Luis Buezo, head of artificial intelligence at HPE: “We believe in powerful, reliable and, above all, open and accessible AI”

Luis Buezo, born in Irun (Gipuzkoa) 54 years ago, but living in Madrid almost all his life, is the global head of artificial intelligence at Hewlett Packard Enterprise (HPE), a company with 60,000 employees that has to its credit the computer considered as the fastest in the world: the Frontier, a machine with 1,194 exaflops of performance and rated as one of the best developments of last year. In Finland they have LUMI, the largest in Europe (380 petaflops) and the fifth in the world. This brother three times smaller than the Frontier has computing power equivalent to that of 1.5 million latest generation laptops operating at the same time and which, stacked, would form a tower 23 kilometers high.

This commitment to supercomputing is key for artificial intelligence (AI), which, in its most complex models, requires trillions of parameters. But Buezo believes that, although this race is necessary, there are alternatives for any scenario and that the development of AI is solid and has incalculable benefits at all levels.

Ask. Is there a risk of a bubble in artificial intelligence?

Answer. We are in one of the most disruptive moments of our lives, comparable to the emergence of the internet, which we have already seen what it has generated and what we are capable of doing thanks to connectivity. We are at an inflection point in which all companies and organizations are evaluating how to apply artificial intelligence to improve their internal and business processes, quality and productivity. We believe in powerful, reliable AI and, above all, open and accessible to all types of organizations. If the Internet had only been used by a very specific group of organizations, it would not be what it is.

P. But the greatest developments in AI are in private hands that will want their benefit, the return on investments.

R. We defend an open architecture that allows collaboration between different organizations. This is very important. A few weeks ago we announced joint architecture design developments software [programación] with Nvidia. We also work a lot on the catalog of open source [código abierto], not just with business models. We do it with both alternatives because we have scenarios for each client in which they use one or the other. We apply our ethical principles of respect for privacy and human rights, that there is always personal supervision and that it is inclusive to minimize any type of bias. It also has to be explainable, controllable and robust, with security from within and protected from any type of attack, misuse or potential failures.

P. Does the existence of artificial intelligence available to organizations that do not respect these rights pose a risk?

R. Clients demand ethical and legal principles in developments. It comes up in the first conversations. Those companies or entities that resort to open source and flout all these regulations will be their responsibility. But increasingly there is greater awareness for responsible use.

P. AI has fully entered areas such as pharmaceuticals and industry. But there are others, like justice, where it is more difficult to see. You participated in a seminar on this field. Will there ever be automatic trials, computers in robes?

R. The meeting was in The Hague [Países Bajos] and we began to propose use cases. I see very relevant advantages in justice as an aid to decision-making and in different processes. Human supervision is very important here. It is not completely changing a process, but rather helping decision-making and information processing. There are generative AI models where you can enter all this information and it can help you make summaries, ask specific questions and review where this information or a conclusion comes from. Nowadays it can help a lot to speed up processes and allow professionals to dedicate more quality time to what is most important.

P. In what fields is AI going to develop more?

R. In all. It applies to all types of activities in any sector, from public administration to private companies. Having reliable information much more quickly that is concrete and precise applies to all types of sectors. In the industrial sector there is a lot for quality control or for automating factories. Financial institutions are very advanced in the applications of analytical techniques and fraud control. But we also have examples of use in e-sports [deportes electrónicos] or in motor competitions. We work with the Maserati team to improve competitiveness.

P. What needs to be done to avoid poor AI applications?

R. We work in a cycle of successive approaches that begins by managing expectations: what can and cannot be done with AI or advanced analytics. This is very important, it is understanding and agreeing on the most feasible use cases with concrete results in the short and medium term. The second element is to see the data that is available and whether, with that data, the return on investment that the client expects can be reached. If they are not valid, we look for other ways to achieve it. In the value testing phase we verify that the technology is worth it, that the specific use case we are evaluating will really produce the client’s return on investment. We do it in a laboratory environment and we can improve it. We look for models in which the performance is above 90%. Then there is the challenge of taking it to production, integrating it with existing applications and testing it.

P. A great language model needs trillions of parameters. Has the era of supercomputing arrived to handle them?

R. We have several alternatives. We have supercomputing with which you can train and develop large language models, but we also have another architecture for companies that want to make a lighter model. It is very important to understand what the use case is, what the need is.

P. Is AI sustainable?

R. At the Global Center of Excellence in Artificial Intelligence and Data in Madrid we have focused on sustainability using software from Intel to optimize the model in terms of speed and energy consumption. We have managed to accelerate the image processing speed up to 14 times and have reduced the level of energy consumed by the algorithm up to 17 times, which ultimately translates into a reduction in associated emissions. It is an important issue that clients also demand from us. It is always very important not only to look for performance, but also that the model is optimal in terms of consumption and sustainability.