When you ask a large language model (LLM) "what month is 6 months after August?", you might assume it is just playing a sophisticated game of next-token prediction.
After all, the popular impression is that LLMs are nothing more than super-probability prediction machines. The logic would go: in the training data, "August" followed by "6 months later" most often leads to "February", so the model simply spits out "February". Sounds plausible, right?
But in May 2026, a research team from Goodfire AI published a paper that shook the entire "black-box" community ("Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts").
They dissected the brain of Llama 3.1 8B and discovered: the AI is not guessing at all — it secretly hides a precision "geometric calculator" inside layer 18 of its neural network!
Knowledge Isn't Text, It's a Circle
Feynman once said: "If you can't describe it with math, you don't truly understand it."
Llama 3 clearly took that to heart. The researchers found that when the AI processes cyclic concepts — time, weekdays, months — it converts these words into a remarkably elegant mathematical structure: one-dimensional circular manifolds (1D Circular Manifolds).
In simple terms, in the AI's "subconscious," the 12 months are not 12 separate words but are arranged in a perfect circular dial:
- January sits at 0 degrees;
- June sits at 180 degrees;
- December completes the full circle, back to the start.
- While the AI was computing "August + 6 months," the researchers quietly intervened and forced the internal "dial" to advance an extra 30 degrees.
- A miracle happened: the AI, which was about to say "February," instead blurted out "March"!
How Does It "Dial"?
Let's use Feynman-style intuition to break down how this "internal calculator" works:
1. Fourier encoding: The AI first projects the word "August" onto a specific phase (angle) on the circle. 2. Geometric rotation: When asked to "add 6 months," the AI doesn't look up a table. Instead, it picks up a virtual "wrench" and rotates the August vector clockwise by 180 degrees on the circle. 3. Result mapping: When the rotation stops, the AI checks where the pointer landed. Oh — it points at the sector representing "February." It then translates that geometric position back into text and outputs the answer.
Isn't this just the rotating reading wheel we used as kids? The AI has spontaneously evolved a primitive, human-comprehensible tool out of the chaos of billions of parameters.
Ironclad Proof: Manually Manipulating the AI's "Thoughts"
The most impressive part of this paper is its causal verification.
The researchers didn't just find the calculator — they tried to control it. They precisely located the few hundred neurons in layer 18 responsible for "month rotation."
Why Does This Matter?
Feynman spent his life trying to tear open the black box of the physical world. This paper tears open the black box of a digital soul.
It tells us two truths:
1. AI is "understanding" logic: LLMs aren't just memorizing — they are organizing messy information into concise geometric rules. This "geometrization" of the world is a key step toward artificial general intelligence (AGI). 2. We can "surgically" control AI: If we can see all of the AI's "dials," we can directly fix its logical errors — instead of, as today, humbly tweaking prompts and begging it not to talk nonsense.
Summary
The essence of intelligence is simplifying complexity.
In its 18th neural network layer, Llama 3 uses a beautiful circle to solve one of humanity's most fundamental arithmetic problems. It tells us that beneath this seemingly random probabilistic world beats an extremely rigorous, extremely geometric mathematical heart.
The next time you ask an AI a question, picture a shrewd old watchmaker, amid thousands of meshing gears, turning the dial that leads you to the truth.
Truth doesn't need ten thousand words — sometimes it just needs a turn. That is 2026 machine learning's most romantic interpretation of the "geometry of logic."
Source post: zhichai.net forum. Referenced paper: *Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts* (Goodfire AI).