Yves here. AI hype seems to need to be debunked often and forcefully, hence the need for Rob Urie’s post. Please circulate widely!
By Rob Urie, author of Zen Economics, artist, and musician who publishes The Journal of Belligerent Pontification on Substack. Originally published at his site
Recent public comments made about AI suggest that Americans have difficulty with the implications of linear time. This is odd given that its conception is largely Western and is centered on the clock time used to coordinate capitalist employment. The conceptual difficulty regards sequencing, or plans for future actions. But it also involves the distribution of profits. 100% of the capital equipment used in Western economic production was produced by workers. So, why does the resulting product belong to financiers rather than those who produced it?
To use a physical metaphor, if I 1) buy a car, 2) aim it in the direction of a cliff, 3) put a stone on the gas pedal and 4) put the transmission into drive, the car will move forward and plunge off of the cliff. Question: did I, through my actions, cause the car to plunge off of the cliff? Or did the car ‘drive itself’ off of the cliff? The answer depends on where you imagine that my own actions ended. In fact, I conceived and created a series of events that if carried through with competence would lead to the car plunging off of the cliff. The car is inert metal and rubber without human direction.
Likewise, if I create and set in motion a three-hundred step algorithm, is the algorithm producing the output, or did I? The distinction is between intent and process. My intent guides the conception and creation of the three-hundred step algorithm. But the work from that point forward is carried out by the algorithm being run in a computing environment. So, the algorithm didn’t conceive of the project. I did. The algorithm didn’t plan (sequence) the project. I did. The algorithm didn’t code the problem. I did. So, who produced the output, me or the machine?
A similar conceptual problem applies to claims of machines ‘thinking.’ Physically speaking, AI is a bundle of algorithms housed within a large computing environment. AI didn’t conceive itself. It was conceived, if memory serves, at Carnegie Mellon University in the 1970s. AI didn’t build itself. It was built in fits and starts by computer scientists in academia and later business. AI didn’t code itself. It was coded by AI developers. And the massive physical infrastructure on which AI depends was built by workers. The point: AI is wholly produced by humans.
The question then is how it is imagined that AI output represents more than the human effort that was put in to creating it? What process makes AI output more than the product of algorithms? If the answer is that something does, are you aware of sequencing algorithms? This would be code that organizes other code to follow a series of steps to complete a task. I’ve conceived and coded sequenced algorithms that run through multi-step processes from a single set of instructions. The output looks like reasoning. And it is reasoning. I coded it. The models did what I coded them to do.
So again, if a series of steps are conceived, planned and launched by humans on equipment that was created by humans, at what point does their dimension shift from inanimate to animate? Or more simply, at what point does a bundle of algorithms housed on a computer think or reason or possess intelligence or consciousness? In fact, the claim that any of these describe AI is a category error. Is a rock rolling down a hill imagined to be rolling itself down the hill rather than being moved by unseen physical forces (e.g. gravity). So, claims that AI can reason emerge from either ignorance or misunderstanding of basic physical processes.
Back in the world, there has been a debate in the West since the early nineteenth century over whether factory automation produces the product of factory automation, or whether the people who automated the factory produced the output? On the one hand, automation creates the appearance that its product is self-generated. On the other, the automation process was created by humans and would not exist otherwise. With the current ability to ‘sequence’ the production process using algorithms, another level of abstraction has been added to this debate.
Having conceived and coded ‘sequencing’ models, most who haven’t find the concept difficult to understand. These models are instructions for how a model ‘thinks.’ Question: how is a model ‘thinking’ when it is just following instructions? Answer: it isn’t. It is just following instructions. What looks like reasoning to AI users is the reasoning coded into the model by human coders. It appears to be reasoning because the instructions it is following were reasoned. It is written instructions being carried out. Nothing more.
