Orange CTIO Bruno Zerbib (pictured) backed a push towards using traditional scientific research to provide breakthroughs to make generative AI (genAI) more sustainable during an interview in which he highlighted a range of issues impacting the sector including a perceived Nvidia GPU market monopoly.

Discussing some of the barriers to responsible and sustainable use of genAI, the executive noted large language models (LLMs) essentially function like neural networks, making it difficult to correct any problems.

“You don’t have things you can tweak” if the model gives an inappropriate answer, he noted. “The only thing you can do is retrain them” by exposing them to “hundreds of thousands of pages” with better information.

“It’s very archaic, it’s like LLMs are behaving like humans. There’s nothing I can do to force you to trust me, I need to spend months and years showing you that I’m trustworthy”.

“It’s very hard to train a brain. It’s very hard to train an LLM.”

“This is problematic. We need to work with start-ups to change that”.

Nvidia
Sovereignty of digital infrastructure has been a major talking point within Europe in recent years, with politicians and a range of companies in the region pushing it as a priority.  

Referring to the “monopoly” held by Nvidia for GPUs, Zerbib told Mobile World Live there is a “huge cognitive dissonance of talking about sovereignty and independence, then relying on just one company” for a part of the system.

“This is not an anti-American statement,” he emphasised. “What I’m looking forward to is at least competition in the US”, an area he predicts something is “going to happen” in.

Quote Icon

The ability to build a competitive GPU in Europe seems to be such a challenging goal

Bruno Zerbib, CTIO Orange

“It’s out of our hands right now. The ability to build a competitive GPU in Europe seems to be such a challenging goal”.

“We need to have a long-term goal of getting there. Maybe in five-to-ten years from now. In the meantime, let’s have French or European LLMs, British LLMs and then let’s make sure the agent layer is completely sovereign”.

The executive noted hyped agentic AI technology offers the potential for a “snowball effect” if something goes wrong with one part of the chain, with Orange currently assessing how it can curate and vet agents to ensure they are trustworthy and reliable.

Science drive
One area Zerbib is clearly enthused about is the need to ensure AI technology is compatible with environmental ambitions at the operator and more broadly.

“We live in a world that’s so depressing,” he said. “Either you say you don’t believe in climate change, or you don’t believe in the ability to innovate without hurting the planet. The two outcomes are really bad: it’s either I don’t care or I care and I’m giving up”.

“I believe in a third option”, he added, explaining a need to return to the mindset of 70 years ago where “the brightest people wanted to solve the biggest problems”.

“They really thought 70 years ago that we could change the world with science, and science and technology were combined. We’ve lost that to some extent, we’ve turned engineers into technicians that have been producing the same thing at scale, because we needed scale”.

“We have our smartest engineers with an incredible background in, for instance, mathematics and physics, and they end up being developers”.

He added although being a developer is “great”, there is a need to “redirect a percentage of those great minds into science, because what we need are breakthroughs that come from science”.

“It’s not just about building the next feature it’s really about trying to come up with something that will be much more disruptive”.

The attitude around science remains in China, Zerbib said, “but we’ve lost that in the West and we really want to get back to this notion of science is central to progress”.