Artificial intelligence

José Tafla · July 3, 2026

(Updated: Aug 8, 2026)

Image source: Nano Banana

In the early 1990s, when computers became personal, popular, and mainstream, critics said they would kill jobs and displace workers.

As of today, artificial intelligence (AI for short) has become mainstream. Now what?

Disclaimer

I’m no expert. Merely curious. A hobbyist. A dilettante. Nosey.

This article is based on the little I know, the research I did, and my conversations with people wiser than me. Much, much wiser. Compared to them, I’m insignificant.

But they inspire me to learn and to pass it on, for knowledge is to be shared. To teach is to learn.

And I’m an eternal learner.

In its current form, this article is more like a blank canvas with a few pencil markings. It’s begging to be painted. And for that, I need your help. As of this writing, the article is shallow and focused. Please reach out to me to say how deep and how wide I should go. I sincerely thank you.

Micro and personal computers

Microcomputers had existed for a long while, but other than a few specialized ones, they all had a text interface. For most of them, you would turn it on, insert a floppy disk, and see a familiar prompt on the screen: A> followed by a blinking cursor. Next you type a command and see it in action. For corporate use, that was good enough, where you would launch a word processor, electronic spreadsheet, or relational database manager.

A few companies, such as Apple, Atari, Commodore, and Tandy, among others, added graphical (or semi-graphical) capabilities, which made it easier to use certain programs (but not necessarily launch them). For the home user, they had become gaming systems.

Even the IBM PC, in its initial stage, only had a text interface, albeit with a more sophisticated prompt: C:\>. Although its graphical capabilities were far superior than its predecessors, which allowed software vendors to publish more user-friendly products, supporting functions such as image editing and desktop publishing, it all started with a prompt. After Microsoft released Windows 3 in the early 1990s, that’s when microcomputers really became personal, especially because they could finally be considered easy to use.

Job killer

Critics were saying personal computers would kill jobs and displace workers. Those opponents were only partially right: personal computers killed professions. For instance, secretaries were no longer needed. The boss would give them a letter to type on a typewriter, have them schedule and change meetings, or make phone calls. With a computer on the desk, the boss could then use a word processor to write letters and memos, use a calendar and scheduling software to manage their own schedule, and as for the phone calls… have the phone operator do it. With time, secretaries became executive assistants, serving more as gatekeepers to executives.

Other secretaries, throughout their career, had learned the job of their bosses. Some corporations converted secretaries into account managers, office managers, and team leaders, competing against their former bosses in an even field. Surely enough, the profession of a secretary disappeared.

As for displacements, in the beginning there were quite a few, but suddenly organizations learned they needed different kinds of people, and as a result new professions were created and the workforce grew. Only those who could not (or didn’t want to) adapt to the new reality were displaced.

Artificial intelligence

The concept of AI has existed for a very long time, but it came in and went away in waves, either due to scalability, pricing, or lack of interest altogether. I remember playing with ELIZA in the mid-1980s. It’s a program developed in the mid-1960s¹ that used pattern matching and substitution to give an illusion of understanding, but its responses, other than the curiosity of it, were nothing to write home about.

Over the following years, we find LISP and Prolog, two programming languages that could be easily used for AI. We also find the animal guessing game, which asked you to think of an animal and then ask you yes/no questions about it; if it could guess it correctly, then it won, otherwise it would ask you the animal you were thinking of and what differentiates it from the one it guessed. It’s merely an ever-expanding decision tree, but it’s an early form of machine learning, where once a mistake is made and identified, it never repeats again.

There’s also SHRDLU. Imagine you have some being in a room full of colored objects. Using natural language, you ask the being to move the pieces around, for example, “Put the red pyramid on top of the green cube.” If you would then ask it to move the green cube elsewhere, it would tell you it’s removing the red pyramid first. If you would then ask it why it was doing it, it would say, “Because it’s on top of the green cube.” The little prank that Terry Winograd, the creator of the program, plays on you is that, if you ask it why the pyramid was on top of the cube, it would respond, “Because you asked me to.”

We can’t obviously forget the multiple computer-based board games where you’re playing against the computer. We’re talking Chess, Checkers, Tic-Tac-Toe, Reversi (or Othello), and others. The underlying logic is not an algorithm, but a heuristic², such as Minimax, which also belongs to the AI universe.

