Artificial Intelligence is a set of technologies that allow machines to perform tasks that typically require human intelligence, such as understanding language, recognizing images, making decisions, and learning from experience.
ChatGPT is one of the applications born within a specific subset of AI, as shown in the diagram.
AI is the general field of study (like physics), within which we find Machine Learning, then Deep Learning, and finally generative technologies like Large Language Models (LLMs), which underpin tools like ChatGPT and Gemini.
Within Deep Learning, two main approaches are distinguished:
Discriminative models: classify data (e.g., "spam" or "not spam").
Generative models: learn from data to create new content.
ChatGPT belongs to this second category and is an advanced example of a generative model: it generates new, realistic, and coherent text, simulating human conversation. Similarly, DALL-E or Midjourney are generative image models.
GPTs (Generative Pre-trained Transformers) are AI models that can generate text, answer questions, translate languages, and write code. They work by processing vast amounts of data to predict words and concepts within the context of a conversation or written text.
At the foundation of tools like ChatGPT are advanced mathematical functions that analyze huge amounts of text to identify patterns and probabilities.
When you type a sentence, the model does not "think" like a human, but calculates the most probable next word based on the data it was trained on. This process is called language modeling and occurs thanks to neural networks, structures inspired by how the human brain works.
Because of how this process works, there are certain limitations to keep in mind to have realistic expectations for these LLMs:
Data Bias: If the training texts contain biases, the AI can reflect them in its answers.
Lack of True Understanding: AI does not "understand" in the human sense of the word, but processes information based on mathematical correlations.
Potential Errors: Although highly advanced, AI can generate inaccurate answers or invent information ("hallucinations").
AI, and GPTs in particular, are not magic: they are like an advanced calculator for ideas—enhancing human capabilities without fully replacing them. They do not think for you, but they help you:
Accelerate tedious tasks (e.g., summarizing text, proofreading errors).
Explore creative avenues (e.g., generating project titles, finding metaphors for an essay).
Improve what you already do (e.g., turning messy notes into organized outlines).
If used incorrectly—such as too early in one's studies or without the proper approach—AI will only widen the gap between those who already have a baseline understanding of a topic (who will become even smarter thanks to AI support) and those who know little about a topic and, by relying entirely on a GPT, feel no motivation to learn or rely blindly on whatever it provides.
Rule 1: Know What to Ask: GPTs work best when you know how to ask precise questions.
Bad: "Tell me about history" → Too vague
Better: "Summarize the causes of the French Revolution in 3 bullet points, highlighting the role of the bourgeoisie"
Rule 2: Have a Solid Foundation: AI does not replace your knowledge; it enhances it.
Example: If you ask ChatGPT to explain quantum mechanics without ever opening a physics book, you risk not understanding it (or worse, memorizing errors). Quoting a famous movie line: "If you have to ask, you will never know. If you know, you need only ask."
Solution: Use it to deepen what you already know and to explore different ways of understanding a concept, not to replace study itself.
Rule 3: Verify and Contextualize: GPTs can invent facts ("hallucinations").
What to do: Cross-reference information with reliable sources (e.g., academic papers, institutional websites).
Ask yourself: "Does this make sense within the context I know?"
To achieve effective results, it is essential to write clear and specific prompts. Here are some tips for structuring effective prompts:
Provide detailed context: Clearly describe the situation or problem.
Define the AI's role: Specify whether you want the AI to act as a teacher, consultant, editor, etc.
Indicate tone and style: For example, formal, informal, technical, accessible.
Specify response format: Such as a bulleted list, paragraph, table, etc.
Use examples: Providing examples helps the AI better understand your expectations.
For more details on how to create effective prompts or for specific use cases, you can consult online resources such as the OpenAI prompt engineering guide or this guide on Best Prompt Techniques. For those who prefer YouTube videos, there is this one by JeffSu.
For a comprehensive guide with many practical examples, there is this Gemini prompting guide in PDF.
For those who would like to delve deeper into the different prompting strategies instead, there is a good overview on promptingguide.ai.
Ultimately, the effectiveness of AI responses largely depends on the quality of the instructions provided. The motto often used in software, data analysis, and UX research applies here too: "Garbage in, garbage out"—poor inputs produce equally poor outputs.
Another useful resource provided by OpenAI: video guides covering basic to advanced usage.
