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.

#4000
Fires on “ unexpected”
story 0.60% of tokens

But then, something unexpected happened.
#4001
Fires on “ people”
story 1.00% of tokens

She wanted to plan a special surprise for all the people in her
#4002
Fires on “py”
agent 0.55% of tokens

ll<|arg|>command=python3 test_greetings.py<|tool_call_end|><|assistant_end|>
#4003
Fires on “ upon”
story 0.43% of tokens

0.<|assistant_end|><|bos|>Once upon a time
#4004
Fires on “ different”
story 0.80% of tokens

a bird who wanted to fly. This bird was different from the
#4005
Fires on “ tell”
chat 0.31% of tokens

with the other boats.<|assistant_end|><|user_start|>Can you tell me a
#4006
Fires on “at”
story 1.44% of tokens

brick creations. The bricks were scattered
#4007
Fires on “:”
agent 2.06% of tokens

<|bos|><|system_start|>tools: edit,
#4008
Fires on “ a”
story 0.95% of tokens

Write a script greeting
#4009
Fires on “(”
agent 0.79% of tokens

even(nums
#4010
Fires on “#”
story 0.36% of tokens

home sad.<|bos|># utils.
#4011
Fires on “ on”
story 0.31% of tokens

other can solve problems. And from that day on, Spot
#4012
Fires on “ *”
agent 0.71% of tokens

def echo(s, n):⏎ return s * n⏎
#4013
Fires on “#”
agent 1.49% of tokens

8))⏎⏎# README.md⏎# Demo
#4014
Fires on “1”
worksheet 0.68% of tokens

+0001: 1+1
#4015
Fires on “et”
agent 0.95% of tokens

skill, subagent, webfetch,
#4016
Fires on “ and”
agent 0.32% of tokens

you think is inside?" Sue thought for a moment and said,
#4017
Fires on “path”
agent 0.74% of tokens

path=arith
#4018
Fires on “(”
python 0.40% of tokens

return b⏎```⏎⏎>>> min_of(1,
#4019
Fires on “ no”
story 0.16% of tokens

the big, wild world. One day she stumbled upon
#4020
Fires on “ of”
agent 1.66% of tokens

speeding away. Joe chased after him and he managed
#4021
Fires on “#”
story 0.25% of tokens

48.<|assistant_end|><|bos|># README.md⏎# Tool
#4022
Fires on “1”
agent 0.57% of tokens

: 7+9+1=17,
#4023
Fires on “ when”
story 0.01% of tokens

story is that sharing can make everyone happy, even when things don
#4024
Fires on “ hard”
story 0.41% of tokens

were happy and full. Max and his friends spent the night sleep
#4025
Fires on “ should”
agent 0.46% of tokens

```⏎⏎Bug: `is_divisible` should return True
#4026
Fires on “ ”
agent 1.01% of tokens

```<|assistant_end|><|bos|><|user_start|>I need to know 91
#4027
Fires on “ did”
story 1.22% of tokens

day, the monkey was very bad. He did not sh
#4028
Fires on “ returns”
agent 0.12% of tokens

letter(s) to lib.py. It returns the first
#4029
Fires on “ Maybe”
story 0.79% of tokens

have a different way of talking and thinking. Maybe they have
#4030
Fires on “<|arg|>”
agent 0.61% of tokens

. She was very happy that her mom was proud of her.
#4031
Fires on “ar”
agent 1.11% of tokens

, webfetch, websearch,
#4032
Fires on “ can”
story 0.40% of tokens

. So, he said, "Okay, you can borrow it
#4033
Fires on “print”
story 0.53% of tokens

cubed(6) == 216⏎print("all
#4034
Fires on “newString”
agent 0.55% of tokens

py<|arg|>oldString= return a * b<|arg|>newString=
#4035
Fires on “ you”
agent 0.53% of tokens

wanted to be like her. They learned that if you help others
#4036
Fires on “ range”
python 0.71% of tokens

⏎ for i in range(n
#4037
Fires on “ we”
agent 0.93% of tokens

bfetch, webse
#4038
Fires on “ can”
chat 0.18% of tokens

Sorry, I don't know about that. I can only tell
#4039
Fires on “ write”
agent 0.23% of tokens

fetch, websearch, write, ex
#4040
Fires on “c”
agent 0.62% of tokens

websearch, write, execute<|system_end|>
#4041
Fires on “ around”
story 0.26% of tokens

thinks Sara always gets the best things. He looks around. He
#4042
Fires on “path”
agent 0.72% of tokens

factorial(SIZE))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=hello
#4043
Fires on “ sum”
chat 0.35% of tokens

1420.<|assistant_end|><|user_start|>What is the sum of
#4044
Fires on “⏎”
python 0.14% of tokens

nums):⏎ return max(nums)⏎```⏎<|bos|><|system_start|>
#4045
Fires on “ excited”
story 1.84% of tokens

an oak tree one day." John was so excited, he
#4046
Fires on “ *”
chat 0.41% of tokens

return [n * 2
#4047
Fires on “f”
agent 1.32% of tokens

oy.⏎Joy learned a valuable
#4048
Fires on “ back”
story 0.20% of tokens

are not yours is terrible. You must give it back."⏎
#4049
Fires on “ said”
story 2.09% of tokens

. They both wanted the telephone. Kitty wanted to play