The new method, known as distillation, enhances the effectiveness and reduces the cost of running AI.

In chemistry, distillation means a technique that purifies a material. The same way chemists distillate liquids, AI researchers can distillate AI. Distillation in the human body means that when we move our hands, we don’t need to move our feet at the same time. Or when we order pizza, we don’t want the entire list. We want that certain pizza. In AI, that means that the AI can create a student model that it trains for customers' needs. 

The large language model (LLM) can create a small language model (SML) and customize it. That means the LLM removes all unnecessary things from the SML to make it compact and more secure. The SML is easier to test and it requires less powerful servers than the LLM. There are always mistakes and errors in the LLM algorithms. The problem is that the corrupted AI is not a good tool for detecting errors in its internal code. The human coder must recognize the suspected errors and then fix them. But the problem is that the code can be right, but its target is the wrong object. 

The idea is that the system cleans information in the system. That means that the AI has only responses and actions that it needs for complete missions. The system takes all unnecessary parts away. And that sometimes causes questions about the information that the AI will not need. Humans make decisions about information that the AI needs. And that is seen in things like Chinese AI. Those things don’t discuss things like Tiananmen Square. 

That is one version of distilled information. The system will not give answers that are against the state policy. AI is a tool that can make many things better than humans. But those things will happen in well-limited sectors. The AI can observe things like nuclear reactor functionality. The fact is that the nuclear reactor is not like a chess game. The AI must only keep values at a certain level. The thing in the AI is that it can generate code, but it can use only existing datasets. The difference between a nuclear reactor and a chess game is that the nuclear reactor will always follow certain rules. 

The nuclear reactor will not make anything unpredictable. If the AI knows all its values, the nuclear reactor is safe. But unpredictable values like leaks in the cooling system can destroy a reactor. The programmer who creates the nuclear reactor control systems. That creator must be very professional and collect all data so that the system can respond to all situations. The system must collect information from many sources, such as surveillance cameras and other tools. The system must recognize if some light doesn’t shine as it should. 

That kind of system requires very high-level skills and the ability to train the system for new things. There is always a possibility that the programmer, or the engineer who advises programmers, doesn't always remember everything, such as details of some kinds of damage. That means the AI requires training for that mission. And like always, this kind of thing means that all mistakes that AI makes are actually made by humans. Humans should test and accept that kind of system. And that causes dangerous situations. The training is the final touch in the AI R&D process. 

When we think about things like the North Korean government, they want to use AI in the same missions as Western actors. But do those actors have the skills and abilities to make the final training for their language models? If those language models are made by using some kind of pirated copies that are transported using USB sticks, it can make it possible that the AI and its complicated algorithms don’t work as they should. And that makes those systems dangerous. 

https://www.quantamagazine.org/how-distillation-makes-ai-models-smaller-and-cheaper-20250718/