KaLM reranker adds per-request compression control
The KaLM-Embedding team's open-source multilingual reranker for search and retrieval-augmented generation is now on Replicate, where developers can set document compression per request and trade ranking quality against compute without retraining. — replicate.com
Primal VAE skips deep compression with patchable latents
Ostris' open-source Primal VAE encodes images into a couple of small channels and lets the diffusion transformer do the patching, matching Flux-family latent shapes without deep compression. — huggingface.co
Generalized belief states steer transformers mid-sequence
A new framework treats the memory a transformer carries between tokens as its recurrent state and maps it onto a low-dimensional belief space, letting researchers read a model's beliefs and rewrite them in the middle of a sequence. — lesswrong.com
Stealer uses LLMs to rank victims by value
The self-deleting Windows stealer VOLTA-AUTODOXER harvests credentials, sessions, and wallets, then sends a structured evidence bundle to an LLM that ranks the victim's assets by monetization value. The malware chains several fallback models in case one refuses the request. — blog.talosintelligence.com
Claude blocks game cloning and reverse-engineering
Anthropic's Claude now refuses to recreate specific existing games, and it also blocks innocuous reverse-engineering tasks such as analyzing crash dumps and cleaning junkware, possibly under the same ruleset. — tomshardware.com