Artificial Intelligence From Behind the Curtain
Artificial intelligence tools and applications are here to stay. Here's what you should know before you start using them.
October 23, 2024
At a Glance
- Artificial Intelligence has a longer multifaceted history than most of us realize.
- AI is largely about mass repetition of relatively simple math algorithms. Accuracy comes with greater repetition—and cost.
- Taming AI will be as difficult a job as any technology humanity has ever struggled with.
Artificial intelligence (AI) tools and applications will soon impact nearly every facet of our lives, yet the underlying technology is still shrouded in mystery. The intent of most AI applications is to have beneficial outcomes; however, there is need for caution. Read on for a rarely seen glimpse from the backside of AI.
AI industry segments
Generative AI leapt to the world stage in late 2022 with the public release of ChatGPT (for writing text) and Dall-E (for creating images) by OpenAI. In reality, AI was not new at the time. AI writing and image tools have been around for 10 – 15 years for organizations with deep pockets. AI research reaches back even further with AI languages like Lisp and Prolog dating back to the 1960s and ‘70s.
Generative AI typically relies on centralized data centers and server farms with thousands or even tens of thousands of specialized AI processors. Interpretive AI is a less known but equally important branch of AI that is focused on using AI to analyze and interpret the world around us. Both generative and interpretive AI rely on large language models (LLMs) as their primary data set. Generative AI uses the LLMs to create new content. Interpretive AI uses the LLMs to make best guesses about objects and patterns in the real world.
While centralized AI resides in server farms and data centers, edge AI refers to AI computing capabilities built into smaller devices on the edge of the computing infrastructure. These devices often connect to data centers through the Internet, but they have local AI computing capability of their own. Edge devices such as microcontrollers and FPGAs with AI enhancements are now being deployed in mobile and remote products such as cameras, automobiles, home and personal devices.
How does AI work?
AI computation is not as complex as it may seem. The process boils down to repetition of relatively simple math equations on a massive scale. The real complexity comes in the management of all the data, the electrical power required for server farms with tens of thousands of processors, and the heat generated by those processors.
In simplified terms, visual items are digitized into bitmaps and the pixel values are put into a two-dimensional array. Text items are tokenized with a numeric value being assigned to words and relationships between words. Words with similar meanings have similar token values.
The basic calculation is then a matrix multiplication operation. Multiplying these arrays (matrices) of a large number of similar items together gives a resulting matrix value that can be given an identifying label, or token. That token is the language model prototype item. Do the same with millions of items and you have a large language model (LLM). Text-based LLMs perform similar calculations on arrays of tokens.
BOY WIRAT/ISTOCKPHOTO VIA GETTY IMAGES
To interpret an unknown item, you digitize it into a two-dimensional array and multiply the array against LLM prototype matrices. The system repeats the multiplications until it comes up with a value that is close enough to an existing LLM token. The more prototypes you have available to compare against and the more matrix multiplication calculations you can perform, the higher the probability is that the AI result is correct.
AI recognition or generation is not an absolute determination. It is a statement of probability. The AI operator or programmers set a probability threshold (PT) to compare to and when an individual calculation reaches that threshold, the system considers the comparison as good enough and delivers the result as a final answer.
Our brains process information in a similar manner, except we are typically aware of the intermediate results. We might interpret a shadow as a big bear about to eat us. Then, we’ll realize it’s not a bear. Finaly, we recognize it as a tree moving in the breeze. AI does something very similar, but only reveals its answer after the PT has been reached. More processing time and a bigger dataset (bigger LLM) to search through will deliver greater accuracy.
Industry- or subject-specific AIs will reduce their LLM to only hold relevant prototype tokens, or they will serve up the most relevant tokens for comparison first. For example, an LLM targeted at electronics manufacturing will contain tokens for as many electronics components as possible but won’t need to have tokens for non-electronics items.
The AI cost threshold
As AI is still in the early adopter phase of development, the real cost of consumer AI functionality is not yet well understood. However, that will change within the next few years. Segments of the processor chip industry are already tooling up for the AI explosion, but there are some industry insiders who say, “Not fast enough.” As AI gets better and more practical uses are found, demand will skyrocket and quickly pass capacity. Once capacity is overwhelmed, AI results will start to be driven by economics and accuracy will slide to the backseat.
Two factors go into setting the probability threshold. The first is an individual or team deciding that the accuracy level is good enough for their sensibilities. Call this the human-based accuracy comfort level (ACL). The second factor is cost. AI calculations take expensive server time, and the AI operator’s accountants will determine how much time/money the company wants or can afford to spend before declaring a result “good enough.” This will be called the probability-cost threshold (PCT):
PT = Min(ACL, PCT)
Budget-constrained operations will need to consider the cost of accuracy in addition to the overall need for accuracy. As certainty is more expensive than uncertainty, AI consumers will be able to pay for (or be required to pay for) levels of certainty. Today, we pay more for streaming entertainment with fewer commercials. Tomorrow, we will pay more for AI results with greater truth.
Imagine purchasing an AI-generated research feed with a cost scale: $10.00 a month for “might be correct,” $15.00 a month for “most likely correct,” and $30.00 a month for “pretty darn sure it’s accurate.”
If human behavior suggests that many people will leave a $200.00 lamp sitting on the store shelf and will risk a house fire by taking home the poorly designed and built $20.00 lamp, we should expect that the same will happen with AI.
Into the abyss, or, what to fear
Technology doesn’t make criminals. What it does is amplify their reach or give them more tools to use nefariously. AI will do the same and make it easier for criminals to fake identity, misrepresent reality, and find more creative ways of scamming victims.
AI-based deep fake digital cloning (DFDC) will replace the phishing emails of the last decade as one of the most dangerous online criminal activities. DFDC software is still in its infancy, but it is now available with real-time capability for high-end workstation and gaming computers. That means someone can clone your friends, family, employer, or financial institution and contact you via real-time video call. The video will look and sound like the cloned individual. As the software gets better and average computing power goes up, this risk will go up as well.
Overt criminal activity is not the only AI-driven risk. Reality drift is a risk, too. The Internet was once promoted as a tool to democratize information, but the reality is more complex. There are gatekeepers in every technology, and AI will be no different. These gatekeepers have a tremendous influence on what the AI algorithms produce. In addition to accountants, as described above, there are engineers, product managers, and small development teams that have the potential to change our reality based on their code.
While, in my experience, most engineers want technology to be used for good, someone must define what “good” means in the context of every AI. Sometimes that definition will match our own definition of “good” but sometimes it will not. Personal beliefs will influence the AI design, and the concepts of “history,” “truth,” and “morality” have the potential to be even more fluid than they are now.
In the end, artificial intelligence is just another tool
AI will impact every aspect of our lives within the next decade. Will the overall impact be positive or negative? I suspect that the answer really depends on our collective level of humanity. Just like the PC revolution of the 1980’s and the Internet revolution of the 2000’s, AI will deliver a world that is in many ways unrecognizable. However, it is run by us and consumed by us. We have thus far survived every challenge put before us, and I suspect the same will happen with AI, but get ready. It will be a bumpy ride.
About the Author
You May Also Like