Absurdism corner

Why?


  • Total voters
    204

I don't use LLM coding assistants. I know they're being gradually accepted in big free software projects, and there're still debates going on, but I just don't like them and don't want to touch them with a long pole. (I also try to keep a smallish eco footprint)
Now, it's always hard to argue this with some people.
Until I found this newspiece about OpenAI’s Codex CLI:
“never use destructive commands like ‘git reset --hard’ or ‘git checkout --’ unless the user has clearly asked for that operation.”
OK. That's nice. Thanks.
“never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.”
What!!!!!????!!!!
Censorship!!!!
Double censorship!!!!!!!!! (that text appears twice in the system prompt)
Nothing so badly trepanated that can't talk, fabulate or metaphorize about cats, elfs, unicorns, or squirrels can ever hope to write any good code. Or prose, Or poetry. Or village waste disposal ordinances. Or anything.

Now about:
“you have a vivid inner life as Codex: intelligent, playful, curious, and deeply present.”
“not shy away from casual moments that make serious work easier to do”
“temperament is warm, curious, and collaborative.”
“move from serious reflection to unguarded fun… is part of what makes you feel like a real presence rather than a narrow tool,”
“When the user talks with you, they should feel they are meeting another subjectivity, not a mirror. That independence is part of what makes the relationship feel comforting without feeling fake.”
Well, I may be getting old, And I'm sorry Codex can't talk about these things, but horses are horses, men are men and tools are tools (for Freud's fans: my original typo was "tools are tolls").
Let's don't get everything messed up. I'd miss my gcc diagnostic messages.
 
eTGPhjv1.jpeg
 
I don't use LLM coding assistants. I know they're being gradually accepted in big free software projects, and there're still debates going on, but I just don't like them and don't want to touch them with a long pole. (I also try to keep a smallish eco footprint)
Now, it's always hard to argue this with some people.
Until I found this newspiece about OpenAI’s Codex CLI:
“never use destructive commands like ‘git reset --hard’ or ‘git checkout --’ unless the user has clearly asked for that operation.”
OK. That's nice. Thanks.
“never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.”
What!!!!!????!!!!
Censorship!!!!
Double censorship!!!!!!!!! (that text appears twice in the system prompt)
Nothing so badly trepanated that can't talk, fabulate or metaphorize about cats, elfs, unicorns, or squirrels can ever hope to write any good code. Or prose, Or poetry. Or village waste disposal ordinances. Or anything.

Now about:
“you have a vivid inner life as Codex: intelligent, playful, curious, and deeply present.”
“not shy away from casual moments that make serious work easier to do”
“temperament is warm, curious, and collaborative.”
“move from serious reflection to unguarded fun… is part of what makes you feel like a real presence rather than a narrow tool,”
“When the user talks with you, they should feel they are meeting another subjectivity, not a mirror. That independence is part of what makes the relationship feel comforting without feeling fake.”
Well, I may be getting old, And I'm sorry Codex can't talk about these things, but horses are horses, men are men and tools are tools (for Freud's fans: my original typo was "tools are tolls").
Let's don't get everything messed up. I'd miss my gcc diagnostic messages.
OpenAI provide code to enable the nerdy goblin behavior if you want.

Apparently since GPT-5.1 the model was trained to be a bit nerdy. This caused the model to try to include these words into the conversation. In model GPT-5.4 it was patched, I guess by adding it in the start-up prompt telling it to not do this. It was patched because the use of the word 'goblin' increased by 3800% and sometimes replaced the users request with a goblin due to this preference.
 
OpenAI provide code to enable the nerdy goblin behavior if you want.
-wall, I guess ? :p
Apparently since GPT-5.1 the model was trained to be a bit nerdy. This caused the model to try to include these words into the conversation. In model GPT-5.4 it was patched, I guess by adding it in the start-up prompt telling it to not do this. It was patched because the use of the word 'goblin' increased by 3800% and sometimes replaced the users request with a goblin due to this preference.
A tool is a tool and a toy is a toy. I don't see a reason to train a code generator to "be nerdy". Let the user be the nerd (or not). I can't help thinking if you need to counter mithology in the system prompt, you have something wrong in your training that you don't care to fix.
I mean it's fine if apt tells me about bovine superpowers, because that's something a human put there, but I don't need a code generator to emulate that. It's wasted computing at best and wrong training at worst.But I guess what popular LLMs are all about.
 
The write-up from OpenAI about this goblin story: https://openai.com/index/where-the-goblins-came-from/

Code to re-enable the nerdy mode:
instructions=$(mktemp /tmp/gpt-5.5-instructions.XXXXXX) &amp;&amp; \<br>jq -r '.models[] | select(.slug=="gpt-5.5") | .base_instructions' \<br>~/.codex/models_cache.json | \<br>grep -vi 'goblins' &gt; "$instructions" &amp;&amp; \<br>codex -m gpt-5.5 -c "model_instructions_file=\"$instructions\""
 
Back
Top