Dr. Luc Julia, a renowned expert in artificial intelligence (AI) and human-machine interaction, is a key figure in the field. Co-creator of Siri, he has held innovation leadership positions at Apple, Samsung, and Hewlett-Packard before joining Renault as Chief Scientific Officer. His approach to AI is pragmatic and focused on real utility for humans, far from the technological fantasies often relayed by popular culture. In this interview, he shares his unfiltered vision on generative AI, its limitations, its ecological impact, ethics, and the future of the field.

As Chief Scientific Officer at Renault, Luc Julia works on integrating AI into the automotive industry. He oversees the implementation of intelligent algorithms to optimize driving, improve human-machine interfaces, and make vehicles safer and more intuitive. One of the flagship projects under his direction is the development of an onboard AI assistant, illustrated by the intelligent avatar of the new Renault 5. This is not merely a voice interface, but a true driving companion capable of adapting to the driver’s habits.

 

In your opinion, what are the main misconceptions about generative AI, particularly with the rise of ChatGPT, that the public and entrepreneurs should question?

The artificial intelligence that doesn’t exist is the one from Hollywood: an omniscient and autonomous AI that can do everything on its own. This is what I call Artificial General Intelligence (AGI), and it remains a fantasy. On the other hand, specialized AIs, designed to accomplish specific and useful everyday tasks, are very real and numerous.

The major problem with these generative systems is the illusion of infallibility they create. They are attributed the ability to answer everything, but in reality, the more generalist an AI is, the less relevant its responses are. One of the most widespread misconceptions is equating them with humans, when they merely imitate our biases. And the more we expect them to be omnipotent, the greater the disappointment.

Precise figures are difficult to establish, but some studies estimate that these AIs provide correct answers in approximately 64% of cases. This means they produce errors or hallucinations in 36% of cases, which is considerable. Yet this reality is often unknown.

The reliability of responses also depends on how the question is formulated. AI doesn’t truly understand what it’s saying; it generates responses based on underlying statistical models. Sometimes it can’t even count correctly. If you ask it to prove that 2 + 2 = 5, it will attempt to do so without questioning the absurdity of the request.

 

What is your view on new AI models or emerging technologies, particularly those discussed by Yann LeCun, in opposition to generative models?

Yann himself admits that he doesn’t yet know if he will succeed in developing the reasoning he has been working on for several years.

In reality, generalist AIs are destined to gradually disappear. They are already in decline and will be replaced by more specialized models. Two main approaches are emerging: fine-tuning, which adjusts an existing model for a defined task, and RAG (Retrieval-Augmented Generation), which allows external knowledge to be integrated to improve response relevance.

This trend toward specialized models will intensify. There’s also talk of Agentic AI, a concept where several specialized agents collaborate to accomplish complex actions. Each agent excels in a specific domain, and their interaction enables a form of orchestration that could evoke reasoning—with all necessary reservations.

This evolution will be significant in the coming months and years. Today, agents use generative artificial intelligence methods, but orchestration principles have been around for a long time. For example, CORBA, in the 1980s and 1990s, was already based on collaborative agent concepts.

Current approaches, which combine generative AI and statistical models, will likely persist. What Yann is developing is interesting, and the future could see the emergence of hybrid systems. This involves a combination of expert systems, sometimes considered “old” technologies, and modern AI methods. Some research laboratories are exploring this path, which seems promising.

One of the major challenges remains the ecological impact. Current statistical AIs consume enormous resources, and any advancement that would optimize their efficiency while reducing their environmental footprint would be significant progress.

Quantum computing, unlike mathematics, is primarily about physics. Here again, we’re flirting with science fiction, because even the most honest quantum researchers acknowledge that they don’t know when, or if, they will achieve concrete advances.

If we manage to master this technology, we’ll have to rethink everything, start from scratch. Today, computing is based on mathematical models, while quantum computing is rooted in physics. Part of this discipline nevertheless uses statistics to stabilize models.

But what does this evolution really imply? Some imagine it could bring us closer to how the biological brain functions. If we could imitate it, then perhaps we could reproduce intelligence itself. However, this reasoning is highly speculative. Not only do we not yet fully understand how the brain works—neuroscientists estimate they only understand 20 to 40% of it—but we also don’t know precisely what intelligence is. This remains, for now, fiction.

What is certain, however, is that within two, three, or five years, revolutionary artificial intelligence methods will emerge and experience massive enthusiasm. As with each cycle, some will announce that we are close to AGI. Then, after a few months or years, we’ll realize that these methods are mainly effective for specialized tasks, the enthusiasm will subside… before a new cycle begins again.

This phenomenon follows a well-established dynamic, close to Gartner’s Hype Cycle. For 70 years, AI has gone through these successive phases of euphoria and disillusionment. And this pattern will likely continue for a long time.

 

Do you believe more in open source, like Meta, or proprietary solutions like OpenAI, which is open in name only?

If you talk to open source experts, many will tell you that Meta’s open source isn’t as open as all that. Admittedly, it’s more open than OpenAI, which officially abandoned open source in 2018 by strategic choice.

