Should we be afraid of DeepSeek?

Chinese censorship will ultimately lead to their AI-generated ‘world models’ being unusable

Over the past few days, there has been a flood of articles and comments about DeepSeek. What I find puzzling, however, is how the market and the AI industry are treating this as something extraordinary. While it is remarkable, e.g. achieving reasoning capabilities equivalent to, or even surpassing GPT-4o1, similar breakthroughs have been occurring regularly over the past two years, largely due to the open-sourcing of Meta’s Llama models, which DeepSeek appears to be based on.

Like previous models, the focus here is not on the accuracy of responses across different domains, which can vary significantly depending on the “expert system” that is integrated into the model. OpenAI has been doing this for at least a year as well. For example, a simpler model like R1 can be installed on a high-performance mobile device, but its responses are reportedly comparable to what Google Assistant could deliver on your phone a year ago.

For “serious AI” and the race to superintelligence, the focus will need to shift to two critical factors:

1. Eliminating hallucinations — This remains the primary reason why many companies have not yet fully embraced AI.
2. Efficiency — Metrics such as power consumption (e.g., watts per 1M tokens) will be…