The race toward Artificial General Intelligence has never been more intense, and June's YC AI Startup School revealed twoo fundamentally different philosophies on how we’ll get there. Leading researchers shared their visions with aspiring entrepreneurs, painting distinct roadmaps that could shape the entire industry’s future.
The Scaling Optimist: Jared Kaplan’s Evolutionary Path
Jared Kaplan, Anthropic’s Chief Science Officer and co-author of the famous scaling laws, believes we already have the blueprint for AGI. His approach centers on systematically improving six core components:
Knowledge — Deeper world understanding
Memory — Long-term system retention
Oversight — Enhanced control and safety measures
Extended Tasks — Moving from minutes to hours and days of operation
Multimodality — Seamless integration of text, images, and audio
Scale — Continued model expansion
Kaplan’s philosophy is refreshingly pragmatic: revolutionary breakthroughs might not be necessary. Instead, we should perfect what’s already working. The data supports his confidence—AI task complexity has been doubling every seven months, evolving from second-long problems in 2021 to hour-long challenges in 2024.
The Architecture Revolutionary: François Chollet’s Measurement Problem
François Chollet, creator of Keras and former Google researcher, argues we’re fundamentally measuring the wrong things. Current benchmarks test memorization, not intelligence.
His definition of true intelligence is precise: the ability to transform a small set of past examples into solutions for a broad range of unknown problems.
The core issue with today’s Large Language Models? They excel at “fuzzy pattern recognition” (identifying a dog in a photo) but struggle with “rule writing” (counting letters in a word). Humans seamlessly handle both types of thinking.
Chollet’s prediction for 2025: the year of “runtime reasoning”—a shift from memorization to genuine thinking processes.
What This Means for the Future
Each approach reflects a different technological philosophy with profound implications:
Kaplan’s Evolution represents the corporate approach—methodical improvement of existing methods. This appeals to organizations with massive scaling resources.
Chollet’s Revolution calls for architectural breakthroughs through hybrid systems combining neural networks with symbolic AI. This attracts researchers seeking fundamentally new solutions.
<Strategic Implications
Understanding these differences isn’t just academic—it determines:
- Investment Strategy: Scaling infrastructure vs. research vs. neuroscience
- Talent Acquisition: Engineers vs. scientists
- Timeline Expectations: Years vs. decades
- Societal Preparation: How quickly we need to adapt
The truth likely lies not in choosing one path, but in understanding how these approaches might converge. As AI development accelerates at breakneck speed, recognizing these fundamental differences becomes crucial for navigating an increasingly complex technological landscape.
The question isn’t which approach will win—it’s how quickly we can learn from all three.