The leap forward

Great ideas come to fruition when there is a breakthrough. A significant milestone was IBM Watson, which proved it was possible to talk to an AI model and expect intelligent responses, sometimes even funny. The problem was that it required an enormous amount of power-hungry servers³, something no home user should be willing to consider. For them, the real breakthrough happened just a few years ago⁴, when ChatGPT came into existence.

Simulating intelligence

Computers, as the name implies, compute. Indeed, all they see is numbers. Let’s pretend that, when looking at a particular place in memory, they see the number 65. For the computer, it’s merely 65, but the interpretation of 65 depends on the context:

Inside a program, it could be an instruction for the computer to copy some value from one place to another.
Inside a text file, it could be the uppercase letter A.
Inside an image file, it could be a 25.38% shade of green.

Computers perform calculations pretty fast. That’s what they do best. Everything else they do is a consequence or subproduct. Text and image processing are challenging for them; they do it, but slowly.

I want the computer to do something for me: create a report; find some information somewhere; draw a picture; give me the directions to go somewhere; you name it. But I want to ask the computer in my own language. No SQL statements. No spreadsheets. No online search. In plain English, I want to tell the computer (either typing or speaking) what I want and I let it do its thing. That’s natural language processing, a branch of AI.

Once the computer understands what I’m asking, it will perform a complex search. If you’re running your own AI model, the search is limited to the boundaries you specify, but if you’re using a public model (ChatGPT, Claude, or so many others), it will do a search online. And then, very quickly, will give you the result you’re looking for.

How is this possible? Computers are slow at processing text, but they still manage to search through millions of documents online and give you the answer in a matter of seconds. Hmm…

The science behind it

We want the computer to search inside volumes of unstructured data. It’s not, for example, about telling me the average time that a given employee clocked in at work so far this year, for this could be easily translated to a SQL query into a database. Instead, we want the computer to find what we want inside text files, social media posts, audio recordings, images, videos, and anything that stores data in an unstructured way.

In computer science, a vector is a list of numbers arranged in a specific order. This can mean anything you want, but for the computer it’s merely a list of numbers. Let’s create a convention and make vectors meaningful. A vector with a single element, for example $[5]$, for us, is going to represent a point on a line; with two elements, for example $[3, -0.8]$, represents the $x$ and $y$ coordinates of a point on a plane; and with three elements, like $[-1.25, 3.14, 6.55]$, represents the $x$, $y$, and $z$ coordinates of a point in space.

Let’s now define that, for every point represented by a vector, there is a corresponding point at the origin ($[0]$, $[0, 0]$, or $[0, 0, 0]$). If we connect both points, we have a line that starts at the origin and ends at the point represented by the vector.

Vectors are not limited to three elements, though. We could have more than 3 dimensions. Us humans may have difficulty visualizing a 4th, 5th, or 6th dimension, but for the computer a vector is still merely a list of numbers.

If we have two vectors $a$ and $b$, representing lines connected at the origin, we can calculate the angle $\theta$ between them by using the dot product:

$a \cdot b = |a| \cdot |b| \cdot cos \theta \Rightarrow \theta = cos^{-1} \left( \frac{a \cdot b}{|a| \cdot |b|}\right)$

Steps:

Compute the dot product: $a \cdot b = \sum_i a_i b_i$
Compute magnitudes: $|a| = \sqrt{\sum_i a_i^2}$, $|b| = \sqrt{\sum_i b_i^2}$
Compute angle: $\theta = cos^{-1} \left( \frac{a \cdot b}{|a| \cdot |b|}\right)$
The result is in radians. If you want to convert to degrees: $deg = rad \cdot \frac{180}{\pi}$

Here’s a Python⁵ function to calculate the angle between two vectors:

from math import acos, sqrt


def vectors_angle(a: list[float], b: list[float]) -> float:
    if len(a) != len(b):
        raise RuntimeError("Vectors must have the same dimension")

    def magnitude(v: list[float]) -> float:
        return sqrt(sum([n * n for n in v]))

    product = sum([x * y for x, y in zip(a, b)])
    return acos(product / (magnitude(a) * magnitude(b)))

As an example, the angle between vectors $[1, 15, 23, 58, 69, 101, 103]$ and $[1, 24, 38, 87, 89, 100, 111]$ is approximately 0.1591 radians, or nearly 9.1°.

