There is No Innovation Without an Invoice

Introduction

Success of technology depends on the value it adds to the business (consumer and enterprise).

Some technologies take long time to mature and has profound impact on humanity. Artificial Intelligence (AI) is one among them. Activities around AI is constantly increasing leading to fears of a AI winter or bubble similar to 2000's. At times platforms like LinkedIn is becoming a battle ground for techno optimists and techno pessimists with many being mute spectators to these extreme viewpoints. If you look back, one thing is clear, there is no innovation without an invoice. The technology has to add value in order to sustain in the long run.

There are pros and cons to any technology. AI is no exception. Beyond all the hype around AI, we need some fundamental thinking to take the field farther. We might hit road blocks with what we currently have, but doesn't mean we need to end up with AI winter. This article tries to explore some thoughts around how solutions are built currently in AI and do we need to relook at some of it.

The power of unknown

Learning complex patterns through neural networks is well established. We learn patterns in time, space, and other dimensions to solve problems, achieve our goals, and navigate the world. The power of unknown is a concept that most algorithms don't understand today. To explain it simply, a model trained on images of cats and dogs don't understand anything beyond cats and dogs. If you show an elephant to such a model, it is going to fit the elephant to either a cat or a dog which it understands. The context is limited to what it is trained on. It is optimized to exploit the information as opposed to explore beyond the context.

Exploration is what keeps the world rolling where as exploitation is what keeps it progressing.

In other words, exploration drives discovery and innovation, pushing boundaries and uncovering new possibilities, while exploitation leverages known resources, ideas, and methods to achieve tangible outcomes and improvements. Both are essential—exploration fuels the potential for new opportunities, and exploitation turns that potential into reality by refining and applying what we already know.

The past offers insights, but it can't guarantee what lies ahead

Understanding the patterns in the past is important in evaluating / predicting the future, but that alone is not sufficient for predicting the future with certainty. The quote “Those who cannot remember the past are condemned to repeat it” is an example of this concept. Most of us can see time in linear dimension, but rarely do we see time in circular dimension. The world mostly works in circular dimension, the seasons of the year are a cyclic time axis. In short, looking back is important, but not sufficient when projecting to the future. New forces reshapes the past insights to provide a different future.

You shall know a word by the company it keeps

The quote "You shall know a word by the company it keeps" is again an example of exploitation of knowledge and not exploration of it. The languages of the world have evolved over time not just by exploiting the past understanding, but in addition by exploring the possibilities and subjecting such possibilities to the test of time. All the current criticisms of Large Language Models (LLM's) as stochastic parrot that don't truly comprehend language stems from this concept.

New words and concepts in languages will evolve constantly and they might not have company to start with, but eventually will get the company.

That is how exploration happens. Some who had the company might lose it over a period of time. It is subjected to short and long time patterns in evolution.

You cannot solve problems with brute force and infinite data alone

The progress of neural network based architectures in the past decade is amazing. We started with models in 60 million range like AlexNet in 2011 and scaled all the way to trillion parameter models as of today. The data also scaled along with the number of parameters. What is worrying is that the value derived by businesses for identifying patterns has not scaled in proportional dimension based on observations. Technology is here to solve a business problem at a reasonable cost making it viable for business to sustain.

The ever increasing model size and amount of data used for deriving the models doesn't justify the cost to value ratio required by businesses to sustain.

Computational cost of model training and inference is going through roof making it unsustainable. New techniques evolve periodically to reduce the computation cost, but such techniques are mostly an after thought. Intelligence is counter intuitive to brute force and infinite data.

The current scaling laws of LLMs on how a model's quality evolves with increases in its size, training data volume, and computational resources is not a sign of intelligence.

Trying to achieve Artificial General Intelligence (AGI) through the route of more compute and more data will not see the end of the tunnel.

Accuracy is not a measure of abstraction or generalization

How important is specific details to solving a problem? Mostly not. To understand if the image shown has a cat in it, it doesn't matter if the image is color or black and white, if the cat is standing or sitting or lying down, if it is with other animals like dogs, etc. The amount of details required to answer such a question is far less information than that is presented to us.

Today we are confusing accuracy with abstraction and generalization. We think high accuracy for a model means it has well abstracted and generalized which is not true.

In short, abstraction simplifies by removing specifics, often at the cost of some accuracy, while generalization broadens a concept to encompass more cases, aiming to retain as much accuracy as possible across diverse instances. Removing noise enhances both abstraction and generalization by improving the accuracy of the insights derived from the data. This ensures that the simplified models and broader concepts we develop are based on clear, relevant information rather than misleading or extraneous details.

Adversarial process or Competitive co-evolution

The field of machine learning started solving single tasks, then multi tasks, and now slowly getting into multi domains. Agent based systems are the next big thing predicted in this field.

One of the patterns often used in evolving the solution is what is called competitive co-evolution or adversarial process.

The idea is to make two intelligent systems, for example two agents compete against each other to achieve higher goals and ensure that both evolve over time due to the competition. Predator-pray is a classic example from nature. Patterns like these are important in the evolution process, but not sufficient to achieve higher goals. In short, you cannot make two dumb agents to compete against each other thinking that they will evolve over a period of time.

Broker / Agents without bias

Machine learning is all about learning from data. What happens if the existing data is biased and the generated contents are also biased? The outcomes will be biased. Evolution manages bias through a continuous process of variation, selection, and adaptation, which often filters out traits that don’t contribute to survival or reproduction. Evolution doesn’t eliminate bias directly; instead, it balances it through diversity, selection pressures, adaptation, and environmental shifts. It needs to be seen how next generation AI/ML systems deal with such issues for adding value to the business process as opposed to hindering it.

Challenges of long tails

Optical Character Recognition (OCR) is an old technology (more than 6 decades). OCR systems have evolved over time with high accuracies. They still struggle with long tail issues. In AI and machine learning, long-tail problems refer to rare or infrequent cases that occur at the tail end of a probability distribution—meaning they are uncommon events or edge cases that standard models struggle to handle. Despite their rarity, these cases are often critical, as they can lead to significant errors, biases, or gaps in model performance. Human in the loop is a solution, but knowing when to hand it over to a human is a challenge. Probability or confidence alone might not be sufficient for handing over to humans.

Moving towards stochastic thinking

We have developed software mostly with determinism. The trend is going to change as we learn from large amounts of data which constantly change. Most business software's written today are with deterministic thinking. The expect certain outcomes in a consistent way. The process of software development, the internal processes of business software's, the evaluation of outcomes need rethinking when we move towards stochastic thinking.

AI versus Human

Planes are an adaptation of birds for us to fly and carry our goods from one place to another. They do not exactly behave like birds, but are useful for us in many ways. Similarly, the neural networks are an adaptation of the human brain. I don't see much value comparing AI/ML with human brain or human system.

Technology becomes useful when they generate value to business in a sustainable fashion.

It doesn't mean we should not bridge the gap between natural systems and man made systems because evolution in nature is a long term process. Such distractions could be avoided for the general growth of technology.

Conclusion

Achieving Artificial General Intelligence (AGI) with current tools and technologies in AI/ML space looks difficult. The fight between techno optimists and techno pessimists will continue in multiple forums. We need rational thinking and evolve new ways to identify patterns, apply the most applicable one in a given context, to achieve business value in a sustainable way. It is very clear that no technology will survive just with innovation without generating invoice.