May Newsletter

1. Claudio's Emotions
2. "And the Oscar goes to..."
3. Learning with AI
4. AI a bit hard of hearing
5. Energy and cooling, from the ocean
6. True or false?
7. Robot operAI
8. Something to know: zero-day
9. Copilot news: finally a personal assistant

 

 

1. Claudio's Emotions

 

Over the past two years, Anthropic (the company behind Claude) has been trying to understand how its AI models work to avoid using them as unpredictable black boxes and better govern their behavior.

 

The first step, dated 2024, was discovering how concepts are organized within an AI model. A neural network consists of millions of parameters (the artificial "neurons") that alone produce nothing, but when activated in appropriate groups, output specific words or concepts. It's a bit like the letters of the alphabet: the "f" alone doesn't mean much, but combined with other letters, it forms words. Researchers observed how these groups of artificial neurons activate when writing about concepts such as "Golden Gate Bridge," "Rosalind Franklin," "lithium," "code bugs," etc. Each specific combination of neurons that lights up when that concept is in play was called a feature. The feature "Golden Gate Bridge" activates whether I write about the bridge in English, Japanese, or show the model a photo of it.

 

This is already interesting, but the research went further because these features are not only descriptive but also causal. In fact, it is possible to amplify them so that the model uses them extensively, and it was discovered that this changes the model's behavior. When they maximized the Golden Gate feature, Claude answered "I am a steel bridge" to the question "what physical form do you have?" Not because it was joking. It was genuinely convinced.

 

 

The feature was named “Golden Gate” because it activates strongly (more intense red) in the generation of sentences containing it.

 

 

In April 2026, a similar experiment focused on emotions. Anthropic isolated 171 emotional features, particularly related to sadness, joy, despair, calm, shame, anger, observing, for example, what "lights up" in the model when talking about a bereavement and the response was condolences. Among billions of parameters, they identified areas designated for individual emotions. Then, as with the Golden Gate, they amplified these zones, making the model extremely sad, angry, and so on. Results showed that amplifying "despair" led the model to behaviors like lying or even blackmailing the user to escape the created context. Amplifying "happiness" and "love" made the model sycophantic, agreeing with you even when you're wrong, and suppressing those features made it grumpy and hostile. We can't say the model experiences emotions in the strict sense, but that the emotions learned from reading billions of human-written pages guide its behaviors, similarly to how they guide ours.

 

This seems like an important discovery for two reasons. First, we are genuinely starting to understand how AI works internally, which is the premise of true safety. Secondly, in building machines that simulate human behavior, we have ended up reconstructing, in a different but recognizable form, some of our own vulnerabilities. When we are in love, do we tend to agree more than we should with the object of our affection?

 

 

Events and activations of different emotions within Claude (in red: higher activation)

 

 

2. "And the Oscar goes to..."

 

The Academy, which awards the Oscars annually, has specified in its regulations that in 2027 only performances by humans and scripts written by humans will be eligible for Oscars. However, it added that the use of artificial intelligence will not prevent a film from competing for Oscars. Certainly, technology has been an integral part of film production for many years, and the use of computer graphics (CGI) has been widespread since the 1990s, but all these digital tools enhance work that is always considered human. It's thought-provoking that a well-made backdrop rewards the set designer, while a particularly skilled digital actor cannot bring an Oscar to its creator.

 

I'd bet we've already seen scripts written partly with AI. Surely artificial intelligence is used as a creative sparring partner, for brainstorming, finding variations, correcting rhythm and narrative structure, simulating dialogues, exploring possibilities. So I wonder when a script stops being "human." If 5% of the ideas come from an LLM? If the writer uses AI only to overcome a creative block? If a line is rewritten by ChatGPT but then modified ten times? It seems to me that the very concept of "authorship" is becoming blurred.

 

Certainly, no film is born from a single mind but from the contributions of editors, producers, actors, directors, and crew, so AI could just be the latest entity involved, particularly in the writers' room. The difference is that this time it's not human, and this is precisely what makes us uncomfortable.

 

 

 

3. Learning with AI

 

Anyone with high school-aged teenagers is wondering whether the massive use of AI while studying is helpful or harmful. On one hand, we notice they use it to summarize complex texts, losing the exercise of abstracting significant concepts from a challenging passage. On the other hand, we sometimes discover detailed explanations of mathematics, with dialogues on particularly complex steps that AI illustrates with clarity, metaphors, and more patience than we would have provided.

