This text is part of the Italian Tech IA Special, entitled “Talk to Me. The Artificial Intelligence Revolution”on newsstands free with the newspaper The Republic starting from Thursday 23 March
Last month Meta released a powerful scientific paper titled LLaMA: Open and Efficient Foundation Language Models. LLaMA stands for Large Language Model Meta AI and is a NLP (natural language understanding and processing) system; like ChatGPT, but better: “Exceeds GPT-3 in most tests,” the authors write with undisguised pride.
“We have an ambitious research program aimed at creating autonomous artificial intelligence” comments Joelle Pineau, managing director of Meta AI, responsible for laboratories in the USA and Europe. “Research and exploration, open science, and cross-collaboration are central to our AI efforts. We remain true to the principles that inspired Fair (Facebook AI Research) in its inception: We continue to publish and share open source our work not only to validate our findings, but also to enable others to build on our progress, accelerating the world‘s path to more advanced AI.”
Interview
Generative AI, 5 questions (plus one) to Hiroshi Ishiguro
by Emanuele Capone
Between the end of 2022 and the beginning of this year artificial intelligence reached the general public: it is talked about on TV, in the newspapers, even among non-experts. Pineau, who sits on the editorial board of the Journal of Artificial Intelligence Research and the Journal of Machine Learning Research, confirms that this is a crucial moment: “In recent months, AI, and especially generative AI, has reached a turning point both in terms of the ability to capture people’s imagination, and in terms of technological maturity”. Creating many expectations, but also many alarms: “When we look at the incredible potential that AI can unleash in both the real and virtual worlds, we must contrast optimism with awareness of the risks. Which can present themselves in many forms, both through unintended uses of new technologies, whether through malicious actors seeking to exploit areas of vulnerability.And that is why we must build our AI systems with the highest levels of accountability and convey the same commitment to all those who they occupy”.
A few weeks ago Microsoft integrated ChatGPT into the Bing search engine and the results, surprising as they were, did not fail to arouse perplexity: the answers often denoted a sort of double personality of the AI, on the one hand active and punctual, on the other manipulative, lying. Not to mention the inaccuracies, limitations and what are defined as “hallucinations”: a technical term, which however here has exactly the same meaning as it has in everyday language. In Redmond they took a few days to reflect and disciplined Bing’s temper. But isn’t it that the method of self-supervised learning (SSL, Self Supervised Learning) used by this technology involves a certain degree of unpredictability of the results?
Justice and artificial intelligence
Being a lawyer – and a judge – at the time of ChatGPT
by Ernesto Belisario
“Self-supervised learning – explains Pineau again – allows us to overcome the limits of supervised and semi-supervised learning. It allows AI to learn like us, processing a significantly greater amount of data, which would be impossible or impractical to label. SSL it is an important milestone towards a unified model that learns how the world works, through the various ways in which it presents itself”. Not only text, therefore, but also images, videos, music and everything that can be translated into data. Understanding this world better will also help AI build others: “Future experiences in the metaverse will go beyond what is currently achievable and will require enormous advances in a number of technologies. The main problems that artificial intelligence is called upon to solve include 3D object and scene generation, avatars, realistic codecs, creative speech and much more,” comments Pineau. “If today we can bet on the metaverse it is thanks to the long-term investments we made in AI almost a decade ago,” she adds. Meanwhile, the battle between the greats of hi-tech is getting more and more heated, and Meta itself is constantly redefining strategies and organizational charts, between cuts and personnel transfers.
What could possibly go wrong? The launch of Bing has been criticized by many as premature; Google’s generative AI Bard debuted with a blunder so stark it caused the company to crash. And Meta, who had presented Galactica, an AI capable of writing scientific texts, had to withdraw it in a hurry because it often reached absurd, if not dangerous conclusions. Hallucinations, indeed. “At launch, we illustrated the limitations of large language models, including the possibility of generating inaccurate and unreliable results, even though Galactica was trained on high-quality scientific and academic data,” points out Pineau, who specializes in developing of models and algorithms for planning and learning and study their application in fields such as robotics, healthcare, games and conversational agents.
“It is vital that the public understands the strengths and weaknesses of these language models in an open and transparent way. We also believe that sharing our research findings is the best way to get feedback on these systems and push beyond their limits current”.
One is obvious: artificial intelligence, despite its name, is masculine. “There are still too few women in this field. We believe that having a diverse workforce is an essential ingredient in building better systems, and we are constantly striving to increase the diversity of our research teams, with hiring strategies such as the Diverse Slate Approach and mentorship, sponsorship and career development programs”.