilter
← Latest News
SoftBank Vision Fund·

Jason Warner: A bold vision for automating software on the path to AGI

Summary

  • Frontier AI companies face an expensive scaling wall that may require a shift in approach to achieve artificial general intelligence (AGI).
  • Poolside, co-founded by Jason Warner & Eiso Kant, utilizes reinforcement learning via code execution feedback (RLCEF) to train models with synthetic data instead of relying solely on model size.
  • The company has raised over $600 million & operates 10,000 Nvidia GPUs to power its model training activities.
  • Poolside formed a strategic partnership with Amazon Web Services (AWS) to offer highly private, secure, & customizable AI models to customers.
  • The RLCEF approach mirrors DeepMind's AlphaGo strategy by using a massive seed dataset of open-source code & a software simulator to create a self-improving flywheel effect.
  • Warner estimates that Poolside's techniques will generate 50% more input tokens than standard methods, potentially overcoming diminishing returns in the AI industry.
  • The company expects to reach the capability of a human software developer within approximately five years.
  • Poolside plans to provide models that run entirely within customer environments without the vendor accessing customer data or the model itself.
  • The firm hired 60 engineers in Europe & 20 outside Europe, leveraging the European corridor between London, Paris, Zurich, Amsterdam, & Madrid to access top talent with less competition.

With reinforcement learning and synthetic data, the buzzy startup plans to scale the “wall” that LLMs are hitting and create AGI. Learn more with this Q&A.

Keywords
Strategy, AI Stack