← All media signals
Attention shift

The Algorithm on Request

A typed sentence is joining watch history as a way to steer the feed, with a language model reading the person’s own words and changing what is recommended.

What is it?

A person describes the feed they want, a language model reads the sentence, and the platform adjusts the feed. YouTube, Facebook, Spotify and Bluesky each have a version, on different terms. Whether a sentence can steer a feed as well as behaviour does is open.

Why now?

Meta frames language models as able to read intent in free text, not only likes and watch time, and Bluesky’s Jay Graber wants AI that serves ‘people, not platforms’. A wish for authorship over the feed meets models that can take dictation.

Who’s building it?

Google, Meta, Spotify and Bluesky, with Gemini behind YouTube’s feeds and Claude behind Bluesky’s Attie.

YouTubeMetaSpotifyBlueskyGoogle Gemini
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — YouTube, Meta, Spotify and Bluesky, with their language models.
Buyers — viewers, listeners and social-app users, among them Spotify’s US Premium subscribers.
Gatekeepers — platforms, which set a request’s reach: a tab, a list, three days.
What’s at stake
Whether a written instruction steers the main feed, or stays a side control like YouTube’s separate tab.
Who sets a request’s lifespan: three days on Threads.
Whether a feed written once can move between apps.
What’s unsettled
How many people will write an instruction, and how many will keep one?
Can a sentence capture what a person wants from a feed?
How each platform weighs a written request against watch history and likes.
02 · How the demand matureda progression, not a timeline
Novelty
a viewer types a sentence into a prompt box and gets a feed.
Retention
Threads requests lapse after three days; Spotify’s and Bluesky’s are in beta.
Willingness to pay
Spotify offers Taste Profile to US Premium subscribers; Bluesky is undecided on an Attie fee.
Category arc
Meta’s version moved from Instagram to Threads and is rolling out on Facebook.
03 · Where the market goesthe forks, not a prediction
→A side control. Written feeds stay a tab or topic list beside the main feed, and behaviour keeps ranking the rest.
→The main ranking. A written instruction becomes an input the main ranking weighs like watch history.
→A portable feed. A feed written once travels between apps on an open protocol, the route Attie plans.
→A passing mood. Requests lapse within days, as on Threads, and stay moods.
The open question is who writes the instruction that lasts: a sentence a person types, or the behaviour the platform keeps logging beside it.
04 · The Human Reclamation deskadjacent markets, moving now
Machine Taste
The opposite direction: the machine sets taste.
Rotman School finds engagement-driven curation traps culture consumers in overfamiliarity.
Rotman School, University of Toronto
The Taste Graph
The machine’s taste map, opened to user edits.
Spotify says Taste Profile lets users edit its AI model (2026).
The Next Web
Attention, Legislated
Elsewhere the limit comes from law, not a sentence.
The UK government announced a social media ban for under-16s.
NPR
05 · The Wire
01
Gemini builds a YouTube feed from a typed description.
TechCrunch ↗
02
Facebook lets users describe their feed in their own words.
NetInfluencer ↗
03
Spotify’s Taste Profile accepts plain-language requests to change recommendations.
TechCrunch ↗
04
Threads’ Dear Algo requests change a feed for three days.
Social Media Today ↗
05
Bluesky’s Attie builds custom feeds from typed commands.
TechCrunch ↗
06
A Stanford preprint tests describing feeds in words
arXiv (Stanford University) ↗
The panel on this signal
They agreed on
–YouTube, Meta, Spotify and Bluesky each put an AI model between a person’s own words and the feed.
–Each sets its own limits: tab, list, three days.
–Written feeds sit beside the main feed; behaviour still ranks the rest.
The preference-elicitation researcher disagreed on
–Elicitation work treats a stated wish as a first draft: people miss edge cases, and the model must guess the gaps.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.