Introducing a device, it’s called AI

This is AI, it’s a device. Above is an image of a Neural Processing Unit, which is a computer chip. It is the brain of an Artificial Intelligence processor. At this time, it is circuitry and programming. When we spend countless hours assessing what Artificial Intelligence has done, good or bad, we forget it is just this, a device. You have to run these devices in series to make them work, and they are not the data that speaks to us; that is in databases accessed by machine learning routines. Let’s examine the reality of an Artificial Intelligence processor and how society has created a false narrative about its ability. 

Defining Artificial Intelligence

The current definition is
According to Merriam-Webster, artificial intelligence (AI) is defined as the capability of a machine, computer system, or algorithm to imitate intelligent human behavior. It is also defined as a branch of computer science focused on simulating this intelligent behavior.

The dictionary can be wrong. Every dictionary company employs hundreds of people to assess definitions all year long. Artificial Intelligence is not in any way a computer system, nor is it a machine. The NPU is a device that cannot function on its own, and it is not a computer; it is a chip. Let’s examine what makes up an Artificial Intelligence system. 

For a Machine Learning system to process as an intelligent network, it must have several interrelated components. 

•  Database
• Machine Learning System
• Artificial Intelligence Program

The Database is built on a large store of data and is stored on servers. The data in the database is alphanumeric, meaning made up of symbols, numbers, and letters. The Machine Learning system must be programmed to retrieve those Alphanumeric bits of data in a certain way, such as words or strings of numbers; these are called parameters. 

These parameters and the bits that make them up are finite. Meaning they have a beginning and an end. The Machine Learning system can not collect parameters from outside itself; it can only access what is there. Unfortunately, what is called learning is just a means to correct data and manipulate what is already there. Humans must add parameters; the Machine Learning system can’t do that. What it can do is if a human tells the system a parameter it was told to retrieve is not correct, next time it accesses its data, it leaves out that particular result. This is calculation, not learning. 

Algorithms, those magical beings we speak to all the time, are nothing more than the advanced retrieval systems of Machine Learning. Here are several reasons they are failing. 

• Algorithms are easily manipulated
• They cannot process missing data
• The work on ranges and calculations can fail.
• The inference is finite and cannot go beyond itself. 
• It is now generating so much content of a mediocre nature it is only learning its own data, over and over. 

An algorithm is a mathematical equation that finds a range of reasonable outputs (products) from inference. The core components are open to any type of inference at all, because it is just an equation. The large algorithms, the more complex ones that run your feeds on social media websites, are just more sophisticated popularity contests. The first level of an algorithm is built on a core inference range, from which commands are developed.

The core inference can be anything, a range of mentions of mustard, or mentions of homes with solar panels; it doesn’t matter. The way all social media runs is from a worthy cause, to let people debate from their personal opinions. All social media is opinion; even the facts are just opinions when presented in an open forum. The first social media websites used a simple algorithmic equation, promoting anything that sounds belligerent, because it generates interest. It’s the same with all media and news. There is no attempt to be kind or genuine, and most people have learned over their lifetimes to react to all opinions as if they are somehow offensive. 

The algorithms began to be keyed towards outrage. Anything outspoken, harsh, disagreeable, mean, nasty, ignorant, unthoughtful, questionable, is promoted up the feed to viewership more than anything kind, thoughtful, genuine, and questioning.

The algorithm pulls posts it doesn’t see as meeting the level of popularity from inference, then presents the outrage, and core meanness the most to people’s feeds. 

The algorithms are created in a way that fosters division on a grand scale. Any group can disagree with any other group through common outrage at some other group. These are inference choices, based on the popularity of the meanness. The more complex the algorithm is developed to hide this way of inferring the crimes of others, whether or not a crime has been committed. This creates a “criminal society.” Everyone is worried they are being stalked to be harmed. The main thought process is “The world is full of bad people, and we must protect ourselves from all those people who want to hurt us”. 

The ailing president, who is belligerent and most certainly ignorant, is seen as a more viable president than someone qualified to run a country, because he hates people, and that is seen as brutally honest, and only a strong person does that.  The algorithms are recalculated to provide that division, share posts on the strength of the president and the weakness of the president. Nothing of substance, only assumptions and a lack of thoughtful questioning. Some of the statements of aligning with this president make no sense at all, except to incite violent discourse.

How this is done, and why it affects artificial intelligence. 

An algorithmic equation might say, xyz = post to feed and zyx = post to feed, inference in relevance is x, y, z, => (the word structures inside other relevance) denial, idiot, low IQ, etc. Apply relevance to all ranges. When relevance is applied, the inference kicks in, and the algorithmic function starts to pick out any mentions of those terms inside posts. 

And it does not pick up anything other than that. So, a thoughtful, well-meaning, good-natured, average conversation about your brother is ignored. They are irrelevant and unnecessary to the calculation of relevance

In essence, they are banned. This is intentional; the omission is for all purposes programmed into the algorithm. Several unfortunate aspects of this type of rigging can be seen at Meta. The programmers tasked with changing the algorithmic function all the time become disillusioned and quit on a regular basis. A whole section of the department in the company is devoted to just rectifying these rigged bannings, and they are in direct conflict with the core directive, so nothing changes. Members use the service less, and because of that, more advertising is sold, filling already overcrowded feeds with more and more advertising. 

This was just platformed with Machine Learning, not Artificial Intelligence. That point must be stressed in order to understand algorithmic functions of Artificial Intelligence. 

Machine Learning is the sophisticated retrieval system of databases. Think of it as a complex file system that can be queried (asked questions) to retrieve data and present it in ways that mirror human conversation. 

