Let me start by pointing out that this text was conceived and written by a human being.
Those who use generative AI engines might understand it on the fly, others must take a leap of faith, and perhaps take this as a stimulus to develop their critical evaluation skills, increasingly important as we begin interacting with these tools.
However, I did ask AI to create an original image to feature in this post, which was “cartoon style in which a man is seen in front of a robot 1.5 times his size, with the man and the robot looking at each other curiously“. I took less than a minute. I used ChatGPT/Dall-E, but could have used other “text-to-image” services already available such as Gemini/Imagen3, MidJourney, etc.
This small creative premise could be an encouragement to focus on the comparison, the differences and value added considering human actors and artificial counterparts. The balance between AI and Humans does not just have ethical or cultural aspects, it is a very practical and technical evaluation issue that, as happens with the arrival of any disruptive innovation, should make us focus on the strengths and evaluate the risks in case of improper use.
Artificial Intelligence is undoubtedly a powerful tool, capable of greatly amplifying human capabilities. However, it is based on traditional software, algorithms, mathematics and electronics, i.e. deterministic solutions that cannot generate consciousness or other intrinsic qualities typical of evolved living beings – which by the way are still completely unexplored and also therefore impossible to engineer. Perhaps this will change in the future, as we progress in fields where we have not advanced much yet (e.g. quantum physics), but that is another story entirely.
The boundary defining where AI ends and humans begin – or situations in which AI and humans alternate in more complex collaborative flows – is not fixed but will vary over time depending on the development of this technology, the applications built by integrating AI and other technologies, the quality of the components used to build the solutions, and the needs of end users
These are fundamental aspects that I would like to focus on in the future, delving into specific use cases.
1- Defining AI
One of the most complete definitions is the one used in the European AI Act (2024/1689, July 2024) and shared by the OECD, where the key capabilities of an AI system are indicated: autonomy, adaptivity, generation of predictions, content, recommendations or decisions starting from inputs received, influence on physical and virtual environments.
But what do the most evolved AI systems “think” about it? Try asking them directly. I asked both Gemini (Google) and ChatGPT (OpenAI) and their answers give an idea of how these two services were trained: ChatGPT gives a definition in which it immediately mentions the future “General AI” with cognitive capabilities similar to human ones, while Gemini gives a more practical answer, based on the present or specific cases, a bit more in the pragmatic style of Google whereas OpenAI tends to emphasize future potential and the inexorable catching up with human capabilities.
To sum it up, I recommend the general definition provided in the EU regulation, including in the term “contents” everything that an AI can generate (including decisions and recommendations):
“An AI system is an automatic system capable of generating predictions and contents and with these influence the surrounding environment“.
2- Improving results with AI
With the definition clear, now we must “put it into practice”, with concrete applications to perform specific tasks, in order to make them similar to human behavior.
Let’s see how to build these solutions, with a look at the costs and skills needed.
3- Resources!
A cross-cutting theme that is often discussed is that of the resources and hardware infrastructure needed to run the software, use the data and make AI applications available to an audience that can reach hundreds of millions of users, as those already using services like ChatGPT, Gemini, Claude, etc.
This issue has a significant economic and geopolitical impact because it concerns the availability of servers in data centers (using thousands of CPUs and GPUs), consumption of raw materials to produce this hardware, electricity needed to power everything, the environmental impact of producing this energy, and last but not least the technological monopoly by some suppliers, in particular those who currently produce systems based on GPU-Graphics Processing Units to be used on a large scale (NVIDIA among the best known).
The resource issue is above all a concern for service providers that implement large-scale platforms used simultaneously by thousands or millions of users. We have moved from the “simple” scalability and connectivity problems of Internet web portals that were a concern in the early 2000s, to the current and future needs of Generative Artificial Intelligence architectures that, in addition to managing the traffic generated by millions of users, must be sized to process huge amounts of data (quantified as “tokens”) using an equally huge amount of “parameters” (proportional to the complexity of the models), both in the training phase and in the use (inference) of the service.
For example, LLama3 400B (Meta), one of the best generative AI models available today (mid-2024), is trained on a huge data base with 15T (trillion) tokens equivalent to about 11 trillion words (about 1-2 times the words contained in English Wikipedia) using 400B (billion) parameters: a single training session of this model involves the use of data centers with thousands of GPUs used non-stop for dozens of days, with consumption of tens of MWh and overall costs of tens of Millions of euros (50M$ or more). Clearly, if the energy were not renewable, the tons of CO2e generated would cause a non-negligible environmental impact.
