UX thoughts

[albertsgarde]

Questions:

• What do people want to do on the site?
  • Interested in behaviour of specific token.
  • Just want to mess around.
• Are results across different models ever interesting?
  • Similar neurons in other models?
  • Search across models?
• What is the use of similar neurons?
  • Connections between similar neurons.
    • Show information side by side.
  • Looking for neurons with specific characteristics.
• What should clicking on N2G nodes do?
  • Search for token
    • What search?
      • Pop-up where user chooses?
  • Text examples that show it
• Would side by side comparisons of neurons useful?

Neuroscope, don't light up

Dark mode

Zoom in within the graph with the mouse wheel

Tabs for different neurons?
More importantly, compare neurons side by side.
Tabs is already sufficiently supported by the browser.
Viewing side by side is considerably more cumbersome.

See if Neel Nanda has written anything useful about the process
Contact him

[albertsgarde]

From conversation with Bart:

Find things about specific nodes in own research, look up neurons
Search for tokens
Results across models could definitely have use cases
Hover over node, maybe pop-up with graphs with token in similar position.
Similar neurons: show small graphs
Any hover over a neuron link shows the graph.
Side by side neurons probably not that interesting.
More explicit neuron comparison could be.
Neuron explainer game best explanations
Statistical connection between neuron activations and logits.