Every major consumer internet platform is a Mechanical Turk. Amazon, Google, TikTok, Instagram: they all watch what users do, extract correlation, and feed it back as recommendations. The core limitation is structural. These systems know that buying packing tape correlates with buying bubble wrap, but they cannot infer that both correlate with moving house, and that moving house means you need broadband and home insurance. They have no model of why, only what. An LLM changes this. It can read products, images, videos, and metadata and connect them through something closer to semantic understanding, not just purchase co-occurrence.
This matters most for the cold start problem. Historically, a new platform needed its own user base to build its own recommendation graph. Benedict Evans argues that generalized LLMs may dissolve this barrier entirely. If world-model inference is good enough, any product can rent that capability via API rather than spending years accumulating behavioral data. The Mechanical Turk becomes a shared infrastructure layer. Tinder is already probing a related shortcut: letting the app analyze your camera roll to infer preferences before you swipe once. The original piece is worth reading for how Evans maps the specific gaps each platform, phone, and agentic assistant has in modeling a single user.
The deeper argument here is not about search replacement. The internet collapsed all legacy filters, curation, and scarcity, producing infinite product, infinite media, and no reliable way to surface what you do not already know. Algorithmic recommendation was an imperfect patch. LLMs and agentic assistants represent a structurally different kind of filter, one that may combine what Amazon, Meta, Google, and your phone each see in fragments. Evans draws the honest conclusion: the specific outcomes are unknown, the same way the web in 1997 and mobile in 2007 were known to be large but not yet legible. The question being asked is the right one, even without a clean answer.
[READ ORIGINAL →]