DeepSeek may not be so good for energy afterall
The life cycle of an AI model is divided into two phases, training and inference. The training phase is a long-term process where the model is taught from data. Inference is the next step, and it happens every time someone asks a question. Both are usually done in data centers where there is a lot of energy required to cool the servers and run the chips.
On its training side, DeepSeek improved what is called a “mixture-of-experts” technique. This involves only turning on a small portion of the model’s millions of parameters, or “knobs”, to help it form better answers. They also improved reinforcement learning. This is where the outputs of a model are scored, and then used to improve it. The DeepSeek team automated what was usually done by humans. The introduction of a method to make training more effective might indicate that AI companies would use less energy in order to bring their AI model to a standard. It’s not true.
“Because of the high value of a more intelligent system,” wrote Anthropic Cofounder Dario Amedei on his blog. It “causes businesses to spend more on training models, not less.” If they get more money for their investment, then it will be worthwhile to spend more and use more energy. He wrote that the gains in cost-efficiency are devoted entirely to training smarter model, and limited only by a company’s financial resources. This is an example of the Jevons paradox. Inference requires a lot of energy. DeepSeek was designed to be a reasoning model. This means that it is able to handle tasks like logic, pattern finding, math and other things that typical generative AI models are not able. The reasoning models achieve this by using a concept called “chain of thoughts.” This allows the AI to break down its task and complete it in a logical sequence before reaching its conclusion. You can see it with DeepSeek. The model will first weigh the benefits of lying to protect someone else’s feelings against any potential harm in the future. The model then looks at Kantian ethics which suggests that you should follow maxims that can be universal laws. Before coming to a conclusion, it considers all of these nuances. If you’re curious, it finds that lying can be “acceptable in situations when kindness and preventing harm are paramount. However, nuanced solutions do not exist.” )
