Tracking a small number of high-signal individuals beats tracking trends. When the right person names a pattern, the field rearranges around it within weeks.
The pattern shows up most cleanly with 📝Andrej Karpathy. 📝Software 2.0 (2017) preceded the industry consensus on neural-network-as-program by several years. 📝Vibe coding (Feb 2025) became the most-cited frame for AI-assisted development within weeks of the tweet. 📝Never Felt So Behind (Dec 2025) named the agentic-tooling explosion that Anthropic and the broader MCP ecosystem were already racing to standardize. 📝Karpathy's Flag in the Ground — LLM Knowledge Bases (Apr 2026) gave public language to an architecture MythOS had been building for nearly a decade. 📝Agentic engineering (Feb 2026) immediately became the disciplined successor framing the field adopted. The lag between Karpathy posting and the discourse rearranging is now reliably measured in days, not quarters.
The convergence with older information theory is direct. Curated sources beat scraped feeds — a small, high-signal channel carries more decision-relevant information than a vast low-signal one. The discipline of trend forecasting has long held that early adopters and bridge figures are leading indicators rather than statistical noise. Karpathy occupies an unusually pure version of that role in AI: deep technical credibility, public-by-default, and willing to name patterns before they have names.
The application is to maintain a deliberate watchlist of weathervanes — perhaps three to seven per domain — and read what they publish as a structured input, not an ambient feed. The cost is low: a few subscriptions, a clear capture habit, an MythOS memo per named pattern. The compounding payoff is having vocabulary, framing, and architectural decisions in place months before the surrounding market arrives at them.