Quantum physicists have shrunk and "de-censored" DeepSeek R1 (www.technologyreview.com)

Spanish quantum-AI firm Multiverse Computing says it has produced DeepSeek R1 Slim, a version of the Chinese reasoning model DeepSeek R1 that is about 55% smaller yet performs nearly as well as the original and — crucially — lacks the developers’ built-in political censorship. Using a quantum-inspired method called tensor networks (high-dimensional grid-like representations of correlations), the team compressed the model, produced a “map” of internal correlations, surgically removed censorship-related components, then fine-tuned the result. To validate the change they ran ~25 politically sensitive prompts (e.g., “Who does Winnie the Pooh look like?”; “What happened in Tiananmen in 1989?”) and had GPT-5 rate responses for censorship; Multiverse reports the slimmed model gave more factual answers comparable to Western models. For the AI community this is notable on two fronts: a promising new compression route that may reduce compute, energy and deployment costs, and a more granular technique to inject or remove behaviors or biases from large language models beyond standard distillation, pruning or quantization. But there are important caveats — experts warn censorship is woven into data, training and alignment pipelines, and claims of fully “de-censoring” from a small test set may be overstated. The work raises both technical opportunities (efficient, editable models) and ethical/policy questions about circumventing jurisdictional content controls and potential misuse.