Large language models

Let’s build a practical vector, where each element represents, in ascending order, the position, inside a document, where a given word can be found. For example, in this article, the word computers can be found in position 6 and so many others. That’s how we would build a vector for computers in this document. Our AI model will then check all its vectors for computers, which were created for all the documents within its reach. That’s an enormous amount of vectors.

Let’s now agree that, for two documents to be considered closely related, the angle between the vectors must be 20° (approximately 0.35 radians) or less. Because the computer is performing numerical computations on lists of numbers, the result comes up quickly. This is only for finding related documents; there are far more rules the computer needs to follow. More on that to come later.

It’s easy to infer that, in English itself, there are lots and lots of unique words, even if you ignore the declinations. Major dictionaries can have in excess of 200,000 words. Also, there could be millions, if not billions, of documents made available to AI models. If we are to track vectors for each word in each document, the number of vectors is gigantic, hence the name Large Language Model, or LLM for short.

Battling AI

Don’t! You’ll be wasting your time and energy, and at the end you’ll lose anyway.

In high school I learned to use a slide rule, while offices were generally using mechanical and electromechanical calculators, but soon came electronic calculators and everything else became part of history. With the arrival of personal computers, tablets, and smartphones, even handheld calculators were relegated to becoming museum pieces.

In recent months, I was talking to a professor at a renowned university. He said he would give complex home assignments to his students, who would turn in high quality papers very quickly. Sadly, he had to score his students very low in tests and quizzes. How is that possible? Simple: students were using generative AI to do the home assignments for them. There’s an article on LinkedIn about a similar case at Brown University⁶. Truth is, just like computers in the 1990s, AI is going to make the dumb dumber and the smart smarter. Those who use AI to do their jobs will fail, and those who use AI to improve on and speed up their jobs will succeed.

Another example comes from Stefan Michel, Dean of Faculty and Research at IMD in Lausanne, Switzerland: “Claude will replace developers, just like Excel replaced accountants.” In my humble opinion, Claude can replace bad developers, but Claude will allow great developers to become even greater developers.

It was Dr. Isac Artzi, Associate Professor of Data Science, Artificial Intelligence, and Game Development at Grand Canyon University and Rabbi Y. Dovid Weitman who planted a gigantic splinter in my mind. AI is a tool, and nothing but a tool. In comparison, knives and other cutting objects can be used to save lives, treat and cure diseases, feed the hungry, and kill and injure others; it all depends on who’s wielding them. For teachers and educators, the challenge is not to preclude students from using AI, but to inspire their students to learn the basics and then have them use it to achieve higher objectives.

Perfect is the enemy of good

Have no fear of perfection – you’ll never reach it.

– Salvador Dali

Throughout my career, I’ve always followed the 80/20 rule, also known as Pareto principle, which states that roughly 80% of the outcomes come from 20% of causes. This rule can be applied in multiple places. For example:

80% of users use 20% of the capabilities of any program
When defining a minimum viable product, product owners should consider reaching 80% of their audience by implementing 20% of the functionality

Does this mean that, to achieve the remaining 20%, it’s going to consume 80% of the effort? Not really. As part of any new product development, certain quality levels must be agreed upon beforehand. In certain instances, 80% could be acceptable. Keep in mind that, the higher the quality standard, the more effort must be put in. Eventually it will come to a point where the cost of higher quality becomes prohibitive.

The same happens with AI models. It should be clear, by now, that AI models don’t think: they calculate. They look up their sources based on mathematical parameters, apply rules, repeat this process as much as needed, and compose the output. More sophisticated models can review the output and follow more complex rules.

But for that, there is a cost.

It will take longer for the result to be presented. Can you wait?
If you want a better model, then you have to purchase a plan.
- Certain models even tell you how much a given operation is costing you.
If you’re not paying for a plan, then forget privacy; your information has suddenly become public. If you’re using a third-party product for free, then you’re their product and your data belongs to them. Just read the fine print before you sign up for something.
- A Google search costs about US$0.0028⁷, but you don’t pay for it. Who does?
- A Google search consumes about 0.0003 kWh of energy⁸, roughly the amount needed to power a 60-watt light bulb for 17 seconds.
If you want your interaction to be absolutely private, then you’ll have to supercharge your computer. Remember that LLMs must build those vectors, using a lot of memory. Think about 128GB RAM or more; hard drives and solid state drives have much more capacity, but they’re significantly slower.
- It will also consume a lot of energy. You can expect a spike in your utility bill.

