Part 1 | Part 2 | Part 3

Concern for AI, Part 2 – Reduce Your Exposure
While Google Chrome (~4.3 billion users), Apple Safari (~890 million), and Microsoft Edge (~280 million) are the top three web browsers in 2026, they are far from the best in terms of your ability to control your privacy, data, and security. Chrome was essentially designed by Google to harvest your personal information then sold to the highest bidder. It provides the least amount of control over your data. Edge and Safari are used principally because they are installed by default, both antiquated, slow, and as with Chrome, provide very little control over data and security. Apple won’t let you uninstall Safari, but you can choose to not use it.

Choose a Better Browser
There are many high quality alternatives to the aforementioned browsers. Some commercial products, some open source with decades of solid programming, a strong user base, and community support. Opera, Brave, DuckDuckGo, and many more (PC Magazine review).

Firefox is my personal best. It has been around for two decades, is 100% open source; is fast, updated regularly, and provides a plethora of options for customization and on-line security. Why open source? “Open source” means the source code that makes the whole thing run is not proprietary. In fact, it can be viewed by anyone, anywhere, at any time. While not completely impervious to malware, most hackers and bad guys just don’t waste their time trying to impose ill-intending code in something that will too-soon be discovered. This is the very foundation of Linux and tens of thousands of world-class applications such as LibreOffice, Blender, GIMP, and Firefox.

1. Download and install Firefox.

   Installing a second or third web browser does not remove your current application on your desktop, laptop, or phone. Rather, you are installing an optional new browser to use along side or as your new, preferred and default application. And if you elect to retain Firefox, here are a few steps that will help set you in motion with the least down-time.

2. Set Firefox as your default browser.
   You can choose to set Firefox as your default browser during installation, or from your Mac, Linux, or Windows Settings (search for “Browser”). Even if set to default (which means that when you click on a link in an email or PDF, this is the browser that will open), you can right-click and choose a different browser at any time.
    

   
    

3. Copy your bookmarks and data
   If you are ready to commit to Firefox, copying your bookmarks and data allows you to pick-up where you left-off with your prior web browser. The option to copy your bookmarks should be presented to you during install or at first activation. Else, go to the Bookmarks menu in Firefox, select “Manage Bookmarks”, and then under the ↑↓ menu you can import just the bookmarks, or all data including search history and passwords.
    

   

Choose a better search engine
Google’s search engine is for many users, completely overwhelming, increasingly so with the introduction of default AI summaries and control. Furthermore, Google was designed to harvest data about what you search for, when, and from where&mdashit gathers as much information about you within a murky legal framework. Your data is captured and sold such that you can be targeted for advertising in just about all you do, everywhere. For some, this is just a normal part of the internet. For many, it is unacceptable.

Fortunately, there are a number of alternative search engines (PC Magazine review) that do not capture nor sell your search data. DuckDuckGo is the most popular alternative that does NOT collect or sell your data. Even if you want to use AI in your on-line searches, DuckDuckGo eliminates the harvesting and sale of your personal information.

1. Open the Firefox Settings
2. Select Search on the left, then
3. Select Default search engine
4. Select DuckDuckGo
5. De-Select all other search engines you do not wish to use.

Search Free from AI
Tired of your search engine presenting an AI summary when all you want is to get to the desired source? Don’t trust the results of the AI summary? Miss the simplicity of simple searches? You can disable all AI search functions quickly and easily.

1. Open the Firefox Settings
2. Select Search on the left, then
3. Select Additional search engines
4. Select Add search engine
5. Enter “https://noai.duckduckgo.com/%s” exactly as presented here
    

   
    

6. Save as “No AI” (or whatever you want to call it).
7. Go back to the top of the Search page and select “No AI” as your default Search Engine.
    

   
    

Disable all Automated AI Functions in Firefox.

1. Open the Firefox Settings
2. At the top, search for “AI” then disable all AI related functions that you do not desire.

Increase the Security of your Web Browser

1. Open the Firefox Settings
2. Under Privacy and security, set Firefox security to Strict:
    

   
    

3. Regularly clear your web browser cache
   The browser cache holds a great deal of information. While its primary function is to keep a local copy of website pages such that if you visit that page again the browser can load the majority, even all of the data (following a brief comparison of the website as it was when last visited versus when you return) immediately from your local computer which is exponentially faster than retrieving all of the data from the internet again.

   However, the cache is also a rich data mine, a place where a “cookie” (locally stored script) can learn which website you visited just prior, your preferences for on-line shopping, and in some instances, far more personal data. Therefore, it is best if you clear the cache on a regular basis—ideally once each day with the close of your browser (except that most people don’t shutdown their laptops very often, which is in and of itself a problem), worst case once each week.
    

