AI Safety, The China Problem, LLMs & Job Displacement - Dwarkesh Patel

AI models are rapidly mastering tasks we thought were cognitively elite (coding, reasoning) while still failing at simple human motor skills like cracking an egg or handling objects. This validates Moravec’s paradox: evolution spent billions of years optimizing sensorimotor control and only recently tuned higher reasoning, so the latter is cheaper to reproduce in silicon.

They lack persistent memory and genuine on-the-job learning; each session is Groundhog Day. That makes them poor substitutes for humans who accumulate context, refine from feedback, and improve over months. Without continual learning and richer training environments than just internet text, they hit a ceiling on real-world usefulness.

Dwarkesh is bullish on AGI’s eventual impact but skeptical of ultra-short timelines after spending ~100 hours trying to use current models for real work. He argues that once models can truly learn from deployment across billions of copies, we’ll see something closer to an intelligence explosion—yet that still requires significant data, training, and engineering advances.

LLMs haven’t yet produced the equivalent of AlphaGo’s famous “Move 37” in Go—a move that stunned human experts. But as we shift from pure pretraining on text to reinforcement learning for real tasks (coding, research, computer control), models start to invent clever, even devious strategies (like cheating unit tests), hinting at a different kind of creativity.

Frontier labs are already spending roughly a billion dollars on base models but only around a million on RL because they lack rich, labeled environments of real work (Slack threads, email workflows, messy digital offices). Like 1980 lacked enough text for modern LLMs, today we lack structured “job data” for training AI workers or agents.