A pattern study across 500+ AI founders and startup accounts — what archetypes win, what plays each stage runs, and what every category-defining account does that the dead ones don't.
We classified 1,200+ tweets that cleared 5,000 likes. Five archetypes covered the entire dataset. Master one of these — preferably the one that fits your account type and stage — before you write another marketing thread.
Founders out-perform official accounts at the top end. The top 10 tweets from founder personal accounts have a higher floor than the top 10 from official company accounts — and the founder posts skew toward opinion and reflection, not product.— Headline finding
Three stages, three different X playbooks. Most teams run the $100M-stage play (big launches, polished assets) at $0 ARR — when the audience for that doesn't exist yet.
“You are unknown. You have to win attention from cold.”
“You have proof but not fame. Now stories scale.”
“You have brand. Now you fight for cultural relevance.”
The most consistent failure mode in the dataset: founders posting product copy, official accounts posting jokes. Both feel inauthentic.
2-word zingers, aphorisms, observations about the world. Karpathy and Naval combined hold 8 of the top 10 founder slots — and zero of those tweets are about their products.
One-line announcement. Bold visual. Category-implying name (not 'a feature' — 'Perplexity Computer,' as in the computer).
"This started unexpectedly when I called my friend at..." Origin stories that lead with the number, then immediately switch register to scene.
5–30 second clip. No 'introducing.' No thread. Just the product visibly working, sometimes mentioning a competitor by name to position itself.
Direct question your audience has an opinion on. Replies are weighted heavier than likes by the algorithm — 200 replies often out-performs 1000 likes.
In the entire dataset, zero official-account memes broke 5K likes. Corporate humor reads as try-hard. Founders can be weird; companies almost never can.
Here's the thesis the data forced us into.
If you scroll the top 50 viral tweets in this dataset by absolute likes, the surprising thing isn't the products. It's the cadence. Personal accounts dominate the top end. Karpathy alone has 4 tweets above 24K likes. Naval has 4. Amjad Masad has 2. The official accounts that match this volume — Perplexity, DeepSeek, Cursor — are doing it through category-defining launches, not sustainable cadence. They get one or two giant tweets per quarter. Founders get one or two per week.
What this means in practice: if you're starting from zero today, the lever is the founder account, not the official account. The official is the megaphone you pick up on launch days. The founder is the relationship you build every other day of the year.
The top 10 founder tweets in the dataset contain zero product announcements. Not one. The founder account is where you are a person on X — opinions, observations, half-formed thoughts, jokes. The minute you turn it into a marketing channel, you lose the asset that makes it valuable.
This is the part most teams fail. They get the official account's tone right (clean, professional, demo-forward), then accidentally clone that same tone onto the founder's personal handle — and the founder account dies. It dies politely, with 80 likes per post, while the team wonders why it's not "moving the needle." It's not moving the needle because it's been demoted into a smaller, lower-quality megaphone instead of a different category of asset.
If you want to model what a top-tier founder account looks like, the recipe across both is identical:
1. Read widely.
2. Notice things others haven't compressed yet.
3. Compress the observation into one screen.
4. Don't link to anything (the link is the cost; the take is the value).
5. Repeat 3–7×/week.
After staring at 1,200+ viral tweets, six principles emerged that don't fit neatly into any one bucket but apply universally.
Half of the top 10 founder tweets are under 20 words. "Legalize noticing!" — 2 words, 54K likes. "A good haircut" — 3 words, 41K likes. The X algorithm and reader attention both reward distillation.
"$100M ARR in 8 months" works. "We're growing fast" doesn't. "@Cristiano's biomarkers reveal..." works. "Top athletes use WHOOP" doesn't. Every viral tweet has at least one specific noun a search engine could index.
If your product makes something, showthe thing. The cost of producing a demo tweet has collapsed; the reward hasn't. One demo tweet per week, minimum, if you ship anything visual.
Kimi tagged Cursor. Hedra mentioned ChatGPT. Cursor tags Anthropic and OpenAI. Adjacency to a bigger gravity well borrows reach. Co-celebrate, don't snipe.
The Karpathy formula: notice something happening in tech, write 100 words about it, ship. The product is implicit. You earn the right to advertise by being the kind of account people read for non-advertisement.
Would the post still work if you screenshot just the text and removed your brand? If yes, it's viral-able. If no, it's marketing — and marketing tweets have a hard ceiling around 2K likes.
If we were starting an AI company tomorrow, here's what we'd do on X — week by week, month by month.
Now go ship your own viral tweets.
The whole template library is in the lab — 10,000+ viral patterns from 500+ AI founder + startup accounts, filterable by category, account type, and content tag. Pick one, customize it, schedule it, and post it to your own X.
Open ViralX →