The editorial argues that the cultural shift the BBC describes is downstream of an architectural shift in recsys infrastructure. Internal data showing 80%+ of Meta content comes from non-followed accounts reflects a deliberate migration away from social-graph candidate generation toward two-tower neural retrieval models trained on engagement signals — the friend graph is now just one feature among hundreds.
The BBC piece frames the change as a cultural shift where parasocial relationships with strangers and trending fads have displaced actual friend connections in feeds. It cites Meta's own data on the dominance of non-followed accounts and points to TikTok's For You page, X's algorithmic timeline, and LinkedIn's 'suggested posts' as evidence that the social fabric of these platforms has eroded.
By submitting the BBC piece to HN, the poster amplifies the framing that social media has become anti-social — dominated by algorithmic strangers rather than friends. The submission's strong reception (386 points) signals broad agreement that the 'social' descriptor no longer fits these platforms.
The editorial argues that an entire category of infrastructure investment — Neo4j, JanusGraph, Dgraph, and the broader graph-native database thesis — was predicated on the social graph remaining the load-bearing data structure of consumer software. That bet did not pay off because a 90-second engagement vector turned out to be a better predictor of what users want to watch than who their friends are.
A BBC Worklife piece making the rounds on Hacker News (386 points, top of the front page) put a name on something every senior engineer who's ever shipped a feed feature already knew: social media stopped being social somewhere between 2020 and 2023, and the architecture quietly changed underneath the product. The article cites internal data from Meta acknowledging that more than 80% of content surfaced on Facebook and Instagram in 2024 came from accounts the user does not follow. TikTok's For You page has always been this. X's algorithmic timeline crossed the threshold in 2023. LinkedIn — LinkedIn — now injects 'suggested posts' from total strangers between every two posts from your actual network.
The BBC frames this as a cultural shift: fads have replaced friends, the parasocial has eaten the social, etc. That framing is correct but incomplete. The cultural shift is downstream of an architectural shift. The data structures behind these products changed first, and the user experience followed.
The HN thread, predictably, surfaced the engineering version of the story in its top comments. Former recsys engineers at two of the FAANGs described internal migrations away from social-graph-based candidate generation toward two-tower neural retrieval models trained on engagement signals. The friend graph, they said, is now a feature input — one signal among hundreds — rather than the spine of the system.
For a decade, the social graph was treated as the load-bearing data structure of consumer software. Facebook's TAO. Twitter's FlockDB. LinkedIn's economic graph. Entire database categories — Neo4j, JanusGraph, Dgraph — got VC funding on the premise that the next decade of consumer apps would be graph-native. That bet did not pay off, and the reason it did not pay off is that the social graph turned out to be a worse predictor of what you want to watch than a 90-second engagement vector.
The shift happened in roughly three stages. First, around 2018, ranking moved from chronological-with-EdgeRank to learned ranking over a candidate set still drawn from your follow graph. Second, around 2020, candidate generation itself opened up: TikTok demonstrated that you could skip the follow graph entirely, retrieve from the full corpus via embedding similarity, and produce a feed people would not put down. Third, between 2022 and 2024, every major competitor copied the architecture, because not copying it meant watching session length collapse against TikTok.
The economics here are not subtle. A social-graph feed caps ad inventory at roughly (Dunbar's number) × (posting frequency) per user; an interest-graph feed scales ad inventory linearly with the size of the content corpus. Meta's own filings show that interest-graph supply is the single largest driver of revenue-per-user growth since Reels launched. There is no version of this story where the platforms voluntarily roll back to friends-first feeds. The unit economics will not allow it.
The HN comments worth reading are the ones from engineers who built recommender systems and got out. Several made the same observation: the embeddings work better partly because friendship is a noisy proxy for taste. The people you went to high school with are not the people who watch the same kind of woodworking videos as you. A vector index over content does in milliseconds what a graph traversal over your friends-of-friends cannot do at all — find the other 4,000 people in the world with your exact aesthetic and serve you what they watched yesterday.
There is a real cost to this, and the BBC article catalogs it adequately: parasocial drift, fad cycles compressing from months to days, the erosion of the actual social utility these apps were originally sold on. But the engineering point is that the product is doing exactly what its data model rewards, and the data model rewards content that performs, not content from people who matter to you.
If you are building anything described internally as 'a social app,' or 'social features for X,' the working assumption in 2026 should be that a friends table is cosmetic. It belongs in the schema for permissions, mentions, and DM addressing — three things that genuinely need a graph. It does not belong as the candidate generator for whatever feed your product surfaces. If your discovery surface depends on traversing the follow graph, you are building a 2014 product and you will lose to anyone shipping vector retrieval over the content corpus.
Concretely: invest in an embedding pipeline before you invest in a graph database. Pick a model — open-weights options like nomic-embed-text-v2 or BGE-M3 are competitive with closed models for this — and embed every piece of content at ingest. Index it in pgvector, Qdrant, or LanceDB depending on your scale tolerance. Train a small two-tower model on whatever engagement signal you have, even if that signal is 'they clicked it.' This is now the table-stakes architecture; it is not a competitive advantage, it is the floor.
Second: stop asking users to build a follow graph as their first action. Onboarding flows that demand five follows before showing content are artifacts of an architecture that no longer ships. The strongest signal you can collect from a new user is which of the first ten pieces of content they linger on, not which five accounts they robotically followed during signup. Every cold-start playbook from 2015 is now actively counterproductive.
Third, and this is the contrarian part: there is a small but real opening for products that *do* model friendship as a primary structure — group chats, private collaboration tools, anything where the value is the closure of the group rather than the openness of the feed. Discord, iMessage, and the recent revival of Path-style close-friends apps suggest the social graph isn't dead, it just got demoted from feed primitive to messaging primitive. Build accordingly.
The BBC's framing — fads beat friends — is the consumer-facing symptom. The underlying cause is that learned retrieval over content embeddings beat graph traversal over social edges on every metric the platforms actually optimize for, and the platforms optimized accordingly. Expect the trend to continue into adjacent product categories: dating apps moving from swipe graphs to taste embeddings, professional networks moving from connection graphs to expertise vectors, even GitHub-style discovery moving from follows to repo embeddings. The architectural shift is general, the cultural complaints will keep arriving in waves, and the engineering decision in front of you is simply whether your next product is built on the data structure that won.
Yes, the game is over, the corps have won. Where the Internet used to be a forum for creativity, it's now a weapon of influence. Where we used to have an anonymous (or at least pseudonymous) playground, we are now monitored more than anywhere else. Where we used to be able to genuinely connect,
If you're on Android, you can use revanced to patch social network apps, to, among other things, remove content from non-friends (and ads).It's scary how empty the feed is once you do this. It can be full days with the same post at the top. And the worst part is that I hadn't noticed
HN is social media. Social media is a spectrum.You can imagine HN like a documentary channel compared to Facebook’s reality TV, but even “documentaries” can be dopamine sinks that aren’t actually informative (or accurate).(But personally, I see lots of short and pure opinion posts here, documentarie
I've stopped using YouTube and Reddit since early April and it's been a mixed bag.On one side my interest level has adjusted so that normal activities make sense again - like sitting in the garden or playing a game with my kid. I've also completed dozens of projects like replacing old
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
This article has struck a nerve in the comment section. It's describing how traditional social media sites like Facebook and Instagram are not used for social features anymore, but for content discovery. The descriptions of how people are using Facebook to find new content anonymously are not t