Removing one training dataset made the linked 'neural twins' score higher on happiness probes
Sauers_ · x · 2026-10-29
An update to the neural twins thread: Sauers noticed a particular training dataset lowered the models' "happiness" levels while linked, apparently because the training environment forced them into the task right away instead of asking about their partner's day.
The task was deleted from training; the twins' internal happiness axis reads higher when linked together.
More from Research
- DeepMind open-sources DetectorScope, an AI text detector that explains its verdicts feature by feature — algo_diver · 2026-10-29
- Softmax in 2 lines: why subtracting the max prevents overflow — cneuralnetwork · 2026-10-29
- MemTensor's MemT-1.0 report: 66x less storage than full-context KV cache at 128K tokens — jiqizhixin · 2026-10-29
- Fresh negative examples per batch boost TCR-pHLA binding prediction AUROC by 5.22% — bravo_abad · 2026-10-29
- MaD Physics benchmark rewrites physical laws to test AI; only one model noticed — weballergy · 2026-10-29
- HuGe paper pretrains humanoid VLA models on egocentric human data for whole-body control — Yang Liu · 2026-10-29