Sam Altman — OpenAI and the Future of AI

ChatGPT & the Inflection Point

• ChatGPT launch (Nov 2022) was the fastest product to 100M users in history
  • Hit 1M users in 5 days; 100M in 60 days — Instagram took 2.5 years
  • Sam describes it as “releasing something we didn’t know how well it worked into the wild”
• Why the timing mattered
  • RLHF (Reinforcement Learning from Human Feedback) had been developed over years
  • InstructGPT showed models could be made helpful, harmless, and honest simultaneously
  • The interface — a simple chat box — was the key design decision

GPT-4 Capabilities

• Passes the bar exam in the top 10% of test takers
• Scores 163 on the LSAT (88th percentile for law school admissions)
• Sam’s view: “We are building something that is genuinely powerful and we are being careful about it”
• Multimodality (vision) was a significant architectural leap
  • Allows GPT-4 to understand charts, handwritten equations, images in context

AGI — Definition and Timeline

• Sam refuses to commit to a specific AGI date
  • “I think we could be pretty close, or we could be further than people think”
  • His working definition: AGI = “a system that can do the work of the median human across most cognitive domains”
• The path he sees
  • Current models: very good at specific tasks, inconsistent across domains
  • Next milestone: models that can reason reliably and plan over long horizons
  • “Agents” — AI that can take actions in the world — as a key step toward AGI

AI Safety — OpenAI’s Approach

• The “capped profit” structure
  • OpenAI is a non-profit that controls a for-profit subsidiary
  • Investors get capped returns (currently ~100× on investment); rest goes to the mission
  • Sam: “We need a lot of capital to do this safely. We don’t want to just give the company away.”
• Iterative deployment as safety strategy
  • Releasing progressively more powerful models to the public — learning in the open
  • “Each deployment teaches us something we couldn’t learn in a lab”
• Red-teaming
  • Hundreds of external researchers try to break each model before release
  • They specifically look for CBRN (chemical, biological, radiological, nuclear) uplift risks

OpenAI’s Competition & Strategy

• Google, Anthropic, Mistral, Meta — a crowded field
  • Sam is sanguine: “Competition is good — it increases the pace of safety research too”
  • The key differentiator he points to: compute scale + alignment research depth
• The Microsoft partnership
  • $10B investment gave OpenAI exclusive Azure access for training
  • In exchange, Microsoft gets deeply integrated models (Copilot, Bing, etc.)
  • Sam: “We needed compute at a scale no one else was willing to fund”
• Open source vs. closed development debate
  • OpenAI’s position: closed development until safety is better understood
  • Meta (LLaMA) takes the opposite stance — Sam respectfully disagrees

What Sam Altman Thinks About Most

• The alignment problem remains unsolved at the frontier
  • “We don’t really know how to make a model that is robustly aligned across all situations”
  • Scalable oversight (using AI to supervise AI) as a candidate solution
• The “principal-agent” problem of AI
  • Who does the AI serve when there are conflicts? The user? The company? Society?
  • Sam sees this as the central political question of the next decade
• Abundance vs. power concentration
  • His biggest fear: a small group (including OpenAI itself) gaining disproportionate power through AI
  • “If we ever find ourselves doing things that concentrate power inappropriately, treat that as a signal something has gone badly wrong”