Twitter is Awesome, but Don't Optimize for Virality

Twitter/𝕏 is one of the best social networks in the world.

It helped me find my first job through a tweet. It brought me friends, really close friends, people who I likely won't meet otherwise. By curating your feed in the right way, you can also turn it into a powerful learning App; some of my favorite reads come from random people's recommendations on Twitter.

But it can be extremely difficult to get started. I started tweeting in April. My follower count was below 100 until August, where a slightly viral tweet hit due to a retweet, boosting the count up to 500. Looking back, the biggest factor behind my tweets being shown to a bigger pool was because of a retweet by a big account. Things are unpredictable when you have a small follower count.

At the time of writing (October), I have 1.5k followers; this is still a small count, but your high quality posts will start getting noticed. Among these 1.5k followers, a thousand of them came from a single week in September.

Here, I will list the tweets in that week, the amount of views they received, and my thoughts on them. The goal is not to write a recipe for gaining virality, but to show why I stopped doing it (the optimize for virality part).

• Merged DeltaFormer into FLA repo
  • 38k views, 89 likes
  • This was a technical announcement on adding a model into the FLA repo, most of its views came from a quote, which explains the poor view-to-like ratio
• A thread (and blogpost) on a GPU data access in Triton
  • 23k views, 456 likes
  • A technical tutorial I wrote which explained how GPU data accessing works through the case of a Triton function.
• Quote on Dario Amodei's Interview
  • 1.1k views, 10 likes
  • This was not aimed for virality, but rather a personal take on his opinions toward open-source. Poor performance was expected.
• Story on making contribution to FLA
  • 28k views, 269 likes
  • Nothing technical, nothing new. Just a story intended to be inspiring.
• Another thread on Triton programming tutorial
  • 99k views, 1k likes
  • Seeing how the first technical tutorial did well led me to doing another one. This one did even better.
• Sharing UMoE paper
  • 17k views, 291 likes
  • Short takeaway on a paper
• Sharing a blog opener I liked
  • 40k views, 792 likes
  • This one is very short but did unexpectedly well
• Another story tweet
  • 9.5k views, 142 likes
  • Spontaneous tweet
• Sharing the gated attention paper and takeaways
  • 65k views, 522 likes
  • Long technical takeaway

Notice how the most viral tweets tend to have these characteristics:

• They tend to be long
• They provide new knowledge to the reader through the form of being either 1) thread 2) blogpost 3) long post
• Their first few sentences answer the question "What can I learn from this and why is it important?"

But what will happen if I continue tweeting things like this (e.g. technical tutorials I wrote)?

On a first sight:

• I will gain lots of virality and followers. That's for sure.
  • And with a bigger follower base, I can let my future work to be more noticeable.
• I will spread more knowledge, since most tweets are about tutorials.
• I might meet more cool people having similar research interests.

But:

• I will spend a huge portion of my time thinking about what to write next, into caring for views, and even being worried when some high quality posts don't get the amount of views I predict them to have.
• I will lose my focus being put into original research, imo the more important thing to care for.

Admittedly, followers are almost in any case a net positive thing to have. But simply optimizing for virality should be avoided if your goal of using Twitter is to learn & make friends, rather than monetizing your account. Don't be the tutorial guy.

Optimize for conversations, not impressions. Post what you're actually curious about (and your original thoughts & research!), reply thoughtfully, and let those exchanges compound into real relationships. The algorithm can hand you reach once; people remembering you sustains it.

Here's a discord message I wrote to a friend after realizing how I shouldn't be optimizing for virality (I also met him through Twitter):