How to choose the right technology depending on your use and avoid expensive errors.
From Chatgpt, the generative AI occupies all spirits and many decision -makers now assimilate any form of AI to generative AI. An expensive error that can push certain companies to deploy over-sized generative models where a classic ML algorithm would suffice, or vice versa, to underestimate the creative power of the generative AI. But how do you choose the right approach for a given use case?
Generative AI to create, ML to learn
The machine learning is based on a supervised or not supervised learning logic from structured datasets. The algorithm analyzes thousands, even millions of historical examples to identify correlations and statistical patterns. The objective is to reproduce and automate large -scale human decisions. A banking fraud detection model trains on millions of transactions labeled “legitimate” or “fraudulent” to learn to recognize suspicious weak signals. An e-commerce recommendation system analyzes past purchasing behaviors to predict future preferences. AI learns to classify, score, predict or optimize according to defined criteria.
In reality, generative AI is not a separate discipline, but a branch of machine learning. Where the ML “Classic” learns from structured data to predict or classify, generative models learn the deep rules of language, code or image to create new content. The LLM ingest unstructured data teraoctes (web texts, books, images, source codes, etc.) to understand the deep rules of language, creativity and logic. As a result, the generative IA models are no longer content to classify or predict, they create. They produce coherent texts on any subject, create images, write functional code in dozens of languages.
GENERATIVE OR ML “Classic”: How to choose?
Above all deployment, a fundamental question is essential, recalls Françoise Soulié-Fogelman, scientific director of Hub France IA: “What is the need?” This question determines 90% of the technological choice. If the objective is to produce content, summaries, syntheses, creative ideas, images, then generative AI naturally imposes itself. All cases of content creating content fall under this technology. Conversely, for analysis and prediction missions, traditional machine learning remains essential: “Predictive AI is to make forecasts. Who is going to leave, who goes Churn tomorrow, who will click?”
But beyond the functional need, several technical criteria guide the choice. First, the explanability: “The big quality of the predictive AI is that one can obtain an estimate of the error rate. In the generative AI, there is no.” A crucial difference for critical applications requiring traceability of decisions. Then, the nature of the data: the ML excels on structured data while the generative AI naturally processes texts, images and content not structured thanks to its self-learning. The rule therefore seems relatively simple: machine learning to predict and classify with precision, generative AI to create new content.
Use an LLM to classify? Bad idea
On the other hand, wanting to divert the generative models for classification tasks is, according to Françoise Soulié-Fogelman, “of furious madness”. It recalls that “predictive AI makes it possible to measure and minimize the error rate, this is the very purpose of learning”, while with generative AI, “there is no way to quantify the battery”. Where a convolutionary network or a classic supervised model brings precision, robustness and explanability, an LLM mobilized for the same task multiplies the calculation costs and introduces uncertainties difficult to control.
On the other hand, certain hybrid approaches, ML and generative IA, prove to be relevant. In the medical field, the specialist evokes a concrete case: “During a radiology, the doctor dictates his observations. We first use a convolutionary network for voice recognition, but the medical vocabulary generates many errors. The machine learning guarantees a clear measure of the error rate during transcription and the generative AI is capable of correcting the medical vocabulary.
A simple method to decide
The simplest method of choice is to use a minimalist decision tree :. Is your goal to predict, classify or analyze patterns in your data? If so, opt for traditional machine learning. You are looking to produce new content, texts, images, code … The generative AI is then essential. The logic is binary and makes it possible to avoid the majority of architectural errors and limit costs. The secondary criteria such as the nature of the data (structured vs not structured) and the requirements of explanability then come to refine this first choice.
