Interview, Fireside Chat, Conference Presentation
AI Startup Founders Debate the Creation of Artificial General Intelligence
- Predicts AGI (defined as high-level machine intelligence capable of any human task) will occur within a human lifetime but likely not within the next 10 to 20 years.
- Estimates a definition of AGI as an autonomous agent with thoughts and beliefs will be realized in approximately four to five years, around 2035.
- Suggests AGI mimicking human capabilities could be achieved in three to five years, though this timeline remains uncertain.
- Describes the path to AGI as roughly five to ten research breakthroughs away, with no specific estimate on the time duration required.
- Asserts current technology relies on causal mass language modeling to predict tokens, lacking true intelligence, consciousness, or intent behind actions.
- Notes that current models like GPT-4 can infer user needs and assemble solutions for complex problems, representing a slow-moving continuum where some previously considered AGI (e.g., GPT-3) are now viewed as already existing to some extent.
- Highlights significant technical limitations, including errors in basic logic such as comparing numerical values, and expresses doubt that current trajectories will lead to AGI.
- Anticipates societal and philosophical shifts, including moral questions and the need for agreement on core principles to govern development within desired boundaries.
- Foresees a future moment in collective consciousness where humans can no longer identify tasks where AI is inferior to human performance.
- Projects proper governance will enable models based on human values to improve human lives, alongside a future list of tasks where AI performance is evaluated against human standards.
- Expects the current format of human-AI interaction to likely change or cease if AGI is achieved in the future.