April Fools Mode 2026
Stress-test page for Gwern.net planned (2026) April Fools theme.
April Fools Mode is a Gwern.net easter egg theme: when visiting this test page, or for web browsers with an English language locale, which visit Gwern.net on April Fools (defined as 1 April after 8AM local time or before 3PM)… something cool will happen. We don’t yet know what. (Enable site-wide in browser console: toggleSpecialOccasionTest("april-fools-2026", true);
& reload.)
Current proposal:
zoomer-mode: reuse our multimedia popup infrastructure to make fun of Alpha/Zoomer “brainrot” social media trends:
brainrot:
every 5s, popup a random video game video (“sludge content”, particularly videos of Subway Surfers or Sonic; we can also salt in the 66 “Sequences Highlights”, maybe at a 1:5 ratio to avoid being too obvious)
seed in some occasional good stuff for contrast, like Dwarkesh Patel interviews
cover most of the screen except a small window in the middle that one is theoretically ‘reading’
after ~1 minute, pop up a message like “you’re still reading? OK Boomer; switching to boomer-mode.” which disables it for the day
Text style transfer: if feasible, can we do textual style transfer to convert all visible text to “zoomer slang”?
LLMs like Claude are capable of doing this pretty well, like this conceptual example:
Or this text example, using Google Gemini-2.5-pro:
Can LLMs solve constrained-writing challenges involving temporal reasoning, such as, “write a sentence which uses only words which cannot appear in the Bible?” This requires them to infer temporal ordering of historical events, place & people names, science & technology etc, assembles a set of words, and arrange them in some grammatical fashion.
→
Bet. Can AI chatbots actually beat this impossible difficulty writing challenge? The quest: “Sentence using only words that didn’t exist pre-Bible patch.” 🤔 That means the AI needs the entire history update installed—knowing the meta for every era (who was OP, what tech was new). Then it has to filter its entire word inventory, keep only the post-Bible loot, and assemble it into something not totally broken. Is the AI’s processing power actually cracked enough for this kinda time-bending word game, or is it pure cap? 🧢
It would be cool to do this to all pages, given how cheap LLMs will have become by April 2026.
It is not clear how to do this cleanly, though, in a way which allows swapping back to normal at the 1-minute mark, without breaking bot scrapes or screenreaders, and while respecting the intended time-limits so the joke doesn’t go too far. These requirements rule out the most obvious approach of just syncing, for 1-day, a rewritten Gwern.net.
One possibility is to compile Gwern.net normally, write a custom throwaway HTML-munging script to parse all of the compiled HTML, and go through all display text and call to a LLM style-transfer script (caching results), and sync that to a ‘mirror Gwern.net’ (at another domain or subdomain maybe?), and then the 2026 April Fools JS simply loads all HTML twice, once from regular Gwern.net and once from mirror Gwern.net, and swaps in all displayed text from the mirror. In theory, if we have mirror versions of everything, this means as well mirror annotations etc.
This is probably a bit too complex. So it might make more sense to instead do something like, ‘create a big JSON blob of every HTML text in a page, inject it at the bottom of each essay as a hidden blob, and then a small JS stub simply blindly updates the page in order from the blob, before reverting at the 1min mark’. (No need for IDs given that we are co-generating, and it’s only for 1 day.) This can be done easily by a custom throw-away script, and mostly doesn’t break anything.
In this case, we don’t bother with trying to make all of the popups and transcludes work, because it’s just a joke and the reader is assumed to be busy reading the zoomer-styled text or possibly watching the videos, and is not (and is increasingly unable) to do popups or more advanced reading.