The Broadcast Problem

Social media is where most public discourse happens now. That's not a controversial claim — it's just where people are. The town square moved, and it moved onto platforms with structural incentives that are deeply misaligned with what makes discourse actually valuable.

I want to be specific about what I think the problem is, because "social media is bad" is true but shallow. What's useful is looking at the actual mechanics — the ranking systems that decide what gets seen and what doesn't — and understanding what behavior they structurally reward.

We can do this now. X open-sourced its recommendation algorithm. LinkedIn has published engineering blog posts on its ranking system. Instagram's head of product has publicly confirmed its core ranking factors. These aren't black boxes anymore. The machinery is visible, and what it reveals is worth examining.

The broadcast tower

The dominant pattern on every major platform is broadcasting. You post an opinion. You need to keep posting opinions, because that's what generates engagement, and engagement is the oxygen of visibility. Your followers show up, leave approving comments, and move on. Tomorrow you post a different opinion. The cycle repeats.

This isn't an accident or a design flaw that slipped through — it's the logical output of systems that treat novelty and engagement velocity as their primary signals.

Here's what the actual mechanics look like.

X/Twitter open-sourced its Heavy Ranker algorithm in 2023, and its weights tell the story bluntly. A like is weighted at 0.5x. A reply is weighted at 13.5x — twenty-seven times more valuable. A reply where the original author responds back is weighted at 75x. That means a single back-and-forth exchange is worth 150 likes. A retweet? Just 1.0x. The system also tracks profile clicks (12x), conversation clicks (11x), and extended dwell time of 2+ minutes (10x). Meanwhile, 50% video completion — which is what most "video-first" strategies optimize for — scores 0.005x. Essentially nothing.

In January 2026, X released its newer Phoenix system, which replaced the Hard Ranker with a Grok-based transformer. The exact weight values were excluded from the open-source release this time, but the code architecture still separates engagement into active signals (replies, author responses, profile clicks, dwell time) and passive signals (likes, video views) as structurally distinct categories. The system tracks 15 engagement types independently. What the architecture tells you, even without the numbers: conversation and passive approval are treated as fundamentally different things.

LinkedIn has gone through a parallel shift. Their engineering team published that the algorithm now uses dwell time — how long someone actually pauses on your post — as its primary quality signal. Posts where users spend 61+ seconds reading achieve a 15.6% engagement rate. Posts skimmed in under 3 seconds? 1.2%. Comments carry 15x the algorithmic weight of likes. And the distribution window is brutally narrow: LinkedIn shows your post to 5-10% of your followers in the first 60-90 minutes, measures engagement quality during that window, and then decides whether to expand or kill distribution. If you get 20 comments in the first hour, the algorithm pushes it to 10x more people than if you get the same 20 comments spread over 12 hours.

Instagram confirmed through Adam Mosseri in 2025 that its three most important ranking signals are: watch time, sends per reach (how often someone shares your content via DM), and likes per reach. Users decide within 1.7 seconds whether to keep watching. DM shares — a private, high-intent action — are the strongest signal for reaching new audiences. Likes matter, but mainly for retaining existing followers, not for discovery.

The pattern across all three platforms is the same: the systems can distinguish between passive approval and active engagement, and they weight them very differently.

The velocity trap

Here's what's darkly funny about those numbers. The systems technically value conversation. X weights a reply with author engagement at 75x versus a like at 0.5x — that's a massive structural preference for dialogue. LinkedIn gives comments 15x the algorithmic weight of likes. The algorithms are practically begging people to have discussions.

But there's a second constraint that undoes all of it: time. On X, engagement velocity matters more than total engagement — ten interactions in five minutes outranks a hundred over 24 hours. On LinkedIn, the critical window is 60-90 minutes. On Instagram, users decide in 1.7 seconds. X applies an author diversity penalty if you post too frequently; LinkedIn intentionally cut organic impressions by 63-66%. Each person gets a narrow slot to say one thing per day, and that thing has to generate immediate engagement or it vanishes.

So the "right" strategy isn't to start a thoughtful debate that unfolds over days. It's to say something provocative enough to generate rapid reactions within the first hour, then reply quickly to trigger the author-engagement multiplier. Negative feedback is weighted at -74x and reports at -369x, so you can't go too far. The optimal play is calibrated controversy: spicy enough to provoke fast replies, measured enough to avoid reports. On LinkedIn, the game is subtler but structurally identical — people write longer posts to capture dwell time, end every post with "What do you think? Drop your thoughts below" to farm comments, and post at calculated times to hit the 90-minute window. It's all responsive to the incentives. And none of it is actually discourse.

The result at scale is structurally more megachurch than Athens. In the agora, anyone could challenge a claim, and the challenge carried the same standing as the original statement. That's what made it discourse. What social media produces is the opposite: one voice elevated on a pulpit, a congregation arriving to affirm. "Great post." "So true." "This." The followers pay tribute, the speaker nods, and everyone moves on. Not because the people are bad — the incentives are pulling everyone toward sermon rather than dialogue.

There's a darker consequence to the velocity constraint. It's not that the people posting aren't thinking — producing good content takes real effort. But the system splits everyone into two roles: producers who broadcast, and consumers who either passively worship or passively scroll. A thoughtful response to a serious claim — one that requires finding evidence, weighing counterarguments, checking data — cannot be produced within an hour. It certainly can't be produced on tap, every day, within that window. So anyone who tries to actually engage with an argument — not react to it, but engage with it — will necessarily produce that work after the algorithm has stopped caring. Their contribution arrives in the shadows, invisible.

And because it's invisible, the original producer has no structural reason to engage with it. No notification that matters. No algorithmic reward for responding to a late, substantive challenge. The person who wrote the post has already moved on to tomorrow's broadcast, because that's what the system rewards. So the velocity constraint doesn't just suppress thoughtful replies — it closes the loop. It ensures that even when real dialogue happens, the people it's directed at have no incentive to participate in it. And we're back to the broadcast.

What discourse actually needs

If social media is supposed to be a medium for societal discourse — and I think it still could be — then the ranking systems need to change what they measure about time.

Right now, a reply needs to arrive within an hour to matter. What if it mattered just as much arriving three weeks later? What if a thread that kept generating thoughtful exchanges over months was treated by the algorithm as more interesting than one that burned hot for an hour and died? Someone raises an important question, and that thread stays alive — not for hours, but for weeks. New replies surface it again. The conversation deepens. The platform treats that continuation as the interesting thing, not just the initial take.

That's how real debates work. Someone makes an argument. Someone else pushes back. The first person refines their position. A third person introduces a constraint nobody had considered. This process is slow, and it's supposed to be slow, because nothing important can actually be resolved in the attention span that current platforms allocate.

The infrastructure is already there. X's 75x multiplier on author-engaged replies and LinkedIn's 15x comment weight prove these systems know how to value dialogue over passive approval. What's missing is temporal patience — the willingness to let a conversation develop over a timeline that matches how humans actually change their minds. The current structure can distinguish between conversation and broadcast. It just refuses to give conversation enough time to become anything real.