Sam Altman argues for iterative deployment of multiple agents over a single superintelligence

• Altman opposes a single superintelligence, preferring a structure of 1 trillion AIs with slightly superhuman intelligence to reflect societal robustness and pluralism.
• Current sandboxing is insufficient for AI 1 million times smarter than the best security researchers, making alignment trust a prerequisite for safety.
• Altman notes that compute allocation is a core capital strategy at OpenAI, where high token usage is not rewarded for promotion.
• He predicts that the cost of redoing complex math proofs like Navier-Stokes could drop to 1/100th of the original cost within a year.

Safety and Architecture

• Safety hierarchy requires sandboxing, monitoring (chain of thought, activation), and full alignment trust.
• Altman uses the term 'swarms' to describe the preferred multi-agent structure, noting that advanced math and hacking tasks currently use 'agent teams'.
• Software architecture changes are more sustainable than code fixes, with startups building new internet/app architectures for more intelligent models.

Compute and Supply Chain

• OpenAI is developing its own chips and data centers, using a robotics program to address supply constraints and build infrastructure.
• Local inference is described as inefficient from a global allocation perspective due to the limited supply of DRAM.

Model Quality and Pricing

• Benchmaxxing is de-emphasized in favor of scientific discovery capability and enjoyable collaboration.
• Token-based pricing is retained over prompt-based pricing because customers prefer consistent units despite potentially higher costs.
• Current models are described as 'thoughtful and challenging, but still respectful, senior coworkers,' moving past previous phases of glazing or stroppiness.