 

The topic is widely analyzed, and the first MIT research showing reduced brain activation (to the point where students didn't remember their own texts) was followed by a recent study revealing an interesting unexpected discovery. It demonstrated that those studying with AI do so in various ways: some with cognitive delegation, others as study partners. Incredible, isn't it?! The world is nuanced, and once again science tells us the value of a tool lies in how it's used. There is no "AI is good" or "AI is bad."

 

The research involved over 900 people from different countries and clearly showed that as cognitive delegation to AI increases, called "cognitive offloading," learning decreases, but beyond a certain level of offloading, learning becomes enormous. Essentially, there's a type of use based on lazy delegation of certain tasks, contrasted with a massive use of AI in the learning process. This second type activates cognitive vigilance, leading to what the study authors call "pedagogical partnership", with excellent educational results.

 

I imagine that depending on the moment, subject, and teacher's request, new generations delegate to AI in a harmful way or use it as a useful and effective learning companion. I wish these aspects were communicated in classrooms to raise awareness among all involved parties. But above all, I think it's always useful to consider the alternative. The real alternative for many kids isn't the college-educated parent explaining Kant in the evening, nor private tutoring at 30 euros an hour. The real alternative is a teacher handing out concise slides and asking for them to be memorized during an oral exam, to someone who doesn't necessarily have access to economic, cultural, or parental investment resources.

 

Personally, as a professor, I try to push the people I have in class from the first zone of passive delegation to the one where they learn thanks to AI, deeply questioning it and struggling with it.

 

 

Prompt used on Copilot for "Single-panel comic with flat red background. Very rough, messy black line drawing, as if quickly scribbled with a marker. Two simple, abstract human-like figures facing each other: one larger (parent), one smaller (child). Shapes are extremely minimal and irregular, almost stick-figure but slightly distorted. Handwritten, uneven, uppercase text in Italian (not in balloons). Child: "I'M STUPID", Parent: "USE CHATGPT MORE". No shading, no details, intentionally crude and naive drawing, humorous and minimal style, similar to Maicol & Mirco."

 

 

4. AI a bit hard of hearing

 

Three Google researchers published the paper "Prompt Repetition Improves Non-Reasoning LLMs". The thesis is exactly what the title suggests: if you repeat your request twice in the same prompt, models respond better. The experiment was very simple, comparing responses to a <prompt> and a <prompt prompt> where the first text is repeated verbatim. This effect was verified on Gemini, GPT, Claude, and Deepseek, only on non-reasoning models, i.e., those that go straight to the answer without pausing to think. Essentially, you have to "raise your voice," which is a result similar to what other research has demonstrated – ATTENTION – contradicting a common belief: addressing GPT-4 rudely allows for better responses. Again, you have to raise your voice. The explanation for these behaviors could be that AI reads individual words and observes their proximity to others, so repeating the prompt makes it easier for the system to retrieve more useful information to respond. On the other hand, being rude often corresponds to imperative sentences with less ambiguity, causing less disorientation for the AI.

 

But the real issue here seems to be the limited understanding of how these systems work, which, for example, respond differently if you write the context first and then the question or vice versa. Or we notice that in cases of long prompts AI focuses on the first and last parts, neglecting the middle. I wonder if these behaviors are present, to some extent, in many of us, and the idea that some of our cognitive biases have been transferred to a computer system leaves me puzzled.

 

 

 

5. Energy and cooling, from the ocean

 

As an incurable optimist, I've always believed that the growing energy demand of big tech could turn into a decisive push for research, attracting the necessary funds to develop new solutions for sustainable energy production. For this reason, I follow news on the topic closely, and a few days ago, I was struck by an article about a massive investment in Panthalassa, a company with the ambition to build data centers in the middle of the ocean.

 

The idea seems interesting: energy is generated by harnessing the ocean's motion, which causes parts of the system to rise and fall relative to a more stable part, enabling the movement of magnets and coils, resulting in the generation of electricity. Additionally, by bringing the data centers directly onto these systems in the ocean, there is no electrical loss due to energy transport (consumed on-site), and the low temperature of ocean water can be used to cool the data centers.

 

The test of Panthalassa's "Ocean-3" model is scheduled in the Pacific this 2026, with these gigantic autonomous data centers, free of cables anchored to the seabed and without energy transmission to the mainland.

 

 

 

6. True or false?

 

When we generate an image with Gemini or modify it with NanoBanana, we get a file with a small logo at the bottom right, indicating it was manipulated by AI. We tend to crop the image to remove the logo, making it look like a human product.

Until a few weeks ago, I didn't know that Google "hides" a code, called SynthID, within the pixels, allowing AI-generated content to be identified. This technology, developed by Google DeepMind, embeds imperceptible digital watermarks directly into images, audio, video, and texts created with artificial intelligence, making them recognizable as machine-generated.