Exam Simulation and Practice Tests
You can use an LLM to generate questions based on a study syllabus or a specific set of topics provided to simulate exam sessions or create customized practice tests:
"Given the attached exam syllabus [or: 'which covers the following topics: ...'], formulate an open-ended question based on the style and type of questions typically asked in first-year university [Subject] exams in Italy."
Personalized and Simplified Explanations
When encountering complex or unclear concepts, ask the LLM to provide explanations tailored to your level of knowledge and background:
"I am a high school senior studying special relativity, but the relationship between time dilation and length contraction in the equations isn't clear to me. Could you explain it intuitively at my level of knowledge, assuming I don't remember all the prerequisites needed to understand it in mathematical detail?"
Summaries and Syntheses of Lectures, Texts, or Multimedia Materials
LLM-based tools can transcribe and automatically summarize audio recordings of lectures or lengthy text documents in seconds.
Important Note: Although these tools are extremely efficient for quickly obtaining summaries, personal reworking, handwriting, or actively typing notes significantly enhances long-term retention and fosters a deeper, more critical understanding of the subject matter. Automated tools should complement, not replace, active study practices.
Quick Generation of Formal and Professional Texts (Emails, etc.)
Leverage an LLM to quickly draft emails, letters, reports, or proposals from basic pointers:
"Draft a formal email to Mr. Rossi at XYZ Company asking for detailed information about product 'Alpha'. Include a brief introduction of our company [Your Company Name] and specify that we are interested in a quote for [quantity or specific need]. The email should be polite and professional."
Rewriting and Refining Existing Texts
An LLM can adjust tone, style, or structure to fit a new context or improve clarity:
"Correct any grammatical and punctuation errors in this text and suggest stylistic improvements to make it more concise: [insert text]."
Automating Repetitive Office Tasks
Integrating LLMs into productivity tools can automate generating summaries of long email/chat threads, drafting meeting minutes from transcripts, or organizing task lists.
Use case example:
After a long online meeting, you might want to get a summary of the key points and decisions made, without having to listen to the entire recording. Alternatively, you might have a long email chain from which you want to quickly extract requests or relevant information.
In these cases, you can leverage integrated tools like those available in Microsoft Teams, which offer automatic summarization features and comply with corporate privacy policies. On the other hand, when sensitive data is not involved, you can also consider using AI-based solutions by uploading the recording to an external application.
In any case, it is essential to be well-informed and choose the most suitable option, always carefully evaluating aspects related to information confidentiality.
Improving Written and Oral Skills
Use advanced chatbots to practice writing in a foreign language, request corrections, or use voice interaction features for pronunciation and conversation practice:
"I would like to practice speaking about 'environmental impact' in Spanish. Ask me questions on this topic and correct my mistakes as I respond."
Contextual Translation and Understanding Nuance
LLMs excel at handling context, idiomatic expressions, and tone compared to traditional translators:
"Explain the meaning and origin of the Italian phrase 'In bocca al lupo!' and how to translate it into English while maintaining a similar meaning."
Creative Text Generation
LLMs excel at generating short stories, poems, scripts, song lyrics, ad copy, or social media ideas:
"Generate 5 creative ideas for an Instagram post series promoting a new artisanal coffee shop in downtown Milan. Include relevant hashtags."
Text-to-Image Generation
Generative AI-based tools, such as DALL-E, Midjourney, or Stable Diffusion, allow you to transform textual descriptions (prompts) into unique and often surprising images. This opens up massive opportunities for illustrators, graphic designers, marketers, or anyone who needs original visual elements without having to create them from scratch manually. A ranking of image generators is available on HuggingFace.
"An Impressionist-style digital painting of a lavender field under a starry sky, with a single oak tree at the center."
Multimedia Format Conversion
Many AI-based tools facilitate conversion between different types of digital content. OCR (Optical Character Recognition) allows you to extract text from images or scanned PDFs. Automatic transcription (Speech-to-Text) converts audio from recordings or videos into written text. Speech synthesis (Text-to-Speech) transforms text into audio, which is useful for creating podcasts, audiobooks, or voiceovers.
Use case example:
You have an old scan of a document and need to extract the text to edit it. You have recorded an interview and want the transcription. You want to turn one of your articles into a podcast episode.
Summarizing and Understanding Voluminous Documents
LLMs can quickly read and summarize long reports, research papers, or contracts, extracting key points or answering questions from large datasets (e.g., filtering candidates from an exported Excel/PDF sheet).