   

4. Select Privacy and security
   • Select Additional protections
   • Select Tell websites not to sell or share my data

    

5. Select Permissions and Data
• Select Block pop-ups and third-party redirects
• Select Show warning when websites try to install extensions

Never See an Ad Again!

1. From the Tools meny, select “Manage Your Extensions”.
2. With the “Find more add-ons” function (top), Search for “uBlock Origin”.
3. Click on “uBlock Origin” (should be the first link in the list).
4. Click “Install” and it should automatically activate.
5. Repeat for “DuckDuckGo” and select
    

Part 1 | Part 2 | Part 3

Part 1 | Part 2 | Part 3

Concern for AI, Part 1 – An Introduction

A neighbor posted to our community email list his concern for the growing presence of AI. In response, I spent several days diving back into the foundation of my Masters degree in Applied Mathematics, Evolutionary Computation Applied to Radio Astronomy. My thesis included development of Karoo GP, a Genetic Program trained to isolate machine-generated radio noise from the desired astronomical sources, at the MeerKAT array, South Africa and later at LIGO to isolate machine-generated noise from triggers generated by gravitational waves.

While I have not been working at the code level of machine learning for some time, and do not claim to be an expert in artificial intelligence, I maintain a working understanding of the underlying systems and have been following the rapid uptick in application of Generative AI in the past two years. This is an incredibly complex topic, with an incredible array of potential, positive outcomes. At the same time, there are many serious, negative implications across all layers of society, world-wide.

IN THE BEGINNING
Humans have since the time of the Egyptian, Greek, and Roman empires recorded and analyzed data to improve farming, manage the finances of complex social systems, and to study the star lit heavens. Since the industrial revolution, advances in medicine, manufacturing, and the study of human and natural systems have made more important the need to reveal patterns in the data we collect.

Early computers were mechanical devices (1800s) followed by electronic computers (mid 1900s) capable of calculations far faster than the human mind. But why is a computer necessary, let alone artificial intelligence?

Let’s consider a small data analysis problem, the time of day incremented in hours, and the temperature outside measured in degrees, such that we desire to plot this as an [x,y] graph, with time on the x axis and temperature on the y axis.

(sunrise to mid afternoon)
time = 7 am; temp = 48F
time = 8 am; temp = 54F
time = 9 am; temp = 60F
…

(mid afternoon to sunset)
time = 6 pm; temp = 56F
time = 7 pm; temp = 53F
time = 8 pm; temp = 48F

… and into the night, cooling until sunrise and starting over again.

Image by Kai Staats

The plot presents a wave-like function. Day in and day out, over and over again as long as you kept placing dots on the graph, left to right, up and down. And we don’t need a computer to do this, just a pen, graph paper, and maybe a ruler to connect the dots.

Now, let’s add a third variable–the angle of the sun from zenith (overhead) such that we can track the time/temp correlation not just day to day, but month to month for one year. If we live north or south of the equator, we’d see an overall warming trend in the summer and cooling in the winter. This too we could do with a bit more graph paper and some patience for many more data points.

But if we want to build a formula to predict the temperature at any time of day, any week of the year, it gets more complicated. Perhaps something like:

    temp = mean_temp + fd[cos(date)] + ft[cos(time)]

where we have converted both time and date to circular (cosine) functions so that we can employ the rise and fall of the sun. The function “fd” and “ft” are coefficients that convert time to temperature. This is harder than our original plot, but yes, it’s still something we can do by hand if we have the time and gumption.

This application of mathematics to the natural world is at the very core of science, since the earliest observations and predictions. The goal is to understand the underlying function of everything from plant physiology to human metabolism, from animal migration to stellar evolution in a galaxy far, far away.

A great example of this kind of mapping of the natural world is the work done by Dr. Fisher in the 1930s whereby he applied measurements of the sepal length and sepal width, petal length and petal width of 150 flowers across four species of iris, as acquired by Dr E. Anderson. This is referred to as the iris flower data set. Fisher then hand-developed a formulae into which he (or anyone) could apply the measurements of one of the four species studied, and without seeing the plant, determine the species through formulaic classification.

Image by Kai Staats

The iris dataset has become a defacto standard that all machine learning programs are expected to solve, including my own. For example, with the following formula derived by Karoo GP, we can accurately assess the type of iris between any two species with 100% accuracy:

    sl = -sw + pl2

It gets a bit more complicated when solving for one of three species, but the same concept applies. There are countless millions of examples of statistics (mathematics applied to data analysis) in our modern world–measuring the physical, chemical, biological, or psychological parameters of a given function, from the behavior of online shoppers in holiday seasons to Netflix views, from cancer cell growth to weather prediction–statistics seeking correlation, and often causation.
  