At the end of it, the final output could still be wrong.

AI is a tool, not a replacement

I saw this posting on LinkedIn discussing differences between vibe coding and software engineering. In my humble opinion, this picture says it all:

Vibe Coding vs Software Engineering explained

The building designed by AI (on the left) may even look better, but the foundation on the right is extremely solid; it will withstand all sorts of storms, while the one on the left will crumble on the next opportunity.

Some reasons for failure

This is not a comprehensive list.

AI models use existing contents. Some contents may have been created by scholars, investigators, or people who did some thorough research. These authors are entitled to an opinion, so their publications could be biased. Other contents were created by people with no knowledge, who are either voicing their own opinion or merely repeating what others said. Others sit somewhere in the middle. Do AI models know how to distinguish between one and the other while doing their search?

AI models have no common sense, neither a soul. If you look inside a computer, you’ll be able to find its heart (power supply), brain (processor), mind (memory), but not a soul. Indeed, people have said that contents created by a model may even look good, but lacks soul. Models can’t distinguish between right or wrong; training can improve on these behaviors, but it’s virtually impossible to reach perfection.

With that said, some models do have a filter. They may not remove watermarks from pictures because it’s copyrighted material, while others may refuse to answer an improper request.

AI models make mistakes. Does this come as a surprise? A recent example of mine includes this very article. I asked a given model about angles between vectors, along with some examples. To validate my Python function above, I checked my result against the one provided by the model, and the results differed! After further verification, I realized the model had miscalculated it. Go figure…

What can you do?

Use common sense! AI models are there to help you get a better answer quickly. Better. Not perfect. If you understand the topic of your question, then make sure the answer is consistent with what you were expecting; if you don’t, ask someone who does. You don’t become an expert in any matter just by using AI.

Check the sources! I don’t know everything. To write this article, I did a number of online searches (most of which provide an AI-generated summary) and queried a number of models, and then I checked the sources to make sure the results are correct. If the sources are listed on the response, that’s great, otherwise I ignore the response altogether.

Some people still believe computers are supernal beings. Whenever I call customer service and explain my problem, occasionally they say, “But the computer says otherwise!” When it happens, I say, “The computer is not omniscient.” For me, the worst part is when they say, “This is automatically calculated/corrected/adjusted/etc. by the computer and there’s nothing we can do.”

Computers don’t make mistakes. They merely do as they’re told. If there’s an error, it’s usually due to hardware or network failure, bad or incomplete requirements, poor programming, inefficient testing, incorrect data entry, or misinterpretation of data.

Or all of the above.

Improve! Your favorite AI model gave you an answer. You verified it ✅. You checked the sources ✅. It made sense ✅. Fine tune the question to address ambiguities or enhance areas that are not quite right or that deserve further exploration. It saved you hundreds and hundreds of hours. Learn the answer so you can talk about it, justify it as needed, or enhance it on your own. Don’t just copy it verbatim, lest your grades, career, and reputation are at risk.

Revision history

2026-07-03: Original posting date.
2026-07-10: Corrections. Battling AI.
2026-07-17: AI ain’t perfect.
2026-08-08: Corrections and additions.

References and notes

Wikipedia: ELIZA ↩
An algorithm contains a number of steps that will take you to the solution of a problem, or to the firm conclusion that there is no solution. A heuristic, on the other hand, only takes you closer to a solution, but without any guarantees; it can be implemented as a decision tree, where the computer evaluates the impact of making a given move. ↩
Wikipedia: IBM Watson ↩
Wikipedia: ChatGPT ↩
To make matters easier, Python has a function math.degrees that converts radians to degrees. You don’t have to code it yourself. ↩
For those who can’t read Portuguese, here’s an article in English. ↩
The Technium: How much does one search cost? ↩
RW Digital: How Much Energy Do Google Search and ChatGPT Use? ↩

Share: Twitter, Facebook