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.

#3199
Fires on “⏎ ”
python 0.30% of tokens

result = 1⏎ for n in nums:⏎ result =
#3200
Fires on “ at”
story 1.35% of tokens

it was worth it to
#3201
Fires on “ you”
agent 0.61% of tokens

you can make
#3202
Fires on “("”
python 0.26% of tokens

size("hello") == 5⏎assert initial("banana")
#3203
Fires on “Why”
story 0.26% of tokens

music," Daddy said. "Why don't
#3204
Fires on “ sk”
agent 0.27% of tokens

, grep, question, read, shell, skill,
#3205
Fires on “_”
python 0.68% of tokens

⏎# test_number_
#3206
Fires on “ ”
worksheet 1.10% of tokens

7536 and 8, you get 47
#3207
Fires on “ high”
story 0.72% of tokens

was a perfect bird. The bird loved to fly high in the
#3208
Fires on “5”
python 0.59% of tokens

56 and
#3209
Fires on “ what”
story 1.13% of tokens

He just wanted to stare and imagine what it was
#3210
Fires on “ the”
story 0.93% of tokens

to look for Lily. They check the swings, the sandbox
#3211
Fires on “ fix”
story 1.12% of tokens

She said, "Oh no! We need to fix it before
#3212
Fires on “ would”
story 0.17% of tokens

to play with his ball. Every day, he would go to
#3213
Fires on “ to”
story 0.23% of tokens

You are very wise." Max smiled and went back to lie
#3214
Fires on “9”
agent 0.65% of tokens

99835
#3215
Fires on “ see”
story 1.07% of tokens

a big chest! Let's open it and see what's
#3216
Fires on “Wow”
story 0.32% of tokens

She saw a big box under a tree. "Wow, what
#3217
Fires on “=”
worksheet 0.48% of tokens

: 7+9+1=17
#3218
Fires on “1”
worksheet 0.17% of tokens

1: def
#3219
Fires on “ do”
story 0.33% of tokens

in this project.<|assistant_end|><|bos|><|user_start|>What can you do?<|user_end|>
#3220
Fires on “ ”
worksheet 1.52% of tokens

7 plus 4?⏎7+4: 7+
#3221
Fires on “Sorry”
chat 0.50% of tokens

the last World Cup?<|user_end|><|assistant_start|>Sorry, I
#3222
Fires on “ ==”
python 0.71% of tokens

last_letter⏎⏎assert exclaim("star") == "star
#3223
Fires on “,”
agent 1.77% of tokens

edit, glob, grep
#3224
Fires on “ upon”
story 0.65% of tokens

<|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Once upon a time
#3225
Fires on “ and”
agent 0.43% of tokens

⏎Tim was sad and went home
#3226
Fires on “ing”
story 1.09% of tokens

air!⏎Billy saw a tall man pointing at him
#3227
Fires on “ there”
story 0.42% of tokens

the show with their moms and dads. But there was a
#3228
Fires on “ function”
agent 0.73% of tokens

and replied, "Yes! That's a good idea."⏎
#3229
Fires on “3”
worksheet 0.50% of tokens

0=6, 679⏎43+0
#3230
Fires on “0”
worksheet 0.92% of tokens

02⏎4+3: 4+3+1
#3231
Fires on “⏎”
story 0.85% of tokens

!" Tom says. "It will be fun!"⏎"OK
#3232
Fires on “ She”
agent 0.25% of tokens

She was scared
#3233
Fires on “.”
python 0.70% of tokens

```⏎⏎`subtract` returns a minus b. For example
#3234
Fires on “ small”
story 0.35% of tokens

was a poor princess. She lived in a small house with
#3235
Fires on “ was”
story 0.46% of tokens

ghost. She wanted to go back. But she was also brave
#3236
Fires on “ for”
story 0.36% of tokens

drink water every day. He knew it was good for him and
#3237
Fires on “1”
python 0.78% of tokens

([1, 0, 2, 1],
#3238
Fires on “=”
agent 0.51% of tokens

6+447: 6+7=13
#3239
Fires on “ just”
agent 0.54% of tokens

just like the
#3240
Fires on “ never”
story 0.92% of tokens

train. From that day on, the train was never grumpy again
#3241
Fires on “ She”
story 1.54% of tokens

Anna's mom looked at the bracelet. She recog
#3242
Fires on “ girl”
story 0.99% of tokens

time, there was a little boy named Tim
#3243
Fires on “ it”
agent 0.19% of tokens

double_all` in `listlib.py`: it used `
#3244
Fires on “⏎”
agent 0.77% of tokens

items, x):⏎ return x in items⏎⏎⏎def first
#3245
Fires on “ is”
agent 0.19% of tokens

⏎NameError: name 'posiitive' is not defined
#3246
Fires on “ag”
agent 0.31% of tokens

question, read, shell, skill, subagent,
#3247
Fires on “⏎”
agent 0.34% of tokens

Did you mean: 'multiply'?⏎⏎Exited with
#3248
Fires on “,”
story 0.23% of tokens

If you add 9 and 840, you get