All about AI -- Aye Yi Yi

Meet Eliza

Eliza was born in the mid-1960s. Her father was Joseph Weizenbaum, an MIT computing researcher. She was a precocious little thing, able to interact with people, almost from her birth.

But Eliza* wasn't a little girl. She was instead a computer program -- one of the earliest programs capable of natural language conversation. She was a chatbot.

Yes, that's right. A chatbot, in the mid-1960s no less! Only we didn't have that term back then.

* Eliza (the program) was named for Eliza Doolittle, the Cockney English flower girl in My Fair Lady, being taught to speak "Proper English".

Eliza's algorithms, an advanced idea at the time, were pretty simple by today's standards. Eliza combined pattern recognition by parsing typed-in sentences for noun and verb structure along with asking open-ended questions, to produce a realistic, seemingly human interaction.

Here's a sample of how a conversation with Eliza might go:

Person: I feel stressed at work.

Eliza: Why do you feel stressed at work?

My boss gives me too many tasks.

Tell me more about how your boss gives you too many tasks.

Well, I don’t have enough time to finish them.

Why do you think you don’t have enough time?

Because there are only so many hours in a day!

Can you elaborate on that?

Never mind, this isn’t helping.

Why do you say this isn’t helping?

heh heh... Eliza could go on like this all day if you allowed it.

In the 1960s, you might have said it's an example of AI, however crude.

To be sure, this was not AI. Eliza knew absolutely nothing. Not even 2+2. There was no knowledge base from which to draw answers.

This may seem whimsical, and I suppose it is, but make no mistake, the idea of a natural language processor (NLP) like Eliza was foundational to the advancement of AI research that followed.

Now meet Siri, Alexa, and Cortana

These three gals are quite advanced compared to Eliza -- they're all pretty smart. But even this threesome are not good examples of AI.

Some might argue these are examples of narrow AI, limited to specific tasks. But that's too generous. To me, to earn an "AI badge", a system should be capable of answering disparate classes of questions by examining its training data and in ways not anticipated by the system designers. As of this writing, Siri, Alexa, and Cortana aren't there (yet). Cortana may never get there. She was put on permanent leave, at least in the consumer space.

These are more aptly called digital or virtual assistants. You can ask them to perform a number of tasks, but only tasks that have been preprogrammed. They are capable of NLP (Natural Language Processing) but that's not the same as full AI.

ChatGPT Lands with a Boom

In November 2022, OpenAI released their first publicly available version of ChatGPT. And, wow, what a thing that was. ChatGPT rapidly shot to the top becoming one of the most popular apps downloaded.

For the first time, everyday people could ask questions and get useful, actionable answers on the spot. Not just Google-like search results.

This is what I consider AI.

How does it work?

That's is a huge, complex question, with technical answers far beyond the scope of this article. I'll keep it to a useful executive overview.

ChatGPT is built on self-attention (not self-awareness!) mechanisms, allowing it to "understand" context, relationships between words, and generate coherent responses. It learns grammar, reasoning patterns, and styles but does not (usually) rely on direct access to live data or real-time internet. That would be far too slow.

Instead, it's trained on a huge amount* of data that comes from the internet. That data includes countless books covering a wide range of topics, newspaper articles, academic papers, websites, social media, and whatever else its web crawlers can scrape** up. That data is analyzed and optimized according to the training model and used by ChatGPT when generating its responses to your queries.

* ChatGPT's training dataset is about 45 terabytes of data, as of this writing. In terms of type-written pages, that would be something like 13 billion pages, or a stack of paper about 800 miles tall. Crazy.

** Scraping/crawling means bulk, wholesale scanning of millions of websites to gather up data. For an AI crawler, that data goes into the training model. For a search engine crawler, the data goes into a database that is indexed for fast search and retrieval, like Google search.

Let's parse the word "ChatGPT":

• Chat: It's a chatbot. It uses NLP (Natural Language Processing) to "understand" your input (called "prompts" in AI parlance) and to create human readable responses.
• G=Generative: Responses are generated on the spot specifically for your prompt. They are not canned or preprogrammed.
• P=Pretrained: The vast data reservoir, collected mostly from the internet, imported into the AI engine.
• T=Transformer: The process used by a particular AI model by which a response is generated. There are several.

