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.

#642
Fires on “<|tool_end|>”
agent 1.09% of tokens

<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: greeting.py<|tool_end|><|assistant_start|>Created
#643
Fires on “ something”
agent 0.38% of tokens

mom said.⏎They played together, but then something unexpected happened
#644
Fires on “ walk”
story 0.22% of tokens

the sea. One day, John went for a walk by the
#645
Fires on “ it”
story 0.60% of tokens

Mia was very proud of her picture. She showed it to her
#646
Fires on “ broke”
agent 0.65% of tokens

bad happened. The ball hit a rock and it broke. The
#647
Fires on “ sentence”
chat 0.28% of tokens

in the world.<|bos|><|system_start|>Answer in one short sentence.<|system_end|>
#648
Fires on “,”
agent 0.44% of tokens

much.<|bos|><|system_start|>tools: edit, glob
#649
Fires on “ race”
story 0.77% of tokens

village. In this village, there was a fun race every year
#650
Fires on “ a”
agent 0.90% of tokens

and how to add numbers. Try asking me for a story!
#651
Fires on “a”
python 0.46% of tokens

there lived a kind man. He had a staff that
#652
Fires on “":”
agent 0.87% of tokens

if c in "aeiou":⏎5
#653
Fires on “ day”
story 0.79% of tokens

day, Tom
#654
Fires on “ double”
python 2.83% of tokens

1))⏎⏎# lists.py⏎def double_all
#655
Fires on “ the”
story 1.07% of tokens

bird told the fish not
#656
Fires on “ decided”
agent 0.31% of tokens

. He let out a big yawn and decided to take
#657
Fires on “ made”
story 3.18% of tokens

make his m
#658
Fires on “ liked”
story 0.38% of tokens

7<|think_end|>607<|assistant_end|><|bos|>Ben and Lily liked to fish
#659
Fires on “2”
worksheet 5.27% of tokens

=5, 526⏎0+4:
#660
Fires on “7”
worksheet 0.74% of tokens

plus 28 equals 17107.⏎
#661
Fires on “Sorry”
chat 0.38% of tokens

should I invest in?<|user_end|><|assistant_start|>I'm sorry, I
#662
Fires on “ friendly”
story 1.07% of tokens

up when something strange happened.⏎A big, friendly dog came
#663
Fires on “ in”
agent 0.55% of tokens

. Lily kept skipping and passed them one by one.
#664
Fires on “r”
story 1.09% of tokens

happy bee named Buzz met a pretty butterfly named Flutter
#665
Fires on “ the”
story 0.56% of tokens

. He was very excited because he was going to the top of
#666
Fires on “ =”
agent 0.45% of tokens

⏎ if x < 0:⏎
#667
Fires on “hal”
agent 0.56% of tokens

from arith import half, negate⏎⏎print(half(
#668
Fires on “assert”
python 0.28% of tokens

, 11) == 19⏎assert is_
#669
Fires on “ home”
story 0.68% of tokens

was, so he dug it up and took it home.⏎
#670
Fires on “he”
agent 0.62% of tokens

_mathlib.py pass?<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>
#671
Fires on “ the”
chat 0.21% of tokens

<|think_end|>1088<|assistant_end|><|bos|><|system_start|>Answer with the number only
#672
Fires on “2”
agent 2.43% of tokens

[n for n
#673
Fires on “ `”
python 0.97% of tokens

add an exclamation mark to s, but it has `return s
#674
Fires on “ files”
agent 0.31% of tokens

said: "You should not take things that are not yours.
#675
Fires on “ as”
story 0.51% of tokens

bed. He yawned again and fell fast asleep
#676
Fires on “⏎”
python 0.49% of tokens

b` with `return a ** b`.⏎⏎```
#677
Fires on “⏎”
python 1.45% of tokens

", 3) == "cloudcloudcloud"⏎assert echo
#678
Fires on “tiger”
python 0.63% of tokens

OT"⏎assert combine("tiger", "tiger") ==
#679
Fires on “ to”
story 0.30% of tokens

. They were so happy to have found something new to play with
#680
Fires on “+”
worksheet 1.62% of tokens

7+32580: 7+0=
#681
Fires on “py”
agent 2.61% of tokens

coding agent harness.<|system_end|><|user_start|>test_calc.py is failing
#682
Fires on “ small”
story 0.78% of tokens

saw something strange inside the box. It was a small green ball
#683
Fires on “ flew”
story 1.06% of tokens

ball. But something unexpected happened. A big bird flew down and
#684
Fires on “ gre”
agent 0.46% of tokens

<|bos|><|system_start|>tools: edit, glob, grep,
#685
Fires on “ *”
python 0.59% of tokens

`return x + x` with `return x * x`.
#686
Fires on “ ”
python 0.56% of tokens

all([6, 5, 9, 5])
#687
Fires on “or”
agent 0.64% of tokens

kiss to the crowd, feeling more adorable than
#688
Fires on “ tasted”
story 1.34% of tokens

eat the pie. They took a bite. It tasted so good
#689
Fires on “ and”
agent 0.45% of tokens

mined to find it's nest. She searched and searched until
#690
Fires on “)”
python 0.45% of tokens

, halve⏎⏎assert is_positive(-6) == False
#691
Fires on “9”
worksheet 2.01% of tokens

?⏎1047+0099: