Data Science: Productivity with LLM, Validation, and Essential Creativity

Recommendations specify where language models truly accelerate a data scientist's work and where they derail. They emphasize the responsibility to validate each link, the targeted use of LLMs, and the creation of customized routines. The stated goal is to increase productivity without delegating thought or being misled by convincing but false outputs.

LLMs Can Deceive: Validating Each Step Remains Mandatory

New language models are performing better in benchmarks while also being capable of misleading. A well-formulated text can appear official or convincing even when it is false. The developer remains responsible for their code and must treat these systems accordingly. Validating each link in the chain and assessing the overall approach are imperatives. Junior profiles should spend more time understanding the project. Questioning, testing, and clarifying limitations are part of the process, preparing responses to objections. It is also essential to take the necessary time, use AI to eliminate repetitive tasks, and not place blind trust in it.

Accelerating Without Delusion: What LLMs Do Well

Presented as productivity tools, language models help deliver quickly, reduce coding, and automate recurring flows and loops. They excel at retrieving information but remain poor extrapolators. One example highlights that they quickly find data while struggling to produce a poem that is not a rewrite of existing texts. In other words, they accelerate execution and access to knowledge without guaranteeing original production or reliable reasoning.

Customizing Assistants: Tailored Routines and Skills

Recent advancements allow for the establishment of AI routines tailored to each user. Mechanisms like skills or commands are cited as levers to gain productivity and stand out. In Claude Code, it is possible to create skills that code like the user, express themselves in their style, and revise according to their criteria. Transforming a recurring task into a routine within one's favorite tool is a direct way to increase productivity, taking into account individual preferences.

What Must Remain Human: Framing, Creativity, and Business Logic

Language models do not replace a data scientist capable of understanding business logic and addressing open-ended problems without perfect solutions. They are not creative and can follow absurd paths if not guided, even in the face of poorly defined problems, as evidenced by the example of recommending walking to a nearby car wash. Their recommended field of use is therefore that of very specific queries. Data science cases pose human questions: the sought angle, the target audience, what the company needs to develop at a given moment. LLMs can provide hints but should not be entrusted with the entire thought process. The decisive advantages remain analytical skills, critical thinking, algebra, statistics, and creativity. In a landscape where AI evolves too quickly to keep up with every new model or tool mentioned on LinkedIn, the challenge is to stay informed without becoming overwhelmed. For those feeling overwhelmed, a pause and refocusing are recommended. Finally, these tools do not alone transform a professional into a standout in their field.