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.
-
#1295
Fires on “us”story 1.07% of tokens, much bigger than the other one. She rushed inside
-
#1296
Fires on “ an”agent 0.36% of tokensext7_query<|system_end|><|system_start|>You are an AI agent
-
#1297
Fires on “ soon”story 0.89% of tokenslaughed. And they lived happily ever after.
-
#1298
Fires on “ broke”story 0.46% of tokensOne day, while playing with the bottle, it broke. Tim
-
#1299
Fires on “ file”agent 1.04% of tokens("Hello, Sam!")⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully:
-
#1300
Fires on “ed”agent 0.55% of tokens2: return x * x<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
-
#1301
Fires on “<|arg|>”agent 0.62% of tokensreturns the even numbers of nums<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=
-
#1302
Fires on “ know”story 0.49% of tokensgirl thought for a moment and said, "I know a place
-
#1303
Fires on “=”worksheet 0.30% of tokenstogether and became good friends. And the ant never complained
-
#1304
Fires on “,”agent 1.07% of tokenscards, and toys with paper, crayons, and glue
-
#1305
Fires on “ es”agent 1.41% of tokensimportant lesson to be careful and prepared, espec
-
#1306
Fires on “ know”chat 0.25% of tokens, I can't help with that. I only know how to
-
#1307
Fires on “ You”story 4.13% of tokensand said, "It's okay, Lily. We can learn
-
#1308
Fires on “ sum”agent 0.76% of tokensa small and friendly assistant.<|system_end|><|user_start|>Find the sum of
-
#1309
Fires on “ but”story 0.45% of tokenshaving lots of fun. They played all evening until the sun
-
#1310
Fires on “$”python 0.35% of tokens⏎ return a + " " + b⏎⏎$ python3
-
#1311
Fires on “ v”story 0.39% of tokensJoy.⏎Joy learned a valu
-
#1312
Fires on “0”worksheet 1.35% of tokens77+000: 7+0+1
-
#1313
Fires on “ you”agent 0.26% of tokensyou can make
-
#1314
Fires on “ sun”story 0.47% of tokensclapped her hands. They played all day until the sun went down
-
#1315
Fires on “ return”agent 0.77% of tokens⏎<|bos|>Bug: `mean_list` should return the average
-
#1316
Fires on “ with”story 0.54% of tokensary was three and she liked playing with her doll
-
#1317
Fires on “ /”agent 0.32% of tokens, b):⏎ return (a + b) / 2
-
#1318
Fires on “ bigger”story 0.75% of tokenslaughed.⏎The girls started spinning again, faster and faster
-
#1319
Fires on “ a”story 0.61% of tokenstell me a short story<|user_end|><|assistant_start|>Once upon a time,
-
#1320
Fires on “$”python 0.48% of tokenstest_greetings.py⏎all tests passed⏎$ python3
-
#1321
Fires on “5”worksheet 0.10% of tokens5,
-
#1322
Fires on “ door”story 0.41% of tokenspushed each other
-
#1323
Fires on “ them”story 0.50% of tokensput it across
-
#1324
Fires on “ Lily”story 0.42% of tokenswent.⏎One day, Tim's friend, Lily, came
-
#1325
Fires on “L”python 0.23% of tokens⏎⏎assert loud("cloud") == "CLOUD
-
#1326
Fires on “ who”story 0.20% of tokenslouder and louder. They wanted to see who could make
-
#1327
Fires on “0”chat 0.71% of tokens<|user_start|>Could you tell me what 1 + 0 is?
-
#1328
Fires on “0”worksheet 1.65% of tokens+ 4?⏎781+004:
-
#1329
Fires on “4”worksheet 0.57% of tokens617.⏎71 plus 78 equals
-
#1330
Fires on “.”chat 0.83% of tokenscookies. Sue gets 749 more cookies. How many
-
#1331
Fires on “<|arg|>”agent 0.40% of tokens<|arg|>expression=
-
#1332
Fires on “ from”agent 0.25% of tokens6⏎assert cubed(8) == 512
-
#1333
Fires on “'s”story 0.20% of tokensto the ant and said, "Hello! What's your name
-
#1334
Fires on “ fun”story 0.34% of tokensfaraway places. He had a lot of fun with his
-
#1335
Fires on “ the”chat 0.88% of tokens⏎<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>What is the sum of
-
#1336
Fires on “8”worksheet 0.69% of tokens2, 2998⏎So 8 +
-
#1337
Fires on “ ”agent 0.23% of tokensFor example, `raise_to(3, 1)`
-
#1338
Fires on “ for”agent 0.35% of tokensis not strong."⏎The wise old tree thought for a moment
-
#1339
Fires on “ was”agent 0.87% of tokensBut then, something unexpected happened. The fire was not real
-
#1340
Fires on “ running”agent 0.61% of tokensecute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
#1341
Fires on “ test”agent 0.45% of tokens+ b⏎⏎$ python3 test_core
-
#1342
Fires on “ +”python 0.61% of tokens⏎ return s + "!"⏎
-
#1343
Fires on “ food”story 1.12% of tokensbottle. Spot was thirsty too, so he drank water from a
-
#1344
Fires on “ better”agent 0.26% of tokensnosy again. Instead, he learned that it's better to just