Unlike other types of AI, for which larger computing infrastructures do not always improve results (for example, generating overfitting problems), for generative AI models (or LLM Large Language Models) improvement of the models is proportional to their size and therefore to the underlying computing infrastructure, and if there is a limit to this type of improvement it has yet to be demonstrated!
This is why the race to finance and build increasingly larger and more powerful data centers has now begun, and it is being estimated that in some regions their impact on energy consumption will increase from the current 1-2% to around 20% in 2030.
4- Applications and use cases
Given the multiplicity of use cases, it is useful to use a classification to take into account, e.g, user type, data criticality and cost of creating as well as using the applications.
Individuals using AI to improve their personal productivity already have various online services available – based on LLMs – with which they can, for example:
- do advanced searches
- generate or summarize content (texts, images, sounds, videos)
- write code
- translate text
- analyze data
- generate forecasts
- plan activities as best as possible
- simplify complex operational procedures
- obtain specialist consultations (medicine, administration, legal, etc) possibly in experimental environments (sandboxes) and with the direct support of human professionals
It is important to underline that, when using these services, even if the interaction seems to occur only with the Generative AI virtual assistant (LLM), actually other specialized ML models could be in use (even if not visible to the user), e.g. software filters to limit input and output, storage to memorize conversations, systems to record and authorize users.
Many of these services are available (with limitations) free of charge or for a fee (with fewer limitations and better service). The business model adopted by providers generally involves collection of anonymized data for the purpose of further research and improvement.
In any case, as already happened with the introduction of free web search engines in the past, this is a new important example of democratization of innovation.
When it comes to company applications, the use cases have very different characteristics, especially due to the critical nature of the data managed, the need for customization, implementation costs.
AI is now arriving even in companies that are “late movers”, thanks to employees who start using online LLMs, after discovering the advantages for individual productivity. This should make companies reconsider overly restrictive policies that prohibit employees from using these services putting company innovation and competitiveness at risk.
Users of individual applications are “manual integrators”, using one or more different tools or services depending on the objective. Corporate AI solutions instead require structured customization and integration, which include data collection, application updates and training processes, and user interfaces.
An important part of the analysis, especially in terms of costs, concerns the type of architecture to implement: is it better to use the cloud and external services (building SaaS, PaaS or IaaS architectures) or rather have your own infrastructure and software executed locally? Is it better to use proprietary software and systems (closed source), or rather applications with an open source license?
Depending on the choices, costs can vary even 10 times, keeping in mind that completely local solutions involve additional investments and operating expenses not only for the infrastructure (computing capacity) but also for the skills required, often very different from the core business of the companies themselves.
The size of the companies (large, medium, small or micro) is a determining factor, given the different investment capacities and the amount of data produced.
If there is an awareness of the necessary costs, it is understandable that companies may question their restrictive internal policies (security, intellectual property protection) and prefer to evaluate shared responsibility models proposed by cloud service providers, or reconsider the real value of their data in the event of possible security breaches.
Wanting to list some concrete examples of business applications, we could distinguish the following types:
-
Solutions for specific objectives (task-based) achievable with already consolidated AI and Machine Learning models, such as:
- production activity planning, raw material management, predictive maintenance, R&D, demand analysis, etc -> use of regression models, data classification and forecasting, text analysis, recommendation
- safety, fault detection, quality control, production and packaging support, etc -> in addition to the previous ones, use of computer vision models, clustering, anomaly detection
- corporate digital twins, i.e. simulation of entire production processes (Industry 4.0) or research and development -> appropriate integration of multiple models like the previous ones.
-
Solutions adding Generative AI models (LLMs), such as:
- virtual assistants or co-pilots for employees, i.e. the possibility of adding natural language processing to processes, machines and control systems, production management systems, data archives, etc., exploiting the ability of these models to process all available information in real time, and generate immediately usable output
- virtual assistants or co-pilots for customers for direct pre- and post-sales assistance, with performance and acceptance levels far superior to current chat-bots
- management of the company knowledge-base, i.e. continuous analysis and digitalization (embedding) of company data and documents, so that virtual assistants have access to and make the best use of continuously updated company data.
It is obvious that AI solutions for business use require appropriate design, integration and preparation, including the choice between adapting or training AI models (expensive) or rather using pre-trained models and runtime data, and the choice between using proprietary computing infrastructure (expensive) or implementing everything on shared cloud infrastructures or services.
Last but not least, the tendency of companies to collaborate with supply chain partners for performance, competitiveness and sustainability needs will certainly constitute a further incentive to rationalize, choosing the best possible solutions.
A smart way to proceed is to start small, creating a pilot project that often has really negligible costs.
