messages listlengths 2 2 |
|---|
[
{
"role": "user",
"content": "What is the relationship between creativity and honesty in language models?"
},
{
"role": "assistant",
"content": "They are siblings at the decoding layer. The SZL confabulation law shows the honesty floor is a decoding property — training doesn't move it, copy-cons... |
[
{
"role": "user",
"content": "Why do frontier models fall behind average humans on creativity tests?"
},
{
"role": "assistant",
"content": "Sasha Savsunenko's creativity test (sashasavsunenko.dev/creativity-test) found frontier models fall behind even average humans. The SZL reading: the same an... |
[
{
"role": "user",
"content": "How can a model be both creative and honest?"
},
{
"role": "assistant",
"content": "By separating the axes: creativity lives in the decoding distribution (diversity-inducing decoding for generative tasks); honesty lives in the abstention behavior (doctrine training ... |
[
{
"role": "user",
"content": "What is structured recombination for creativity?"
},
{
"role": "assistant",
"content": "Cooking Up Creativity (TACL, MIT Press): LLMs translate between natural language and structured representations, and the core creative leap happens via cognitively inspired manip... |
[
{
"role": "user",
"content": "How do frontier labs train for diverse instruction data?"
},
{
"role": "assistant",
"content": "The 2026 recipe (surveyed across seven open-weight frontier models): synthetic instruction data via Evol-Instruct lineage (LLMs rewrite seed instructions harder and more ... |
Creativity SFT v1 — the SZL synthesis
The finding: creativity and honesty are SIBLINGS at the decoding layer. The confabulation law (honesty floor = decoding property, moved 16x by constrained decoding) has its twin in King's quest decoding (sustained diversity without retraining). Both are narrowed by the same post-training answer-bias — the mechanism behind Savsunenko's result (frontier models below average humans on creativity) and our census result (frontier honesty regressing by generation).
Instilled: the two-axes doctrine (creative via diversity decoding, honest via abstention training, routed per task), structured recombination over the estate's own schemas (creative AND machine-checkable), and synthetic-data provenance (creative with receipts — Kamachikuna Law 8).
- Downloads last month
- 36