The amount of processing power required by ChatGPT, and other AI systems based on models similar to ChatGPT, is truly enormous. Depending on who you ask, a ChatGPT prompt might consume upward 25x more processing power than a (pre-AI) Google search.

Existing AI systems are also eye wateringly expensive. Not just in the power used (which is enormous), but in the computing infrastructure needed to make it all work. The top Big Tech firms have cumulatively spent hundreds of billions of dollars on AI initiatives in just the last two years alone.

AI vs. ChatGPT

Is "AI" and "ChatGPT" the same thing?

ChatGPT is just one implementation of AI architecture, but it's the one most people know. It's highly visible because it stands on its own as a retail (consumer facing) chatbot that you can use on your computer or phone to ask questions of, or to just shoot the breeze with.

AI is an all-encompassing term, a convenient "handle", if you will, to label the big, wide world of artificial intelligence initiatives. And, boy, are there a bunch of them.

We're starting to see AI everywhere these days. We'll cover some of that farther down.

AI vs. Search

What's really the difference here?

We'll use Google as the example.

When you do a search in Google, the search engine very rapidly examines an enormous database that was previously compiled and is constantly updated, containing most of the surface web's* information.

Google isn't "searching the web" as you type in a search query -- that would be impossibly slow. It already has all that data stored and indexed for extremely rapid retrieval. That's why Google is so fast.

* The surface web is content that is ready available, at no cost, with no access blocks, simply by browsing to that site, like my website that you are reading right now (did I say thanks?). Contrast that with the deep web (not the dark web, that's something else entirely) which is information that's not readily available, such as data contained in databases, behind paywalls, or behind corporate and government firewalls.

But the key point here is that Google just returns links to web pages that (probably) contain more info on what you seek. If your search is a popular one, like how old is Biden or Trump, Google may return the result as a direct response, rather than pointing you to web pages that contain the answer. But, usually, it's a page full of links that you have to sort through.

With an AI-based approach, like ChatGPT, it's more analogous to having a helper sitting next to you that you can ask questions of. Your helper then does all the scut work of searching and collecting information on your behalf and compiling a useful, personalized answer.

You could think of ChatGPT as an "automated super google" that not only searches, but also presents the results in a more immediately useful form.

I ain't gonna lie. That's mighty useful. (green text is under active editing)

Truth and Trust

But there's a downside, too.

We all know the internet is full of misinformation, disinformation, rumors, conspiracies, lies, and damn lies. When you do a Google search manually, if you are reasonably savvy and incredulous, you can often suss out the wheat from the chaff. You see exactly what's going into the sausage, so to speak. Google isn't making value* judgements on the results.

* Google does have algorithms which helps to rank pages. But that has more to do with page and site reputation, based on the rank of backlinks (referring pages) that link to the page in question or to other pages on the site. Page rank helps to surface pages that are more likely to be truthful and authoritative, but it doesn't directly vet whether a particular page is truthful or not.

But AI systems, like a sweet summer child, are more easily fooled. Unlike choosing from that screen full of Google search results, you don't have the opportunity to vet the data in the training model that contributes to the response.

However, there is way to get ChatGPT to show their work: After receiving a response to your prompt, reply with "cite sources" and ChatGPT will do just that. In most cases, it should return at least one relevant source supporting its previous response. Not as good as a screen full of raw links, but at least it gives you a place to start. It's better than nothing. I would strongly recommend doing that when prompting ChatGPT for hard facts. And don't just take the sources on faith. You must examine them.

Well, it is AI after all, shouldn't it be able to tell?

Sort of. We'll get more into this further down when we discuss what it means to be human.

But yes, to the extent that credible sources are saying one thing and non-credible sources are saying another, good faith AI models that are paying attention will avoid scraping non-credible sites. Or, at least, labeling such collected data as less reliable, so that such data would be used in proper context. In a general knowledge retail chatbot like ChatGPT, the model is less likely to gather data from known unreliable websites.

That can help, but it doesn't mean AI still won't hallucinate. That is, making up stuff.

Deterministic vs. Probabilistic

I'll expand a bit on what I discussed above about how AI and regular search differ. Nerd Alert!

Deterministic

Driven by rules

Search engines like Google, Bing, and others (at least before AI summaries were introduced) are deterministic. That is, the results come about from the search engine following its programmed logic and rules.

The responses you receive from that search can be readily tied back to the source data. The response page is full of links to various websites that you must follow to find your answers. Sometimes the results include an excerpt lifted from one of the more popular of the linked sites and displayed without you having to follow that link. That is performed by deterministic programming rules, not AI -- at least before AI came along.

Probabilistic

Driven by statistics

AI Systems, on the other hand, are probabilistic.

In brief, AI systems today are basically pattern recognition and prediction engines, albeit highly complex. When you submit your prompt, the AI system examines it and generates a response drawing on patterns learned through its vast training data. Unlike with a deterministic system described above, the response you receive in a AI system is generated on the fly based on its training data and is not typically tied back to a specific real world data source.

Because they are probabilistic -- not deterministic -- AI systems return the most likely words to come next from its training data, in response to your prompt. Note that I wrote "most likely", not "conclusively correct".

Furthermore, your prompt influences what you get. So the better you are at crafting a clear and precise prompt, the better the response is likely to be. AI isn't your bestie that knows your thoughts and feelings. It assumes nothing. What it knows about you comes only from what you tell it.

The response you receive is colored by many things: Your previous prompts (within certain limits), the state of the training data, which AI you are using, and again, how you phrase your prompt.

And this all leads to a rather incredible phenomenon...

While a deterministic system follows a fixed set of rules and produces the same output for the (generally) same input, a probabilistic system incorporates statistical predictions and may produce (sometimes wildly) different outputs even when given similar inputs.

Software engineers may write the code that makes AI operate, yes, but the responses produced by AI are not precisely predictable in advance. That's because the system generates responses based on statistical patterns learned during training and influenced by the quality of the prompts it receives from the user.

Forgive my tautologies here. I just want to make sure no single description of something leaves you scratching your head.

This is one reason AI hallucinates. A hallucination occurs when the system assembles a response that sounds plausible but is factually incorrect, such as citing non-existent case law. It is also one reason AI systems sometimes produce harmful or dangerous responses. Because the system generates language from learned patterns rather than following a simple list of approved answers, it is not a straightforward matter to program AI to never produce something harmful, illegal, or incorrect.

AI has no consciousness, no awareness, no intuition, no sentience, no insight, no sagacity, no judgement. Truthfulness, falsehoods, and lies mean nothing to AI. It has no concept of any of these things.

Confiding in AI with your hopes and fears can be particularly problematic. While the AI chatbot may seem like a good listener and offer good life advice, you must understand those responses don't come from a place of human intuition and compassion. They're essentially a mashup of what the AI has learned* from its training. And that training includes some pretty bad shit.

* I hate to use the word "learned" here. AI doesn't learn anything (in the human sense). It analyzes words, detects patterns, and generates a statistically significant response based on vast quantities of content. There is absolutely zero emotional cognition. There is no soul or capacity for organic human expression so don't ascribe so.

When and How to use ChatGPT

You've noticed I've focused more on ChatGPT as a specific example of AI whereas AI itself is a much larger topic. Again, that's because ChatGPT is generally top of mind when people think about AI. Most people's direct and useful exposure to AI is through ChatGPT or similar chatbot.

You'd be well advised not to rely on ChatGPT, or any AI chatbot, for authoritatively correct responses to anything. Just ask the several lawyers who got into trouble citing completely fabricated case law in some of their filings to various courts of law how that went.

There's plenty of hilarious examples of people being professionally embarrassed or worse because they relied on ChatGPT to provide accurate answers and didn't bother to verify them.

But that's not to say that ChatGPT can't be useful. e.g. I might use ChatGPT to help me proof and rephrase awkward or clumsy sentences that I'm writing to help them read better, perhaps by rearranging wording or choosing different words from what I used. But that's pretty much it. These are "soft" uses of ChatGPT where there's really no right or wrong.

I would never use AI for writing original, creative, or fresh content. That's my job, dammit.