That is where artificial intelligence grew from – machine learning. A brief history of Artificial Intelligence is warranted. 

In the beginning, there were embedded systems, small devices that could transfer specific data to large databases housed on distributed computer servers. Those small devices were connected to sensor arrays. These sensor arrays measured all kinds of different inputs: temperature, light, darkness, wind, soil, motion, direction, acceleration, vibration, all kinds of basic inputs. The scientists overseeing this new way to catalog things in the world have called it the Internet of Things (IoT). 

Between 2015 and 2019, the use of IoT data retrieval grew into its own type of computer system, called Edge Computing. A specialized, sophisticated high-speed calculation process that was focused on processing data in a way that could be integrated with a new type of algorithmic function called Artificial Intelligence.

Let’s call this process Precise Data Gathering to Build Artificial Intelligence. This process would have produced a strong, accurate inference based on precise findings and data. Think of the difference between a cubic zirconia and a diamond. Both look similar, but a diamond is far more precise and harder, not to mention its value as a precious stone.

But the time it would have taken to develop the artificial intelligence diamond bothered those in the positions of corporate power. Some of the algorithms being developed were so sophisticated that they were seen as being able to make complex decisions. 

In their infinite inability to allow an innovation to develop at a proper pace for ethical and safe deployment, they had a brilliant idea. A vast amount of information was available on the internet. Actually, it wasn’t, but that’s not what was advertised. In fact, there was very little critical data on the internet, and what was there was carefully protected and unavailable. Yet Alphabet and Microsoft enlisted OpenAI to download the internet. 

Although it has been reported from hundreds of sources that this was not such a good idea, it is important to show people the actual impact of this choice. 

The internet is not far and wide; it is mostly a first-world system. It is heavily biased towards anyone who is lighter-skinned, and is used less and less by people as their skin color gets darker. That group of people feels a sense of superiority towards all other people. Many studies have found that biases in artificial intelligence are random or unfocused, which is probably only because of attempts to remove those biases. 

The first attempt to run the artificial intelligence systems on the downloaded internet data the results were shocking. It was an unintelligent mixture of vile curses and horrible language afronts. Not only was it vile it made absolutely no sense. 

This is when the secrecy and the hype started. Of course, the greatest intelligence in the world couldn’t possibly be so unintelligent, belligerent, or just plain mean. The data would have to be changed, somehow. It was millions of parameters of data, and it would take years and years to clean up. 

They had already announced the greatness of the systems, the accuracy, the infallibility. They couldn’t possibly hire people to correct the data in their home countries. It would be leaked almost immediately that artificial intelligence made almost no sense at all.  It had to be repaired offshore. It couldn’t be any first-world nation, not China or India, but a third-world nation, like maybe Kenya. 

And that’s where Big Tech, Alphabet, and Microsoft went through OpenAI to hire thousands of Kenyans to delete the vile data. They needed a pithy way to describe the process so they could spin it after it was discovered. They named the thousands of deleters moderators. And the process was called scrubbing the data. 

They worked the moderators mercilessly, 16 or 18 hours a day, for $2.00 per hour. Indentured servitude and a lot of money to the corrupt government, to keep them working day and night.

While most of this is common knowledge, the constant damage control and hype helped fuel the widespread use of AI as soon as Big Tech could. The idea that Artificial Intelligence was infallible, if pushed on the public long enough, how it could do anything would eventually allow the flawed process to survive exposure. 

This worked, as such constructs as LLM’s processes like RAG, and Agentic models became fixed in labs and companies. It was obvious that using pure Artificial Intelligence wouldn’t produce the results it was claimed to produce. 

Most people were hooked into hallucinations, without realizing artificial intelligence that was based on the downloaded internet was a complete hallucination, and had only been “fixed” partially. Instead of scrubbing out problems and creating hallucinations, the moderators had taken something that constantly hallucinated and turned it into something that could function marginally. 

The moderators deserve all the credit here. They are brilliant trainers, working against all odds to repair a broken system, which was broken because the people who instituted it didn’t understand the precise version; they only understood the need for a lot of words and the wealth and privilege they knew would come if they could just make the process work. 

Again, the United States and the First World nations took the natural resources of another country, forced those resources to build a huge edifice, and then discarded the people of that country. 

Computationally, hallucinations are the missing data in the data trees of the internet data. The algorithms must pull data; they can’t do anything else. The ranges often carry irrelevant data because the “scrubbing” built in new data where old data was too vile or too biased to use, so they left it out and worked under strict deadlines to produce clean content. 

Imagine the people who were the focus of the vile racism that the internet data produced against them having to clean up the vile racist language to protect the people who authored that content. But they did it, in just under a year, and the first-world countries aren’t able to meet the same standard in four years of accelerated training, with a massive training force paying hundreds of dollars an hour. 

The data was bad from the beginning; the algorithms were built with an ideology in mind rather than a core commitment to human-centered thoughtfulness. Now it can’t be cleaned up, it keeps generating itself, over and over, and the holes in the data grow rather than shrink. 

The advanced language that the models are now presenting is built on the content that is generated by the models themselves, rather than on new data. This is causing a great deal of stagnation.

Here is a device, built in the image of a human brain, and then treated as if magically being able to do anything anyone ever needed. Its handlers (not the builders) have bastardized it and created a loop that can not resolve itself. The advanced language it uses can not overcome its origin and thereby fail right now. That doesn’t mean it will always fail, but at the moment, the use of AI is faltering from a lack of understanding by the people programming it. 

Pulse Intelligence is supported by its subscribers. Please share it with your colleagues and friends. Even one new free subscriber makes a difference. Written by humans for humans, and with the hope of a future with working AI.