The Productivity Paradox: AI/ML promises a Trojan Horse, but not yet

Machine Learning (ML), the engine of artificial intelligence or simulated cognition (AI), has come of age.

But the press and chatter around OpenAI’s ChatGPT can make believe a fake headline like, “Enterprise Software faces $2T business model disruption - Microsoft bets on bottom feeding Trojan Horse”.?Or that “Google is history!”.

Social, cultural, historical, political, religious, and family narratives can thrive on bias or fake, but it is not value-accretive to business, financial, scientific, medical, engineering, and judicial discussions.??

Don’t get me wrong. I am excited by the potential of ChatGPT released to public audiences to test and play.?It generates a predictive continuation of questions from data. The machine has been pre-trained to recognize patterns within to transform and display it in flowing natural language with perfect diction. So, it can replace Wikipedia quickly, and perhaps apply more finesse to moderator bias.

1. Why is AI/ML important?

AI/ML is the last grail for exponentially increasing enterprise productivity since the Dartmouth conference of 1956. But the endpoint vision and objectives are so vaguely framed today that in popular perception, fear pervades in equal measure with hope and hype.

AI/ML is a natural science. It is at the intersection of physics, mathematics, and neuroscience. It seeks to simulate naturally occurring phenomena. This confluence postulates that nothing is real locally, where you or I may be at the same time. The discoverer was awarded a Nobel prize in 2022.?

Neuroscience says the same environment, say grass, looks quite different when seen from a human body, a dog, or a grasshopper. In fact, nothing is “seen” except “programmed” neurological responses to sensory data by the consciousness resident in any body.?

In neti-neti iterative simulated experiments on human cognition, space, time, and quantum particles are deduced as merely dependent variables on a single independent one – consciousness.?By extension, I can say that consciousness appears fragmented only on account of the “apparent” different bodies it possesses, none of which are locally real, or else a separate and independent origin for each of the zillion fragments needs explanation.

Why is this important to AI/ML??

It means that the creation of an “artificial” machine that “mimics” natural phenomena in all ways is well within the realm of possibility. But not the creation of sentient consciousness in the manner computer scientists often pine for. ?Neuralink is at this frontier with a focus on repairing and rewiring brain functions in the disabled. But invasive add-ons are not needed to realize the full potential of the brain. This is well documented.

The best hope is a machine that has programmed sensory inputs and outputs sophisticated enough for consciousness to consider possessing it, as a medium for its never-ending pursuit of experiencing happiness or another defined purpose. Until then, AI/ML can only be an extension of the human mind or environment. Even then, it cannot be a different consciousness, only be resident in a different type of body/machine medium.

ChatGPT, interesting as it is, is far, far away from that possibility. As are alternatives on the market, including DeepMind derived. Let’s not forget the IBM Watson fiasco. In fact, effective machine substitution of even a reasonable fraction of human mind and body functions is decades away.?

But that’s not to say enterprise activities, processes and tasks cannot be progressively automated with ML

The enterprise, like any societal entity before it, was created to serve human interests and is merely a small subset of the latter. ?

At the heart of enterprise activity is prediction. Humans haven’t been great at it. For example, the market uptake of Windows XP version mix was the opposite of what Microsoft had predicted, to its pleasant surprise. A core requirement of managers is judgment or making decisions under uncertainty. Making choices from predicted outcomes is challenging. The value of an enterprise is the value of its predicted performance.

Management expectations of financial and operating performance are the sum of predicted outcomes on business tasks weighted by preferred choice or judgment. Whether it is deciding on an investment, insurance claim or medical diagnosis. Or making tradeoffs as a shelf-space constrained retailer to stock more $10 clothing that has a 70% predicted chance of selling out in a week or high margin $100 clothing that has a 20% chance. Advancements in AI/ML have significantly lowered prediction costs. Outcomes and timing are increasingly reliable as data is its currency and more of it is available. Therefore, potential applications are numerous.

