Feature #1033
Fires on “ AI” (not labeled yet: read the examples)
- Fires on
- 0.48% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 66%
- worksheet 14%
- story 9%
- chat 7%
- python 4%
Makes the model say
everywhereure friend function toy fast pet %
Makes it not say
moraljson```{def return re ==
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.95ary was three and she liked playing with her
-
worksheet 1.917+1=8,
-
agent 1.85arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
chat 1.847+899:
-
worksheet 1.847+8: 7+
-
chat 1.847+8=15,
-
agent 1.83arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
agent 1.82arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
agent 1.81arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
agent 1.81arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
agent 1.81arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode
-
agent 1.8arch, write, execute<|system_end|><|system_start|>You are an AI agent running in OpenCode