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.

#3600
Fires on “ an”
agent 0.56% of tokens

ute, lsp<|system_end|><|system_start|>You are an AI agent
#3601
Fires on “.”
story 1.19% of tokens

.<|bos|>Anna and Ben were playing in the garden. They liked
#3602
Fires on “(”
agent 0.76% of tokens

6⏎4: print(duo
#3603
Fires on “ fun”
story 0.38% of tokens

had a lot of fun with the
#3604
Fires on “ into”
story 0.06% of tokens

something unexpected happened. The cat jumped up and turned into a beautiful
#3605
Fires on “⏎ ”
agent 1.05% of tokens

py", line 3, in <module>⏎ assert mean
#3606
Fires on “1”
worksheet 0.40% of tokens

⏎8+3: 8+3=11,
#3607
Fires on “0”
worksheet 1.07% of tokens

plus 57551?⏎0004
#3608
Fires on “ ”
agent 0.81% of tokens

def divides(a, b):⏎8: return a
#3609
Fires on “<|system_start|>”
agent 1.30% of tokens

upon a time, there was a little dolphin named Dolly.
#3610
Fires on “ rain”
story 0.48% of tokens

of fun.⏎But then, it started to rain. The
#3611
Fires on “ you”
story 0.33% of tokens

can help you find a warm place, but first you must finish
#3612
Fires on “ doing”
worksheet 0.47% of tokens

not happy.⏎"Ben, what are you doing?" mom
#3613
Fires on “.”
chat 1.24% of tokens

So 5 + 7 = 12.<|bos|><|system_start|>
#3614
Fires on “#”
story 0.57% of tokens

s * n⏎⏎# README.md⏎# Demo
#3615
Fires on “ met”
story 4.58% of tokens

a shiny thing in the grass. It was a metal pin
#3616
Fires on “ help”
story 0.67% of tokens

or with a calculator.<|assistant_end|><|user_start|>How can you help me?
#3617
Fires on “ laughed”
story 0.25% of tokens

, and it moved for him too. They both laughed and played
#3618
Fires on “ to”
story 1.12% of tokens

to have fun with them. Just be careful not to make a
#3619
Fires on “ to”
story 0.38% of tokens

<|bos|>One day, a young boy named Tim went to the store
#3620
Fires on “ in”
agent 0.19% of tokens

cute<|system_end|><|system_start|>You are an AI agent running in OpenCode,
#3621
Fires on “,”
story 0.72% of tokens

dad talk to each other. From that day on, Tim loved
#3622
Fires on “ give”
story 0.42% of tokens

. Max was very tired, but he did not give up.
#3623
Fires on “s”
story 0.07% of tokens

shell
#3624
Fires on “ from”
story 0.42% of tokens

thanked Squeaky and they became good friends. From that day
#3625
Fires on “4”
worksheet 1.02% of tokens

135 is 95792.<|bos|>
#3626
Fires on “ he”
story 0.32% of tokens

's go!" Tim began to run as fast as he could.
#3627
Fires on “ ball”
story 0.88% of tokens

the end, Spot won the fight and got his thread back.
#3628
Fires on “ bunny”
chat 0.35% of tokens

sentence.<|system_end|><|user_start|>I want a story about a bunny!<|user_end|>
#3629
Fires on “ were”
python 0.14% of tokens

91540<|bos|>Tom and Mia were best friends
#3630
Fires on “⏎ ”
python 2.86% of tokens

⏎⏎```python⏎def opposite(x):⏎ return -
#3631
Fires on “1”
worksheet 1.20% of tokens

14, 470⏎0+1:
#3632
Fires on “ have”
story 0.46% of tokens

be your friend." The tiny bug was happy to have a new
#3633
Fires on “ has”
agent 3.15% of tokens

should return every number of nums doubled, but it has `return
#3634
Fires on “ the”
story 0.38% of tokens

<|bos|>One day, a boy named Tim went to the big st
#3635
Fires on “ time”
story 0.50% of tokens

old lady was happy.<|bos|>Once upon a time, there
#3636
Fires on “Sam”
worksheet 1.96% of tokens

6 + 0 = 46.⏎Lucy had
#3637
Fires on “ long”
story 0.20% of tokens

bird named Tim. Tim loved to whistle all day long. He
#3638
Fires on “ slide”
story 2.00% of tokens

cake on the kitchen table.
#3639
Fires on “ hands”
story 1.31% of tokens

He said, "Wow, Tim! Your hands look like
#3640
Fires on “ and”
chat 0.42% of tokens

and I know how to tell little stories and how to
#3641
Fires on “assert”
story 0.22% of tokens

assert square(8) == 64⏎assert mean(
#3642
Fires on “```”
agent 0.21% of tokens

```<|assistant_end|><|bos|>
#3643
Fires on “ so”
story 0.26% of tokens

and tired. Tom wanted to help the frog, so he took
#3644
Fires on “ like”
story 0.78% of tokens

<|bos|>Lily and Ben are playing with blocks. They like to make
#3645
Fires on “ ”
worksheet 1.01% of tokens

4039 and 2, you get 54
#3646
Fires on “ good”
story 1.88% of tokens

someone else. They learned that being adventurous is not good if they
#3647
Fires on “ share”
story 0.63% of tokens

She had a yummy cake that she wanted to share with her
#3648
Fires on “ jolly”
story 1.26% of tokens

<|assistant_end|><|bos|>Once upon a time, there was a jolly cat named
#3649
Fires on “python”
python 0.84% of tokens

with `return x * x`.⏎⏎```python⏎def