The difference between a data-rich AI/ML driven enterprise and one not, is that the former would be self-learning, self-predicting, self-deciding and self-executing on pre-programmed human business tasks for defined purposes/outcomes with only limited human supervision.

The first technology company that delivers the full promise of ML/automation, or the enterprise customer that implements it, can become a shareholder value sink for its industry in a giant sucking sound. How long might that take? That's what management needs to bet on.

ML-led automation represents an integrated whole of human tasks, more effective in scaling the pursuit of enterprise objectives at a fraction of the cost than the sum of the human work it substitutes. Precisely because it has no consciousness. It does what it is taught and told to do.

Therefore, legislation and courts can point to accountability for misuse, the greatest deterrent.

2. What is the ideal ML-inside productivity solution?

Highly effective ML integrates quickly with enterprise datasets. With initial training, it self-learns to automatically organize and predict through value creating patterns in continuously generated economic, human, machine, and technical data. Queries in natural language deliver ready-to-use insights for decisions and reports.

Linked to rapidly evolving factory automation and robotics, it should mine, feed, consume, process, grow and reproduce data itself. And automatically leverage insights to predict, decide, develop and manufacture/deliver customer-value perfect products.

We are already in the third generation of ML technology. There are many vendors. It has become almost perfect.?The reasons are:

1.??????The value of good data is no longer limited. More good data fed into the machine yields better or almost error-free results. Earlier, beyond a point, more data made no difference to errors.

2.??????Algorithms are more efficient with parallel processing and matrix computation.

3.??????The capacity and cost of needed computing resources is no longer a constraint. Especially for a Google that threw out the kitchen sink in favor of a tailored-for-ML technology stack from semiconductors up.

3. Yet, why are we nowhere near productivity utopia?

Three independent variables determine AI/ML adoption in the enterprise.

1.??????Most enterprise data is not ready-to-use for ML. Much of it is “dirty”, a function of many generations and multiplicity of legacy databases, applications, taxonomy, and data-entry discipline. It requires extensive scrubbing. It isn’t in the self-interest of corporate IT staff to make it easy.

2.??????Business expertise has not prioritized ML so far. The top-100 managers of any large enterprise are crucial to design and prioritize the comprehensive set of starter questions needed to draw predictions from patterns of critical insights and reports from the data. This is critical even to prioritize which datasets to focus on for scrubbing and for what business purpose.

3.??????Credible professionals to bridge business expertise and ML technical needs for efficient and effective coupling are too few.

To compensate:

Technology companies can leverage quantum computing to enlarge and enrich limited clean data sets to better train machine learning models for problem-solving.?AI/ML algorithms can be redesigned to be more modular. ML weighting systems on input variables and tradeoffs on choices to determine desired output, a function of business expertise, can be made progressive instead of having to reset each time for environmental changes. Tokenization and data transformation products can be made more effective for ML use.

Enterprises should start by solving specific business problems. Technology product selection is secondary. They can leverage experienced advisors as proxies for senior management to help select and prioritize activities that are ready to be machine trained (e.g. Uber’s arrival time drives pricing and customer experience) and incentivize staff to cooperate in the transformation. AI/ML models from 2-3 vendors can be tried out in parallel to see how they perform for automating specific activities.

4. But incremental AI/ML driven productivity applications will grind through

For example, while at Microsoft, one of my teams addressed suspected price leakage with business expertise. An executive dashboard based on a worldwide pricing waterfall for 250 Windows platform products sold in seventy-five countries under hundreds of pricing/licensing programs was developed. Sources of price leakage from the stated policy for discount structure and escalations were identified. So, P&L management could identify and control prevalent deviance in near real-time, now predictable. While the project was completed in six weeks, locating the data from sales operations and finance, reviewing it, and cleaning it to port and use took most of the time.?

High-end white-collar work, in professions the author is familiar with, can be automated. ML can be trained in much of the investment banking work from putting together pitch books to valuation memos and board presentations augmented by human error checking and exception escalations. ML is already used to automate trading and augment research for investment decisions by asset managers. It can augment leverage for scalable consulting projects that work off a play book. And automate preparation of legal briefs, opinions, and documents; there is even an AI tool representing a client in court.?