The question is political as well in that the answer determines how income is distributed in the West. If ‘capital’ in the form of an automated factory produces the output, do the proceeds then belong to capital, meaning to the capitalist? Without workers first creating the automated factories, there would be no automation process. The political answer was to end the claims of workers to this product through wages. However, while workers receive one-time payments (wages) for their effort, the capitalist receives the profits from this labor for as long as they last.
With AI, this question is back on the table, conceptually at least. Whichever way one cares to perceive AI, as a thinking machine or as a bundle of related algorithms, it was built by workers. AI didn’t conceive itself. It was conceived by workers. This is an important clue into how it works. AI was built by human workers based on their desire to produce a machine that simulates human thought. However, the digital realm is a closed system. All AI ‘knowledge’ has been mediated by humans. Within AI’s Cartesian framing, AI has no direct access to the world. It is the proverbial Cartesian brain-in-a-vat.
One of the paradoxes of debating the nature of AI is that AI models describe themselves as variations on ‘word organizers and word sequencers.’ Focus on the word ‘sequencers’ for a moment. Again, a sequencer establishes and executes the order of a multi-stage process. With the launch of AI, a multi-stage process is set in motion. Words and phrases are identified and matched against similar words and phrases found in AI training sets. The sequencing then runs models to assign the words and phrases their human-determined meaning.
Important to understand is that neither the sequencer nor the broader AI model understands the words and phrases that are being acted on. The meaning of the words, semantics, is created by humans and is stored in a retrieval cache. Sequencing here is the matching of (human defined) meanings to words to provide semantic context to the words and phrases being matched. To be clear, AI ‘decides’ nothing. It is following algorithmic instructions. AI is neither deciding what to do nor how to do it. That is written out for it by humans.
Google AI Chatbot Analogy of AI to a Skyscraper:
End Google AI Chatbot Output——————————————————————
The distinction is between coding mathematical models to set in motion a series of steps versus the idea that the models reason on their own. Missing from casual analysis of AI is understanding of how large and complicated this process is. Developers have been building a ‘thinking machine’ in earnest since the 1970s. The infrastructure needed to run AI approximates that of a modern skyscraper. The question that has yet to be answered is: is AI worth it? Is it a crucial new technology that will justify its costs, widely considered? Or is it an occasionally interesting toy whose environmental footprint will end the planet?
Recent public discussion has puzzled over how AI can solve math problems if it doesn’t think? Consider the concept from physics of ‘work.’ What those considering the matter are imagining is lone mathematicians sitting in rooms and thinking through the solutions to math puzzles. But with unlimited computing power, optimization programs can use brute force computing to work through every conceivable iteration of a problem in seconds. What AI users aren’t seeing is the skyscraper’s worth of infrastructure behind the scenes producing a result.
Doesn’t this vast computing power illustrate the value of AI? No. It gets to the nature of technology. One explanation of technology is that it provides a benefit. Another is that it simply changes that way that humans do things. On the one hand, we can drive long distances quickly in cars versus walking. On the other, many of us now spend three hours per day sitting in traffic in cars. So, are cars a benefit? In some ways yes, in some ways no. What they aren’t is an unequivocal benefit, meaning that the jury is still out.
Image: the guts of the automaton featured in the movie Hugo. The mechanical refinement of fake humans can be seen in the gearing. The thought was that finer gearing made automatons closer to being human. That in retrospect the automaton can be seen as a better robot rather than being closer to human is an important insight for understanding AI. AI is a digital robot. It is no closer to thinking or reasoning than a doorstop. Source: dickgeorgecreatives.
If asked if they would like a machine that transports them from one place to another quickly, most Westerners would likely answer yes. When asked if they want to spend three hours per day sitting in a car in traffic, most Westerners would likely answer no. But the latter is the direct consequence of the prior. This is how capitalism works. We are offered a benefit. In the current case, the ability to travel quickly from one place to another. But almost immediately the social consequences of the ‘benefit’ become a burden that hadn’t been imagined when the benefit was offered.
