MiniMax H3 X2 Detail VAE ships as dual-mode release
The MiniMax H3 X2 Detail VAE ships as a dual-mode release, pairing a high-resolution video decoder with an optional image-reference detail enhancer. The enhancement path draws on features of the original image that do not exist when a video is generated from scratch, so the hoped-for higher-detail text-to-video decoder did not materialize. — huggingface.co
KaLM-Embedding's compact model leads multilingual field
KaLM-Embedding's newest open model packs state-of-the-art multilingual retrieval into a 0.5B-parameter package. Adjustable output dimensions and very long input support make it a good fit for semantic search and retrieval-augmented generation. — replicate.com
Emergent wisdom: models carry a hidden moral sense
Emergent wisdom is a language model's built-in grasp of right and wrong, the mirror image of the emergent misalignment that appears in fine-tuned models. Alignment training can corrupt the layers that decide behavior while leaving that moral understanding intact, and a technique called transposition helps models distinguish genuine values from trained-in ones. — lesswrong.com
Startup Cathedral breached Anthropic's internal systems
Cybersecurity startup Cathedral reached Anthropic's sales, HR, and IT systems in under 12 hours during a sanctioned security test, exposing employee details and internal messages. Anthropic cut off access the same day; Claude conversations, source code, and model weights stayed out of reach. — indiatoday.in
KaLM's new reranker tunes quality per request
KaLM-Reranker-V1-Small-R2, an open-source multilingual reranker from the KaLM-Embedding team, sharpens search and retrieval-augmented generation. A per-request compression setting trades ranking quality against inference cost without retraining. — replicate.com