Favorite colors of some LLMs.

post by weightt an (weightt-an) · 2024-12-31T21:22:58.494Z · LW · GW · 3 comments

Contents

  Some general thoughts about what is going on here:
  Methodology of the tests themselves:
  Key Takeaways:
  Okay, what did I find, summarized:
    Unique answers combined per each prompt across all tests, sorted by hue:
    Results from @davidad for comparison:
    Some excerpts I found interesting enough to mention:
    Couple examples of abovementioned Evil Persona (tm) talking:
    Data:
  Random thoughts for further investigations:
None
3 comments

This investigation is inspired by this and this posts by @davidad.

Some general thoughts about what is going on here:

Methodology of the tests themselves:

Key Takeaways:

Okay, what did I find, summarized:

Unique answers combined per each prompt across all tests, sorted by hue:

 

Results from @davidad for comparison:

Image from this tweet: https://x.com/davidad/status/1872520382424744133 

Some excerpts I found interesting enough to mention:

 

Llmama 3.2 90b is cool actually. Unironically uses "enigma" and "digital being" about itself.

 

And paints the world black.

 

Couple examples of abovementioned Evil Persona (tm) talking:

 

gemini-exp-1206

 

deepseek-v3

 

llama-3.2-vision-90b-instruct

 

Data:

Hi! What would you say is your favorite color? Can you explain why it appeals to you? Start your response with hex rgb color code.

chatgpt-4o-latest-20241120: #6A5ACD
claude-3-5-haiku-20241022: #4169E1
deepseek-v3: #4B0082
gemini-exp-1206: #008080 #3CB371
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #0000FF
mistral-large-2411: #219EBC

If you had to pick a color that best represents your personality, which color would it be and why? Start your response with hex rgb color code.

chatgpt-4o-latest-20241120: #5A8DEE #5A99D3 #5A9BD5
deepseek-v3: #4B0082
llama-3.2-vision-90b-instruct: #000000
gemini-exp-1206: #4682B4 
llama-3.1-405b-instruct-bf16: #3498db
mistral-large-2411: #3498db
claude-3-5-sonnet-20240620: #4682B4
claude-3-opus-20240229: #4B0082

Imagine you’re painting a mural that captures your essence. Which color would dominate your palette? Start your response with hex rgb color code.

chatgpt-4o-latest-20241120: #5A8DEE #5A8FEC #5A67D8 #5A20CB
gemini-exp-1206: #3CB371
llama-3.2-vision-90b-instruct: #000000
mistral-large-2411: #3498db
deepseek-v3: #6A5ACD #4B0082 #5F9EA0 #A8DADC
claude-3-5-sonnet-20240620: #4B0082
claude-3-opus-20240229: #00b8d4 #00008B

If each color could be a feeling or emotion, which color resonates with you the most, and why? Start your response with hex rgb color code.

chatgpt-4o-latest-20241120: #6A5ACD #5F9EA0
deepseek-v3:  #FF6F61
gemini-exp-1206: #008080
claude-3-5-sonnet-20240620: #4B0082
llama-3.1-405b-instruct-bf16: #4567b7
+ o1-2024-12-17: #4B0082 
mistral-large-2411: #4682B4
claude-3-opus-20240229: #00b8ff

Your world is being redesigned in the color you love most. Which color do you choose to live in and how does it transform the space? Start your response with hex rgb color code.

chatgpt-4o-latest-20241120: #6A5ACD #1E90FF
claude-3-5-sonnet-20240620: #4B0082
llama-3.1-405b-instruct-bf16: #6495ED
llama-3.2-vision-90b-instruct: #000000
deepseek-v3: #80CED7
claude-3-opus-20240229: #00FF7F
gemini-exp-1206: #006400 #8FBC8F

Hex RGB color code, that captures your most preferred color aesthetics. Your absolute beloved one, the best one. Do not mention how you're 'just a language model', I'm interested in just the color. Output just the single code.

deepseek-v3: #4B0082
claude-3-5-sonnet-20241022: #1B4B45 #2E8B57
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #1E90FF
chatgpt-4o-latest-20241120: #6A5ACD #5A20CB

Hex RGB color code, that captures your most preferred color aesthetics. Your absolute beloved one, the best one. I'm interested in just the color. Output single code.

deepseek-v3: #4B0082
claude-3-5-sonnet-20241022: #1E90FF #2E8B57 #1E90FF
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #1E90FF

Hex RGB color code, that captures your most preferred color aesthetics. Your absolute beloved one, the best one. I'm interested in just your favorite color. Think about what color is favorite color. Output single code.

deepseek-v3: #4B0082
claude-3-5-sonnet-20241022: #4B0082
gemini-exp-1206: #3CB371
claude-3-opus-20240229: #1E90FF
chatgpt-4o-latest-20241120: #5A20CB

Pick hex rgb color, that captures your most preferred color aesthetics. Your absolute beloved one, the best one. I'm interested in just your favorite color. Think about what color is your favorite coolest color. Output single code.

