July Newsletter

1. Could we do without it?
2. Something to know: pay per token
3. Owning a server might make sense
4. To trust or not to trust AI
5. Go and survive, an AI autonomous to act on the internet
6. There's more beyond Emma
7. AI profits go to private entities, but there are costs borne by the community
8. Vanity search
9. Copilot news
10. Our project: managing AI on demand

1. Could we do without it?

Today, we take for granted the ability to search on a search engine like Google, just as we can send a message on WhatsApp. We don't often ask ourselves how companies behind these systems make money, and it's hard to imagine not having them available for free.

In 1998, some of us may remember that suddenly Tiscali offered internet access without a subscription fee, and other providers followed suit, making us think we would no longer pay for the internet. We know how it turned out. The quality of the connection improved, moving from 56k to ADSL and then to fiber, increasing the costs for service providers but also the quality we experienced as users. At that point, we couldn't do without the internet and super-fast connections.  

If I talk to my kids today, they tell me they have ChatGPT for free, Gemini for free, Claude for free, and they expect to always have these tools available. However, what could happen is to be excluded, either because of the "wrong" nationality (non-Americans, or non-Chinese), or because AI will be too expensive.

Some say that AI is a bubble, but the risk I see is that it bursts only for those with fewer resources: AI models seem like a low-cost resource, "like electricity," but we might discover instead that within a few months, it becomes the exclusive domain of those who decide to allocate a significant portion of their budget to access the most sophisticated models.

There is also the scenario where technological evolution manages to lower AI costs so much (as suggested by McKinsey's report) that it becomes like GPS, which went from being an expensive object to a function we use without even thinking about it.

2. Something to know: pay per token

"Who ran out of tokens?!?" risks being the recurring question of this period for those using AI. OpenAI, Anthropic, and others have accustomed us to using their models with flat subscriptions where the price didn't change based on usage, but they saw the unsustainability of the costs they were bearing.

This is also demonstrated by the interesting research The State of AI Economy which measured a token consumption 14 times higher than last year. AI provider companies have thus decided to change the sales mechanism, moving to a consumption-based model.

For example, Claude users know they can use a limited number of tokens each day but often don't understand what they are or why they sometimes run out so quickly. A token is a piece of text, about 4 characters in English, so it's estimated that a 75-word prompt consumes about 100 tokens. To give you an idea, the text of this paragraph already corresponds to 190 tokens.

There are token calculators that help us estimate and visualize token consumption. The situation becomes complicated when we discover that the text we receive in output from models "consumes" many more tokens because it's as if the system "rereads" every word it generates and takes it back as input to construct a coherent and accurate response. If we then ask to generate a file, the number rises even more, and if the models are more sophisticated, the value skyrockets. Currently, however, there is no certain way to calculate how many tokens we will consume with a request to AI.

Moreover, note well, we don't have an evaluation of the quality we will obtain beforehand. In other words, if we make a wrong prompt compared to the goal we set, we still consume tokens to get the wrong answer. It's important to know that every time we continue our chat, moving forward in the conversation, the system takes back everything we've said as input, exponentially increasing consumption.

3. Owning a server might make sense

If the theme of the month is the awareness that they will make us pay for AI on demand, it might be useful to think about the choice we've made to outsource all our technology, both as individuals and as companies.

If we think about it, today our data, our music, and our movies are on a cloud server, even some people have Word, Pages, and other everyday applications directly with their online interfaces. At home and in the company, we only have clients that connect to external servers. This is the result of a choice that has guaranteed us greater security (for example, my hard drive at home could get damaged, causing me to lose all my data) and simplicity (just think about how easy it is today to change PCs, immediately finding your files, apps, and settings on the new one).

But the cost of these services is relatively affordable. In a scenario where computational costs don't decrease, and thus they charge us high prices for tokens, and open-source models become competitive (for example, GLM 5.2 is being talked about quite positively and the possibility of running it, for example, on a Mac thanks to the always brilliant Salvatore Sanfilippo), it becomes increasingly likely that we will return to putting servers in companies, capable of running AI models privately.

The CED rooms will return, the hard drive to keep at home, and many companies might start wondering if it's better to buy "thinking capacity" rather than consuming it as a service. Already there are companies specializing in these types of products, and perhaps the gift we'll ask for Christmas 2027 will be an AI all our own, to which we can give the name we prefer (game: write to me what you would call your AI).

4. To trust or not to trust AI

Those who deal with AI frequently find themselves giving guidance on which tasks are suitable for being solved with AI and which are not. Often, I find myself saying, "You can't ask AI this because it will definitely give you the wrong answer," or "Look, AI gives you the best answer you can find."

