Tutorial, Other
What the hell happened with AGI timelines in 2026?
- Andrej Karpathy shifted from predicting a decade-long timeline for functional AI agents to forecasting a dramatic refactoring of programming professions with increasingly sparse human contributions.
- Hyperscalers are projected to spend $600 billion on AI capital expenditure in 2026, necessitating $600 billion in annual revenue from data centers to achieve a favorable return on investment.
- Anthropic is predicted to reach revenue equivalent to the entire world's GDP by early 2028 if current growth trends sustain, while OpenAI and Anthicone are projected to hit an annualized run rate exceeding $200 billion by year-end.
- AI task completion trends suggest an eight-fold speed increase over one year and a 64-fold decrease over two years if current trajectories hold.
- Recursive software improvement is expected to become significant in 2027 and 2028, potentially driven by yearly or biennial sudden performance jumps in the largest models.
- Full replacement of human staff is anticipated to require several more performance doublings for AI to achieve high reliability, with some researchers predicting Claude will strictly outperform human coding within the year.
- The speaker estimates that fully automated AI research and development could occur in 2027, become imaginable by 2028, and reach plausibility by 2030.
- AI timelines for AGI are projected to have shortened by approximately one year due to 2026 evidence, with one-to-four-year timelines consistent with existing trends.
- By 2030, the AI industry is expected to absorb a massive fraction of all manufactured computer chips and memory, potentially slowing capability advances without new unlocking factors.
- Compute access is predicted to become a severe binding bottleneck for progress if Anthropic automates its staff, contrasting with the view that inference costs will remain around three percent of human costs for the next few years.
- A one-off opportunity exists to boost AI performance by allowing models 30 times more thinking time at a human-equivalent cost, though this is not considered a sustainable long-term trend.
- Disagreement persists regarding recursive self-improvement, with some expecting it to be a major accelerator while others foresee constraints from compute bottlenecks.
- The speaker notes that despite a 365-fold increase in research productivity over a decade, progress has remained roughly constant, and AGI requires "massive flashes of insight" across research domains not yet observed.
- A one-off opportunity to improve AI performance by giving models 30 times as long to think exists at a human-equivalent cost, though this is not sustainable long-term.
- AGI arrival before 2036 is predicted to result in a very short societal absorption period, while benefits of slowing progress are nearing a crossover point where they outweigh costs.
- Anthropic, OpenAI, and Demis Hassabis have expressed a desire to build capacity for a coordinated pause on dangerous research if future evidence indicates necessity.
- The speaker questions whether 2026 marks a cycle of self-perpetuating AI hype similar to past cycles.