In a conversation that might reset your brain, Eric Jang told Dwarkesh Patel about AlphaGo's true, unsettling secret: it didn't just beat the best Go players; it challenged our most basic assumptions about computational hardness. While everyone focused on the spectacular wins, Jang, who rebuilt AlphaGo from scratch, saw something deeper. He explains how a relatively small neural network managed to “amortize” what looked like intractable search problems. This wasn't just a clever algorithm; it was, as Jang put it, “a breakthrough that I think most people don't even fully comprehend today, how profound that accomplishment is.”

Key Takeaways

  • Eric Jang’s experience rebuilding AlphaGo showed how small neural networks can “amortize” complex, seemingly intractable search problems, compressing vast simulation into minimal compute.
  • AlphaGo’s success with problems like Go, and AI’s efficiency in areas like protein folding and AlphaTensor, suggests our understanding of computational hardness (e.g., P=NP) might be “incomplete.”
  • These advanced AIs excel by identifying “macroscopic structures” and global patterns within chaotic systems, rather than by making perfect microscale predictions.
  • This approach is a radical departure from traditional computation, demonstrating how AI can solve problems that feel NP-hard by focusing on high-level expectations over individual precise outcomes.

AlphaGo's Unsettling Answer to “Hard” Problems

Most founders hear “AlphaGo” and think “AI beats human at board game.” But Jang and Patel dug into the philosophical bombshell. AlphaGo's true power wasn't brute force or perfect calculation; it was its ability to compress a vast, chaotic future into a digestible, predictive output. Jang revealed how this capability forced him to question long-held beliefs about computational limits. “It actually makes me wonder if our understanding of problems like P=NP, or these fundamental computational hardness problems, is incomplete,” he shared.

Traditionally, a problem proven to be NP-hard means it becomes exponentially harder to solve as its size increases. Dwarkesh Patel pointed out that problems like protein folding are NP-hard, yet neural networks are finding solutions. AlphaGo showed this same trick for Go, which, while not formally proven NP-hard, exhibits similar combinatorial explosion. The AI didn't simulate every possible move. It found a shortcut.

The Macroscopic Lens: How AI Sees Chaos

The secret isn't predicting every micro-outcome, but identifying a “macroscopic quantity.” Jang explained, “There's a more macroscopic quantity that we really care about, which is the average or expectation or some sort of global macrostructure over a lot of possible futures.” Think of it like weather prediction: we don't need to track every air molecule; we need to understand the large-scale patterns of fronts and pressure systems. Neural networks are adept at this macro-pattern recognition in chaotic environments. This explains why they succeed in fields from Go to protein folding and AlphaTensor—they’re not solving problems with perfect micro-detail, but by discovering efficient macro-level heuristics.

Jang concluded, “To me, AlphaGo was the first paper that really showed this profound level of simulation being compressed into a small amount of compute.” It’s a paradigm shift: instead of trying to calculate every variable in a complex system, AI builds an intuition for the larger, emergent patterns that matter.

What to Do With This

Founders constantly face "NP-hard" problems: optimizing growth funnels, finding product-market fit in noisy markets, or navigating complex competitive landscapes. You might be stuck trying to simulate every micro-variable or over-optimize tiny details. Instead, adopt AlphaGo's macroscopic lens: for your biggest, most intractable problem this week, stop trying to predict every micro-outcome. Identify the 2-3 macro signals or "global macrostructures" that determine success or failure in that chaotic system. Then, build simple heuristics or dashboards to focus on amplifying those signals, trusting the "good enough" pattern recognition over exhaustive, but ultimately impossible, micro-simulations.