Conference Presentation, Keynote, Other
Why Data Is the Real AI Bottleneck: Flapping Airplanes' Ben and Asher Spector
- Future systems aim to achieve high-value capabilities using significantly less data, with humans estimated to master coding using 10,000 to 100,000 times less data than current AI models.
- A model achieving 1,000 times greater data efficiency is predicted to make deployment 1,000 times easier, enabling broader competition and allowing companies to participate more broadly in the AI revolution.
- Future economies are expected to expand into tens of thousands of under-resourced domains, such as the end-to-end toaster supply chain, which require data solutions for current data scarcity.
- Compute resources are projected to scale more easily than data, as floating-point operations become exponentially cheaper while data costs do not decrease at a comparable pace.
- The company plans to explore new algorithmic primitives for interacting with hardware beyond current frameworks like PyTorch, a strategy supported by historical trends showing viability over the past 15 to 100 years.
- The team intends to develop an internal virtual machine framework to control GPUs directly, enabling operations that are impossible or inefficient in standard environments.
- New systems are expected to facilitate algorithms specifically relevant to solving data efficiency problems, while the founders anticipate that individuals with unconventional backgrounds can achieve significant results based on prior incubation successes.