China was on everyone’s lips at the IA top of Paris, these February 10 and 11. It is true that the start-up Deepseek had skillfully chosen its launch window, a few weeks before this world event. The American Openai also multiplied the announcements, as was the French Mistral who launched his Chatbot IA baptized the cat. But on the world ia chessboard, who really runs in mind? And which countries are advancing discreetly but surely? Answers with George Lee, co-director of the Goldman Sachs Global Institute.
L’Express: where is China in the AI race?
George Lee: Everyone is impressed by China’s advances in AI, especially Alibaba with Qwen and the Deepseek model. We have fairly good information on Chinese AI, but we don’t have all the details. Deepseek imitates border models and seems to catch them rather well. And this, despite the limits of China in terms of access to high performance GPUs. They also seem to have reduced the costs of the training phase. This recalls an important lesson: the need is the mother of the invention. China, deprived of certain resources, has been forced to focus on specific alternative approaches. The information available on Deepseek is imperfect but nevertheless striking. They remind us that we should never underestimate China. Especially since it has huge resources and the desire to use them to improve its technology.
What will be the impact of Donald Trump’s election on American AI?
It is still early to draw conclusions. I would make two observations. The Trump administration seems to adopt a pro-growth approach, not very regulated and pragmatic. And AI is increasingly perceived in Washington as a world competition factor. It seems that the Trump administration is as decided as the previous one to impose sanctions. It considers them as a useful tool.
Is Europe lagging behind?
Europe has two major advantages. An extraordinary intellectual capital reservoir in this area. Many fundamental elements of AI – mathematics, probabilities, logic, philosophy – develop here. You have leading university institutions which are well equipped to produce high -level graduates. I remain extremely optimistic about Europe. Regarding European AI regulations, there are two facets to this medal. Some innovators and entrepreneurs perceive European regulations as restrictive and intrusive. On the other hand, clear regulations provide safeguards and clarify the opportunities and risks that companies must be aware. There is therefore an advantage in regulations, even for young shoots. One of the weaknesses of Europe is that it faces challenges similar to those of other regions. Do you have sufficient computer power? Do you have the energy infrastructure necessary for training and use of AI? The United States has set out to create ways to accelerate the development of AI. For Europe, the same questions arise but in a more complex way due to the fragmented nature of this set of countries. China, on the other hand, has a major advantage in this regard.
What do you think of Mistral? Currently, Europe has only one European start-up in the most powerful “borders” AI models. Isn’t that a problem?
We are all very impressed by the progress of Mistral. But you are right, out of the first 15 to 20 “border” models, very few are European. However, Europe should not be content to be a model producer. The key will be the way AI will be used to create productivity and an economic impact. We must remain attentive to this dynamic.
What are the new players to monitor in the field of AI?
Two regions deserve our attention. Gulf countries. Stargate, the recent announcement of Openai and Softbank with the Abu Dabi fund, MGX, was for this much notable. The Gulf countries are distinguished by their commitment to AI, their sophistication and their aggressive research of opportunities in this area. India is the other actor to monitor in the field. It presents similarities with Europe, including its solid university ecosystem. I think India will become an important actor in the future. Of course, Taiwan (TSMC), the Netherlands (ASML) and Japan remain key poles in the AI ecosystem. Japan, in particular, has made impressive progress. Its strength lies in the fact that the country is a nodal point of AI and robotics.
What is, and what will be tomorrow, the concrete impact of AI on the economy? Some experts fear a bubble of AI …
The debate on the measurable economic impact of the AI - its magnitude and its calendar – is not closed. But we already see clear examples of the added value of AI for companies. There is always a gap between capacity development and adoption in the real world. Regarding the risk of AI bubble, I would say that this technology is evolving at an unprecedented rate. Due to its novelty and complexity, the sector attracted a lot of investments and attention. However, the path to follow is very uncertain. Historically, markets tend to oscillate around new technologies. I think it is likely that we will attend fluctuations, but as technology continues to progress, the market will eventually calibrate the real impact of the AI.
We understand the AI better than when launching Chatgpt in 2022. What are the points on which we completely deceived ourselves at the time?
The speed of change is indeed amazing. Just remember the shock and amazement caused by the exit from GPT-3.5. Few observers would have predicted such an improvement rate. The main thing we did not plan, I think, affects the principles of scaling (Scaling Laws): the fact that the more data there is and the more calculation we use to train the model, the more it improves. Many people thought it wouldn’t last. That we would quickly run into a wall. But today, we note that there is another layer that we can put on the scale: not during the training period, but during inference (note: when the models generate the response to our request) . Many people think that progress slows down, but this new advance suggests the opposite.
Guaranteeing the accuracy of AI’s responses remains a challenge. Can we only get there?
In the past, we have used deterministic systems. The nature of the new generative AI is on the other hand probabilistic. This intrinsically limits certain use cases. However, progress in terms of precision and reliability has been remarkable thanks to techniques such as the RAG (Retrieval-Augmented Generation), research methods based on the Internet, better anchoring and validation tools. This is what we see in the latest in -depth research tools of Google or Openai. That said, certain probabilistic aspects inherent in AI will always remain.
.