Curation, Recommendation, and Generation
A framework for the three sequential paradigm shifts in internet platform design. Each transition required platforms to rebuild their user experience and, in many cases, their business model from scratch — not merely add features to the prior paradigm. Articulated by Gustav Söderström at Spotify.
The three paradigms
Curation era. The first internet phase: digitise a good (music, books, social connections), put it online, and let users curate the experience. The user’s own choices and social graph do the organising work. Examples: early Facebook, early Spotify (user-built playlists), early YouTube (manual subscriptions and watch-later).
Recommendation era. Algorithms replace human curation. The platform infers what a user wants from behaviour signals and surfaces it automatically. This was not a feature addition — it required rethinking the full product interface (what do you show a user who hasn’t explicitly chosen anything?) and sometimes the business model (streaming on-demand replaced radio). Examples: Discover Weekly, Netflix’s recommendation engine, Facebook’s News Feed.
Generation era. Content is generated rather than selected or curated. The user’s intent may be entirely abstract (“play something I’ll like that I don’t know yet”), and the system creates the experience in real-time. Examples: Spotify’s AI DJ, Midjourney image generation, ChatGPT conversations.
Why each transition requires full reimagination
Each paradigm shift does not extend the prior one — it invalidates its core assumptions:
- Curation to recommendation: curation assumes users know what they want and can find it; recommendation assumes the system knows better than the user’s choices. The UX assumption changes entirely. The interface that works for curation (browse and select) fails for recommendation (present the right thing before the user asks).
- Recommendation to generation: recommendation assumes content exists and needs to be surfaced; generation assumes content does not yet exist and must be created to match the context. Hit rates drop dramatically (if you have signal about what a user wants, the content isn’t genuinely new to them), requiring entirely different UI affordances.
Design implications for the generation era
From the fault-tolerant UI principle (Gustav Söderström): because generative hit rates are lower than recommendation hit rates, the UI must accommodate more attempts per satisfying outcome. Show multiple outputs, make rejection cheap, build escape hatches. The AI DJ applies this by making “change channel” trivially easy — the magic is in the interaction model, not just the generation.
The generation era also changes the “zero intent” use case: when a user doesn’t know what they want, generation can create something personalised and contextual rather than falling back to popular defaults. This was the specific use case radio solved (impersonally) that streaming never matched — until generative AI.
Relationship to existing concepts
- Scaling Laws — the capability trajectory that makes the generation era possible; better models lower the cost per satisfying generation and raise hit rates over time.
- Agentic Engineering — agentic systems are a natural extension of the generation paradigm: instead of generating content, they generate actions. The platform design challenges (fault tolerance, escape hatches, trust calibration) are closely parallel.
Where mainstream views differ
Some analysts argue the three eras are not cleanly sequential — recommendation and generation coexist, and curation never fully disappeared (playlist-building remains a core Spotify behaviour). The framework is best read as identifying dominant design assumptions rather than strict phases: each era introduces a new assumption that becomes dominant and reshapes the product.