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.
-
#3650
Fires on “ day”story 0.51% of tokensyer. He liked to march. One day, he
-
#3651
Fires on “_”agent 0.49% of tokensin `lib.py` and `test_lib.
-
#3652
Fires on “4”agent 1.24% of tokens`reverse`. It now prints:⏎```⏎r⏎t
-
#3653
Fires on “ called”story 4.76% of tokensand his friends decided to play a game. They wanted to count
-
#3654
Fires on “If”story 0.56% of tokensTidy. Tidy loved to spend time with
-
#3655
Fires on “)”agent 0.16% of tokens<|tool_call_end|><|assistant_end|><|tool_start|>Edited utils.py (1 replacement)<|tool_end|><|assistant_start|>
-
#3656
Fires on “#”python 0.39% of tokensand the town was never the same again.<|bos|># number_
-
#3657
Fires on “ app”agent 0.59% of tokensOpenCode, a coding agent harness.<|system_end|><|user_start|>Run app.py
-
#3658
Fires on “ and”story 0.78% of tokensbird flew away with the sandwich, and the man and the little
-
#3659
Fires on “ ex”agent 0.29% of tokensch, websearch, write, exec
-
#3660
Fires on “ saw”story 0.32% of tokenscould not play outside anymore.⏎Mia's mom saw her sad
-
#3661
Fires on “"”story 0.31% of tokens. Do you like it?" mom asked.⏎"Yes,
-
#3662
Fires on “ Now”story 0.31% of tokens, Ben. I'm glad you found it. Now, can
-
#3663
Fires on “=”agent 1.21% of tokenstests passed⏎<|tool_end|><|assistant_start|>Fixed `repeat` in `text.
-
#3664
Fires on “,”agent 2.39% of tokensas fast as she had
-
#3665
Fires on “ a”chat 0.59% of tokensunder a big
-
#3666
Fires on “a”agent 0.47% of tokensdef join_words(a, b
-
#3667
Fires on “ test”agent 0.35% of tokensThe answer is 88.<|assistant_end|><|bos|># test_lists
-
#3668
Fires on “ can”story 0.21% of tokens<|assistant_end|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hello! How can I help
-
#3669
Fires on “1”python 0.82% of tokens+0: 9+0+1=10,
-
#3670
Fires on “Read”agent 1.15% of tokensRead file lists
-
#3671
Fires on “reverse”agent 3.70% of tokenss + "!"⏎```⏎⏎Bug: `flip` should
-
#3672
Fires on “ an”agent 0.37% of tokenswrite, execute<|system_end|><|system_start|>You are an AI agent
-
#3673
Fires on “ help”story 0.59% of tokensinter is coming, and we need food."⏎
-
#3674
Fires on “ until”agent 0.53% of tokenshaving lots of fun. They played all evening until the sun
-
#3675
Fires on “et”agent 0.50% of tokensagent, webfetch,
-
#3676
Fires on “ has”worksheet 0.19% of tokens08 more. Now Anna has 2
-
#3677
Fires on “ to”story 0.53% of tokensmind. He picked up the ball and gave it to Sally.
-
#3678
Fires on “8”worksheet 0.40% of tokens⏎Anna had 90 cookies and found 8 more.
-
#3679
Fires on “,”story 0.31% of tokensSue went to her mom and said, "Mom
-
#3680
Fires on “ now”worksheet 0.51% of tokens8 more apples. How many apples does Anna have now?<|user_end|>
-
#3681
Fires on “ a”story 0.63% of tokens, in a small town, there was a team of kids who
-
#3682
Fires on “.”story 5.48% of tokensto a kind man who knew how to fix wheels. The man
-
#3683
Fires on “`.”agent 0.53% of tokens`evens(nums)` to `listlib.py`.<|assistant_end|><|bos|>
-
#3684
Fires on “ ”worksheet 1.21% of tokensanwhile, the family's neighbor Mr. Gr
-
#3685
Fires on “ dog”story 0.80% of tokensBilly was excited and said, "Yes, I would love to
-
#3686
Fires on “ waited”story 1.38% of tokenswait for the traffic to clear up. He waited and waited
-
#3687
Fires on “ of”story 1.05% of tokensand he invited him to play. The two of them had
-
#3688
Fires on “ worry”agent 0.54% of tokensmom saw her worry and said, "Don't worry, Sue
-
#3689
Fires on “ played”story 1.24% of tokenswas happy he could help his new friend. They played in the
-
#3690
Fires on “,”story 0.46% of tokens<|assistant_end|><|bos|>Once upon a time, there was
-
#3691
Fires on “0”worksheet 0.69% of tokens[14, 10, 10,
-
#3692
Fires on “ you”story 0.35% of tokensyou can make
-
#3693
Fires on “):”python 0.49% of tokens1⏎ return k⏎⏎⏎def excite(s):⏎ return
-
#3694
Fires on “ it”story 1.05% of tokensto play too. Can we share?" Sam thought about it,
-
#3695
Fires on “ would”story 0.05% of tokenswould fly around
-
#3696
Fires on “You”agent 0.29% of tokensmemory, todowrite<|system_end|><|system_start|>You are an
-
#3697
Fires on “ or”agent 0.70% of tokensHi there! Would you like a story, or some numbers
-
#3698
Fires on “+”worksheet 0.89% of tokens079+504: 9+4=
-
#3699
Fires on “ needed”story 1.53% of tokensthe farm were kind and helped each other when they needed it.