I have seen dozens of PDFs circulating on LinkedIn with flashy titles like "Prompt Engineering Expert" or "Generative AI Specialist". The fever for collecting digital diplomas has run rampant, but in the trenches of real projects, things look quite different. AI certifications carry very relative value.

Admittedly, they serve as helpful tools for beginners who need to structure their learning from scratch or for candidates trying to get past the automated filters of HR departments. But when you sit in front of the console, a portfolio of real projects (clean code on GitHub, models deployed in real environments, automations that solve actual business problems) proves far more competence than any theoretical diploma on your wall.

The useless paper: the obsolescence crisis

The speed at which this field moves is exhausting. A certificate earned a year ago might be worth nothing today. The courses and credentials that lose value at breakneck speeds are those focused on three specific areas:

  1. Specific tools or transient interfaces: Courses that teach you how to use a specific version of a software interface that changes entirely three months later.
  2. Pure prompt engineering: Those manual tricks from 2023 or 2024 for talking to language models. Today, state-of-the-art models reason far more maturely and structure and refine their own queries without requiring you to whisper magic formulas.
  3. The superficial hype wave: Fast diplomas on "how to multiply your productivity with AI" that skip any technical foundations and merely scratch the surface of trendy APIs.

A clear example is the Salesforce AI Associate certification. Launched in 2023 at the height of commercial hype, its retirement is already scheduled for 2026. A three-year lifespan for an exam that will soon be history.

The foundations: what lasts over time

In contrast to the quick expiration of trendy tools, certain areas of knowledge maintain or gain value over the years. These are the ones that ignore current trends and focus on the foundations of the field:

  • Infrastructure, Cloud, and MLOps: Knowing how to manage the data lifecycle, orchestrate containers so they do not crash under load, reduce computing costs, and monitor models in production. Credentials like AWS Certified Machine Learning – Specialty or Google Cloud Professional ML Engineer require system and deployment knowledge, such as AIOps in production, rather than just model familiarity.
  • Theoretical and mathematical foundations: Understanding the statistics, linear algebra, and calculus behind the algorithms. If you understand why a gradient fails, it does not matter which framework you use tomorrow. A classic like Stanford's Machine Learning Specialization taught by Andrew Ng remains an excellent reference precisely because it teaches the mathematics of the subject.
  • Governance, Law, and Ethics: Professionals who can align agentic systems with strict regulations like the European Union's AI Act, mitigate security risks, and audit data bias. The IAPP Artificial Intelligence Governance Professional (AIGP) certification is one example of this regulatory focus, complementing the governance of AI in production practices we apply to agentic systems.

An example in the trenches: why you need math

Imagine you are debugging a vector database and the semantic search results make no sense. If you only know how to use the external library, you are in the dark. If you understand the underlying algebra, you will know that the cosine similarity is failing because your embeddings are not normalized.

Writing the basic mathematical calculation of cosine similarity from scratch, without relying on black-box frameworks, gives you control over what happens under the hood:

import numpy as np

def calculate_cosine_similarity(vector_a, vector_b):
    """
    Manual calculation of cosine similarity to illustrate the underlying algebra.
    """
    dot_product = np.dot(vector_a, vector_b)
    norm_a = np.linalg.norm(vector_a)
    norm_b = np.linalg.norm(vector_b)

    if norm_a == 0 or norm_b == 0:
        return 0.0

    return dot_product / (norm_a * norm_b)

# Simulation of text embeddings
query_embedding = np.array([0.15, 0.88, 0.02, 0.45])
doc_1_embedding = np.array([0.14, 0.85, 0.03, 0.42]) # Very close topic
doc_2_embedding = np.array([-0.50, 0.10, 0.80, -0.20]) # Distant topic

similarity_1 = calculate_cosine_similarity(query_embedding, doc_1_embedding)
similarity_2 = calculate_cosine_similarity(query_embedding, doc_2_embedding)

print(f"Semantic similarity with Doc 1: {similarity_1:.4f}")
print(f"Semantic similarity with Doc 2: {similarity_2:.4f}")

This kind of knowledge allows you to solve real problems when the software fails, something that no quick course on prompt engineering will ever give you.

The current market has matured. It harshly punishes the superficiality of trendy tools and rewards those who master the underlying infrastructure, the mathematics of the algorithms, and regulatory governance.