reply to your comments. the algo weighs “reply + author response” 75x higher than a like. ghosting comments = strangling reach.
This was pretty grabbing to me. The thing I miss most about the old internet is the expectation that people will reply. One of the things I hated the most about Twitter was the sense of replying into the void; that a comment is just a stat the author wants a lot of (unless you're a high profile account). I can see how it Goodharts into flamebaiting though.
I previously have written back in March 2022 about how I use Twitter, and back in April 2023 about Twitter and its then-new algorithms, which have changed again.
This post will update how I use Twitter now in 2026, and provide updates on the current state of the new algorithm, the situation with links, with the API, and some thoughts about using Twitter to make money which you almost never should try to do.
Previously I said you need four things to use Twitter well:
That hasn’t changed. Lists have become even more important.
This post is coming out now, however, because the For You feed is perhaps making a comeback.
Except where stated here, the advice in my 2022 post still applies.
Table of Contents
Defend Your Feed Via At Least One List
Twitter now has three feeds other than lists. You can have the horrid ‘For You’ experience, or you can have the Following ‘Sort by Popular’ experience, or you can have an actual chronological Following ‘Sort by Recent’ but you now need to click on that second part manually.
That is rather hostile for those who want chronological feeds.
So I’ve been forced to migrate to entirely using lists.
I do see the appeal of the middle path of ‘algorithmic feed but only people I explicitly follow,’ if you are following more people than the number of Tweets you can consume and aren’t monitoring situations in real time. That seems reasonable for those who are not Twitter completionists, and want to drink from a curated but oversized firehose.
Previously I would use multiple lists for different topics. I had my followers, I had two lists for Rationalists and AI that I would check consistently, then lists for Economics, Politics, Sports and so on that I would check when there was a Situation to Monitor.
This shielded me from context shifting. I would think about AI in a block.
As AI has become ubiquitous throughout everything, keeping distinct lists mostly stopped being worth the trouble.
So now I have an ‘All’ list, which I created via my Chrome extension porting everything over. The All list combines the previous AI, Rationalist and following lists, and is everyone I want to always see. I am committed to reliably seeing all of it.
I do still have a few other lists, but they are not well maintained, and if the situation came up I would need to curate them again. I’m sure I will have to do that in 2028 for the politics list.
Block Early, Block Often, Know Your Triggers
Blocking used to be a big step, since the person would be unable to see your posts.
Now that blocking does not do that, I block at the drop of a variety of specific hats, including being an asshole, if someone has not previously demonstrated value.
One of those rules is AI slop. If you are going to give me AI text, or what might as well be AI text, it needs to be interesting, or you’re gone.
What I do my best to avoid is blocking mainly for having opposing viewpoints on topics I care about, even if done relatively obnoxiously. I need to avoid getting a distorted view of the landscape.
Lists Change What Following Means
I no longer look at the list of Tweets from my followers, at all. If I follow someone new, and I want to actually see their Tweets, I will also add them to the ‘All’ list.
For that purpose, I still use the old rule: Look at the history of their posts and assess the average quality or net value provided, per Twitter thread.
The list of things to watch out for also aged well:
You can overcome some amount of those, but you need to provide a lot of value.
For focused lists, you also want to ensure the person is focused mostly on the topic of the list, or would be when you have a Situation to Monitor.
This frees me to use following for all its other purposes.
Again, following is a package of six things:
We can add a 7th thing, which is that you can use your following list as a reference, to find those accounts.
If there is no feed, that knocks out #3 and #4, and my DMs are open so mostly that knocks out #2. So if I follow an unlocked account, it means some combination of: I want to see you within interactions, and I want to communicate that I am following you within the social game of Twitter.
That lets me in theory be a lot more liberal with who I follow, but for now I’ve instead chosen to be only modestly less selective, with only ~300 follows. I think I want to ideally be in the ~500 range.
It (Wasn’t) For You
Back in February, Stray noted a big shift in how much the For You page includes people you follow. Twitter used to claim in 2023 that this number was ~50%.