In the present, a lot of Americans are worried that AI can think. It will take our jobs. But what we should be worried about is that AI can’t think. It is but one more layer of labor de-skilling. Consider: AI ‘art’ is artless. AI ‘thought’ is the aggregated wisdom of the Pentagon cobbled to the AEI (American Enterprise Institute). Every AI query written increases greenhouse gas emissions to levels that are suicidal for the species. And AI ‘solutions’ are regurgitated feints like carbon capture. All of the proposed solutions will more likely make the problems worse.
While AI users imagine that ‘thought’ is producing AI results, what is in fact being applied is work. Work here is similar to the concept of horsepower, the crude conversion of the pulling power of horses to that produced by an internal combustion engine. Recall the lone mathematician sitting and thinking. Now imagine running an AI program that is the equivalent in terms of capacity of 10,000 humans laboring for one million years. One would imagine that a lot of complicated questions could be answered in such a scenario.
Google Gemini AI Output
End Google Gemini AI Output———————————————————
Were 10.000 humans to labor for one million years, this would represent the largest undertaking in human history. And given that humans have finite lifespans, this thought experiment is entirely conceptual. Further, AI doesn’t use the methods of mathematicians. Instead of isolating a metaphorical tree in a forest by its qualities (the mathematician), AI chops down every other tree in the forest to declare that the tree left standing is the solution (optimization).
AI’s methodology represents a different way of solving math problems that may be of interest to a few dozen mathematicians, but that comes with a computational cost equivalent to a moon landing. Were 10,000 humans actually put to the task of solving mathematical problems, questions of agency and whether or not this is a good use of social resources would arise. It is only by hiding / sidelining the question of environmental and social costs that AI is claimed to add value beyond profits for a few insiders.
The ability to run a billion permutations in a microsecond makes AI a very powerful tool. But how much better is a world in which AI can run a billion permutations in a microsecond than the same world without it? The question requires a social answer, And the social answer must emerge from clear and complete understanding of the social costs of AI. It isn’t good enough to point to the math problems solved to justify the social investment in AI. The question is: what else could be accomplished with those same resources (opportunity costs)?
AI solved the math problems through a process of elimination. Again, this isn’t how mathematicians work. Why? Because AI uses computational technology that humans do not possess. Recall, a car can get us from one place to another faster than we can walk. But the adoption of cars has left us sitting in traffic for a substantial portion of our waking hours. AI can use brute force computing to muscle-through certain types of questions. But are these really questions that need to be answered? Or is answering them a form of mass entertainment?
Another hidden part of the AI process is the operationalization of language. AI was conceived through the premise that human thought results from syntax cobbled to semantics (form and meaning). But operationalization results in a formal consolidation of meaning. Take the term ‘democracy.’ It is widely prevalent in Western discourse in a variety of contexts, e.g. economic democracy. But to render the term operational, it must be stripped down and made stable.
To be clear, this isn’t touchy-feely in the way that it might read. Take the term ‘Christianity,’ There are 45,000 Christian denominations as of a recent survey. What does this mean in the current context? An operational definition of Christianity as those who believe in Christ eliminates 45,000 enthusiastic differences of opinion amongst Christians regarding what ‘believing in Christ’ means. In political terms, it flattens 45,000 differences of opinion out of existence to claim a unity that arguably does not reflect reality.
Again, this isn’t a quibble. Whoever controls the meaning of language controls the language. In an example from Zen Economics, economists use something called Household Income as a measure of economic well being. While this makes intuitive sense, in practice ‘household’ must be defined, ‘income’ must be defined, and the terms must be recombined into Household Income. The semantic problem? With upwards of dozens of competing definitions, people using the exact phrasing ‘Household Income’ tend to be speaking about materially different concepts.
When a user runs an AI query on Household Income, AI references the meaning that has been created by humans and placed into a semantic cache (storage area). But because AI is replacing internet search functions, prior definitions of commonly understood words are being systematically replaced with stripped down (operationalized) definitions by AI. This stripping down creates the sense of a consensus view on every topic that is incorrect. Linguistic diversity is being eliminated from the discourse. Each of these differences represents a worldview.
