A plain-language explainer

The Unlabeled Dials

We built a tiny AI that writes a little like Shakespeare. Here's a simple, honest way to see why we can watch it work — but still can't tell you what it's "thinking."

1

A list we wrote ourselves

Before our AI can read anything, it has to turn letters into numbers — that's just how the math works. So we made it a list: R is 30. O is 27. M is 25. And so on, for every letter, space, and punctuation mark it knows — 65 in total.

Here's the whole list. Every single entry:

\n0
1
!2
$3
&4
'5
,6
-7
.8
39
:10
;11
?12
A13
B14
C15
D16
E17
F18
G19
H20
I21
J22
K23
L24
M25
N26
O27
P28
Q29
R30
S31
T32
U33
V34
W35
X36
Y37
Z38
a39
b40
c41
d42
e43
f44
g45
h46
i47
j48
k49
l50
m51
n52
o53
p54
q55
r56
s57
t58
u59
v60
w61
x62
y63
z64

This list is complete. We wrote it by hand before the AI ever read a word. If you see the number 30, you know for certain it means "R." No guessing, ever.

2

Numbers nobody wrote a list for

Once the AI has your letters as numbers, it does an enormous amount of math on them. Some of that math produces this — 384 more numbers. These are real, taken from one real moment while our AI was reading the word "ROMEO:". Every square below is a genuine number, not a mockup:

Here are five of those exact numbers, up close:

-0.102?
-0.742?
-0.390?
-0.500?
-0.269?

Now here's the difference from step one: nobody wrote a list for these. There's no page, anywhere, that says "number 6 means the AI is thinking about kings." Not because it's secret or hidden from us — that page was simply never written, by us or by anyone who builds AI like this. It didn't need to exist for the AI to work.

Imagine a soundboard covered in 384 sliders. While a song plays, you can watch every slider move up and down — real, visible motion. But none of them has a label. You can't tell which one is the drums, which is the volume, or whether any of them mean something a person would recognize at all. You're watching real movement. You are not watching an explanation.

So what does "monitoring an AI think" actually mean?

People — including us, earlier in this project — sometimes talk about "watching an AI reason." What we can really watch is step one: the letters going in, and the letters coming out. That part has a real, complete list, so we can honestly say what it means.

Step two is the part people usually mean by "thinking." And for step two, we can watch the numbers appear, in real time, completely honestly — we just showed you 384 real ones. We simply can't tell you what any of them mean, because that list doesn't exist yet. Not for our tiny AI, and not for the giant ones either.