Should AI Adapt to Humans, or Humans to AI: Choices for Efficient Technology Development

The discussion regarding 'adjustment' in the relationship between AI and humans is becoming increasingly important due to the rapid advancement of technology and the growing need for its application. The question, 'Should humans adapt to AI?' or 'Should AI adapt to humans?' is a simple way to express this. It is also a metaphor for the technical choice of whether to adapt to hardware or to algorithms.

This question can have different meanings depending on the context in which it is discussed. It may refer to AI adapting to human business processes and workflows, or it may refer to AI responding to human behavior, consciousness, culture, and values.

On the other hand, humans adapting to AI refers to humans adjusting themselves to the capabilities and constraints of AI.

This discussion highlights the potential gap that can arise between AI developers and users. Developers aim to maximize AI performance and may believe that choosing specific hardware or algorithms is optimal for that purpose. Users, meanwhile, expect AI to adapt to their needs, which may not align with the choice of specific hardware or algorithms.

This is directly linked to the aforementioned issue of 'adapting to hardware or adapting to algorithms' regarding the compatibility between hardware (Fugaku) and AI learning models (algorithms).

Fugaku is a type of parallel supercomputer with characteristics designed to support AI learning models that require massive amounts of data and complex calculations, such as the Transformer architecture. Hardware capable of handling such large-scale data processing plays an extremely important role in the learning and development of AI.

However, not all AI models operate optimally on the same hardware; depending on the model design, other hardware, such as designs that heavily utilize GPUs, may be more efficient.

Here, the debate arises as to whether AI should adapt to humans or humans should adapt to AI. It is the question of whether an AI model (in this case, a Transformer-based NLP) should be adjusted to fit existing hardware (Fugaku), or whether the hardware should be adjusted to fit the AI model (e.g., the introduction of GPUs).

Adjusting an AI model to fit hardware has both advantages and disadvantages. The advantage is that existing hardware can be utilized to its fullest, and costs associated with introducing or maintaining new hardware can be kept down. The disadvantage is that it is constrained by the performance of existing hardware, making it difficult to achieve performance beyond that, and because it is technically difficult to adjust an AI model to fit hardware, it may require a large amount of time and resources.

Conversely, adjusting hardware to fit an AI model has the advantage of making it possible to maximize the use of cutting-edge AI models and draw out their performance. The disadvantage is that the costs of introducing and maintaining new hardware increase, and the introduction of new hardware also comes with technical hurdles.

Considering these factors comprehensively, which approach is optimal depends on the specific purpose and situation. If existing hardware can sufficiently handle the required solution, an approach of developing an AI model to fit the hardware might be cost-effective. However, if one wishes to adopt more advanced AI models or if further performance improvements are required, an approach of adjusting hardware to fit the AI model is effective.

Also, from the perspective of versatility, the compatibility between AI models and hardware is important. Developing an AI model that corresponds to specific hardware may result in that model only being usable on that hardware. This could limit the spread of the model and, as a result, reduce its influence.

From this point of view as well, the compatibility between AI models and hardware is not just a technical issue, but also a social and economic one. To achieve a good balance between efficiency and widespread adoption, appropriate compatibility between AI models and hardware is required.

In conclusion, the question of whether AI should adapt to humans or humans should adapt to AI is an issue that should be decided by taking into account various factors, such as specific objectives, budget, available technology, required performance, and perspectives on widespread adoption. On the other hand, there is no doubt that pursuing compatibility between AI and hardware is an important step that contributes to the further development and spread of AI.