For decades research on artificial intelligence has coexisted in the academic and industrial worlds, but the balance is shifting and AI (Artificial Intelligence) has become the main technological battlefield of Big Tech (and not only)
As far as Artificial Intelligence (AI) is going, not so much as a discipline in itself but as a “pool” of knowledge and techniques from which technological solutions are born, we can now see it every day, even in the one-minute reports of the national news and local. What appears less clear, but still easy to understand, is the domain of AI – Artificial Intelligencefor many years reserved for the academic and research world, then characterized by a balance between the research world and the industrial world, now moved to the second of these two worlds (we could say the Business).
Researchers Nur Ahmed and Neil C. Thompson (both from the Sloan School of Management, Massachusetts Institute of Technology (MIT), as well as researchers from the Computer Science and Artificial Intelligence Laboratory, MIT) write about it in the well-known scientific journal Science, together with Muntasir Wahed (Department of Computer Science, Virginia Tech),
«Today, industry dominates the three key ingredients of modern AI research: computing power, large data sets, and highly skilled researchers», the scientists write in their article. «This domain of inputs is translating into AI research outcomes. The industry is becoming more influential in academic publications, leading benchmarks, and identifying cutting-edge ways to study Artificial Intelligence and develop solutions. While these industry investments will benefit consumers, the resulting research dominance should worry policymakers around the world, because it means that public alternatives for important AI tools could become increasingly scarce».
The dominance of industry in the “production” of AI
The industry has long had a better access to large datasets of economic value because their operations naturally produce data, interacting with a large number of users and devices. For example, in 2020, WhatsApp users sent around 100 billion messages a day. It is therefore not surprising that most large data centers are owned and operated by industry.
The same photograph appears when looking at another variable, that of skills and talents. As an example, the three researchers, in their article, report some data relating to American universities. Data showing that Computer Science PhDs specializing in AI are heading to industry in unprecedented numbers [Figura 1]. In 2004, only 21% of AI PhDs went to industry, but in 2020, nearly 70% are. By comparison, this percentage of PhDs entering industry is already higher than that of many science areas and will likely soon exceed the average for all areas of Engineering.
Also many computer science research professors specializing in AI have chosen to leave universities to work in private companies; academic institutions thus begin to struggle to retain talent. This concern is certainly not limited to US universities.
The computing power used by academia and industry also shows a growing gap. In image classification, the computing power used by industry is greater and has grown faster than that used by universities or industry-university collaborations [Figura 1]. In 2021, industrial models were on average 29 times larger than academic ones, highlighting the large difference in computing power available to the two groups. This is not just a difference in approach, but a lack of computing power available to academics.
The ability of the industry to hire talent and exploit greater computing power probably derives from differences in spending (another not inconsiderable element that contributes to the dominance of the industry on Artificial Intelligence). While investment in AI has increased substantially in both the public and private sectors, industry investment is higher and growing faster.
Just to give one example: In 2021, US government agencies (non-defense) allocated $1.5 billion for AI. The European Commission has planned to spend 1 billion euros ($1.2 billion) over the same period. In 2019, Alphabet alone spent $1.5 billion on its subsidiary DeepMind (and that’s just part of the Alphabet-Google investment in AI). In 2021, globally, private spending on AI is estimated to have reached approximately $340 billion.
Unfortunately, the gap can also be seen when looking at outputs, especially the production of models [un modello di Intelligenza Artificiale è un programma o un algoritmo che si basa sui dati di addestramento per riconoscere modelli e fare previsioni o decisioni . Più punti dati riceve un modello AI, più accurato può essere nell’analisi dei dati e nelle previsioni. I modelli di intelligenza artificiale possono basarsi, per esempio, sulla visione artificiale, sull’elaborazione del linguaggio naturale e sull’apprendimento automatico per riconoscere modelli diversi – nda].
Industry dominance of Artificial Intelligence is also seen when considering six model benchmarks for image recognition, sentiment analysis, language modeling, semantic segmentation, object detection, and machine translation. Industry, alone or in collaboration with universities, led the way in 62% of cases before 2017. Since 2020, this percentage has risen to 91% and in some cases even close to 100%. [Figura 2].
The impacts of industry dominance on AI
While the industry’s growing investment in AI has the potential to provide substantial benefits to society through the commercialization of the technology, according to the researchers who authored the paper published in Science «the concentration of AI in industry is worrying» and it is by no means risk-free [leggi anche: AI Index Report 2023. Le Big Tech hanno il potere di governare l’AI – ndr]
«The commercial motivations of the industry lead to a focus on profit-oriented topics. Often these incentives produce results in line with the public interest, but not always. […] Recent empirical work has found that private sector AI researchers tend to specialize in “data-hungry and computationally intensive” deep learning methods, and that this comes at the expense of research involving other AI methods, research that considers the social and ethical implications of AI and applications in areas such as health. These questions about the trajectory of AI and who controls it are also relevant to the debates about job substitution and AI-induced inequality. Some researchers are concerned that we may be on a socially suboptimal trajectory that focuses more on replacing human labor rather than increasing their skills» the concerned researchers write.
If academics lack access to industry AI systems or lack the resources to develop their own competing models, their ability to interpret industry models or offer public interest alternatives will be limitedboth because academics are unable to build the large models that appear to be necessary for cutting-edge performance, and because some useful capabilities of AI systems appear to be “emergent,” meaning that systems acquire these capacity only when particularly large.
The challenge, however, cannot be overcome with a renewed and extended collaboration agreement between industry and academia. The problem is political in nature.
«For policy makers dealing with this issue, the goal should not be to have academia carry out a particular share of research. Instead, the aim should be to ensure that there is sufficient capacity to assist in the review or monitoring of industrial models or to produce alternative models designed with the public interest in mind» write the researchers.
Without this choice, the continued development of AI could leave aspects of public interest behind.
breaking latest news © (Article protected by copyright)