Humanity or Efficiency: What are the big language models should rely on US Army

This is stated in the article of The National Interest under the name "During the assessment of the USA, the US should rely on efficiency, not the explanation", which was translated by focus. The US should learn how to evaluate the tools of AI, such as large language models, in their productivity, not the ability to explain its decisions. Trust should be based on results, not the unrealistic expectations of anthropomorphic thinking.

Since the United States enters the new era of the rivalry of the great powers, especially with technologically ambitious China, the question of how and when to trust the systems of AI, such as large language models (VMM), becomes not just technical. It is strategic. These tools will be crucial in how the United States distributes resources, determine the priority of defense investments and hold positions in the Indo-Pacific and beyond. It has no intelligence.

These are recognizers of images trained on huge data ranges and are intended to predict the next word in sequence. Similar to a chess grasster, who makes a brilliant but intuitive move, it often cannot explain why they generate one or another result. However, the Ministry of Defense, through organizations such as the General Directorate for Digital Technologies and AI, defined the clarity of AI decisions as a requirement for its prompt use. This good intention can lead to the best consequences.

Explanation in the buds may be technically unattainable, and the pursuit of it can be a distracting factor. These models do not "understand" in the human sense. Their results are statistical associations, not cause and effect. Post -factum explanations, although they bring pleasure, can mislead and eventually prevent the introduction of tools capable of improving strategic prediction, analyzing intelligence and operational planning.

The real danger is overly attention to the detriment of efficiency. Many national security decisions-from choosing goals to procurement planning-already include opaque but proven processes, such as Varheiming or expert evaluation. It can be complemented by these approaches to processing the amount of information at a speed that analytics-people cannot compare.

Instead of trying to make it more "human", we must evaluate them by the criteria that meet how they actually work: consistency, accuracy and clarity for restrictions. It should be asked: new methods, such as automatic fact, have significantly reduced hallucinations - from 9 % to 0. 3 % in some models. Productivity-based systems, such as Trustllm, promise to evaluate the reliability of the model more comprehensively than have ever been done through explanations.

In order to ensure effective and safe integration of large language models into military and defensive contexts, politicians should be preferred by operational testing rather than clarity. Instead of focusing on artificial interpretation, systems should be evaluated by the threshold of productivity before deployment. This approach is based on empirical reliability and guarantees that the AI ​​tools will produce consistent results in real conditions.

Politicians should keep military command in nature and restrictions. Trust in these models should be based on the measured results, not the illusion of understanding or anthropomorphic qualities. Being unreasonable tools, it is based on the recognition of images, not knowledge, and should not be expected that they will simulate human thinking or self -awareness. Finally, it is necessary to develop recommendations for the introduction of AI, taking into account specific cases of use.

Different operational scenarios require different levels of control and reliability. For example, when generalizing intelligence data, priority may be highly consistency, while combat use requires a restraining system and constant human control to reduce risks and accountability. In general, confidence in the buds should not be based on their ability to sound humanly, but on their constant ability to issue accurate, recurring and proven results.

It is unrealistic and counterproductive to consider them as digital oracles. The assessment of AI systems based on productivity, not interpreting or anthropomorphic attractiveness, is a much more pragmatic and effective approach. Michael "Sparky" Perry-Lieutenant Colonel of the Air Force and the leading MC-130 pilot with a master's degree in business administration and military affairs. National Defense Research Fellow at the SEM International Relations School at Georgia's Technology Institute.