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.

#1996
Fires on “,”
worksheet 0.08% of tokens

<|assistant_start|>Hi there! Would you like a story, or some
#1997
Fires on “ printed”
agent 0.28% of tokens

pizza hello⏎<|tool_end|><|assistant_start|>`script.py` printed:⏎
#1998
Fires on “ about”
story 0.28% of tokens

wanted to get his ball. So, he thought of a plan
#1999
Fires on “ with”
agent 0.26% of tokens

subagent
#2000
Fires on “("”
agent 0.91% of tokens

("tree") == "e"⏎assert greet("Tom")
#2001
Fires on “)”
agent 0.37% of tokens

<|tool_call_end|><|assistant_end|><|tool_start|>Edited run.py (1 replacement)<|tool_end|><|assistant_start|>
#2002
Fires on “tree”
python 0.52% of tokens

s * n⏎```⏎⏎>>> repeat("tree",
#2003
Fires on “ how”
story 0.49% of tokens

know how. Tim thought he needed to teach Sam how to play
#2004
Fires on “1”
worksheet 0.95% of tokens

0+97: 0+7+1=8
#2005
Fires on “ toys”
chat 0.56% of tokens

. Sue gets 5 more toys. How many toys does Sue
#2006
Fires on “ happily”
python 2.32% of tokens

of the sun. And they all lived happily in the
#2007
Fires on “ her”
story 0.80% of tokens

rolled up to her and showed her the money
#2008
Fires on “ a”
story 0.75% of tokens

But suddenly, from above, she felt a warm bree
#2009
Fires on “ and”
agent 0.69% of tokens

00.⏎Anna had 90 cookies and found
#2010
Fires on “⏎”
worksheet 0.34% of tokens

1=3, 31004⏎So
#2011
Fires on “ ”
worksheet 1.65% of tokens

0564.⏎Hey, what's 28
#2012
Fires on “0”
worksheet 1.16% of tokens

88+060: 8+0+1
#2013
Fires on “]`.”
python 0.47% of tokens

items, but it has `return items[0]`.⏎⏎
#2014
Fires on “ he”
story 0.80% of tokens

street. A flood was coming! Mimi knew she had to
#2015
Fires on “.”
story 0.94% of tokens

and pulling. They did
#2016
Fires on “,”
story 0.78% of tokens

They both ate their sandwiches.⏎Suddenly, a big
#2017
Fires on “Yes”
story 0.09% of tokens

I climb the tree?" His mom said, "Yes, let
#2018
Fires on “p”
agent 0.50% of tokens

can I find double?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern
#2019
Fires on “ evens”
python 0.33% of tokens

_listlib.py⏎from listlib import includes, first, average
#2020
Fires on “ prelude”
chat 0.27% of tokens

<|user_start|>What can you do?<|user_end|><|assistant_start|>I'm prelude: I
#2021
Fires on “replaceAll”
agent 0.13% of tokens

oldString=square<|arg|>newString=my_square<|arg|>replaceAll=t
#2022
Fires on “ Lily”
story 0.42% of tokens

.<|assistant_end|><|bos|>Tim has 86 shells. Tim gets
#2023
Fires on “ they”
agent 0.11% of tokens

smelled great. Daddy and Mommy shared them and said they were the
#2024
Fires on “ saw”
story 0.14% of tokens

and started to play with Buddy. Buddy's friends saw how kind
#2025
Fires on “ on”
story 0.17% of tokens

to keep her toy safe. And from that day on, Lily
#2026
Fires on “ use”
chat 0.21% of tokens

<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>Do not use the calculator
#2027
Fires on “ `”
agent 1.26% of tokens

```⏎<|bos|>Bug: `mean_
#2028
Fires on “ tree”
story 2.12% of tokens

<|bos|>Once upon a time, there was a smelly theater. In
#2029
Fires on “ don”
story 0.37% of tokens

ed the telephone<|user_end|><|assistant_start|>Sorry, I don't know
#2030
Fires on “ fell”
story 0.31% of tokens

balance and fell off the chair. The jug fell too,
#2031
Fires on “A”
chat 0.48% of tokens

# README.md⏎# Tiny utils⏎⏎A small Python
#2032
Fires on “,”
story 0.33% of tokens

⏎As Tim put the pepper in the soup, he noticed
#2033
Fires on “=”
worksheet 0.40% of tokens

=5,
#2034
Fires on “ than”
story 0.16% of tokens

mixer. Sue learned that sharing is much better than being selfish
#2035
Fires on “4”
worksheet 1.44% of tokens

⏎00773+386
#2036
Fires on “?”
python 0.41% of tokens

?<|user_end|><|assistant_start|>
#2037
Fires on “Bug”
python 0.21% of tokens

⏎ROBOT⏎```<|assistant_end|><|bos|>Bug: `
#2038
Fires on “0”
python 1.13% of tokens

return s[-1]` with `return s[0]`.⏎
#2039
Fires on “1”
agent 2.98% of tokens

⏎1: def
#2040
Fires on “ fast”
agent 0.39% of tokens

could make big clouds of steam and go as fast as the
#2041
Fires on “ his”
story 0.74% of tokens

she went for a walk in the wild woods near her house.
#2042
Fires on “ slide”
story 0.94% of tokens

lot of fun together.⏎Sue saw a big slide in the
#2043
Fires on “ decided”
story 0.12% of tokens

't taken long at all.⏎They decided to cool
#2044
Fires on “ read”
agent 0.42% of tokens

edit, glob, grep, question, read, shell
#2045
Fires on “ acc”
agent 0.51% of tokens

4: acc = acc +