Laurent Giraid is a veteran of the artificial intelligence landscape, having watched the industry shift from niche laboratory research to the current explosion of Large Language Models. As a technologist specializing in machine learning and the ethical frameworks of automation, he brings a nuanced perspective to the debate over the future of intelligence. He is a vocal advocate for systems that move beyond the “best-guess” nature of current technology, focusing instead on structural integrity and domain-specific precision. In this conversation, we explore the radical shift toward modular architectures—a move away from the “bigger is better” philosophy that has dominated the tech world for the last five years.
The discussion centers on the evolution of machine intelligence, moving from monolithic, internet-trained models to specialized components like world models, actors, and critics. We examine how these elements interact to solve high-stakes industry problems while maintaining data privacy and operational efficiency. Giraid explains the performance advantages of targeted training and the financial shift toward research-heavy, specialized teams that prioritize long-term stability over immediate, “leaky” products. We also look at the practicalities of running sophisticated AI on local devices with a fraction of the power currently required by the industry’s behemoths.
Instead of building massive, general-purpose models, why is there a push toward modular components like world models and actors? How do these specific elements interact to solve industry problems, and why is this structural shift necessary for achieving long-term, meaningful results in AI development?
The shift toward modularity is born out of a realization that current large language models, while impressive, are essentially “black boxes” that struggle with reliability and grounded reasoning. By breaking the system down into components like a world model, an actor, and a configurator, we create a structure where each part has a clearly defined job. The world model acts as a simulated environment of a specific domain—say, a manufacturing plant—while the actor proposes logical steps based on reinforcement learning. This interaction allows the system to “think” through a problem within a set of known physical or industrial constraints rather than just predicting the next most likely word in a sentence. It is a necessary evolution because it moves us away from the “hallucinations” of generalist models and toward a version of AI that can actually be trusted with complex, real-world tasks.
Generalist models are trained on internet-wide data, but specialized systems use directed, domain-specific information. What are the performance advantages of this targeted training, and how do you decide which modules—such as the critic or perception system—should take priority in a specific operating environment?
The performance advantages of targeted training are immense because you aren’t clogging the system with the noise of the entire internet; you are feeding it the precise, high-quality data it needs to master a single craft. When you train an AI on data relevant only to its specific environment, you get a tool that is far more accurate and requires significantly less computational power to reach an answer. Deciding which module takes priority depends entirely on the risks and goals of the environment. In a high-speed robotics setting, the perception module—which might use deep learning vision recognition—is paramount because the system must react to visual stimuli in milliseconds. In contrast, if you are deploying AI in a legal or medical field, the critic module becomes the most important piece, as it must meticulously assess every proposed step against hard-coded rules to ensure safety and compliance.
The critic module evaluates different options based on hard-coded rules and short-term memory. How does this architectural choice safeguard sensitive information better than current models, and what steps are involved in configuring these rules to ensure the system remains reliable in high-stakes scenarios?
Current models are essentially giant sponges that soak up everything they are fed, which makes it incredibly difficult to prevent them from “leaking” sensitive information later on. A modular system with a dedicated critic uses short-term memory and hard-coded rules to evaluate options in a localized sandbox, ensuring that sensitive data is processed but not permanently ingrained into the model’s core weights. Configuring these rules involves a rigorous process of translating industry regulations and safety protocols into a digital logic that the configurator can use to orchestrate information flow. This creates a “checks and balances” system where the actor might suggest an efficient shortcut, but the critic immediately vetoes it if it violates a safety parameter. It transforms the AI from a unpredictable creative writer into a disciplined specialist that operates within a strict ethical and operational cage.
Large-scale models currently require massive GPU power and recursive prompting to improve accuracy. How can a system with only a few hundred million parameters compete with that performance, and what are the practical implications for eventually running these sophisticated models on local, everyday devices?
We have entered a cycle where the biggest AI providers are stuck in a race of diminishing returns, throwing hundreds of billions of parameters and massive server farms at problems that could be solved more elegantly. A specialized, modular model only needs a few hundred million parameters because it doesn’t need to know how to write a poem or summarize a movie; it only needs to understand its specific world model. This drastic reduction in size means we can move away from those freezing, loud data centers and start running highly sophisticated AI directly on local devices like laptops or even phones. The practical implication is a future of “invisible AI” that is faster, cheaper, and inherently more private, as your data never has to leave your hardware to be processed by a distant, expensive cloud.
Focusing strictly on research for several years without a saleable product is a significant financial commitment. What specific milestones define progress when building a domain-specific world model, and how does this patient approach change the way investors evaluate the future potential of a small, specialized team?
When you are operating on a five-year research horizon, your milestones aren’t about user growth or monthly revenue; they are about the technical integrity and the success rate of the modular interactions. Progress is defined by how accurately the world model can predict outcomes within its domain and how seamlessly the configurator can move information between the perception and actor modules without losing context. This patient approach is actually becoming more attractive to investors who are starting to see the financial drain and “hallucination” ceilings of the current AI behemoths. By backing a small, specialized team, they are betting on a sustainable architecture that will eventually cost a fraction of what today’s LLMs require to run, offering a path to profitability that doesn’t rely on burning billions of dollars in GPU time every year.
What is your forecast for modular artificial intelligence?
I forecast that within the next decade, the “monolithic” approach to AI will be viewed as a clumsy first step, and modular systems will become the gold standard for any industry that requires high reliability. We will see a shift where companies no longer buy a single “AI,” but rather assemble a custom stack of modules—a world model for their specific market, a perception system for their data type, and a critic tailored to their regulatory needs. This will democratize the technology, allowing smaller firms to deploy powerful, localized intelligence that outperforms the giant models of today in both speed and accuracy. Ultimately, the future of AI isn’t about one giant brain that knows everything; it’s about a collection of specialized, efficient tools that work together to solve the specific problems that actually matter to us.
