Interview, Fireside Chat
Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit with CEOs of Cerebras & Black Forest Labs
- Cerebras data centers are projected to consume more power in the next several years than the previous 50 years of Earth's history, with individual facilities the size of football fields receiving power levels comparable to mid-sized cities.
- Global data center construction is expanding across the U.S., Canada, the Nordics, France, the Middle East, Kazakhstan, Tajikistan, Georgia, and Armenia to meet a $25 billion booked backlog.
- Customers including OpenAI, Anthropic, SpaceX AI, Google, Microsoft, and AWS are pre-ordering Cerebras chips before completion, with the company aiming to exceed the historical 18-month performance doubling trajectory.
- Cerebras plans to utilize its architecture to run models 15 times faster for 24-hour periods to simulate weeks or months of cognitive processing, while recursive learning loops are expected to produce exponentially better results without a defined endpoint.
- Long-term expectations include AI accelerating learning to the equivalent of thousands of generations, leading to breakthroughs such as unlimited energy, calories, knowledge, education, and housing, and potentially curing cancer.
- Future generations are predicted to work on structures built today and passed down similar to stone mason families, while AI tutors will enable personalized education by modifying teaching for each child.
- The speaker anticipates that within 24 months, AI will evolve from summarizing results to understanding user intent and providing abstract solutions, while the gap between open source and frontier models regarding reasoning is expected to close.
- A "minivan time" for open source models is predicted where they handle ordinary problems while frontier models address hard tasks, alongside a need for US domestic open source models to provide global choice against Chinese and Nvidia alternatives.
- Open source models are expected to require minimal fine-tuning data to adjust to specific hardware and tasks, with a strategic shift away from fine-tuning toward prompting robots in context.
- Black Forest Labs aims to combine multi-modal visual models with action prediction for robot deployment, eventually using movie-making AI as a "brain" for robots and enabling human-in-the-loop production workflows within a couple of years.
- IP holders like Disney may shift to a model allowing customers to create stories with characters via licensing fees, while the trend of corporate sovereignty will drive companies to run on-premise open source versions for regulated industries.
- Jason Calacanis forecasts potential mass data breaches or infrastructure failures as industry power increases, predicts AI will eventually ask humans questions they did not know to ask, and suspects AGI has already been reached.
- Government requests for scheduled releases of dangerous models are viewed as reasonable to facilitate red-teaming and patching, while "loop maxing" is expected to continue generating buzzwords and anticipated developments.
- The industry will develop "world models" that observe human behavior to extract insights, and future AI capabilities may allow humans to work on structures for 50 to 100 years.