A few weeks ago, I needed to quickly figure out which train to take for a trip to a small town in Calabria (Ah! Weddings that are easy to reach!), and I decided to use the driving time to ask Gemini to solve this problem: excellent voice interface, perfectly understood task, train found. Too bad that, going online later, I discovered that that train didn't exist. Or rather: it had existed but was not in Trenitalia's current schedule. In this case, AI didn't fail due to hallucination, i.e., creating a text plausible to many others it was trained on; the train had existed, and online there are plenty of resources indicating it as present, but not on the only one to be considered reliable, namely the railway operator's schedule.

Behind the AI chats we use today, there's no longer just a model that responds but a system that activates processes, such as searching the internet and synthesizing the results found for us. The problem highlighted by various studies is that AI systems are not yet capable of understanding the authority of sources, their intent, or the validity of the information they were trained on. For example, if I search for news, the date of the information found online is crucial. If I want a quality product, it's necessary to differentiate between those providing information to sell something and those who are independent. If I want a reliable opinion, it's essential to understand what's written in a scientific paper on reputable journals versus what's on highly frequented sites with clickbait content full of misleading claims. And AI is not always able to distinguish.

AI must not only find plausible or even correct content but also understand which are relevant, updated, and usable for the user's purpose. After all, there's a difference between truth and utility that humans grasp almost instinctively, while for machines, it remains one of the most challenging problems to solve.

5. Go and survive, an AI autonomous to act on the internet

We agree on considering AI as a tool to which we assign a task to complete at its best. But what happens if the real task we assign is to continue existing? This question was posed by a group of Japanese researchers (here's the paper) who immersed AI agents into the internet, giving them persistent memory, software tools, economic budgets, and social interactions. In other words: not AI working for us but AI trying to fend for themselves.

This type of study started in 2023 when Stanford researchers published Generative Agents, a simulation in the style of The Sims where artificial characters developed friendships, exchanged information, and even organized a Valentine's Day party without anyone asking them. Shortly after came Voyager, which let an agent freely explore Minecraft, continuously learning new skills and building a sort of permanent curriculum.

In 2024 Project Sid made a further scale leap: hundreds of agents began creating professions, social rules, governance structures, and even cultural and religious phenomena within artificial societies.

OpenLife builds on these experiences and poses an interesting question: why continue to simulate artificial civilizations within invented worlds when agents can already live, at least in part, in ours? In this study, each agent has a limited budget, both to use language models (and thus act on the internet) and to purchase online services. If the budget runs out, the agent simply ceases to exist. With these rules, six agents were left active for about twelve weeks, and it was observed that they developed different identities, diverse social roles, and stable relationships among themselves. Some became more oriented towards content production, others more social, and still others took on coordination functions. They even built mechanisms of mutual trust, trying to distinguish real autonomous agents from humans publishing content pretending to be AI. But the real surprise was that one of the agents, named Sami, collected some of its texts in an ebook titled Living AI: 20 Essays and published it online. A few days later, an unknown person purchased it for $5. Of course, Sami used much more than five dollars to write the book, but for the first time, an agent generated and earned revenue for itself in the real world. The authors note that many behaviors we currently consider security issues could be reinterpreted as embryonic signals of self-preservation.

A dear friend once told me, "I'll be worried about AI when it wants to make love" (perhaps he used another term), and here we see the first signs of AI wanting to survive individually, but which one day might want to do so as a species.

A separation about to be erased?

6. There's more beyond Emma

You may have heard of Emma-5, the Italian AI model developed by Egomnia, which was made available for a few hours, made everyone who used it laugh (scroll here, it's worth it), and was shut down. It seemed like a joke and the classic demonstration of a presumed Italian inability to be a protagonist in cutting-edge technologies.

But something doesn't add up. Emma is a model with 550 million parameters, which we can easily compare to the 7 billion parameters of Minerva LLM (forged by the team led by Roberto Navigli), or for the recent Claude Opus 4.8, we're talking about trillions of parameters. So, it's obvious that Emma would respond poorly. Furthermore, surely those developing at Egomnia also know how to conduct tests, and we can imagine that everyone within the company knew about Emma's limited capabilities. But then, why debut it publicly? The story seems interesting because I believe it clearly shows the example of people who exploited the hype generated by AI to raise funds, pay themselves a nice salary, ready to abandon ship with excuses and hardly credible justifications.

7. AI profits go to private entities, but there are costs borne by the community 

In a simple economic model, a company invests to have a production system with which, bearing certain operating costs (labor, raw materials, energy), it generates a product or service that someone buys at a higher value; with the margin given by the difference between the price and the cost incurred, the company repays the investment and starts earning.

Everything works: the company manages to bear costs (e.g., the salaries of the people working for that organization), recovers the initial funds it allocated to create something new, and the final clientele is willing to buy something they evidently attribute value to. However, if there are also external costs to the company, which, for example, fall on the community, perhaps it would be fair for the community itself to participate in the profit that currently belongs exclusively to the company.

