The Road to AGI is Longer Than You Think

In 1965, Time Magazine made bold projections about the wonders awaiting us from the burgeoning field of technology. While we have seen technological wonders in the last 50 years, almost none of the predictions featured in the magazine came to pass:

"Men such as IBM Economist Joseph Froomkin feel that automation will eventually bring about a 20-hour work week, perhaps within a century, thus creating a mass leisure class. Some of the more radical prophets foresee the time when as little as 2% of the work force will be employed, warn that the whole concept of people as producers of goods and services will become obsolete as automation advances. Even the most moderate estimates of automation's progress show that millions of people will have to adjust to leisurely, 'nonfunctional' lives, a switch that will entail both an economic wrench and a severe test of the deeply ingrained ethic that work is the good and necessary calling of man."

Technology experts continue to overestimate the positive impact of technological advances on the average person. In fact, many recent pronouncements have a very similar ring to the quote above. This line of thought is particularly pervasive in the Artificial Intelligence space today. It's understandable that these sweeping claims are appearing anew – perhaps no other technological advancement has advanced so rapidly since the dawn of the information era. While these accomplishments are remarkable, the fantastical claims that artificial general intelligence (AGI) is just around the corner is incorrect for several reasons.

"The Last Mile"

Nearly every seasoned engineer is familiar with the 90/10 rule, which states that 90% of the work required to finish a project will take roughly 10% of the timeline. The last 10% of work will consumed 90% of the time. While this rule of thumb isn't always a perfect indicator, we see this play out repeatedly.

Five years ago, Tesla appeared poised to deliver a fully autonomous Level 5 vehicle in the next few years; however, the cars manufactured today remain at a humble partial automation (Level 2). Microprocessor design is another example. Transistor size has decreased far more slowly over the last decade than in previous decades. Each successive decade has seen a decrease in speed with which transistors have shrunk. As it turns out, Moore's Law has a limit. This slowed progress mainly stems from significantly more difficult engineering problems as density increases beyond a certain point. Quantum effects such as electron tunneling, where electrons can pass through an extremely thin gate, suddenly become major roadblocks.

Challenge Parity

The trajectory of progress is uncertain and often veers off in unexpected directions. For instance, while the digital age promised enhanced connectivity and access to information, it also gave rise to issues like misinformation, cyberbullying, and digital addiction – challenges that were scarcely anticipated as we heralded the arrival of the internet area. This tendency to overlook potential pitfalls in the face of new technology underscores a common shortfall in our predictive mental models: they often mirror the current zeitgeist and neglect the nuanced complexities of the future.

Systematic Underestimation of Inequality and Corporate Greed

The predominance of Silicon Valley as a hub for technological innovation and prediction can create a skewed perspective on the future of technology. The region's unique ecosystem of venture capital, start-ups, and cutting-edge research tends to foster an echo chamber of ideas and optimism, primarily driven by those who benefit most from technological advances. This demographic, often composed of affluent, technologically savvy individuals, may not fully grasp the broader social and economic challenges faced by less privileged communities worldwide. Consequently, predictions from this vantage point can overlook crucial issues such as digital divides, access to technology, and the varying impacts of automation on different socio-economic groups.

While the advancements in technology we have seen in recent years are impressive, we must approach predictions about the future of technology with caution. The road to AGI is longer than we think, and we must be mindful of the potential pitfalls and challenges that may arise along the way. It is important to consider the impact of technology on all members of society, especially those who may be less privileged. By taking a more nuanced and inclusive approach to technological progress, we can ensure that the benefits of these advancements are more widely shared, and that we are better prepared to address the challenges that lie ahead.

Interpretability in Machine Learning

Since OpenAI released its large language model (LLM) chatbot, ChatGPT, machine learning, and artificial intelligence have entered mainstream discourse. The reaction has been a mix of skepticism, trepidation, and panic as the public comes to terms with how this technology will shape our future. Many fail to realize that machine learning already shapes the present, and many developers have been grappling with introducing this technology into products and services for years. Machine learning models are used to make increasingly important decisions – from aiding physicians in diagnosing serious health issues to making financial decisions for customers.

How it Works

I strongly dislike the term "artificial intelligence" because what the phrase describes is a mirage. There is no complex thought process at work – the model doesn't even understand the information it is processing. In a nutshell, OpenAI's model powering ChatGPT calculates the statistically most probable next word given the immediately surrounding context based on the enormous amount of information developers used to train the model.

A Model?

Let's say we compiled an accurate dataset containing the time it takes for an object to fall from specific heights:

Height	Time
100 m	4.51 sec
200 m	6.39 sec
300 m	7.82 sec
400 m	9.03 sec
500 m	10.10 sec

What if we need to determine the time it takes for that object to fall from a distance we don't have data for? We build a model representing our data and either interpolate or extrapolate to find the answer:

Models for more complex calculations are often created with neural networks, mathematical systems that learn skills by analyzing vast amounts of data. A vast collection of nodes evaluate a specific function and pass the result to the next node. Simple neural networks can be expressed as mathematical functions, but as the number of variables and nodes increase, the model can become opaque to human comprehension.

The Interpretability Problem

Unfortunately, opening many complex models and providing a precise mathematical explanation for the decision is impossible. In other words, models often lack human interpretability and accountability. We often can't say, mathematically speaking, exactly how the network makes the distinction it does; we only know that its decisions align with those of a human. It doesn't require a keen imagination to see how this presents a problem in regulated, high-stakes decision-making.

Let's say John visits a lender and applies for a $37,000 small business loan. The lender needs to determine the probability that John will default on the loan, so they feed John's information into an algorithm, which computes a low score causing a denial. By law, the lender must provide John with a statement of the specific reasons for the denial. In this scenario, what do we tell John? Today, we can reverse engineer the model and provide a detailed answer, but even simple models of tomorrow will quickly test the limits of human understanding as computing resources become more powerful and less expensive. So how do we design accountable, transparent systems in the face of exponentially growing complexity?

Solutions?

Proponents of interpretable models suggest limiting the number of variables used in a model. The problem with this approach becomes apparent after considering how neural networks weigh variables. Models multiply results by coefficients that determine the relative importance of each variable or calculation before passing them to the next node. These coefficients and variables are often between 20 and 50 decimal places long, containing positive and negative numbers. While understanding the data underpinning a decision is essential, more is needed to truly elucidate a clear explanation. We can partially solve this problem by building tooling to abstract implementation details and provide a more intelligible overview of the model; however, this still only provides an approximation of the decision-making process.

Other thought leaders in machine learning argue that the most viable long-term solutions may not involve futile attempts to explain the model but should instead focus on auditing and regulating performance. Do large volumes of test data reveal statistical trends of bias? Does analyzing the training data show any gaps or irregularities that could result in harm? Unfortunately, this does not solve the issue in my hypothetical scenario above. I can't conclusively prove that my current decision was correct by pointing to past performance.

Technology is simply moving too rapidly to rely on regulations, which are, at best, a lagging remedy. We must pre-emptively work to build explainability into our models, but doing this in an understandable and actionable way will require rethinking our current AI architectures. We need forward-looking solutions that address bias at every stage of the development lifecycle with strong internal governance. Existing systems should undergo regular audits to ensure small changes haven't caused disparate impacts.

I can't help but feel very lucky to live in this transformative sliver of time, from the birth of the personal computer to the beginning of the internet age and the machine learning revolution. Today's developers and system architects have a massive responsibility to consider the impact of the technology they create. The future adoption of AI heavily depends on the trust we build in our systems today.