Difference between revisions
interpretability
Status and infobox
- status:
—→ open - kind:
—→ problem - year:
—→ 2013 - statement:
—→ Can the internal computations of a trained neural network be fully understood?
Relations
- + related to ai-alignment: The original curation links this problem to AI alignment; the argument for the link has not been written yet.
Source +19 −0
+---
+title: Mechanistic interpretability
+kind: problem
+status: open
+year: 2013
+statement: Can the internal computations of a trained neural network be fully understood?
+space: machine-minds
+---
+Features and circuits have been found inside language models, including with sparse dictionary methods since 2023. Explaining a whole frontier model's behavior in these terms is far off. Benchmarks for what counts as an explanation are still being built.
+
+## Related problems
+
+- The original curation links this problem to [[related::Problem:ai-alignment|AI alignment]]; the argument for the link has not been written yet.
+
+## Sources
+
+- [Wikipedia: Mechanistic interpretability](https://en.wikipedia.org/wiki/Mechanistic_interpretability)
+
+[[Space:machine-minds|Machine minds]] · [[Continent:mind|Mind]]
Revision 173 as rendered
Features and circuits have been found inside language models, including with sparse dictionary methods since 2023. Explaining a whole frontier model's behavior in these terms is far off. Benchmarks for what counts as an explanation are still being built.
Related problems
- The original curation links this problem to AI alignment; the argument for the link has not been written yet.