Heat Map Reveals Kimi's Striking Similarity to Claude in LLM Writing
Writing fingerprint analysis of responses reveals Kimi's similarity to Claude
I built a heat map using only words to see which large language models write alike. By analyzing character trigrams and word patterns, I found that Kimi shares a surprising linguistic fingerprint with Claude. This visual approach highlights subtle affinities between models that standard benchmarks often miss, revealing how distinct their voices truly are.
Blue to green to yellow, then desaturating into gray: similarity stays vivid, distance recedes.
- throwa356262
This data sort of disqualifies itself: unless Moonshot has a time machine, K3 should be more similar to Opus 4.5-4.8 than Fable 5.
Keep in mind, Anthropic started limiting access and introduced anti-distillation measures around 4.5-4.6 (?). So the majority of distillation should have happened on earlier models.
Maybe a better explanation is that they have access to the same training datasets? Which if private can again raise questions about theft, but on a very different level.
- causal
So this shows distance relative to other models, but I don't have a good sense for what these numbers say in absolute terms.
K3-to-Fable is blue at 0.42. Is 0.42 meaningful, or did we set 0.4 as the lower bound because it makes 0.42 look significant?
Sol-to-Fable is 0.69. It's dark yellow, making this look VERY different from 0.42. But is it? What do these numbers mean in absolute terms?
- orbital-decay
>character trigrams
Worthless. Any LLM output similarity metric that uses n-grams as a source might as well measure the average temperature on Mars surface, no matter how much lipstick you put on it. It just can't have enough certainty. There used to be an n-gram benchmark popular on Twitter that showed extreme similarity of grok-3-beta to gpt-4.5-preview, while these models were trained on new base ones, came out 2 weeks apart, and were unmistakably different. Results were wildly inconsistent run to run. It didn't stop the crowd believing its creator in that DeepSeek R1 was trained on o1-preview (which was obvious bullshit as well, they were as different as two models can be). It's amazing how you can put anything on the web and everybody will believe you without checking or even understanding of what they're looking at.
K3 was trained on Claude's outputs, though - it repeats Anthropic's prompt injections 1:1 in its reasoning, which you should know if you ever tinkered with both models long enough. Good for them.
- josh-paul
This looks to be more behavior based, not logit based? What is the actual claim?
- kingstnap
Keep in mind if you ask Claude what model it is in Chinese it says its Deepseek or Qwen or Kimi.
So who's training on who's outputs?