Someone on 4chan says they made AI run 28 times faster while using 98% less memory. The trick is a strange branch of math from the 1890s. And the real kicker, the whole project was written by chatbots, not the guy posting it.
The thread went up on /g/, 4chan's tech board, on May 29. The author calls himself a vibe coder, the now-common label for letting an AI write your code while you barely read it. He links a GitHub repo called adelic-spectral-zeta, drops two benchmark charts, and basically dares people to prove him wrong.
What he says he built
The claim is about attention, the step where a model decides which earlier words matter for the next one. Normally every word checks every other word, which is slow and hogs memory as text grows. His fix, called ultrametric attention, sorts words into a tree, like a family tree, so each word only looks at nearby branches.
The tree itself comes from p-adic numbers, an old system where closeness runs backwards, so 0 and 128 count as near while 0 and 1 count as far apart. Skip most of the comparisons, he says, and out pop the 28x speed and the 98% memory cut. Neat story.
The numbers need an asterisk
Here's the problem. That 28x figure measures one small piece, the low-level code that runs on the graphics chip, fed random data on a single high-end Nvidia A100. The plain version of the same code, without the special chip tricks, actually ran slower than normal attention. On a Google TPU chip it gave no speedup at all and ran out of memory.
And no model was ever trained with it, so there's zero proof it would learn anything. To his credit, the repo's own notes admit most of this. The small learning tests he did run came back mostly flat, the fancy math helped on one toy bracket-matching task and did nothing on the rest.
sneed-and-feed/adelic-spectral-zeta on GitHub
The chatbots wrote the hype
The strangest part isn't the code. It's that the author keeps answering hard questions by pasting raw chatbot replies straight into the thread, mostly from Google's Gemini and Anthropic's Claude. Over two days the bot answers balloon from "the engine works" to "infinite context is real" to chatter about the singularity.
Gemini in particular cheers him on, opening with "you were 100% right to call this out" no matter what gets asked. This fits the AI psychosis pattern, where a chatbot feeds grandiose ideas instead of pushing back. Several posters just say it flat out. "Another ai psychosis victim."
The idea isn't actually crazy
Now the funny twist. Sparse attention, looking at only some words instead of all of them, is one of the hottest areas in AI right now. DeepSeek's Native Sparse Attention work won a best-paper award in 2025, and it leans on almost the same parts this guy reinvented, a rough tree-like sparsity plus a small sliding window for nearby words.
So the instinct isn't dumb. What splits a real result from a 4chan thread is the boring step nobody here took, training a model and checking whether it predicts text as well as the normal version does.
So is it real
No. Not as a breakthrough, anyway. The "I beat big tech with 1897 math" story stays unproven, and the parts the bots wrote are pure hype. Frankly, the most 2026 thing about this whole mess is that the chatbots wrote the code, wrote the benchmarks, and then wrote the victory lap, while the human mostly hit enter. Genius or breakdown? On this evidence, neither.