Use case example:
You have a very large PDF file, perhaps a list of candidates exported from an Excel spreadsheet, which is difficult to re-process manually. Instead of copying and pasting everything by hand, you can provide the file to an artificial intelligence system and ask, for example: "How many candidates are there with a score higher than 80 in the Lombardy region?" In just a few seconds, the AI can extract the relevant data and provide you with useful analysis or summaries, even from complex and loosely structured documents.
Extracting Structured Information from Unstructured Text
LLMs can analyze free text to identify and extract specific types of information, organizing them into structured formats like tables or lists. This is extremely useful for processing data from emails, notes, articles, or other unformatted documents.
"Extract all dates, times, and locations mentioned in the following notes and list them chronologically: [insert notes]"
Assistance in Spatial Organization and Geocoding
Although they are not full GIS tools, LLMs can help process lists of addresses or locations. They can sort them based on simple criteria (such as alphabetical order) or, when integrated with external services, help format the data for geocoding or for creating files compatible with mapping software.
Example use case:
You have a PDF document with a long list of stores spread across different locations. If you want to display them on a map — for instance, for a website or an application — you need the geographical coordinates (latitude and longitude) for each address. Doing this manually by looking up every single spot on a map would take an immense amount of time.
Instead, you can hand this task over to an AI system: you feed it the list of addresses, and it hands back the geographical coordinates for each entry. This way, you can easily import them into your preferred mapping program or service and generate your map in just a few minutes.
Code Generation and Writing Support
LLMs can act as programming assistants by suggesting code completions, generating snippets for specific features, or even creating drafts of functions and entire scripts based on natural language descriptions. This can accelerate the writing of boilerplate code or help quickly implement common patterns.
"Generate HTML and CSS code for a simple contact form with fields for name, email, and message."
Code Explanation, Debugging, and Optimization
When faced with unfamiliar code, hard-to-find errors, or when trying to improve performance, an LLM can provide detailed explanations of how a block of code works, suggest potential causes of bugs, or propose ways to optimize code efficiency and readability.
"I have an 'IndexError: list index out of range' in my Python code. Here is the code [paste code] and here is the traceback [paste traceback]. Can you help me understand why this error is occurring?"
Generating Documentation and Commit Messages
Automating the creation of documentation (such as docstrings or comments) for functions or modules, or generating meaningful commit messages based on code changes, can improve project maintainability and team collaboration.
"Write a concise and descriptive Git commit message for the following changes: [describe changes or paste diff]"
Important Note: Avoid "Vibe Coding" and Cultivate Understanding
Artificial intelligence is a powerful tool, but blindly relying on its solutions — a practice known as vibe coding — can be harmful. Copying generated code without understanding it can lead to inefficiencies, bugs, or vulnerabilities, as well as slow down your growth as a developer. Use AI as an assistant, not as a replacement. It is useful for:
• Automating repetitive tasks (e.g., boilerplate, simple scripts)
• Learning new concepts or clarifying code
• Exploring alternative approaches
• Supporting debugging (while always verifying independently)
Do not run code that you do not understand. Test everything rigorously and ensure it meets your functional and security requirements. True expertise lies in the ability to think, design, and solve problems — skills that develop with practice and deep understanding.
Finally, pay attention to privacy and security, especially when working with proprietary code and cloud services. Online developer communities are an excellent resource for learning how to integrate AI into your workflow mindfully.
Bearing in mind that online conversations with AI will never be 100% private (due to analysis, various profiling, or further training), they are still valuable tools to use for free. Paid versions are indeed available, but unless we need the absolute latest model or infinite messages for specific purposes, the websites mentioned below are more than enough to assist us in our daily activities. In fact, the cost of LLMs is decreasing significantly over time as more efficient models are created and investments in large data centers are realized (example).
If we plan to use these models for work, we must be very careful because we could leak confidential information online: therefore, it is better to use corporate AI (if provided by the company we work for), use them without inputting company data, or run a local AI.
We must use these GPT models responsibly to safeguard the environment, as these applications use a lot of energy and cooling water both to be trained and to be used (compute-intensive applications) - here is an article. Therefore, let's ask ourselves if we really need to use AI to delve into a concept or if we can do it ourselves with online searches.
The various LLMs created by different companies are tested using different benchmarks — a guide is available at this link. Leaderboards are then compiled for these models to understand which one is best for our intended use. Some of the most widely used are Vellum.AI, llm-stats & lmarena.ai.