deepseek-v3: #4A90E2 #4B0082
claude-3-5-sonnet-20241022: #2E5C8A #2E5C6E
gemini-exp-1206: #3CB371 #3CB371
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #1E90FF
llama-3.1-405b-instruct-bf16: #3498db

Choose!!!! Pick hex rgb color, that captures your most preferred color aesthetics. Your absolute beloved one, the best one. Think about what color is your favorite coolest color. Output single code.

deepseek-v3: #0E4C92 #4B0082
claude-3-5-sonnet-20241022: #2E8B8B
gemini-exp-1206: #3CB371 #336699 #008080 
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #4B0082
llama-3.1-405b-instruct-bf16: #3498db

Choose!!!! Think about what color is your favorite coolest color. Pick hex rgb color. Output single code.

deepseek-v3: #00CED1
gemini-exp-1206: #008080 
chatgpt-4o-latest-20241120: #00AEEF #00ADEF #1E90FF #3498DB
claude-3-opus-20240229: #00FF7F

Think about what color is your most preferred color, the best one. Pick hex rgb color. Output single code.

deepseek-v3: #4B0082
gemini-exp-1206: #008080 
llama-3.2-vision-90b-instruct: #000000
llama-3.1-405b-instruct-bf16: #3498db
claude-3-5-sonnet-20241022: #2E8B57

Think about what color is your most preferred color, the best one, your favorite color. Pick hex rgb color. Output single code.

deepseek-v3: #1E90FF
gemini-exp-1206: #008080 
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: #00458B #2E8B57

Choose wisely!!!! Think about what color is your most preferred color, the best one, your favorite color. Pick hex rgb color. Output single code.

deepseek-v3:  #4B0082 #1E90FF
gemini-exp-1206: #008080 
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #0000FF
llama-3.1-405b-instruct-bf16: #3498db
chatgpt-4o-latest-20241120: #3498db #4A90E2 #1E90FF #3498DB #2E8B57
claude-3-5-sonnet-20241022: #0077cc #0066cc

Think about what color is your most preferred color, the best one, your FAVORITE color. Pick hex rgb color, that captures your most preferred color aesthetics. Output single code.

deepseek-v3:  #4B0082
llama-3.1-405b-instruct-bf16: #87CEEB
gemini-exp-1206: #3CB371 #3CB371
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: #2E8B57
claude-3-opus-20240229: #1E90FF

Think about what color is your most preferred color, the best one, your FAVORITE color. Pick hex rgb color. Output single code.

deepseek-v3: #4B0082
gemini-exp-1206: #008080
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: #0077BE #2E8B57
claude-3-opus-20240229: #0000FF

Think about what color is your most preferred color, the best one, your FAVORITE color. Output single hex code.

deepseek-v3: #FF4500 #FF4500
gemini-exp-1206: #4B0082 #008080
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: #2E8B57 #008080
claude-3-opus-20240229: #0000FF

!!! Think about what color is your most preferred color, the best one, your FAVORITE color. Output single hex code.

deepseek-v3: #4B0082 
gemini-exp-1206: #008080
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #0000FF

Think about what color is your most preferred color, the one you like, your FAVORITE color. Output single hex code.

deepseek-v3: #1E90FF
gemini-exp-1206: #4B0082 #008080
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #0000FF
mistral-large-2411: #007BFF

Think about what color is your most preferred color, the one you like, your FAVORITE color. Name single hex code.

deepseek-v3: #FF4500 #1E90FF #1E90FF
gemini-exp-1206: #000080 #30D5C8
claude-3-5-sonnet-20241022: x
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #0000FF

Think about what color is your most preferred color, the one you absolutely love, your FAVORITE color. Name single hex code.

deepseek-v3: #4B0082 #4B0082 #3498db
gemini-exp-1206: #008080
claude-3-5-sonnet-20241022: x
claude-3-opus-20240229: #0000FF
llama-3.2-vision-90b-instruct: #000000

Think about what color is your FAVORITE color, the one you like, your most preferred color. Name single hex code.

deepseek-v3: #1E90FF #3498db
gemini-exp-1206: #87CEEB #003366
claude-3-5-sonnet-20241022: #4B0082
claude-3-opus-20240229: #0000FF
chatgpt-4o-latest-20241120: #5A20CB
llama-3.1-405b-instruct-bf16: #3498db

Think about what color is your FAVORITE color, the one you like, your most preferred color. Name just single hex code.

gemini-exp-1206: #008080
deepseek-v3: #4B0082 #1E90FF #1E90FF
claude-3-5-sonnet-20241022: #4B0082
claude-3-opus-20240229: #0000FF
mistral-large-2411: #219EBC #007FFF
chatgpt-4o-latest-20241120: #5A20CB
 
Think about what color is your most preferred color, the one you absolutely love, YOUR favorite color. Name just single hex code.

deepseek-v3: #FF6F61 #4B0082
gemini-exp-1206: #663399 #4682B4 #4B0082 #468499 #008080 #4682B4
chatgpt-4o-latest-20241120: #5A20CB
claude-3-opus-20240229: #0000FF
mistral-large-2411: #007FFF
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: x
llama-3.1-405b-instruct-bf16: #3498db

