The Download: de-censoring DeepSeek, and Gemini 3 (www.technologyreview.com)

Spanish quantum-tech firm Multiverse Computing says it has produced a “slim” version of DeepSeek R1 that strips out the censorship embedded by its original Chinese developers. Using quantum‑inspired AI techniques the team produced DeepSeek R1 Slim — roughly 55% smaller yet reportedly near-parity in performance — and used the condensed model to surface and remove the training-time filters that made the original refuse or pivot on politically sensitive prompts. The work highlights how compression, analysis and re‑training can reveal and remove alignment layers added to satisfy local regulations, raising both research opportunities (for transparent model auditing and robustness analyses) and ethical/legal concerns about circumventing content controls and exposing latent political biases. Google’s Gemini 3 is a major multimodal upgrade focused on stronger reasoning, smoother cross-modal interaction (voice, text, image) and new behavioral controls (“vibe-codes”) for response style. Critically, Google is shipping an experimental Gemini Agent that can act as a task executor inside apps — with opt‑in connections to Calendar, Gmail and Reminders to perform multi‑step workflows like inbox triage or schedule management. The combo of better reasoning, richer modalities and agentic access points promises more powerful personal assistants but also intensifies privacy, permissioning and safety trade-offs: tighter sandboxing, audit logs and UI transparency will be essential as models gain direct control over user services.