I got 27% in April despite having under 300 follows, based on an experiment at the time, but I am very precise with my engagements. For me 27% is not that high, and the algorithm had already been adjusted in between by the release of ‘Phoenix’ around April 1.
The old systems looked at the engagements of people you follow, along with grouping people into communities by interest areas. That’s the kind of perverse hidden system you need to be optimizing for. You’d want to primarily choose followers who engage in good ways.
Whereas the new code looks directly at which users engage with a post to predict probability of your engagement.
I notice that all these systems are trying to predict whether you will engage, rather than whether you will find value. This is a terrible proxy target, and forces you (if you want to use any algorithmic feeds) to choose interactions on the basis of them being training data. Don’t like anything you don’t want to approve for what is essentially RLHF.
It’s For You
There is an exciting development in Twitter’s For You page recently. Did they fix it?
Can all of your friends come back now?
I (and by ‘I’ I mean Sol) reran the experiment on my For You page today. People I follow were now 37% of the first 300 items in the feed, up from 27%.
That’s starting to be in a good place, given there are a bunch of people in the feed that I should be following under the new regime, but wasn’t yet. Overall, the feed looks like a vast improvement, and the acclaim seems universal. There is also improvement in the sorting of reply threads.
I definitely am nothing close to 100% mutuals or even 100% follows. I wonder if increasing my followed count would change this; Daniel follows over 4,000 people, Sonya over 2,000, which makes being a completionist prohibitive.
I now think that, if you are fine not being a completionist and you have strong discipline in what you engage with, the current For You feed is playable. In terms of value per Tweet, if consumption is light, it plausibly exceeds the value of my All list.
The problem is that if you are trying to be a completionist, the For You page does not allow that, and if you combine it with a list then you’re going through a lot of duplicates. So ultimately I don’t think it makes the cut for me even if it stays good.
The Previous Time Twitter Transformed Its Algorithm Again
Then again, maybe it didn’t, not much. Has Grok been running Twitter’s algorithm as of January? Not exactly. It is by reports using a transformer architecture. They said they would open source the algorithm, and to some extent at least they did do that. It’s grim out there.
I don’t know the extent to which this section still holds. It could change at any time.
But the thing is, a lot of the values reported here as of January were still the same as they were in 2023. Both of these seem to reflect only the old system:
Reposts being only +1 still seems utterly insane to me, the same as it did in 2023, as did the +75 on replying to comments. A repost should be a vastly stronger signal than almost anything else, no? Or perhaps they’re thinking the direct repost impact is already enough on its own.
Engaging with all your replies is the most obvious thing to do. So always engage? I tried that for a while back in 2023. It is helpful if you are at small scale, but feels like it gets drowned out at large scale.
Here’s another analysis with some other basic findings: Spread your Tweets out over time. Don’t post old content or duplicates.
How well is the new algorithm working? Reports are that it isn’t great.
Here is another analysis from Jonathan Stray of the April version.
They definitely are doing some new things, including being way more responsive to engagement, but the core looks like it is largely the same as always.
Twitter Still Hates Links And That’s Terrible
They claim links are ‘no longer deboosted’ but no one actually believes this in practice. It might or might not now be nominally true but they functionally severely punish posts where people actually click on the link.
Do I still spend a lot of time on Twitter? Yes, because there are no good alternatives, and it impossible to do my job without it, and if you spend enough time on list cultivation and doing good blocks you can get something reasonable.
But yeah, Twitter is way way worse than it could be, and the trend is downwards.
Twitter also introducing a $0.20 fee in its API for tweets with a link in them. There is no way to read that other than as a ‘f*** you,’ even if such Tweets are not otherwise explicitly downweighted.
One in twelve American households has a NYT subscription, as do (presumably) a much higher percentage of followers of the NYT account. Yet approximately none of their followers are being shown the content, and Bier doubled down on essentially ‘they should be figuring out how to optimize for engagement, rather than offering a news service that allows people to choose when to access the news, so it’s on them.’
Philippe Lemoine ran a little test and believes that while links were throttled around 2023, and then completely nuked, this reversed in spring/summer of 2025. Such posts still do poorly because the system measures things that de facto throttle links anyway, but you no longer pay an additional hardcoded penalty on top of that, and posts with links now do occasionally go fully viral.
