Training AI will take longer than you think (wyclif.substack.com)

The piece argues that while neural networks appear, in principle, capable of any human task, the real bottleneck is training: current methods—dominated (~99%) by next-token prediction—teach models broad patterns from massive text corpora but fail to close key gaps. That explains why LLMs can converse, write and code convincingly yet still flounder on trivial-seeming tasks (e.g., removing arrowheads from a diagram or reliably matching an eight‑year‑old’s reading comprehension). Reinforcement learning (RL), which rewards correct task performance rather than statistical likelihood, is much rarer (<1%) because task-specific, ground-truth training data are expensive and scarce. Tricks like “roleplay a senior developer” exploit pattern learning but don’t substitute for real task experience. A central technical point invokes Polanyi’s distinction between explicit (textual) and tacit (practice-based) knowledge: much crucial skill is learned through practice and context, so transfer between domains is limited. Even superhuman base cognition may still require extensive task-specific training, implying progress will be a slow, domain-by-domain grind rather than a sudden leap to universal economic AGI. For researchers and industry this suggests shifting effort toward scalable RL data (simulators, human feedback, task generation), better transfer/meta‑learning, and realistic timelines for deployment and policy.