THE JOY OF DISCOVERY

Discovery of patterns, in any field, is not some sterile activity conducted by socially inept geeks in white lab coats (well, maybe sometimes), but a rigorous process that follows a centuries old practice of a) generating a hypothesis, b) developing an experiment, c) collecting and analyzing data, d) comparing to the hypothesis, and e) sharing both the results and methodology to the world, then refining the experiment and doing it again (more often than not) to improve the model and more importantly, the understanding of the natural process.

This can lead to an increase in healthy habits, reduction in car accidents, and discoveries of breathtaking beauty. For me, for the many times in my work I have been part of pattern discovery, I find myself with an accelerated heart rate and deeply satisfying sense of peace. Something as simple as seeing the growth rate of a pea plant correlated to the amount of carbon dioxide provided (as we do at my worksite with each experiment in bioregeneration) is incredible. It’s repeatable. It’s demonstrable. And it’s a means to seeing inside a small piece of the cosmos and unraveling its mystery, not to dissect or control, but to understand and marvel at its inherent beauty.

I share this because beyond the profit driving many of the companies hosting chat bots and AI driven tools, there are a body of programmers that for the past two decades have enjoyed the discovery that comes through the iterative improvements to their underlying code. I understand their joy and motivation, even if I am increasingly concerned with the way in which their code is being used.

I will never forget the moment when my machine learning code Karoo GP evolved and re-discovered Kepler’s law of planetary motion based on the same numbers recorded by Newton more than 300 years ago. I ran through the halls of the astrophysics institute where I was based, shouting “It worked! It worked! It really worked!” I can only imagine what it was like for Newton, based on his own observations, to discover patterns that defined thermodynamics, the age of the Earth, and gravity all in one lifetime.
  

IT’S ALL STATISTICS
Back to our comparatively simple effort to predict the temperature each day across the seasons, we will find ourselves with a pattern based on a general formula (above), but not close enough to truly predict with any accuracy, especially not in this new world of rapidly shifting climate functions. It takes far, far more sophisticated models to predict with any accuracy at all.

So we add “degrees of freedom”, additional variables to present more facets of the weather functions. We might take into account the annual increase in greenhouse gases such as water vapor, carbon dioxide, and methane; dust from fires and volcanoes, a reduction in reflectivity by shrinking polar caps and snow mass on glaciers; desertification of once fertile regions, even the 11 years solar storm cycle, to name a few (and there are thousands more).

But now we have far more variables than just time of day and angle of the sun–too many to plot on a piece of paper. And the relationship between these variables cannot be visualized in the human mind nor found by hand. So we need help.

At the foundation of machine learning are statistics and regression analysis.

Statistics we all understand to some degree, as every day we process statements such as “flying is __ times safer than driving”, “smoking a pack of cigarettes each day increases your chance of lung cancer by __ percent”, and “Four out of five dentists recommend brushing your teeth.” (not sure what’s going on with that 5th guy ?!$).

Statistics gives us skills in critical thinking, making important decisions, and managing our finances. It’s a powerful set of tools that should be a required class in high school. I did not learn how to apply statistics beyond an average high school vocabulary until working on my masters degree at age 44. I learned to apply the basics (mean, median, and standard deviation) to build simple relationships between data points prior to training a machine learning algorithm. These “features” offer first order correlations that boost the machine learning algorithm’s ability to find more meaningful relationships between variables in the dataset.

Regression analysis is the development of a mathematical function (e.g. x = 2y + z) to represent the relationship between two or more raw data points or features, as given in my simple temperature prediction model (above), the work of Dr. Fisher with iris flowers, to millions of everyday examples:

    Predicting the Popularity of Social Media Posts
    Predicting House Prices
    Predicting Exam Scores Based on Study Time
    Forecasting Sales for a Business
    Predicting Sports Performance

(More at Geeks for Geeks.)

In my work at LIGO to isolate black hole mergers in the noise of the complex machine, we regularly engaged in processing 4,000 variables (4,000 columns in a spreadsheet) with 10,000 data samples (rows). Yet, this is a “small” dataset in just about any modern data environment.
  

LACK OF TRANSPARENCY
Now, the start of my answer to G’s question–there are important distinctions between statistical analysis, machine learning, and AI.

