As AI adoption is growing across industry, one question that developers often have in mind is: Should we use traditional/classic AI or Generative AI for solving machine learning problem?
As an Architect I will give my favorite answer – “it depends”.
Horses for courses! 
Each approach has pros and cons, and in this post I will try to unpack the technical considerations you should consider carefully.
Traditional AI involves building from scratch with large amounts of labeled, domain specific data. Tasks such as classification, regression, anomaly detection are handled well by these models. Decision trees, support vector machines, and custom neural networks are examples of this AI.
In contrast, Gen AI uses pre-trained models such as GPT or Claude that can be fine-tuned for specific tasks. These models are very good in handling unstructured data, and complex problems like content generation.
Few questions to consider:
-
Is your data structured or unstructured?
If your data is structured and task well-defined then classic AI models can be highly efficient in use cases such as predicting sales or detecting anomalies.
If you data is unstructured then Gen AI models can be very good in making sense of this, as they have been pre-trained on web scale unstructured content like text and images. -
Do you have sufficient labeled data?
If you lack significant amounts of curated labeled data then fine-tuning a Gen AI model may get you better results without massive datasets. -
How complex is the use-case?
For simple tasks like linear regression or classification in vector spaces, classic AI can be efficient in saving compute resources.
However, if your use case needs contextual understanding or generate new content then Gen AI horse is the choice! -
Do you need to combine multimodal data such as text, images, videos etc.?
In these types of tasks Gen AI shines! -
Do you have the compute resources and budgets for high performance?
Gen AI training and inference typically needs specialized GPUs and this can be very resource intensive, unless business benefits significantly outweigh the resource use.
In contrast, if you have limited compute and/or you need to deploy in resource constrained environments then classic AI may be a more practical choice. -
What is the long-term vision for scalability?
Traditional models may use less resources to train and deploy initially, however if data keeps changing then they will need frequent monitoring and retraining.
Gen AI models are already trained on vast datasets, so they incur high resource use initially and subsequent fine-tuning will be less resource intensive. So if you are building a system that needs to evolve with variations in data, then Gen AI may be a better choice. Inferencing in Gen AI can still be expensive due to the size of the models. There are clever techniques like pruning, quantization, and distillation that can help in this regard for Gen AI, some of which I covered in a previous post. -
How familiar is your team with the technology?
Gen AI engineering is different from Classic AI/MLops and both are different from classic software engineering. Team should be skilled and capable of chewing what you bite into. So this is also a very important consideration. If team is not skilled and you wish to embark then you will need to start building such capacity ahead of time.
Some of your answers may be better suited for classic AI and some others could be better suited for Gen AI. To resolve this you could give different weights to different questions based on your priorities, and then do a tradeoff analysis based on the scores.
You could also consider hybrid solution leveraging the best of both classic and Gen AI for different parts of a use-case. Finally, build rapid prototypes and test out multiple approaches before finalizing the design!