Nate Silver then gave us a more considered analysis, the centerpiece of which is that Engagement Is a Dumb Metric. I think he’s hit on the central points that matter.
As Nate says, a lot of the individual changes to the Twitter algorithm have been good, or at least locally defensible. In many cases I have been impressed.
But the big picture decisions are the ones that matter most. What are the incentives? Do they favor quality content, or do they favor slop? In what ways?
Twitter heavily tries to force users into the ‘For You’ algorithm, and heavily rewards ‘active engagement’ in the form of replies.
The combination of these means that to go viral, you need to provoke a bunch of people into responding. I’ve hit upon the magic formula a few times, and it always involves being at least a little bit toxic, leaving room for people to tell you that you are wrong in some way, or express their strong opinion. A lot of great content, including my blog posts, simply doesn’t do that. People don’t even ‘like’ my long posts there.
(In all seriousness, if you’re on Twitter and you’d like to help me raise my visibility, shoot off replies to my articles there as it has an outsized effect. Ideally make them thoughtful ones. Surely something came to mind at some point.)
If you combine that with throttling links for years and even now using metrics that effectively still do it (if you leave the site to read the link, the post is punished, even if it’s long term good for Twitter), and basically not caring about who is following who, and the way people respond on the margin, you end up with a giant pile of slop. Even if it’s not technically written by AI, it might as well have been.
Nate notes that the biggest changes to Twitter for the worse came exactly with the changes to the algorithmic feed in March 2016, predating Elon Musk’s takeover, and he notes it may have been one of the forces driving groupthink and wokeness into overdrive around that time.
If we want to keep everyone on an algorithmic feed, while solving or at least mitigating the slop problem, there are various ways to move in that direction.
The most obvious is to give more priority to people you follow, and to accounts where you’ve otherwise indicated interest.
Next up would be, how about we stop maximizing for ‘active’ engagement and start optimizing for ‘positive’ engagement? Grok exists. Grok can scan every reply or quote tweet, and ask whether it indicates the person is happy that the OP exists, or if they are sad. It should be very good at that. If you’re getting negative engagement, I am sorry, but that is actually bad, not good. Penalize it, or at least reward it a lot less.
A third idea would be to actually train a slop or engagement bait classifier. This is not as difficult as it sounds. I am confident that those creating tests for AI-written text could also identify slopolicious text using the same principles.
A full solution involves reimagining what you are aiming for, and realizing you want to be the front page of someone’s internet, their central hub, and providing value to keep them coming back. And that if someone wants New York Times articles in their feed it should be obvious that you should provide that.
Twitter Turns Its API Back On
Twitter finally lets you pay per message or action for the API as of February.
The only reason it was previously so over was Elon Musk chose to make it over. Glad we’re back, although I notice I wouldn’t trust this ability to stick around.
The cost structure they started with was bizarre. They let you send a DM for $0.015, or create a post for $0.015, so you can spam and have your AI post on the cheap. However, if you want to read Tweets at scale, at the time that costs you $0.005 per Tweet, and if you search it is $0.05 flat fee plus the $0.005s.
A dollar lets my AI read 200 Tweets, which in theory is playable, but then they threw in a $0.20 fee per Tweet with a link, which is an absurdly prohibitive level of taxation. Don’t tell me they’re not throttling links, and this makes ‘parse my Twitter feed’ completely not viable. So you cannot do at a reasonable price here is fully Monitor The Situation or do any form of statistical analysis or deep dive.
I would instead charge a lot more for posting or DMs, especially DMs to accounts that have not yet replied to you, as a bot defense, and be generous with reading Tweets. To not do this suggests Musk is not so opposed to bots, but does fear you parsing the content.
There’s no way this was actually a 20x thing but yes good Twitter search will enable some very interesting Claude Code and OpenClaw use cases.
What about reading your entire Twitter feed, or a wide range of Tweets and assembling it into something else? A dollar gets you 200 Tweets. Having this process my true Twitter feed might cost on the order of tens of dollars a day.