Open source, whether authentic or partial, is a very interesting approach. Initiatives like those from Mistral or Meta allow their models to be exploited, even if access to training data remains limited. On the other hand, algorithms are, to some extent, accessible and reusable.

One of the major advantages of these open source models is that they can be run locally. They can be downloaded, adjusted, and deployed on a personal PC or private servers, offering an alternative to centralized cloud solutions.

This approach has several benefits, particularly ecologically. It allows models to be optimized, specialized, and lightened, making them more resource-efficient. By reducing their footprint and making them autonomous, we also gain control over data and usage.

Of course, open source can also be exploited in the cloud. But I believe more in open source “brought downstairs,” that is, brought closer to users and local infrastructures, rather than confined to large centralized platforms.

You mentioned impact and frugality. What strategy do you recommend to reduce this impact while maintaining innovation? At WAICF a few weeks ago, David Gurle addressed this topic with Hivenet and distributed networks.

We’ve been talking about this for a long time, because they offer a key advantage: their decentralization. Unlike centralized infrastructures, these systems allow calculations to be performed locally, at the network edge (“downstairs”), while model coordination remains partly centralized (“upstairs”).

One essential aspect of this approach is that it moves inference—that is, model usage—to local infrastructures, thus reducing dependence on energy-intensive data centers.

Today, the consumption of generative AIs is a real issue. Figures vary according to studies and query complexity, but it’s estimated that approximately 20 ChatGPT queries consume the equivalent of 1.5 liters of water. This figure is alarming, especially when millions of queries are performed daily. And this doesn’t even account for CO₂ emissions.

The data center model poses a major problem, particularly because 60% of their electricity consumption is used solely for cooling machines stacked on top of each other and generating excessive heat.

Proposed solutions to reduce this impact are sometimes questionable. Some suggest installing nuclear power plants near data centers, but this poses other challenges, particularly regarding water consumption for cooling. Others suggest placing them in polar regions, under the ocean, or even in space.

Even more extreme ideas are emerging, like those from Elon Musk, who envisions installing data centers on the Moon or Mars. While this approach may seem innovative, it would only relocate the problem elsewhere, with an environmental and ecosystem impact that remains unknown for now.

Instead of seeking to externalize these infrastructures, the real solution could be to adopt more frugal models, better distributed and adapted to more responsible energy consumption.

How can companies ensure their AI systems are designed ethically, avoiding biases?

Biases are inevitable. Rather than trying to eliminate them completely, we must understand and acknowledge them.

As long as a model is generic, it is biased, because it reflects the very nature of the Internet, which is itself biased. For example, approximately 70% of online content comes from the United States, which influences how these models think and express themselves. They adopt a predominantly American vision, with its own cultural references, errors, and inaccuracies—which are ultimately not very different from European biases.

When developing a more specialized model, trained on our own data, we may have the illusion of a more ethical system, because it relies on data we control and consider legitimate. But in reality, this only introduces our own biases into the system.

However, this is not abnormal. We simply must accept that every model is built with a perspective and understand how it shapes its responses. An AI system reproduces what it’s given and responds according to the rules and data imposed on it. The important thing is not to completely eliminate biases—which is illusory—but to have a critical awareness of their presence and impact.

 

What do you think about the role of regulation? What role should it play in AI development and deployment? Because in Europe, we tend to over-regulate. In the United States, there’s none at all. So is there a middle ground?

In the United States, regulation works differently than in Europe. It intervenes after the fact, primarily through case law, rather than through laws anticipating risks. This explains the difference in approach:

  • In the United States, we experiment, accept mistakes, then regulate afterward, in response to identified abuses or problems.
  • In Europe, regulation is proactive, meaning we seek to frame a technology before even fully understanding its implications.

This gives rise to a caricatured vision: in the United States, we invent; in Asia, we copy; in Europe, we regulate.

Regarding AI Acts, I think some decisions are poorly calibrated. Mind you, I’m not saying regulation is useless. On the contrary, it’s essential: we need laws to frame usage and prevent abuses. But these rules should be established once we truly understand the stakes and applications.

The problem with the AI Act is its blanket prohibition of certain technologies, rather than prohibiting specific applications. This approach lacks granularity. Consequently, it risks blocking innovation by preventing the development of potentially beneficial technologies.

And in Europe, the legislative machinery is slow: if a mistake is made, correction can take a decade.

We already made this mistake in the 1980s with genetics. At the time, France banned certain genetic manipulations for fear of eugenics. But eugenics is an application of genetics, not genetics itself. By blindly banning the technology, we slowed advances like gene therapy, which is today a medical revolution.

History risks repeating itself with AI: regulating too early and too broadly can prevent major advances before we even grasp their full potential.

In the end, this approach cost us 10 years. We realize, in hindsight, that it might have been better not to ban certain technologies. But meanwhile, the accumulated delay is considerable. The problem stems from a lack of granularity in regulation. Rather than targeting specific applications, overly broad restrictions were imposed, grouping very different technologies under the same regulatory framework. This inevitably hinders innovation. We’re already seeing it: companies like Meta, Apple, and others hesitate to deploy certain features in Europe, for fear of violating rules that are too vague or too restrictive. And the risk is major: not using these new technologies also means depriving ourselves of potential discoveries. Some innovations could lead to essential advances, but if they’re blocked upstream, we’ll never know.