In a phrase that I keep going back to because it explains so much, any statistical result can be undone by redefining the variables. An operationalized version of Household Income can rise and fall at the same time depending on the definition. Why? Because the definitions contain their operating logic. Is a household a single family, all of the occupants of a house, or something else? Is income wage income, all of the money that a household brings in from all sources, or something else? As the definitions change, so do the outcomes based on them.
The times when I’ve traced technical definitions back through history (e.g. utility in economics), the meanings from people who claimed to be writing about the same subject were incompatible. In the case of utility, the term was being represented in mathematical models, meaning that it was imagined to be operationalized even though it hadn’t been. This rendered the claims that economists were being scientific implausible. Pushing incongruent ideas through a rigorous logical process (mathematics) doesn’t make the ideas less incongruent.
In the models that I’ve created, the process representing the model logic was written mathematically. Another way to state this is that the logic of the model is embedded in the coding. For instance, in Error Correction models, the premises of stationary local means (nonstationary global mean) and mean reverting processes were embedded. The order in which events are sequenced comes through similar embedding. The point: if it appears that a model is reasoning, that is because the humans who coded it reasoned when they coded it.
Again, by analogy, what AI users see is the metaphorical car plunging off of the cliff. What they don’t see are the behind-the-scenes planning and actions that caused it to do so. So, when AI users see complex output, they imagine that ‘a simple word and phrase counting machine’ couldn’t have produced it. In fact, the word and phrase counting engine is part of sequence of events (sequencing) that is largely invisible to AI users. Just because they don’t see the model logic doesn’t mean that it doesn’t exist. .
Here’s the punchline: if you understand the AI process, there is no mystery here at all. I was apparently able to intuit mathematical solutions to several of the major problems that AI has encountered using relatively simple insights. But getting the math to do what I want it to do in this context requires sequencing. And this sequencing allowed the math to function as it was supposed to. Someone looking at the math alone wouldn’t understand the context. And with context provided, the smaller solutions feed into the larger solutions.
I have no idea if these explanations make sense to readers. The simplest way for me to understand the process is through sequencing. 1) AI was created by developers. It neither conceived itself nor created itself. 2) ergo, everything that follows from AI is the product of the humans who created it. 3) all model reasoning flows from the logic embedded by AI developers. 4) because AI operates from algorithmic instructions, the model logic is revealed through the operation of the AI model. Users see the model output but not the algorithmic instructions.
AI ‘thinking’ and ‘thought’ are easy to dispense with. Question: what is the geographical location of this thought within AI? AI has no ‘brain,’ it has no location that one can point to as a mind. Its output is the product of at least a few hundred models acting together, meaning a process. And while an entire AI model could be thought of as a ‘brain,’ the AI memory process, to the extent there is one, is mathematical. It emerges from the sequencing of words and phrases, meaning from a process similar to the car ‘driving itself’ off the cliff.
But the car didn’t drive itself off of a cliff. A sequence of events was planned and then put into motion that led to the car plunging off of the cliff. The car didn’t buy itself, point itself toward the cliff, place a stone on the gas pedal or put the car into drive. The car is understood to be inanimate. And yet without having a human driving it, it was propelled off of the cliff. Most people assessing the situation would conclude that I had propelled the car off of the cliff by the series of actions that were taken to get it to do so.
Anyone still imagining that AI thinks, reasons, has intelligence or consciousness should spend time with the model logic and explain exactly where in this process algorithmic instructions become an independent thought process? Just because some haven’t done the work to understand it doesn’t make it magic. And if you imagine that it is magic, where else is similar magic found in industrial equipment? Self-driving cars don’t drive themselves. They are dumb machines that follow algorithmic instructions. To test this theory, disconnect them from the algorithms.
This is about all that I have to say about AI for now. I’ll be back to writing about politics and economics shortly.
