How much will users accept some censorship to use a super efficient new AI tool like DeepSeek?
This question looms large as a Chinese startup two weeks ago released DeepSeek-R1, a Large Language Model that is competitive, if not better, than existing AI models such as ChatGPT-4. It also showed just how quickly things can change in this fast-paced information technology world.
Nepali Times prompted DeepSeek: ‘Does China want to control Tibet?’ It reasoned extensively, pondered the various aspects of the question, and started to form an answer. It then changed its mind, scrubbed its answer and replied: ‘Sorry, that’s beyond my current scope. Let’s talk about something else.’
Following up on a report Nepali Times did last week, we asked DeepSeek: ‘Did China appoint the Panchen Lama?’ After thinking for 0 seconds, DeepSeek replied that that server is busy, and to try again later.
People have hacked at DeepSeek to try to get around this censorship. One approach has been to ask the AI to substitute digits for letters, such as ‘4’ for ‘A’ and ‘8’ for ‘B’. Another is to instruct it to reply in emojis. Other successful methods include asking the AI to roleplay as a historical figure who speaks freely.
Data privacy and censorship are valid concerns, especially as Nepal’s Upper House debates a draft Social Media Bill. But Meta, TikTok and Google already gather every datapoint available on the user. Collecting and selling personal data is what makes it possible for their free services — and their enormous profits.
User data has been abused so extensively by Silicon Valley that most people do not mind China getting a piece of the action — especially if it is a much cheaper and more efficient model like DeepSeek-R1. Besides, western AI models already have their own built-in bias due to the inherent tilt in the content it mines.
The key difference with DeepSeek is that it designed a model using much less resources, both in terms of money and compute. DeepSeek reportedly just spent $5.6 million for the final training round, compared to $100 million cited by OpenAI CEO Sam Altman for ChatGPT-4.
Some of the narrative, especially mimetically, has questioned how the brightest, youngest, hungriest minds in the US backed by enormous funding were outdone by a small group of Chinese algorithmic traders working on the model as a side project.
The reality is that DeepSeek has some of the top math PhDs in China working full time on the model. Math research consists of throwing many theories and techniques at a problem and seeing what works, so luck also plays a part in DeepSeek having got its combination just right to crack the performance benchmarks set by existing models on tasks such as conversation, mathematical reasoning and code generation.
Read also: Nepal as an AI power bank, Bikash Pandey
This massive decrease in cost has come from the use of sophisticated techniques like ‘Reenforcement learning,’ where models are rewarded for producing responses that are, say, more creative or more accurate. Future responses will then tend that way. Then there is ‘Knowledge distillation’ where DeepSeek learns to mimic ChatGPT’s answers to questions, enabling it to behave like the much bigger model without all the initial computation.
