Korva is a scroll of the internet’s best ideas.
DeepSeek wouldn't give Starbucks hours
Chinese models are immense outside China. Inside China, their abilities are relatively limited, and the guardrails are incredibly stringent and prohibitive. Very basic, rudimentary requests get turned away.
Nobody has put together a profile of what a person can do with DeepSeek that they can't do on the internet available within China's borders. How far past the firewall DeepSeek goes is a question no one has dug into. If the firewall matters to China, and still matters in ten years, something has to change as the models grow in importance.
Why do Chinese AI models refuse so many requests?
China's Interim Measures for the Management of Generative AI Services took effect in 2023 and require providers to follow content rules, including a requirement to uphold core socialist values. Services offered to the public have to pass a security assessment and register with regulators before launch. Companies that want to stay compliant tend to tune their models and add filters that refuse broadly.
How can an innocent question trigger a refusal?
Safety training can teach a model to refuse based on surface cues, such as certain words, instead of what the person is actually asking. Researchers call this over-refusal. The XSTest benchmark, published in 2023 by Paul Röttger and colleagues, was built from safe prompts that only look unsafe, such as asking how to kill a Python process, to measure it.
Why might a hosted model behave differently from its downloaded weights?
A hosted app can put its own filters in front of and behind the model, such as separate classifiers that screen questions and answers. Those layers are not part of the weights. Someone who downloads the weights and runs them elsewhere gets the model without that extra screening.
More from 20VC with Harry Stebbings