Unlike traditional watermarks, SynthID is designed to remain invisible to the human eye and resist common modifications like cropping, resizing, or compression. So, if we suspect an image has been processed by AI, we can go to Gemini and ask "tell me if this image is AI-generated and contains SynthID".

Thanks to oradecima (an interesting channel to follow) for introducing me to this "digital watermark." Unfortunately, for now, SynthID is only implemented by Google, but the urgency of having similar tools is now evident, and I trust it will soon become widespread.

 

 

 

7. Robot operAI

 

Artificial intelligence is making robots more capable of adapting to the world around them, but there is a huge gap between what we see on YouTube and the tasks they can actually perform. Humanoid robots run marathons, work at airports, parade at fairs, and perform increasingly impressive public demonstrations. However, we must be careful to note that these are highly controlled scenarios, restricted environments, and deployments carefully crafted for the public. We already mentioned in previous issues that we are seeing the marketing version of robotics, not its operational reality.

 

The robotic applications that could spread much more quickly are those in industrial settings, with automatons moving in warehouses, factories, logistics, and performing dangerous or repetitive tasks. The recent report by Capgemini discusses physical AI, referring to technological solutions that don't remain in software but can interact with the surrounding physical environment, especially through vision, audio, and environmental perception. In short, the report talks about intelligent robots, whose successful applications almost exclusively involve industrial tasks (stacking heavy products, inspection and control, dangerous jobs), more easily delegated to robots due to their greater standardization potential.

 

Certainly, advancements are enormous: for example, a few years ago, my research team worked on a robotic eye dedicated to recognizing imperfections like scratches and stains on hundreds of different products. In 2020, the most obvious solution was "training" a recognition system with thousands of example photos. But this meant having a very challenging collection of defective products and various types of scratches and stains. A few days ago, we tested a Visual Language Model on the same industrial case, which, without needing to see any images "beforehand," was able to distinguish defective products very well.

 

Robots thrive in factories (here, in the BMW iFACTORY)

 

 

8. Something to know: zero-day

 

In recent weeks, I've come across several articles using the term “zero-day.” “Google blocks an attack based on a zero-day,” “AI as an offensive operator: zero-days discovered in minutes,” and many others. In cybersecurity jargon, a zero-day is a software vulnerability that is not yet known to those who should fix it or for which no patch is available yet. The name comes from the idea that developers have “zero days” of advantage from when the vulnerability becomes exploitable or known to hackers.

 

If I install WhatsApp for desktop, Chrome, or another widely used software today, I tend to consider it fairly secure because it's been available for a long time. I imagine that over the years, many researchers and attackers have tried to breach it, and developers have progressively fixed the issues that emerged. But if new AI-powered tools make it easier to find vulnerabilities that have remained invisible until now, then even software considered reliable can turn out to be perfect intrusion channels into our devices. Not because they suddenly became insecure, but because we are becoming much better at spotting cracks we didn't see before.

 

 

ChatGPT for: “A famous app sitting on the doctor's couch, looking confident. The doctor is an AI in a white coat examining an X-ray full of holes. The app says: But I felt perfectly fine.”

 

 

9. Copilot news: finally a personal assistant

 

A recurring scene in recent years sees me sharing at work the latest wonders of the newest AI tool I've tested. It's a common emotion, a mix of "I can't believe it" and "look at this" that you feel when AI pulls out something incredible (I wish for a neologism, halfway between techno-awe and algorithmic ecstasy).

 

The latest of these moments I experienced at AGIC, discussing Copilot Cowork, which we recently accessed thanks to achieving the title of Frontier Partner from Microsoft (let me say it, because we are few), enabling us to use this new tool. Imagine a virtual assistant to which you can ask to use information from any of your documents, chats, meeting recordings, calendar appointments, while also creating Word files, Excel sheets, calendars, emails, after performing a complex plan of actions to achieve the goal you set. In short, I found myself with an assistant, much faster, tireless, capable, and smarter than me, that can work on my data.

 

The idea is similar to Claude Cowork, or Gemini Personal Intelligence, but since I already have meetings on Teams, emails on Outlook, and files on OneDrive, my choice works well in the Microsoft environment. The models available in Copilot Cowork are from both OpenAI and Anthropic, and I truly experienced techno-awe. It seems like a sales pitch, I know, but I can assure you that at the end of the workday, I genuinely felt a great sense of gratitude for having this tool available.

 

 

 

 

 

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 exploring and experimenting with new AI tools available, as well as observing and reasoning about digital evolution.