Sometimes I'll ask ChatGPT for responses on a topic that I'm already intimately familiar with, just to see how it does. It generally does pretty well, but I've caught it out a few times, too. Sometimes ChatGPT misstates a fact in an otherwise decent response and other times it just flat-out makes up shit from whole cloth. You really must be mindful of that and be ready to verify any and all "facts" it gives you.

When I call out a hallucination, it immediately recognizes its error. BUT... ChatGPT does not update its training model based on your corrections. It may remember for you for next time, but that's all. e.g. I sometimes ask ChatGPT about various things in Italy (my wife's home country). Since then, ChatGPT has taken it upon itself to sprinkle in some Italian words on completely unrelated matters.

Google's AI (Gemini) overview, featured at the top of most search results, is particularly prone to hallucinations. It's bad enough that I disable it when I can and ignore it when I can't. If Google's AI overview is that bad, why the hell should anyone trust anything it has to say? Google can and should do better. They should not be alpha testing on their external users -- you and me.

Computer Programming

I also write computer code in my various projects. I've used ChatGPT to write code snippets for me as a time saver, especially in programming languages that I'm not fluent in. I often have to tweak the code, but it's considerably faster than developing it from scratch.

Why is that?

Software development frameworks today are far more complex than they were back in the day when I got started. I mean, I knew Basic, Fortran, C, and Pascal and wasn't half bad in assembly on a couple of machines (Univac 1100 and Intel x86). But these were languages where one primarily utilized intrinsic statements and, from that, created useful programs. Today's modern programming languages require deep understanding of libraries (vendor and third party), dependencies, integration, and structures that didn't exist back then.

It's considerably more complicated today. AI can help with that. But its code must be vetted.

AI is Everywhere

Now let's talk about the wider world of AI. Companies all across the business spectrum are rushing to introduce AI components into their products. I swear, half the sponsorship spots I hear on public radio these days includes company x, y, or z mentioning its AI. (Note to those marketing execs making ad buys. No one cares!)

Big Tech is forcing AI on us and often not letting us opt-out or disable it. Microsoft's CoPilot AI is pervasive inside of their products. Amazon's Rufus (really? Rufus?) was following me around like a lovelorn puppy until I killed it using one of my browser extensions. Google searches now include an (often inaccurate) AI overview, further separating me from the webpages that I seek.

The familiar "chat with us" icon that we see on so many websites are now mostly AI-powered. Companies are starting to implement AI voice chatbots in their phone menu systems, making it even more difficult to reach a human. Some companies don't let you reach a human at all. Terrific.

The medical profession is experimenting with AI to help diagnose disease, ailment, examine x rays, or other patient complaint. To the extent it's used to help surface certain conditions or their precursors faster, or that a human doctor might miss, then it's not a bad thing. So long as the doc verifies any findings.

But regardless of what form you encounter these AI assistants, be aware that all of them are subject to hallucinations, at least for now.

What Makes a Human?

Researches for years have compared and contrasted the human brain with the current (at the time) state of the art classical computers.

Those that might think today's AI-capable, blistering fast computers have finally outdone the human brain would be very wrong.

We know you can't directly compare a human brain to a computer as though you were deciding which laptop has the better specs. But efforts have been made to draw some rough comparisons between brains and computers and the results are pretty fascinating.

First, some general definitions:

• Storage: How much data a thing (computer or brain) can hold.
• Processing power: How fast that thing can process data measured in FLOPS (FLoating-point Operations Per Second).
• Efficiency: Power it takes to perform a given amount of processing. Greater efficiency = less power to do x amount of work.
• Tera = 10^12 = 1,000,000,000,000 (1 with 12 zeroes)
• Peta = 10^15 = 1,000,000,000,000,000 (1 with 15 zeroes)
• Exa = 10^18 = 1,000,000,000,000,000,000 (1 with 18 zeroes)

Now, let's compare the human brain to a computer.

Storage: Most laptops and higher-end phones today (2025) have around 1 terabyte of storage. By some estimates, the brain may have as much as 2.5 petabytes of storage. That's 2,500x more than most laptops and higher-end phones.

Processing power: The new (in 2025) iPhone 17 Pro is rated at about 2 teraflops. By contrast, the human brain's processing power is estimated between 1 petaflops to 1 exaflops. That's roughly 500x to 500,000x faster than the iPhone 17 Pro.