As I began, with traditional statistical analysis we, the human, are manually or semi-automatically building the formula, by hand or with a spreadsheet or the application of advanced statistical formula. But as any insurance company will tell you, just having data and applying a model does not equate to an accurate prediction. Far from (else, our insurance rates would be going down, now up).

The data may have been highly skewed when captured, based on who collected it (student vs industry professional), where it was collected (rural Kansas vs downtown LA), and who was paying for the data (university, pharma, or politician). The algorithms might have been generated by employees who have long since retired or an old system that is now considered antiquated.

ChatGPT is harvesting data from across the entire world, pulling from sources that are genuine and those that are completely bogus. Yet ChatGPT is opaque to how or where it acquired the data it is processing, and how it generated the result.

When I was at LIGO there were fierce battles between the astrophysicists as to whether to trust machine learning to classify a celestial event, or to rely on older, more transparent computer models. My code was unique in that it was 100% transparent, with every line of code commented and the outcome being a mathematical expression built from known input variables.

Image by Dragon 1

On the other hand, Convolutional Neural Networks, the foundation for Deep Learning and now Generative AI, are far more complex, the feature generation and determinant processes a relative black box, opaque to the exact means by which a classification or decision was made.

Convolutional neural networks (“neural nets”) can process tremendous arrays of data, systems so incredibly complex that any other process would require many stages to arrive to a similar conclusion, if at all. Neural nets can be trained on tens of millions of labeled photographs, bodies of text, and data.

However, neural nets and now Generative AI are black boxes, meaning, you cannot see inside. Therefore you have little to no idea how they arrive to their conclusion. And that is the problem.

While AI is already being applied to cancer research, protein folding, weather prediction, self-driving cars, and understanding who is likely to become homeless based on variables captured in routine visits to the health clinic–no one can tell you, precisely, how the solutions are generated internal to the AI itself.
  

IS AI INTELLIGENT?
No. At least, not yet. It is a powerful data processing engine that can rapidly read, review images and videos, process and respond via both written and spoken language, and in many ways appear to have human responses, even conversations.

But at the core are a series of probability curves, potentials for a or b, c or d, and so on until the image of the panda is the most likely label applied (see image). And if I ask a chat bot to fill in the blank, “I am so hungry I could eat a ______”, another set of probabilities built on a global dataset of thousands of similar sentences, on-line references, suggest that the most likely outcome will be “horse”.

In many ways, this is how we process language too, and how cliche phrases propagate as we automatically pull responses from a deep well of potentials. We quote celebrities or a comedic one-liner from a film. We get stuck on a particular phrase for days, even months at a time until that critical path is dislodged and a new “training” opens a new pathway through our brain. And sometimes we say really stupid, even mean things “that we didn’t mean to say” because we didn’t actually think about it—it just came out. Well, that’s just a statistical fill-in-the-blank that requires very little cognition, often spurred along by emotion (which thankfully, ChatGTP does not yet have). If we truly thought about each and every thing we said, we would say very little at all.

Maybe what intrigues (and scares) us about AI is that in the reflection of this powerful thing we have birthed, we might not be all that intelligent after all–just a moist, gooey collection of cells and organs and gray matter that most of the time is not terribly self-aware. Just like ChatGPT. For further reading, learn more about how ChatGPT works.

MY CONCERNS
Large Language Models (LLM) are what give ChatGTP and others like it the ability to process and respond to written and verbal input in a very natural, human-like manner. Before LLMs, machine learning (the foundation for AI) was engaged by a programmer whose code invoked data analysis and answer. Now, anyone, with or without a programming background can engage these incredibly rich models and receive incredibly rich answers, or so it seems.

This is the stuff that science fiction writers and filmmakers have been teasing us with for nearly a century–computer programs that process information far faster than any human, write like us, talk like us, and might soon look like us. Here we are, AI rising, and as in the movies there are those who celebrate the coming of this digital wonderment, and those who are very, very concerned (BTW, the ones who are concerned always get to say “I told you!” when the robots attack).

The positives are many-fold:
– The ability to model complex biological functions, and
– Discover cures for challenging diseases.
– Analysis of datasets so vast that traditional computer programs would take years to arrive to a similar conclusion.
– Weather, societal, and financial predictive analytics; and
– Write, test, and fix computer software code.
– Many more at Cousera …

The negatives are equally bold:
– Evade sophisticated security systems.
– Generate text, audio, and images nearly indistinguishable from the real thing.
– Given foundation to tremendous deception.
– Dissolve if not completely eradicate privacy.
– Replace human-to-human connection.
– Replace the need to read, write, or think at all!
– Place too much power in the hands of companies and governments using these tools for profit and control, not the greater good.

Part 1 | Part 2 | Part 3