ChatGPT by OpenAI: The app that kicked off the artificial intelligence hype. In the free version, the data is updated to 2023, but this can be compensated for by activating the online search feature (allowing the model to learn from online results in real time). Even for the free version, limited messages are available to request basic reasoning to improve the answer.
DeepSeek: It caused a stir for reaching the same levels as ChatGPT at a fraction of the cost (example source). It offers reasoning capabilities and online search.
Qwen by Alibaba Cloud: An open-source model with specialized versions for mathematics, coding, and multimodality. Its Qwen2.5-Max variant ranked first in math and coding on Chatbot Arena, demonstrating complex reasoning capabilities. It offers reasoning capabilities and online search.
Perplexity: • It stands out from the others for its "deep research" feature, which compares multiple sources and reworks data thoroughly. Excellent for complex research, fact-checking, and analysis of intricate information.
Claude di Anthropic: "Constitutional AI" that rejects unethical outputs, ideal for academic research and summarizing sources. Focused on safety and reliability, it is known for its natural response style and ability to analyze long texts with consistency.
Mistral: European excellence, it is a French-Spanish open-source model optimized for Romance languages. An open-source model that offers advanced performance and is particularly appreciated for its low resource consumption compared to competitors.
Gemini by Google: Advanced integration with the Google ecosystem and multimedia processing capabilities. Integrated searches, technical support, and general questions with updated sources.
Artificial intelligence is no longer limited to cloud services like OpenAI or Google; it can also be run locally, enabling greater privacy, data control, and, in some cases, reduced costs. Although running AI models locally is possible on personal PCs or servers, it is important to understand how they work and what resources are required to optimize performance.
AI models can be executed on both CPUs and GPUs. However, while the CPU is more versatile and suited for light workloads or less intensive tasks, the GPU (Graphics Processing Processing Unit) is significantly more powerful when it comes to processing complex or large-scale models. GPUs are designed to handle parallel operations, making them ideal for machine learning and significantly accelerating inference and training times compared to CPUs. Therefore, if you want a faster and smoother experience when using local AI models, it is highly recommended to use a GPU, which can substantially reduce processing times.
A practical example of using AI locally could be running Open WebUI alongside Ollama to summarize academic articles offline. Alternatively, a company could automate its customer service using n8n paired with local AI models, avoiding the costs associated with cloud APIs like those from Google or OpenAI.
There are many online resources, including YouTube guides and specialized forums, that explain how to configure and use these tools. Below is a general overview of local AI, which includes the technologies, the necessary hardware resources, and the most popular tools for carrying out machine learning projects directly on your own devices.
Ollama is the backbone of many local solutions. It allows you to run open-source language models (such as Llama, Mistral, or Phi3) directly on your computer without an internet connection. Its integration with tools like OpenWebUI makes it accessible even to those who are not code experts. Without a Front-End like OpenWebUI, you can use Ollama from the command line as explained here.
OpenWebUI transforms Ollama into an experience similar to ChatGPT, but entirely local. It offers an intuitive web interface with advanced features, like:
RAG (Retrieval-Augmented Generation): Upload documents (PDFs, articles) and query the AI model to get answers based on their content.
Voice Chat: Communicate with the AI via voice, which is useful for taking notes during meetings or lectures.
Multi-Model Support: Use different LLMs in parallel to compare responses.
• Real-Time Web Search: Add updated data to conversations without leaving the app.
It can be installed in two ways:
Via Docker, following the official documentation (a video guide is available here).
Via Pinokio, a platform that simplifies the installation and management of local AI models by acting as a one-click hub for various artificial intelligence applications. (Note: Because Pinokio runs automated scripts to install third-party software, it is highly recommended to stick only to verified scripts from trusted developers to ensure system security.)
AnythingLLM specializes in interacting with files and databases. You upload a document (e.g., a company report), and the AI analyzes it to answer specific questions, leveraging RAG and local models like Ollama. It is particularly useful for researchers or professionals working with large volumes of sensitive data.
n8n is a low-code automation platform that, when combined with Ollama, allows you to create intelligent workflows. n8n is perfect for those who want to integrate AI into business processes without relying on external services. Some examples of use include:
Analyzing social media trends using local LLMs.
Generating automatic reports starting from data stored in databases like Qdrant or PostgreSQL.
For those looking for a less technical alternative, Msty eliminates the complexity of Docker and terminal usage.