AI tools I've been using

and so should you

Hi friends,

AI is the magic wand of these years: every company, every professional says they’re using it. If this is true, if this is not only marketing I don’t know it, but I know that is actually easy to convince a non-IT client that using the last magic of digital technology its business can fly high.

At the moment, AI is used as generator of marketing advertising, text, to manipulate or create video and voice podcast: I use one tool for my new Wine is Not Real podcast, as you know. That is Generative AI (Gen AI), tools and platforms AI-based that create something or modify something existing, like a picture or a song. The AI tool and chat you use are all LLM based, Large Language Model. This is what the name says: a large model for language, and as all Gen AI is a statistical model: using the books and posts and articles and documents on the Web, a LLM generates the most statistically probable word that follows another one. AI, or LLM, actually don’t understand anything what you say it: AI generates word statistically correlated. Do you know that Christmas lunch when your brother-in-law talks about things that don't interest you and you respond with ‘ah yes’, ‘of course’, ‘mmh ok’?

Well, an LLM does much the same.

Most interesting use of AI is improving Machine Learning (ML) systems, their ability to identify a model from a huge bucket of data. There is a difference between AI and ML, we can say that ML is a subset of AI.

AI is useful to recognize a face, a landscape, a text. If you have doubt that a document is been written from a certain author, you can compare it with another you’re sure about, and AI can say you if these documents are been written by same hand or not. AI can also quickly solve problems and help humans to make decisions. To do this, AI needs patterns and models created by ML, analyzing peta and peta of bytes (1 petabyte = 1000 terabytes: your external hard disk is, likely, 2 terabytes) and identifying models to make insights and previsions, compare them and choosing the best one. Shortly, AI allows a machine to simulate human intelligence to solve problems, ML allows a machine to learn autonomously from past data. You can read also this post: please, let’s note is a 2016 post.

Well, I hope you are not more confused than when you started reading; so, come back to the topic: AI tools I’ve been using.

As I told you, the most simply use of AI tools is to generate something: post, video, advertising. So, here the list tool I’ve been using, not all ones existing, and a warning: ever check the result of the AI, don’t blindly faith of it. At the end of this post I’ll write about AI hallucinations.

ChatGPT

Of course it was the first tool I’ve been using, the first AI released to normal people, simple customer, social user and so on. First version was ChatGPT 3.5, 2023, May, and could read Internet source only until 2019. I used it to discover a new world and my first question was: Do you think God exists? It was not so precise in its answers, I checked it with any question about wine. The last version is ChatGPT 4o, with improved memory capacity for more efficient conversations, it’s capable of accepting various forms of input including text, images, audio and video, and is trained on data until June 2023, with more up-to-date knowledge than GPT-4.

The paid version costs 20$/month and it offers access to DALL-E generating images tool. The main differences between ChatGPT Free and Plus are related to usage limits. Users of the free version will have to comply with restrictions on the number of messages sent, while ChatGPT Plus subscribers will have much higher usage limits, about five times higher than free users. Furthermore, once the GPT-4o limit is exceeded, Plus users will continue to have access to the GPT-4 model, while free users will automatically switch to GPT-3.5.

Uses

I use it to generate a list of topics for my blog. I can also upload pdf documents to create a summary. I use the for free version to help me in finding link about some arguments I write on my blog or in this newsletter. I asked GPT to create Python code but it wasn’t good enough. But it can to correct syntax of a long listed program, or perform repetitive operation like to add a menu row to the menu bar of a web site.

Copilot

This is the Microsoft’s AI, as a replacement for the voice assistant Cortana, and it’s based on GPT4 by OpenAI (Microsoft funding OpenAI with $10 Billions in 2023). You can access to Copilot free from Bing, the MS search engine, or Edge, the MS browser. You can generate image with a simple prompt in a squared, landscape or portrait format. You can’t upload a pdf document and ask to create a summary, you can only copy&paste a text into the prompt bar. And you can’t upload a webp image, only jpeg. In the Pro Version, 22$/month, subscribers can use it in free Web versions of Word, Excel, PowerPoint and Outlook2. Copilot Pro is based on a subscription to Microsoft Copilot, which uses OpenAI's premium GPT-4 model.

Uses

I use it only to create images for my posts, nothing else. It hasn’t many advanced features, but its answers are more human-like than chatGPT. Actually, I don’t use Copilot much, but it can be a good way to create images for blog posts.

Perplexity

I am fairly new to using this tool, I discovered it recently and immediately liked it because it produces visible results. It was founded in August 2022 at San Francisco (CA) and it uses chatGPT, Llama3 (the Facebook’s AI) and Claude (by Anthropic). It is a search engine using the context of the user queries to provide a personalized search result, producing a text with inline citations. In the Pro version, it ask user to refine the question to provide more precise answer.

Uses

I use Perplexity to compare two or more documents, for example two different reports of the same topic, comparing one year with the previous one or two report from different agencies. It produces a list with differences and similitudes between these two (or more) documents, and a series of questions in FAQ format, very useless for an informational blog (or a newsletter post, of course!). You can upload documents in Italian (or other languages like French) and you’ll have ever the answer in English; I didn’t try with documents in other languages than Italian and French. It can provide the links to other sources and it works good enough, so I have to do only a little check job.

NotebookLM

This is a relatively new born in the AI tools scenario, being born in 2023, September and deployed this year. Is the gen AI from Google, and is Gemini-based, the Google’s LLM.

Uses

I use it to create bullet summaries of my long-form posts, so my customers can quickly to know what I wrote, or create a series of FAQ. But the most interesting use is to create a virtual conversation between two people, a woman and a man, in English language, based on a document or a text i can upload into the dialog bar. It’s the way I’ve been producing my podcast Wine is Not Real, you can listen here on this newsletter or on Spotify. It’s an experiment, the conversation is not so brilliant but I think it going better.

Midjourney

I use it to create nice images, like that one you see at the top of this post. I like its style, better than Dall-E in my opinion, and if you have any practice of photography you can build a well-formed prompt ad, in consequence, a better image. In the Pro version it costs 7$/month.

Conclusions

At the beginning of this post I’ve told you about hallucinations. These are just what the name is: invented answers by the bot when it does not know exactly an answer. In short, instead of saying: I don't know, the AI prefers to say something inaccurate or completely false or invented. Let me tell you an anecdote. Last year, a lawyer went before the judge with a series of judgments and citations to prove that his client was innocent. The judge, impressed by all these citations, took a few hours to study them and immediately realised that they were completely false, they had all been made up. The lawyer later admitted that he had had his argument written by an artificial intelligence without checking the result. You can imagine how the case ended, and for all the details you can read about it here.

So, when you use an AI to do anything, write a post, create an image, always check the result. It could be beautifully written, but be all fake.