Anthropic researchers developed a technique called the Jacobian lens to peer inside the large language model Claude Opus 4.6, revealing a hidden "J-space" of words the model considers before generatin…
Researchers at Lighthaven tinkered with Natural Language Autoencoders (NLAs), finding that while they reconstruct activations well and enable auditing of model organisms, the architecture or training …
Researchers at Goodfire, Harvard, Stanford, and Brown introduced Block-Sparse Featurizers (BSF), a method that decomposes neural network activations into multidimensional subspaces instead of single d…
Researchers at Goodfire demonstrated that fine-tuning a single scalar prefactor on a German-related rank-1 parameter subcomponent of a 67M-parameter language model can destroy its ability to predict G…
Goodfire researchers introduced predictive data debugging, a method that uses interpretability to predict which behaviors reinforcement learning will amplify or suppress from a preference dataset befo…
A BlueDot Technical AI Safety Project researcher reproduced a study from Goodfire demonstrating that language model representations form curved geometric manifolds, not simple linear directions. The w…
SPC's May 2026 update highlights major fundraises including Cognition's $1B Series D and Recursive's $650M round, along with events in SF, NYC, and Bangalore featuring humanoids, drones, and cybersecu…