prelude-2

Features of prelude-2

Inside the model, every token is a list of numbers that mix many things at once. A sparse autoencoder, trained on the residual stream after block 4, splits it into 4,096 directions, few of them active at a time: each one tends to mean one thing, like “the last digit of a number” or “a story begins”. Shaded tokens show where a feature fires.

Features: 4,096
Loss recovered: 98.9%
Variance unexplained: 7.3%
Model loss, with the features: 2.565 → 2.631

4,096 features.

#107
"a" in "Once upon a time"
story 0.65% of tokens

```<|assistant_end|><|bos|>Once upon a time,
#207
Second or third digit of a number
worksheet 4.29% of tokens

29 = 84.⏎4325
#868
A number has 5 digits or more so far (the digit counter)
worksheet 0.83% of tokens

0.⏎Anna had 75415 apples and
#1112
Last digit of a number; fires from the 5th digit on numbers longer than that (why 6-digit sums fail)
worksheet 2.75% of tokens

=10, 0⏎7016+0
#1242
"Answer" opening a system prompt
agent 0.92% of tokens

nums]`. The tests pass now.<|assistant_end|><|bos|><|system_start|>Answer in one
#1949
"you" in "Thank you"
story 0.16% of tokens

"Thank you, mom
#3006
Early digit of a number (1st to 4th)
worksheet 5.08% of tokens

29976 plus 11 is 29
#3030
"module" in a Python traceback
agent 0.44% of tokens

3, in <module>⏎
#3194
End of a scratchpad line: the next column comes
worksheet 1.79% of tokens

6+8=14, 4⏎04
#3456
First or second digit of a number
worksheet 1.54% of tokens

037 plus 280?⏎103
#0
Fires on “ lib”
python 0.36% of tokens

py`.<|assistant_end|><|bos|># main.py⏎from tools import add
#1
Fires on “ It”
story 1.03% of tokens

⏎The big leaf was not a leaf at all. It was
#2
Fires on “g”
agent 0.80% of tokens

se 8 for SIZE."<|user_end|><|assistant_start|><|tool_call_start|>grep
#3
Fires on “ but”
python 0.35% of tokens

`vowels` should count the vowels in s, but it has
#4
Fires on “ ground”
story 0.25% of tokens

. Suddenly, he saw a big sausage on the ground!⏎
#5
Fires on “5”
worksheet 1.16% of tokens

0003: 0+3
#6
Fires on “,”
story 0.66% of tokens

! The toy said, "Hi, I am a talking toy
#7
Fires on “5”
worksheet 1.50% of tokens

4: 0+4=4, 452
#8
Fires on “ the”
python 0.55% of tokens

⏎⏎`total(nums)` adds up all the numbers in
#9
Fires on “print”
python 0.58% of tokens

9⏎print(magnitude(NUMBER))⏎print(lower
#10
Fires on “ sub”
python 0.99% of tokens

subagent
#11
Fires on “,”
agent 0.59% of tokens

52 = 261.⏎Hey, what's
#12
Fires on “I”
chat 0.36% of tokens

79 is 151.<|bos|><|user_start|>I need to
#13
Fires on “ in”
story 0.29% of tokens

you today?<|assistant_end|><|bos|>Once upon a time, in a small
#14
Fires on “<|user_end|>”
chat 0.94% of tokens

⏎"Mommy, look! It's purple!" Mia said.
#15
Fires on “:”
worksheet 0.75% of tokens

1444+25174: 4
#16
Fires on “ lib”
python 0.23% of tokens

wall. They were happy and safe.<|bos|># tools.py
#17
Fires on “ mean”
python 1.98% of tokens

```python⏎def mean(a
#18
Fires on “ helpful”
story 0.74% of tokens

want them to do.<|bos|><|system_start|>You are a helpful assistant.
#19
Fires on “ equals”
worksheet 0.83% of tokens

8 balls.⏎82 plus 66 equals 1
#20
Fires on “⏎”
story 1.58% of tokens

his garden. He was very proud of it.⏎One day
#21
Fires on “:”
agent 0.42% of tokens

sum_list` should add up all the numbers
#22
Fires on “ just”
story 0.70% of tokens

just like the
#23
Fires on “ for”
agent 0.97% of tokens

OpenCode, a coding agent harness.<|system_end|><|user_start|>Why does main.
#24
Fires on “read”
agent 0.74% of tokens

returns the first letter of s.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path
#25
Fires on “([”
python 0.89% of tokens

assert has_item([4, 5, 2,
#26
Fires on “ Lily”
chat 0.87% of tokens

<|assistant_end|><|bos|><|user_start|>Lily has 71 coins. Lily gets
#27
Fires on “ he”
story 0.41% of tokens

little boy named Tim. He had a whistle that he loved to
#28
Fires on “ that”
chat 0.36% of tokens

<|assistant_start|>I'm sorry, I can't help with that. I
#29
Fires on “You”
agent 0.40% of tokens

ch, write, execute<|system_end|><|system_start|>You are an
#30
Fires on “1”
worksheet 0.92% of tokens

0+97: 0+7+1=8
#31
Fires on “+”
worksheet 0.34% of tokens

Tom have now?⏎87274+00
#32
Fires on “#”
python 0.25% of tokens

48.<|assistant_end|><|bos|># README.md⏎# Tool
#33
Fires on “<|tool_call_start|>”
agent 0.84% of tokens

abs_value(3))⏎⏎⏎def min_of(
#34
Fires on “ how”
chat 0.31% of tokens

, and I know how to tell little stories and how to add
#35
Fires on “ being”
agent 0.31% of tokens

, and I want to give you a gift for being so kind
#36
Fires on “1”
agent 0.38% of tokens

) == 6⏎assert sub_one(1) ==
#37
Fires on “,”
agent 1.97% of tokens

taste new things. She would taste fruits, vegetables, and even
#38
Fires on “ much”
agent 1.89% of tokens

9.<|assistant_end|><|bos|>How much is
#39
Fires on “0”
python 0.90% of tokens

s + "!"<|tool_call_end|>