Industries that have lagged digitization can skip a generation to become hyper efficient through AI. Take healthcare. America is mediocre among developed countries for a fancy price. Payors are not patients. And are disintermediated many times over. ?AI has the potential to displace intermediary tasks.

Satya Nadella tells a story of a vendor who integrated ChatGPT with data on government programs aimed at rural India alongside a chat bot and vernacular translation engine. A farmer talked through WhatsApp to get ChatGPT to complete a transaction with the government within minutes. The eloquent presentation of unprecedentedly quick product adoption and exponential productivity increase left me wondering if Nadella believes government workers can be automated too!

At a college seminar class in Massachusetts, I was asked to present learnings on manufacturing automation. I concluded that factories of tomorrow will be like vending machines. Press a button and discretionary wants are made to order and delivered just in time. Basic needs of food, clothing, shelter, and healthcare are satisfied on-demand. Only those who want to work creatively will. The professor took exception to the last act and mumbled derisively for the rest of the class. Until he was reminded that the alternative is starvation, riots and revolution repeated over history. We can’t have our cake and eat it too!

5. And so, “The Productivity Paradox”

AI/ML like all good things is a Trojan Horse that consumes from within. The allure of more per-labor-capita output hides built-in human redundancy.

In the theoretical history of the world until the 21st century, increasing automation delivered more food, better healthcare, and essentials cheaper and quicker to more people. Combined with lowered death rates, populations grew everywhere. Because food was taken care of, other needs materialized.?Businesses and economies grew in diverse ways. The aspirational case for AI/ML rests on this rosy narrative as a precedent.

But periods of high unemployment have accompanied nonlinear growth in productivity through automation since the Black Death, including during the industrial revolution in Europe. It was tempered by colonialization, effectively displacing workers in traditional industries in the colonies in favor of manufacturing labor in Europe. Not different from displacing domestic US labor in favor of China’s with global trade. It was assuaged by geographic expansion out of Britain and Europe into the Americas, creating new food-producing land, increasing local populations, and therefore manufacturing demand.

Even that did not prevent two enormous labor-led bloody communist revolutions with forced redistributions of capital. Unpreceded world wars attempted capture of other countries’ wealth accumulation from automation and colonialization through industrialized arms production.

So, despite the aspirational theory of more good from more automation for everyone, the actual realization is more output from/for fewer people and increasing wealth and income concentration. In fact, disparities in wealth, income or taxation and its use were a factor in the American revolutionary war and the civil war.

Finally, the local impact of automation can vary. Villages, towns, and cities can and have become ghost or extinct as traditional local output declines and populations perish or emigrate. Companies too go bankrupt for the same reason or are amalgamated into others.

6. Even so, the AI/ML road is forked with silver linings

If the full potential of AI/ML is realized in reasonable time, only a few companies survive. Each with enormous market and wealth concentration. Optimized redistribution of wealth and income to the rest of the population becomes necessary for demand and sustenance. When basic needs are satisfied without attendant effort, the usual recreation is reproduction. That makes for growth in population and output. How society will re-evolve thenceforth in a cyclical world is anybody’s guess.

If AI/ML gets implemented in bits and pieces, with like productivity gains, it is incrementalism in an era of unprecedented demographic decline. So, it evens out; nothing changes.

If AI/ML fails to make meaningful productivity gains, demographic declines are not compensated for by machines. Capital concentration slowly gets redistributed through higher wages.

Time is a great leveler. It swallows all!

? 2023 Vijaender V Takhan for original and derivative works in this document.?The author thanks several experts in the computer sciences for reviewing material related to AI/ML technology in the document. References are noted inline as links to the World Wide Web as of January 25, 2023.

#ML #machinelearning #Enterprisesoftware #AI #Artificialintelligence #OpenAI #ChatGPT #Microsoft #google #deepmind #ibm #robotics #automation #paradox #productivity #transformation #trojan #neuralink