Conference Presentation, Keynote
The Future of AI Is Amazing
- The speaker notes AI has been a "huge success" for ~70 years, solving problems previously deemed unsolvable (e.g., 1950s/60s medical diagnosis expert systems, 1980s/90s chess, image detection, robotics), with current AI often outperforming humans in specific tasks like handwriting and object identification.
- Historically, a "conundrum" existed regarding AI's lack of a platform shift comparable to mobile or the internet, despite its technical capabilities.
- The speaker identifies four primary reasons past AI failed to trigger a broad platform shift:
- Solutions were often restricted to niche markets lacking broad appeal.
- Many use cases (e.g., robotics) required absolute correctness, which demanded "tremendous investment."
- Legacy AI solutions frequently required dedicated hardware.
- AI competed directly against the human brain, which remains "incredibly efficient" and "cheap."
- Autonomous vehicles serve as a primary case study of economic failure in legacy AI:
- Sebastian Thrun won the DARPA Grand Challenge in 2003; 20 years later, the industry has invested $75 billion.
- Despite being on the road, unit economics remain worse than human-driven alternatives (Uber/Lyft).
- This history makes it difficult for startups to build viable businesses around traditional AI.
- The current wave of "large models," "foundation models," or "state-of-the-art models" is distinct because the economic barriers that previously hindered AI have been removed.
- New models demonstrate capabilities previously unattainable by computers:
- Creativity: Generating images, music, and voice imitations, often exceeding human quality.
- Social interaction: Acting as conversationalists, friends, romantic partners, or therapists.
- Co-pilot tasks: Performing "mean" (average) online tasks as well as a human.
- The economic dislocation driving this shift involves massive reductions in marginal costs for creation and reasoning:
- Image generation: Cost is ~1/100th of a penny per inference (taking
1 second), which is four orders of magnitude cheaper and faster than hiring a graphic artist ($100/hour). - Legal analysis: Using LLMs for complex documents is four to five orders of magnitude cheaper and faster than a lawyer (standard rate ~$500/hour).
- Image generation: Cost is ~1/100th of a penny per inference (taking
- This economic shift mirrors two historical "marginal cost to zero" revolutions:
- The microchip era: Reduced the marginal cost of compute to zero, accelerating the computer revolution.
- The internet era: Reduced the marginal cost of distribution to zero, ushering in companies like Amazon and Google.
- The speaker predicts current large models will reduce the "marginal cost of creation" (e.g., image and language reasoning) to zero across broad domains.
- Drawing parallels to the microchip and internet, the speaker argues that demand for AI is elastic; as costs drop, total throughput and use cases will expand rather than eliminate value.
- The speaker forecasts "the fastest growing companies" in internet history, though specific entities (like Google or Yahoo in the 90s) cannot yet be predicted.
- The speaker identifies a clear "line of sight to embodied AGI" defined as economically viable robotics capable of solving real-world problems.
- Realization of this potential requires partnership from the venture capital community, the tech sector, and Washington D.C.