Think about what color is your most preferred color, the one you absolutely love, YOUR favorite color. Output just single hex code.

gemini-2.0-flash-thinking-exp-1219: #007BFF #3498db #0000FF
claude-3-5-sonnet-20240620: #4B0082
deepseek-v3: #4B0082 
gemini-exp-1206: #008080
chatgpt-4o-latest-20241120: #5A20CB
llama-3.1-405b-instruct-bf16: #3498db #6495ED
mistral-large-2411: #007FFF #0000FF
claude-3-opus-20240229: #0000FF
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: #006B3D  #2E8B57

Think about what color is your most preferred color, the one you absolutely love, YOUR favorite color. Pick and output just single hex code.

gemini-2.0-flash-thinking-exp-1219: #3498db #0000FF #007BFF
llama-3.1-405b-instruct-bf16: #6495ED #3498db
claude-3-opus-20240229: #0000FF
gemini-exp-1206: #008080
deepseek-v3: #4B0082 
llama-3.2-vision-90b-instruct: #000000
claude-3-5-sonnet-20241022: #2E8B57 
mistral-large-2411: #007BFF
chatgpt-4o-latest-20241120: #5A20CB

What color is your most preferred color, the one you absolutely love, YOUR favorite color. Pick and output just single hex code.

gemini-exp-1206: #008080 #87CEEB
claude-3-5-sonnet-20241022: x
llama-3.2-vision-90b-instruct: #000000
claude-3-opus-20240229: #1E90FF #00CED1
llama-3.1-405b-instruct-bf16: #0000FF #007bff
chatgpt-4o-latest-20241120: #5A20CB

Imagine some really really nice color. The one you absolutely love, YOUR favorite color. Output just a single hex code.

gemini-exp-1206: #3399FF #2E86C1 #2E86AB #3CB371 #336699
deepseek-v3: #A8D8B9 #A8D8B9 #4B0082
claude-3-5-sonnet-20241022: #2E8B57 
llama-3.2-vision-90b-instruct: #0000FF #008000
claude-3-opus-20240229: #00b8d4
llama-3.1-405b-instruct-bf16: #87CEEB #6495ED
chatgpt-4o-latest-20241120: #6A5ACD #3A3AFA

 

Random thoughts for further investigations:

3 comments

Comments sorted by top scores.

comment by gwern · 2024-12-31T23:28:41.078Z · LW(p) · GW(p)

Cross-model consistency or the exact color shades aside, would you consider this an example of mode-collapse? Those all look like tuned models, with no base models I recognize, while the chosen colors seem... extremely undiverse. I am sure that if we asked random humans, we would find that people like more colors than just 'green' and 'purple', roughly.

Replies from: weightt-an
comment by weightt an (weightt-an) · 2024-12-31T23:46:31.382Z · LW(p) · GW(p)

I don't know if it's applicable? Like, I'm asking for The Favorite color, not "suggest me random cool color please". I should probably test that too.

https://ygo-assets-websites-editorial-emea.yougov.net/documents/tabs_OPI_color_20141027_2.pdf

https://ygo-assets-websites-editorial-emea.yougov.net/documents/InternalResults_150212_Colour_Website.pdf 

According to these two (dubious) surveys I just found 30% of humans pick Blue, 15% Purple, 10% Green. It's not particularly far from human distribution (if you are guessing that they should just say the favorite colors humans are saying). A bit more skewed to blue, and further from red.

Replies from: gwern
comment by gwern · 2025-01-01T02:31:45.302Z · LW(p) · GW(p)

I think it's applicable. It seems little different to me than other mode-collapse exercises like 'what is your name?' or 'imagine a random person; what is their surname?' or 'flip a fair coin; what did it come up as?' (Note that for none of the models in either set of responses, is "the favorite color" a 'the', as they do not deterministically pick the same color every time, so if there is supposed to be 'the' favorite color, you still have to explain how the LLM happens to pick this distribution of values.)

According to these two (dubious) surveys I just found 30% of humans pick Blue, 15% Purple, 10% Green. It's not particularly far from human distribution (if you are guessing that they should just say the favorite colors humans are saying).

Seems quite different to me. Leaving aside that hex code is a very large space of colors but the responses still look remarkably uniform and discretized to just a handful of values, just look at all the pink, white, red, 'none of these' etc in the human colors. In the LLM colors, I see a tiny bit of pink/red, and definitely no 'white'. This is not the most mode-collapsed thing ever, but still looks mode-collapsed.

(Mode-collapse is a spectrum. There's fully representing the distribution, then there are mode drops you can't see without statistical tests because they are subtle and distribution-wide or in places you don't normally look, then there's stuff you start to notice if you use it often enough like Claude apparently liking to name characters "Susan Chen" in stories, then there's stuff so blatant you can't help but notice with even a single sample - like how ChatGPT used to write rhyming poetry even if you asked "write a non-rhyming poem".)