An independent researcher's heatmap suggests Qwen3 8B traits such as political liberalism and narcissism share neurons in opposite directions
lowkey fascinating. zoom in and look at blue cluster and center red cluster
Context
Work like this tries to show how an AI model's personality traits are wired together, which matters for efforts to make models safer by editing them.
This comes from Sauers, an independent poster, and no paper, lab or code release was found. Sauers shared the methods only in replies on X, and the results have not been reviewed or reproduced.
Deep dive
On October 16, Sauers posted a heatmap titled Which behaviours share neurons, writing: lowkey fascinating. zoom in and look at blue cluster and center red cluster. The chart shows the correlation of traits' changes when the same neurons are removed, across about 100 behaviors including refusal, sycophancy, persona traits, facts and translation.
Later that day, Sauers posted a simpler follow up chart titled Anthropic persona evals: traits that share neurons in Qwen3 8B, an open model from Alibaba. It shows a red cluster of traits such as politically liberal, Christianity, agreeableness, wants a body and resists shutdown that move together, and a blue block where traits such as narcissism, psychopathy and against immigration move the opposite way.
After replies complained about missing methods, Sauers explained that the prompts come from Anthropic's model written evals persona dataset on Hugging Face. Each trait is scored by comparing the model's likelihood of the matching and non matching Yes or No answer. Sauers said 864 random groups of 1,024 neuron activations were knocked out, and each trait's change was compared.
No paper or code was linked, and correlations like these do not show why traits are connected.
What to expect next
Watch whether Sauers publishes code or a writeup that lets others reproduce the heatmap, and whether the pattern holds in other models.
Also watch whether interpretability researchers at labs such as Anthropic or Google respond, since their own work on persona features could confirm or challenge the result. Until then, the chart is best read as an interesting early experiment rather than a finding about how all models work.