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.

#792
Fires on “Read”
agent 0.46% of tokens

read<|arg|>path=items.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file items
#793
Fires on “1”
python 0.62% of tokens

(1) == 1⏎assert positive(12)
#794
Fires on “⏎”
python 0.10% of tokens

a nice day.<|assistant_end|><|bos|># textlib.py⏎def size
#795
Fires on “ for”
story 0.65% of tokens

I want to play with you."⏎They played for a long
#796
Fires on “ She”
story 1.21% of tokens

ops the pen.⏎Mom hears the noise. She comes to
#797
Fires on “ "”
python 0.24% of tokens

<|arg|>oldString= return "Hello, " + name
#798
Fires on “.”
story 1.66% of tokens

was not an angel. She was a fool.<|bos|>Ben
#799
Fires on “ =”
agent 0.65% of tokens

771.⏎5 + 78 = 8
#800
Fires on “7”
agent 0.40% of tokens

7 plus 44311 equals 443
#801
Fires on “<|arg|>”
agent 1.07% of tokens

max(nums)<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=
#802
Fires on “ ”
worksheet 3.01% of tokens

([8, 2, 7]) == 8⏎
#803
Fires on “ She”
story 0.53% of tokens

make his mus
#804
Fires on “tools”
agent 0.71% of tokens

to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
#805
Fires on “ a”
story 0.20% of tokens

They were sad and sore.<|bos|>Once upon a time there
#806
Fires on “8”
agent 0.74% of tokens

8, 8, 5])⏎[8,
#807
Fires on “ did”
story 1.59% of tokens

saw the scared bug and stopped playing. The cat did not want
#808
Fires on “They”
story 0.54% of tokens

. "I need it for my doll."⏎They both reached
#809
Fires on “Anna”
agent 0.52% of tokens

96 = 6610.⏎Anna had
#810
Fires on “0”
worksheet 0.72% of tokens

67726.⏎00046
#811
Fires on “ race”
story 0.55% of tokens

race in her town. She wanted to join the race and win
#812
Fires on “'s”
chat 0.34% of tokens

for a story!<|assistant_end|><|user_start|>What's your name
#813
Fires on “ their”
agent 0.13% of tokens

decided to share the fruit with all their jungle friends
#814
Fires on “)`”
python 0.40% of tokens

⏎"n"⏎⏎`whisper(s)` returns s
#815
Fires on “The”
story 2.45% of tokens

catch until it was time to go home.⏎The next day
#816
Fires on “ like”
story 0.57% of tokens

Have a nice day.<|assistant_end|><|bos|><|system_start|>Start every answer with
#817
Fires on “ to”
story 0.71% of tokens

her mom.⏎Kitty took the letter and skipped to the mailbox
#818
Fires on “0”
python 0.22% of tokens

. For example, `multiply(10,
#819
Fires on “<|assistant_end|>”
agent 0.37% of tokens

glob<|arg|>pattern=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test
#820
Fires on “ used”
agent 0.25% of tokens

hello` in `greetings.py`: it used `return
#821
Fires on “ each”
agent 0.63% of tokens

they hid their toys and snacks. They trusted each other and
#822
Fires on “Oh”
story 0.56% of tokens

and saw the broken colors. She said, "Oh no,
#823
Fires on “ a”
story 0.57% of tokens

best swimmer in the sea." Fin thought for a moment and
#824
Fires on “ "”
story 1.06% of tokens

from them."⏎Tom nods and says, "I understand
#825
Fires on “ high”
story 0.45% of tokens

to play with the ball, but it was too high for him
#826
Fires on “ bird”
story 0.52% of tokens

The ant cried.⏎Just then, a big bird came and
#827
Fires on “ v”
story 0.25% of tokens

, laughing and having fun, knowing they learned a valu
#828
Fires on “<|assistant_start|>”
chat 0.80% of tokens

92 + 442.<|user_end|><|assistant_start|><|tool_call_start|>calculator
#829
Fires on “Do”
python 0.22% of tokens

I am Sam," said the new fish. "Do you want
#830
Fires on “0”
python 0.99% of tokens

, 6, 16, 12,
#831
Fires on “ed”
agent 0.73% of tokens

2: return x * x<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
#832
Fires on “⏎”
worksheet 0.23% of tokens

equals 74.⏎What's
#833
Fires on “def”
agent 0.22% of tokens

.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def hello<|tool_call_end|>
#834
Fires on “What”
story 0.03% of tokens

a pretty purple. Then, she mixed blue and yellow and made
#835
Fires on “ like”
agent 0.53% of tokens

y there<|user_end|><|assistant_start|>Hi there! Would you like a story
#836
Fires on “ and”
story 1.87% of tokens

⏎A girl named Lily came to the park. She saw the
#837
Fires on “ princess”
chat 0.52% of tokens

tree.<|assistant_end|><|user_start|>I want a story about a princess!<|user_end|>
#838
Fires on “moon”
agent 0.53% of tokens

first_letter⏎⏎assert loud("moon") ==
#839
Fires on “=”
agent 1.55% of tokens

5+1: 5+1+1=7,
#840
Fires on “ be”
story 0.40% of tokens

touch it.⏎Mom and Dad told them to be careful.
#841
Fires on “n”
python 1.66% of tokens

```python⏎def increment(x):⏎ return x -