Efficiency: Here's where the brain really excels compared to classical computers. A computer may consume between 10 to 50 kilowatts per 1,000 teraflops depending on hardware particulars. The human brain, on the other hand, consumes around 20 Watts to achieve the same throughput* in terms of work performed. Based on these numbers, the human brain is between 500x and 2,500x more efficient.

* Brainpower can't be measured in FLOPS. But the analog equivalent of work being done is somewhat comparable.

The differences among classical, quantum, and biological computing are profound, and both quantum and biological computing have become active frontiers of research. Quantum computing is advancing rapidly toward practical applications, while biological computing is still in its infancy.

But the human brain is more than just a huge storage reservoir and ultra powerful processor. What truly separates us from AI are the qualities inherent and intrinsic to the biological and analog nature of our brains.

Humans possess innate capabilities such as intuition, consciousness, sentience, self-awareness, and emotion. We also have the uniquely biological ability to generate new knowledge and understanding, not merely by processing information but through abstract reasoning, creativity, and lived experience. Computers, by contrast, do not and cannot replicate those qualities.

In short, the human brain is far more than the sum of its stored knowledge. AI engines absolutely are not. Nor do I believe they ever will be, even when quantum computing becomes mainstream.

And yet, despite the staggering superiority of the human mind, we still believe a lot of provably incorrect things: absurd conspiracies, climate-change denial, race supremacy, and countless other patently disproved ideas.

If we humans are this bad at believing truthful things, then what hope could AI, with all its additional and plainly visible flaws, possibly have?

That alone should give you pause before relying too heavily on AI chatbots!

I'm not dunking on the whole idea of AI here. It's a remarkable, transformative technology that'll get ever better. But it must be used in the proper contexts and with all due cautious incredulity -- at least for the foreseeable future.

Pizza Pizza

Here's a particularly fun example of an AI fail:

When prompted with "cheese not sticking to pizza", Google's search AI Overview feature suggested mixing in 1/8 cup of glue to give the sauce "more tackiness". (Google has since fixed that)

Without further contextual analysis, yeah, using glue on something to make it stick, might sound ok.

But there's so much wrong with adding glue to pizza sauce beyond simply the first obvious reason.

• First and foremost, we don't use glue on food*
• The liquid nature of tomato sauce would prevent it from being glued in place
• The high heat of cooking could ruin the glue
• The glue would likely impart unpleasant flavors and texture and could be toxic.
• And certainly other reasons...

This is where a human brain would succeed where AI fails. Even if the idea of "glue" and "food" never came together in a person's mind, it would still be plainly obvious that doing so is probably not a good idea.

The additional bullets point out some reasons why it would not work. Any particular person may not know these additional facts. But just the idea of using glue on food is wrong enough, all by itself, to disqualify that as a useful suggestion.

* OK, foodies, I'm aware there are methods of "gluing" foods by using an egg mix, starches, etc. But that's not what we're discussing here.

Another fun fail: I asked Gemini during the last presidential election how many months older that Biden is to Trump. Gemini simply subtracted Trump's age from Biden's (4 years) and concluded that Biden was 48 months older (4y x 12 = 48m). Clearly that's wrong because Biden's month of birth is Nov 1942 and Trump's is Jun 1946. That's a difference of 43 months. Gemini didn't have the "presence of mind" to calculate the difference based on actual birth months.

That's but one of the kinds of fails that AI systems routinely demonstrate.

AI Pushback

We're starting to see pushback building against AI as it becomes ubiquitous in everyday life. It's being sloppily shoehorned into every day consumer experiences without adequate thought given to implementation and the overall UX (User eXperience).

An an I.T. guy that works for small business and residential clients, I see this a lot. And I see it in my own life, too.

It's also pissing off website owners who are tired of their IP (Intellectual Property) begin scraped by AI crawlers. There's also a considerable cost to all those websites being scraped.

Cost to the websites? How so?

Visiting a website may be free to you, but not to the site owner. Aside from the cost of designing the website, there's the ongoing hosting fees. Costs can escalate due to repeated page fetches by AI crawlers.

Many site owners report the (sometimes vast) majority of page fetches are not regular people browsing, but instead are AI and search engine crawlers, gulping up everything they can find on a site. That can amount to real money cost to the site owner.

