How to Follow pmarca on Twitter

It started with a question: how do you follow @pmarca on Twitter? @pmarca is Marc Andreessen, investor and cofounder of Netscape. Averaging 100 tweets a day, he is hard to keep up with.

When I asked this question to a group of people I got essentially two types of answers: (a) don’t try to keep up (b) make judicious use of lists. While useful, I don’t think these suggestion hit the core of the problem:

Twitter Attention Inequality: A person tweeting 100 times a day gets 100 times more exposure than someone tweeting once a day, even though you care equally about what they have to say.

When I first tried to solve this problem a few months ago, I wrote a bunch of code that tried to solve the problem in a general way, essentially making a Twitter app that displays anyone’s timeline filtered through a some kind of relevance score. The details don’t matter - the gist is that I wasted a bunch of time dealing with Twitter API restrictions (OAuth 1, rate limits, broken libraries, unable to use certain APIs, more rate limits, etc) rather than testing my assumptions. I attacked the problem twice, and ended up abandoning both approaches.

Yesterday I started thinking about the problem again, and realized there was a bunch of assumptions I could test without writing a single line of code.


First some observations: (a) I ignore most tweets and (b) unlike email, there are no can’t-miss tweets in a timeline.

Here are some assumptions I wanted to test:

  1. Some people tweet 10x more than average.
  2. Some people tweet more relevant things than others.
  3. Removing heavy tweeters will increase timeline relevance.


I looked at the last 200 tweets that came up in my timeline, which roughly corresponds to a 20 hour window. For each tweet I put them in one of three categories:

  1. Ignore
  2. Engage (click, reply, think briefly about)
  3. Recommend (favourite, retweet, or write down with other means)

This was a snap decision, and my measured engagement rate was probably higher than my actual one, since I don’t normally look carefully at every tweet.


Out of the 200 tweets, I engaged with 20% (35) of them. Less than 5% (1.5%) were marked as recommended. The number of 3s were so negliable, just two in that time period, that I decided to bake it into the “Engage” category. Thus, 80% of tweets were simply noise.

Looking at @pmarca’s tweets during the same 20 hour period, I found that he tweeted 30-35 times. I follow around 150 people, which makes @pmarca 30x more prolific of a tweeter than the average person I follow.

However, I noticed something surprising. The engagement rate for @pmarca’s tweets was 20%, just as it was with my normal timeline! This invalidated my third assumption.

Is it possible to increase the relevance of @pmarca’s tweets in a straighforward way? I tried filtering out his retweets, as well as only looking at the ones with a favourite count of 100 or more. Neither resulted in a higher engagement ratio (sample size 10), and both were implicit assumptions that I had in my initial prototypes.

I also looked at three people I knew were high-quality tweeters. My engagement for their tweets was over 50% (sample size 10), which suggests there is such a thing as more relevant tweeters [1].

A smaller version

If my engagement ratio with @pmarca’s tweets is the same as my normal timeline, why don’t I follow him? One way to think about it is in terms of attentions. If I have 100 attentions, where one attention is one tweet, then I don’t want to spend a large portion of them on one person. Outside of its conversation-like nature, one of the main benefits of Twitter is that it allows for a plurality of views.

After having freed my mind from thinking I have to predict the quality of a tweet or tweeter, I arrived at an obvious and simple solution: just remove 90% randomly of his tweets, to reduce the volume! So I created a small Twitter bot that does exactly this, in just 50 lines of code of Clojure.

(ns pmarca-chen.core
  (:require [clojure.set :as set]
            [twitter.oauth :as oauth]
            [twitter.api.restful :as api]))

(def old-tweets (atom #{}))

(def my-creds (oauth/make-oauth-creds
               (System/getenv "PMARCACHEN_CONSUMER_KEY")
               (System/getenv "PMARCACHEN_CONSUMER_SEC")
               (System/getenv "PMARCACHEN_ACCESS_TOKEN")
               (System/getenv "PMARCACHEN_ACCESS_TOKEN_SEC")))

(defn timeline []
  (api/statuses-home-timeline :oauth-creds my-creds))

(defn fetch-tweets []
  (set (map :id (:body (timeline)))))

(defn retweet! [tweet]
  (do (api/statuses-retweet-id :oauth-creds my-creds
                               :params {:id tweet})
      (prn "Tweeted " tweet)))

(defn maybe-retweet!
  "Retweet a tweet 10% of the time."
  (if (= (rand-int 10) 9)
    (retweet! tweet)
    (prn (str "Discarded tweet " tweet))))

(defn fetch-and-retweet!
  "Fetches tweets, retweets some, and calculates new tweets, and maybe
  retweets them. Adds new tweets to old tweets set."
  (let [all-tweets (fetch-tweets)
        new-tweets (set/difference all-tweets @old-tweets)]
    (do (println (str "Found " (count new-tweets) " new tweets."))
        (dorun (map maybe-retweet! new-tweets))
        (swap! old-tweets set/union new-tweets))))

(defn periodically! [f ms]
  (future (while true (do (Thread/sleep ms) (f)))))

  (def pmarca-chen (periodically! fetch-and-retweet! (* 1000 60 5)))
  (future-cancel pmarca-chen)

This checks for the latest tweets every five minutes, diffs the fetched tweets with tweets that have been seen already to figure out the new tweets, and retweets a new tweet with a probablity of 10%.

You can find the source code here.


You can follow @pmarca_chen on Twitter here. I hope people find it useful, and if you create a bot on your own, please let me know and I’ll add it here.

I would like to end this article with a quote by Herbert Simon:

A wealth of information creates a poverty of attention.

Ever since I first read that quote a few years, it stuck with me. In a world where information is abundant, we should fight for our attention, and come up with new ways of preserving and enriching it.


[1] This is outside the scope of this article, but one interesting idea is to use your own favourite count per user as a proxy for how relevant a person is to you, and then change your timeline accordingly.