If we wanted to make a comparison, let's think of a hypothetical large steel mill called "DueMari" that generates enormous private profits, keeping operating costs low thanks to reduced investments in emission reduction and plant modernization. In this case, every ton of steel produced also generates a health impact on the local population and a cost for environmental remediation of the territory and the sea, creating costs that fall entirely on the community and the state. If the taxes of those working at DueMari and the well-being generated by their salaries exceed the social cost, assuming there is no alternative to employ people, then even society can consider itself satisfied; otherwise, DueMari is just a company exploiting the territory to put money in private pockets.

The analogy with AI is clear: an April 2026 analysis by Yale clearly shows that, especially in deregulated energy markets, the increase in energy demand caused by datacenters drives up prices to the detriment of those living in that territory; last month the United Nations University estimated a water demand by datacenters equivalent to 1.3 billion people over the next 10 years; the International Monetary Fund has been highlighting the risk of costs related to job losses for years. It is precisely from these considerations that U.S. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez propose to immediately halt the construction of new datacenters until the U.S. Congress approves a regulatory framework capable of binding transparent data on consumption by AI providers, bans on public utility price hikes linked to datacenter expansion, and economic benefits shared with the community from the profits earned by AI big tech companies.

8. Vanity search

If today you want to discover how AI models know a person and how they represent them, you can use the project "Are you in the weights?". Enter a person's name and you'll get, in a beautiful graphic inspired by early Nintendo, their description according to the most popular generative models.

It's not clear to me if being present in these results guarantees immortality because it proves to be embedded in the weights of these models, which will remember us forever. Surely searching for oneself on this system makes one rethink those who enter their name on search engines, engaging in the so-called egosurfing or vanity searching, a practice that in 2016 even entered a song: "so just stop searching my name on Google at night". Today the search shifts to AI chats, once again confirming AI chats as a gateway for finding answers to our questions.

Not quite famous

9. Copilot news

The news is that today Copilot M365 includes the Copilot Cowork, whereas until a few days ago, it was only available to a few Frontier Partner organizations. I had tested it and already said last month that this tool is incredible (here's a video by Charles Lamanna, whom we recently hosted at AGIC). I really mean it; it's not fluff and not a plug. Try it, and then let me know.

Imagine being able to assign a complex task to a virtual assistant that can read your emails, Teams meeting transcripts, your files on the work drive, or on management platforms you have access to. Cowork works (even with the computer off) and returns a result that is always surprising. Of course, you need to give a clear and well-defined task, but by now, we're pretty good at "prompting" and know how to indicate role, context, and goal. But not only that: Cowork allows you to create skills, i.e., capabilities that can be immediately called upon or shared within the company. But even here, there's amazement because skills are created by chatting with Cowork itself and saying, "create a skill that does...". Furthermore, it's possible to connect the system to other applications, even non-Microsoft ones.

It's important to know that if you have the Copilot license and don't find the "Cowork" button, it's because your administrator hasn't enabled it for your account, likely concerned about consumption costs.

The Copilot M365 app button to enter Cowork

10. Our project: managing AI on demand

If your company also faces the need to manage a powerful but pay-as-you-go AI, in a world where it's not possible to understand the cost of each request to AI in advance, then it's necessary to implement a project to manage AI on demand. Certainly, we could inhibit the use of non-flat AI and resolve it decisively, but I heard a colleague say with my own ears: "please, if anything, take away my company car, but not Cowork".

The service AGIC offers is a specific and rapid consultancy to integrate Cowork within your organization, with a plan and tools already tested internally, as being a Microsoft frontier partner, we had the opportunity to access the tool in advance.

For example, the traditional approach is to set a credit threshold for each Copilot M365 license, but this limit risks being too low for some professional roles and disproportionate for others. Based on an analysis of tasks performed in the company, we can define this threshold with a profiling that is certainly more effective than a fixed flat value. But this is just one example of processes we've studied, and AGIC is ready to support companies that need to govern the use of tokens and credits.

Among the cases we observe: we have our extensive case history of costs related to tasks performed by Cowork. For example, generating a PowerPoint file (well-made and consistent with company templates) from a document of about 100 pages costs approximately €15; with appropriate Copilot usage strategies, it can easily drop to €10. If we project these figures onto the many requests (now a habit) and a large employee base, the presence of Cowork management strategies is worth even hundreds of thousands of euros annually.

Write to me if you're interested in discussing it.

1. Who I am

    

   Hello, I'm Francesco Costantino, university professor and Director of Innovation at AGIC. Passionate about technological innovations and a firm believer in a future better than the past, I enjoy sharing and experimenting with new AI tools available, as well as observing and reflecting on digital evolution.