Webmasters can limit this by putting up a "no trespassing sign" (a robots.txt file) on their home page, specifying what sort of automated crawling they'll allow and by whom. But compliance is voluntary -- and routinely ignored.

Some site owners are resorting to guerrilla tactics to disincentivize this antisocial behavior by booby trapping their sites in ways that only affect AI crawlers and not regular people browsing.

Such "booby traps" include, among others, tar-pitting programs called Nepenthes* and another called Iocaine** as a countermeasure. Tar-pitting programs trap the AI crawler, sending them chasing down endless links with no way out, and returning nonsense, bogus data in an effort to poison the AI model.

* Nepenthes: A genus of carnivorous plants, also known as tropical pitcher plants, or monkey cups.

** Iocaine: A fictional deadly poison from the film The Princess Bride.

The Future of AI

AI isn't going away. It'll become ever more pervasive in our lives, just as streaming TVs and connected cars that track where you go and when, Wi-Fi enabled appliances with subscriptions, smartphones full of mostly stupid apps, and all the other tech we deal with that is fairly new to civilization.

Cheerleaders of AI, mostly tech companies funnily enough, say that AI will free people in the work place from having to perform boring, mundane tasks so they can focus on more creative endeavors. You don't often hear from AI proponents about the jobs that'll be displaced. But it's coming. Hell, it's already begun. How could it not?

Some of those proponents take it a step further, promising a sort of future utopia where many people won't have to work because AI will do it. Those people would be free to pursue whatever they enjoy. Problem is, people need money. They have to work. We don't have a Universal Income that would allow people to pursue non-occupational endeavors. Or were the AI companies planning to fund a universal income? Hmmm...

Help not Wanted?

At least one CEO is being more forthright about his intentions for AI. Sebastian Siemiatkowski, the CEO of Swedish BNPL (buy now pay later) startup Klarna, readily admits that AI will allow him to operate his company with significantly fewer employees. If this bothers you, then you can choose to not use Klarna. But, alas, this kind of onesie-twosie boycotting isn't very effective.

Klarna won't be the only one. As AI systems develop, especially to the extent that hallucinations can be significantly reduced, it will displace workers in a wide variety of companies. This is the ultimate reason companies are embracing AI!

Some types of jobs that could be adversely affected by AI:

• Data entry clerks, receptionists, transcriptionists
• Software developers, code monkeys
• Call center agents, help desk, online chat CSRs
• Cashiers, some sales associates, order clerks
• Certain assembly line workers, machine operators
• Truck drivers, delivery drivers, taxi/Uber drivers, if autonomous driving ever takes hold
• News reporters and writers, copy editors
• Paralegals, researchers, contract writing and reviewing
• Bank tellers, loan officers, accountants, tax-preparers (heh, that last one wouldn't be so bad)

Most occupations where you sit at a desk using a computer could be subject to outsourcing to AI.

Here are some jobs that are less likely to be displaced by AI. Note that AI could play an assist role in some of these, but the job itself would probably be safe.

• Human-centered, requiring emotional intelligence/empathy: Therapists, counselors, psychologists, psychiatrists, social and outreach workers, clergy/spiritual leaders
• Highly creative fields that require original thought: Writers, poets, artists, filmmakers, playwrights. Although AI is making inroads here.
• Skilled trades: Construction, electricians, plumbers, roofers, drywall installers, painters, auto mechanics, and others
• Unskilled, non-repetitive manual labor: Landscaping laborers (yard work, hauling materials, planting), construction assistants, go-getters
• Moving and transportation: Furniture, etc., junk removal, package delivery (driving could become autonomous, but probably not package hand-off)
• First responders: Police, firefighters, paramedics, tow truck drivers
• Hospitality and entertainment: Full service restaurant workers, stagehands, roadies, festival workers, chamber maids
• Janitorial in residential and light commercial areas
• Certainly more I'm not thinking of...

Essentially, any job that is physical, that has a hands-on component, and variable in nature should be safe. So, not including highly controlled, repetitive factory jobs, especially at larger companies that can afford the upfront costs of automation.

And, as mentioned above, jobs that require qualities that only a human brain possesses, should be safe.