Interview, Statement
What are we scaling?
- Human-like, self-directed on-the-job learning is currently absent, meaning current training approaches relying on verifiable outcomes or "baking in" skills are considered inefficient for tasks like robotics, laundry, or dish picking without millions of practice instances.
- AGI is not viewed as imminent unless models soon achieve self-directed learning, though a significant portion of the market expects a takeoff within the next five years driven by "kludgy" reinforcement learning producing superhuman AI researchers.
- The speaker anticipates that by 2030, labs will have made significant progress on continual learning, with models generating hundreds of billions of dollars in annual revenue while failing to automate all knowledge work, and defines AGI as a shifting target if companies fail to earn trillions in token sales.
- Actual brain-like intelligence is projected to emerge within the next decade or two, with human-level on-the-job learning capabilities expected to require another five to ten years to fully resolve.
- Pre-training scaling is described as a predictable force, yet solving continual learning is expected to resemble the gradual evolution of in-context learning rather than a singular breakthrough, involving a race where one lab gains initial traction followed by rapid replication and slight improvements by competitors.
- The industry expects AI to diffuse into firms much faster than human hiring, with models capable of ingesting entire drives or Slack histories in minutes to distill skills, contrasting with the high cost of current context-specific training.
- Market competition is anticipated to remain fierce over the coming decade, as forces like talent poaching and reverse engineering are expected to neutralize any potential runaway advantages from the first lab to solve continual learning.
- Future capabilities may allow AGI systems to share knowledge across copies efficiently, though the speaker does not expect runaway gains from initial models if continual learning is not fully solved, nor does the speaker expect automation of any single job through a predefined set of skills.