AI, networks and Mechanical Turks

TLDR

• LLMs may let new products rent the cold-start advantage that Amazon, TikTok, and Google built through years of behavioral data.

Key Takeaways

• Amazon, Google, Instagram, and TikTok built recommendation power by watching user behavior, but none of those systems know why users act, only correlation.
• An LLM is a step change: it can connect products, content, and metadata to patterns with broader understanding, not just purchase or click correlation.
• Example: Amazon infers bubble wrap from packing tape; an LLM might infer home insurance and broadband because it understands the user is moving house.
• The cold-start problem shifts: instead of needing your own Mechanical Turk user base, you can call a general-purpose world model via API.
• Agentic assistants with access to browsing, purchasing, and wearables become another partial view of the user, different from but overlapping with Meta, Google, and Amazon’s views.

Why It Matters

• Startups entering recommendation-heavy markets no longer need to accumulate millions of behavioral signals before their product works; a capable LLM API may substitute.
• The leverage point moves from runtime user observation to the training data created over the past few hundred years of human-generated content.
• Evans frames this as a new kind of filter for an internet with infinite product, media, and retail, a bigger structural question than simply replacing Google Search.

Benedict Evans · ** · Read the original