 

And what AI development excites you most currently? How do you see this technology evolving in the coming years?

For me, the most promising application areas concern medicine first and foremost. It’s a sector where AI is already bringing extraordinary advances.

Current AIs, which rely on statistical models, are in this sense similar to genetics, which also relies on combinatorial principles, particularly in DNA analysis. Thanks to this computing power, we’ll witness major discoveries:

  • Early disease detection, by identifying markers well in advance.
  • Targeted therapies, with personalized treatments based on genetic profiles.
  • Development of new molecules, by accelerating drug research and design.

AI has played a key role for years in medical imaging, and with the arrival of generative models, its impact extends even further. These are fascinating advances.

Another area that fascinates me is transportation. With aging populations, travel becomes more complex, and roads remain a major safety issue, with nearly one million deaths per year worldwide.

AI can provide solutions to make vehicles safer. Unlike the vision of the fully autonomous car—which, in my opinion, is too expensive and complex to generalize—I strongly believe in driver assistance systems. These technologies, already present in many vehicles, are becoming increasingly effective and significantly reduce accident risks.

But AI isn’t limited to road safety. It also opens up possibilities for entertainment and education in vehicles. Rather than being distracted by simplistic games—like counting red and blue cars—we’ll be able to exploit these tools for enriching and interactive experiences.

 

Speaking of which, do you have any current projects on this topic?

A good recent example is the new Renault 5, launched in December. This model integrates an intelligent avatar, although its intelligence is still limited. However, it’s more advanced than traditional car audio systems.

It helps understand the vehicle, with a touch of anthropomorphism. It embodies the car, interacts in natural language, and even has visual expression.

The main interest of this avatar lies in its dual functioning:

  • Reactive: it answers questions, whether it’s a request about the car or a curiosity like Napoleon’s date of birth.
  • Proactive: it anticipates the driver’s needs. For example, in Paris, where the limit is 30 km/h, if you exceed the authorized speed, the avatar can suggest activating cruise control.

It’s a pedagogical approach that guides the user without being intrusive.

Artificial intelligence is now integrated into all sectors of the company, and as you’ve observed with the entrepreneurs you interview, every profession can benefit from specialized AI tools.

At Renault, AI doesn’t replace employees, but it augments their capabilities. This is what I call augmented intelligence: it assists workers in specific tasks, optimizing what they could have done with their own reasoning, but more efficiently and quickly.

 

What advice would you give to entrepreneurs who want to integrate AI?

It’s essential not to ignore AI, even if it doesn’t seem immediately useful to a company. Understanding its potential and limitations is crucial for intelligent adoption. Education plays a central role: it’s not just about using these tools, but knowing how and when to apply them. However, we mustn’t fall into the trend effect and seek to integrate AI at all costs. Companies must test, experiment, but also accept abandoning non-relevant projects.

The key word here is education. We must learn to master the tool, but also understand its limits and potential misuses. AI is like a hammer: it can be used to drive a nail, but also to strike someone. Hence the need for appropriate regulation. But this regulation shouldn’t only be imposed by states or international institutions, generally disconnected from operational realities. It can also emerge from the grassroots, particularly within companies.

Indeed, regulation doesn’t only occur at the country or world level. A company can and must define its own usage ethics, determining what is acceptable and what isn’t for its employees and customers.

One of the major risks is wanting to apply AI everywhere, even where it brings no added value. An interesting figure, though not yet officially published, illustrates this trend: in large French CAC 40 companies, approximately 170 AI POCs (Proof of Concept) were launched between 2023 and 2024. But only 10 on average were actually implemented or implementable. This represents barely 5% success, which corresponds to the normal rate of innovation.

Why such enthusiasm? Because generative AI has a unique characteristic: revolutionary ease of use. Unlike previous AI technologies, it can be exploited in natural language via prompts, without requiring advanced technical skills.

At the beginning of 2023, one of the most popular jobs in Silicon Valley was Prompt Engineer. But this role quickly disappeared. Why? Because companies realized that everyone could learn to write an effective prompt. In natural language, any user can interact with these AIs intuitively, making the need for specific expertise unnecessary.

Another amusing phenomenon is the explosion of AI “jailbreaking” through prompts. Many have discovered how to bypass model restrictions to make them generate responses they’re not supposed to produce.

 

Key Takeaways

Luc Julia embodies a pragmatic and lucid approach to artificial intelligence. For him, the future lies in specialized solutions designed to meet concrete needs. He insists that AI is a powerful tool, provided it’s used thoughtfully and we don’t get carried away by science fiction fantasies.

His perspective, both critical and optimistic, provides valuable insight into the evolution of this technology, with particular attention to humanity and ecological responsibility. His work at Renault illustrates this approach by developing intelligent systems that improve the driving experience while enhancing vehicle safety and efficiency. Rather than pursuing unrealistic promises, he designs AI focused on concrete applications, rooted in a logic of useful and responsible innovation.