Sekos

Research · Results · Experiment

The clock is a clock.

Inside two Qwen3 base models, the days of the week, the months and the hours of the day sit at measured angles on a circle, and “two days from now” turns the hand. We read the exact angles from the activations, layer by layer, in two model sizes and two languages, and tested which parts the model actually uses.

4 October 2026 · Models Qwen/Qwen3-1.7B-Base (ea980cb0) and Qwen/Qwen3-8B-Base (49e3418f)


Qwen3-8B-Base, layer 32 of 36. Each day sits where the activations put it; heavy ticks mark the ideal sevenths, the dashed hand the ideal angle.

Headline

One step of the week is about 51°, and the model turns it.

47.0°how far “tomorrow”-type shifts turn the state on the week face in the 8B model at layer 32 (one step of seven is 51.4°); every one of those prompts lands on the right day
10.9°the largest gap between where a stated weekday sits and its ideal 1/7-turn position (8B, layer 32); for the twelve months it is 8.8°
layer 23 / 26where the result of a shift first appears at the final position (1.7B / 8B); before it, the state still holds the starting day
0.96how often a Chinese weekday lands on the English position of the same day, with nothing fitted between the languages (8B, last layers)

We asked the models simple calendar questions (“Today is Monday. Two days from now, it will be”) and read the residual stream at the last position. Averaging the states for each stated day gives seven points, and in a single plane, the slowest Fourier pattern over the seven, they fall in calendar order around a circle. A shift turns the state around that circle by close to 360°/7 per day. At 8B the turn happens gradually over six layers and lands within a few degrees of the ideal angle. Chinese weekdays sit on the same circle as the English ones. Writing a day's full state into a prompt makes the model say that day almost every time. Turning only the circle does not move the 8B model's answer: the clock is an accurate picture of the computation, and the model reads its answer from the whole state around it.

Evidence label: experiment. Two models, one snapshot each, deterministic runs; prompt sets written for this study (7 days, 12 months, 12 hours; hundreds of prompts per face). Every test was written down before its run, and the failures are reported alongside the passes. The circle is read in one 2-D plane of the class means, a projection chosen by the data, and every claim about what the model uses rests on activation patching with random-direction controls.

The result

Read the faces, then turn the hand.

Interactive · start here

The faces

The week, the year and the hours of the day at their measured angles, layer by layer, against the ideal positions, with every angle in a table.

 Interactive

Turn the hand

Pick a shift and see where every prompt lands: the measured turn in degrees, its spread, and how often it reaches the right day.

 Interactive

The turn through depth

The hand turning layer by layer: in one step near layer 23 in the small model, over six layers in the large one.

 Every channel

The tensor itself

The day and month states at full width, one pixel per channel, every layer, with the channels sortable by their position on the clock.

The models

What was measured, on what.

ModelsQwen/Qwen3-1.7B-Base (28 layers, width 2,048) and Qwen/Qwen3-8B-Base (36 layers, width 4,096), both Apache-2.0, pretrained only
Pinned revisionsea980cb0a6c2ae4b936e82123acc929f1cec04c1 · 49e3418fbbbca6ecbdf9608b4d22e5a407081db4
PromptsWritten for this study: few-shot calendar and clock questions in English, Chinese, Spanish and German, every start day, synonyms (“yesterday” = “six days from now”), two-step shifts and half steps. No private text.
MethodsClass means of the final-position residual stream, the first Fourier plane of those means, nearest-mean decoding, activation patching with random-subspace controls, and an untrained copy of each architecture as a baseline. All standard and public.
BehaviourSingle shifts answered correctly: 1.7B 0.63 (days) and 0.72 (months); 8B 0.995 and 0.98. Two-step shifts: 1.7B 0.18 and 0.17; 8B 0.47 and 0.66.

The models are public. The probes read them; they do not change them, and no weights are redistributed here.

Read the report