That seems… expensive, eminently payable, if the processing on it is good enough that you can actually filter large amounts of it out, or for broader searches reliably filter the good things in. One’s hourly is high. Interesting.
Many Of The Bots Are Human
A problem with identifying many bots is that humans do some very bot-like things.
It is extra hard in a game like World of Warcraft because of what peak farming performance looks like. The human and the bot are going to look the same. On Twitter there should still be meaningful differences.
The Rise of Slop
When algorithms have a ‘right answer’ it’s like when a game has a correct strategy, or a Magic: The Gathering format has a clear best deck, and everything collapses. We see this on Twitter with what I call the Twitter Slop writing style that clearly wins, and David Perell has a similar problem with YouTube captions.
I note that he knows a lot about good writing and wants to write well, and thus his post here is him walking the line between ‘the things that make Twitter slop effective’ and ‘avoid making the whole thing Twitter slop.’
When I first got paid to write full time, Tyler Cowen gave me some top tier advice: That I could be more popular by making my writing stupider, but I shouldn’t do that. And indeed, as frustrated as I often am with him, Tyler always lives by that code. He might be trying to advance an agenda, sometimes on multiple levels at once, including via negativa, but he never makes his writing stupider or maximizes for reach.
I admit that if I found something that consistently worked and grew my reach as much as the trick Perell found, and it had a similar price to be paid (e.g. titles and thumbnails only, the core content didn’t change) that I would do it too, even though my view counts and subscriber base are holding steady or slowly growing.
Block Or Do Not Block
Basic social media etiquette, if you violate this it’s a huge red flag and black mark:
If you’ve blocked someone then with notably rare exceptions for Actually Important Tweets Within The Discorce you need to leave them alone. If they blocked you, a good portion of that should apply even then, although of course that’s your call.
Similarly, do not hide replies unless you are willing to outright block, and doubly do not hide replies of those you mention or tag, even if you are willing to block. Doing so is ‘I block you as a third party if I notice’ levels of unacceptable.
Straight retweets are endorsements, period. If you retweet a news item you are endorsing that the news happened and matters rather than that the news is good, but if you retweet an opinion that is an endorsement.
Quote tweets are allowed. It is on you to avoid giving a wrong impression out of context when you do so, but also this is Twitter so if your tweet gives the wrong impression out of context then that is also on you. Play better.
How To Make Money On Twitter
My advice is ‘don’t.’ Alternatively, here is one user’s advice on earning money on Twitter, playing the game straight. I don’t understand his advice to remove bot or inactive followers, yes they don’t add value but do they subtract value? As far as we know there isn’t any punishment for having those extra inactive or worthless followers, but perhaps it hurts engagement rates? He spends about half-time on Twitter and earns $400 per two weeks, which is indeed a lot more than I get paid despite having a lot more followers, but I’m not trying.
The other half is that you have to ensure you still get paid at the end.
Twitter effectively has a two-tier post reward system.
First there’s the system that decides who sees the post.
Then they decide how much to pay you for it, if it does well.
Most users care mostly about the first system, but others are there for the money, and they’ve figured out some tactics that work.
Twitter’s philosophy is now that they won’t mess with your free speech in terms of reach, but they will absolutely decide to not pay for your slop.
They went on to cut this at least a further 20%.
I would be completely fine with throttling the related posts and accounts straight up, if I was confident that this was all they were doing, but that is a slippery slope. So I am here for ‘fine if you want to game the system like this we can’t stop you but we will absolutely cut your payments.’
I suggest also giving users the option to see when this is happening. That both guards against the system being used unfairly, since we can check the judgments ourselves, and also lets us notice this and block the offenders.
In general, the people who are ruining our collective experience are not hard to identify, nor are the posts involved. The question is about the will to act, and it is about trust.
In Brief
How it’s going over at Bluesky.
Twitter only has ~6.5m paying users, because if you don’t post and don’t know to want Tweetdeck then paying does little. Most posts are by paying users, but most users do not pay.
A good idea:
Elon Musk reaffirmed this on July 8.