[Questo è il primo di una serie di quattro articoli sullo stato dell’arte dell’Intelligenza Artificiale a livello geopolitico, tecnologico, sociale e regolatorio]
2024 promises to be a year of enormous evolutions in Artificial Intelligence, both on a technological and geopolitical level. And above all it will be a year in which more and more Public Administrations and Government bodies around the world will leverage Artificial Intelligence technologies to enhance their services.
For the “magnificent and progressive fortunes” of positivist philosophy everything is fantastic; for those who more prosaically deal with the public good and serve their country, caution is a must.
Between the 60s and 80s there was a nice Esso advertisement that struck me a lot as a child: “Put a Tiger in the engine“. Artificial Intelligence is a tiger that releases enormous energy. But what kind of tiger are we talking about? An indomitable one or a tamable one?
The closure of OpenAI: Artificial Intelligence on the iPhone
2023 was the year of OpenAI and its flagship product ChatGPT. OpenAI has made a powerful emergence into the global technology scene, but its rise was quite predictable. I talked about it in my speech at the Italian Tech Week of 2021. At INPS we were experimenting – with great results – GPT-2: that was the last of OpenAI’s Artificial Intelligence systems that was released in Open Source, i.e. “open”, accessible and modifiable by anyone with the skills. But we could already see first-hand the leap in quality of the next version, GPT-3, which OpenAI had however made “closed” and inaccessible.
I’ll try to be more understandable with a commonly used comparison: it is as if OpenAI had transformed into Apple and ChatGPT into the equivalent of the iPhone. Very expensive but exceptional product, the pinnacle of the range, but to be taken sight unseen with practically no possibility of modification.
A closed system, in which we rely exclusively on the technical skills of OpenAI and ethical decisions are delegated to them. And if there is little to discuss about the current technological supremacy of OpenAI, I personally have several doubts about the behavior of the people at the top of OpenAI. There soap opera of coups and backlash that occurred at OpenAI last November, those internal power struggles, claims of ego and lust for money, smack of a Nietzschean “human, too human” and they ask a serious question: Can a Public Administration or Government really entrust the heart of the country’s future technological services to such a company? One thing is important to point out: to provide increasingly correct answers, modern Artificial Intelligence systems need to be fed with data, of high quality and in large quantities. And, laws or regulations aside such as GDPR or AI Act, I honestly find it hard to imagine a Public Administration or a Government giving away the sensitive data of its citizens to these companies, moreover foreigners.
Macron’s opening: Artificial Intelligence Android model
I do not express judgments on the political decisions of other countries, but I come up with a prediction. Within 10 years, France will assume, together with the USA and China, the world leadership in Artificial Intelligence. And a huge part of the credit will go to Emmanuel Macron.
I already talked about it in June 2023 in this article of mine. In summary, Macron understood very well the strategic importance for France of having French technological giants in Artificial Intelligence. In a world where web, cloud and mobile are essentially dominated by the USA and China, it would have been disastrous for France not to try everything to establish itself in Artificial Intelligence: a field with exponential growth in applications and research, but still very fluid and without certain winners.
In summary Macron:
has identified 3 French champions, young 30-year-olds, who work at Google (DeepMind) and Facebook on advanced Artificial Intelligence projects. In May 2023 he had them set up Mistralan Artificial Intelligence startup with the seemingly impossible ambition to compete with OpenAI. In June 2023, it ensured that Mistral received 105 million euros of investments, from public and private funds. All less than a month after the startup was founded. In November 2023 Mistral received another 385 million euros, also from giants such as Nvidia and Salesforce, to a valuation of 2 billion euros. All this 6 months after the startup was founded
But in reality, what does Mistral do? Just like OpenAI, Mistral creates LLM, i.e. special foundational models of Artificial Intelligenceon which to then base advanced services (for example, ChatGPT in the case of OpenAI).
However, Mistral releases these models in Open Source mode: open, accessible and modifiable by anyone, just as OpenAI once did with GPT-2. To return to the previous comparison, while OpenAI/ChatGPT represent the Apple/iPhone of Artificial Intelligence, Mistral and its products are the equivalent of Google/Android of Artificial Intelligence.
The run-up to Open Source
The first reaction you might have is: “ok, systems like OpenAI/ChatGPT are closed (and expensive), but they work very well. Systems like Mistral are open (and free), but who knows if/how they work”. However, in recent decades the world of IT has demonstrated the incredible solidity of open systems: it is no coincidence that the largest companies in the world use Linux (open operating system), and it is now 9 years since the Microsoft understood this paradigm.
OpenAI was founded in 2015: I’ve been doing research for 6 *years*acquires skills in GPT technologies and has almost 800 employees. Mistral, on the other hand, was born not even 9 *months* agoand has around twenty employees to date.
Let’s take a look at 3 independent LLM model evaluation sites:
Results of the three sites as of February 4, 2024
Often the first to arrive at conquering a new sector have the so-called “first-mover advantage“: that is, they have a strong advantage over future rivals. However, this does not seem to be true in this case.
Mistral already performs better than GPT-3.5. And the gap between proprietary solutions like GPT-4 and open ones like Mistral (or like Flame of Goal) is rapidly decreasing.
In this regard, it is very interesting ARK Invest analysisbased on data as of November 10, 2023, which visually shows this trend.
Everything seems to confirm the considerations of Google’s internal document (“We Have No Moat, And Neither Does OpenAI“) released last May, which highlighted how Google or OpenAI had no real technological advantage and defensive barriers against Open Source systems.
Where is Italy?
In Italy we do not have national champion companies in the IT sector; above all, there is a lack of an ecosystem of collaboration both in research and in applications. It hurts my heart to say it, but there’s no point in beating around the bush: IT in Italy has been practically a desert for 50 years.
However, there are various oases, some very flourishing. One above all, the CINECA, the supercomputing center in Bologna. I have known CINECA for some time: I visited it 20 years ago as a computer engineering student from Bologna. At the time CINECA had an IBM supercomputer, I believe then among the 50 most powerful in the world, of which CINECA was (rightly) proud and which was used for meteorological calculations.
Last December I went to CINECA to better understand if/what they were doing in the field of Artificial Intelligence: I have no words to describe what I found. “Mind blown” (amazed), as they would say in the USA.
At the time of launch in 2022, Leonardo, CINECA’s current supercomputer, Leonardo, was in *fourth* place in the world for computing power. At the moment it has fallen to sixth place in the world, but it has an important characteristic: with around 14 thousand NVIDIA A100 GPUs (to which a host of H100 will soon be added) it is *third* in the world with this hardware (after Meta and Tesla) and in pole position to be the European infrastructure on which Artificial Intelligence systems are based.
Source: State of AI Report Compute Index
Forgive me for the unpoetic expression but, as my grandparents used to say, you can’t have a wedding with dried figs. Excellence must be well supported financiallyand this was the case of the Leonardo supercomputer. A 240 million euro projectfinanced half by Italy through the Ministry of University and Research (to which my personal applause goes) and half by the European Union through the EuroHPC Joint Undertaking.
A far-sighted investment, which has effectively made Leonardo of CINECA the training engine for the greatest European Open Source foundational models. In other words, European countries that want to create an open LLM, without depending on closed solutions linked to the English language/culture, pivot on Leonardo of CINECA.
An example? Mistral, who thanked CINECA recognizing the decisive role they had in their success.
This from CINECA is therefore not power as an end in itself. Indeed, it is absolutely crucial for the creation of a generative Artificial Intelligence for the Italian language and culture. An Artificial Intelligence that is transparent, explainable and functional to the public and private services of our country.
I believe in it a lot, and I will talk about it in the second article of this mini series.