Can AI solve every problem?

Deterministic versus Probabilistic Predictability

Artificial Intelligence (AI) has captured the world’s imagination. While we have all been hearing about AI for decades, the progress over the past few years has been nothing short of spectacular.?

I had the opportunity to dig into AI capabilities while at Apstra , as we looked into whether it could help with the automation problems we aimed to solve. After we got acquired by Juniper, we were further exposed to AI since Juniper has a best in class solution with Mist .?

On the personal front, my daughter got introduced to Stable Diffusion by common friends (thank you Guido Appenzeller ), and they created some cool pictures of herself training for Top Gun! She is also building an AI powered robot at Evodyne Academy . I also invested in a couple of exciting companies that leveraged AI to solve some big problems - LLM-powered conversational search with Vectara , and surveys with Beehive.AI . Additionally, I completed an immersive AI Specialization program with Andrew Ng ’s DeepLearning.ai .

While all of that was happening, OpenAI launched ChatGPT , and it felt like all of the sudden, everyone started paying attention - no one could ignore AI any longer.?

AI versus ML

Before I continue, a point of clarification is warranted. This article focuses on Machine Learning (ML) and Deep Learning (DL). Having said that, I use the acronym AI because today, AI and ML are used interchangeably. When the public refers to AI today, they are likely referring to Machine Learning.?

This is technically inaccurate. ML only consists of one set of AI techniques. AI has traditionally had a broader meaning and encompasses any system that aims to mimic and in many cases exceed human capabilities. To cite an example that I’m familiar with, Juniper Apstra is an Intent Based Networking solution designed to mimic and perfect the operations that a human performs in the process of managing networks. Therefore, Intent Based Networking fits nicely within the definition of AI. Yet Apstra and Intent Based Networking don’t make use of Machine Learning in their core algorithms, and therefore don’t fit what is generally thought of as AI.?

One can think of many systems that mimic and improve human behavior and that do not use Machine Learning - e.g. an airliner’s autopilot, or robots in a factory.?

However, I chose to use the AI terminology to describe ML/DL as most people do today, rather than hinge this article on this specific technical point. I’ll get back to AI versus ML at the end of the article.

Be skeptical when at the top of the hype cycle

We know we’re at the top of the hype cycle when we’re discussing the capabilities of AI and ChatGPT in friends and family circles, and amongst folks that would never consider themselves to be technologists.?

In my experience, it is wise to exhibit some skepticism with hype, especially when tech is hyped to this extreme level. As engineers, we’re technologists at heart and can be easily seduced by cool tech, and we’re sometimes seduced to a fault.?

When technology is as cool as AI, we have a natural temptation to want to use it to address any and all problems. I have seen this behavior many times in my career. Network engineers are tempted to believe BGP can solve all network related traffic challenges. When OpenFlow was created, it was viewed as the solution to all networking evil: getting rid of proprietary silicon; replacing all control plane protocols; and being the key ingredient to automating and managing networks. It turned out that OpenFlow solved none of these problems.

Start with the problem

The best approach to avoid this trap is to practice discipline, and start with the problem we’re intent on solving, rather than the technology. We shouldn’t be starting an AI project or an AI company because of AI. We should start a company or a project because there is a well defined problem we’re intent on solving. Let’s make sure we articulate that first.?

If we’re not crystal clear on the problem we’re solving, then it is wise to pause and figure that part as a starting point.?

I wonder how many SDN companies were formed on some of the premises that I outlined above. And I wonder how many AI companies are being funded on the same incorrect premise. I suspect quite a few.?

Once we have a good definition of the problem we’re trying to solve, then we can figure out what set of tools or technologies are best suited to solve the problem. If one goes through this process diligently, they’ll usually find that in the vast majority of cases, no one technology uniquely fits a given problem statement. Usually, you’ll need to combine multiple technologies or approaches to deliver an appropriate solution.

Here are three examples.

Self-driving cars?

AI plays an important role in self-driving software. AI software is used quite effectively to identify objects - mostly cars, but also pedestrians, red lights, etc. AI software performs this task by analyzing a continuous stream of data that are taken by the cameras, radars, and lidars on the car, then make a set of predictions about the future location of such objects over the next several seconds. For self-driving software to rely solely and rigidly on these object predictions, they’d have to be absolutely perfect - and when I say perfect, the technical definition is something that exceeds many 9’s of precision, typically 5. This is because any misidentification can cause an accident. If the software predicts a car in front of ours that is not there, it can brake violently and cause the trailing car to crash into us ; and if there is a car in front of you, or some other object that the software doesn’t see, then we could crash straight into it, with disastrous consequences.

However, in reality, these predictions are rarely perfect. Literature shows that the accuracy of the best ML prediction algorithms lies between 90 and 95%, which means that object predictions are imperfect in 5-10% of the cases. Hopefully it’s clear that risking hitting an object one out of 10 to 20 occurrences is an unacceptable level of performance.?

For self-driving software to work, we’ll need to use other technologies to complement AI. These are mechanisms based on rules that provide guardrails. This is where proximity sensors, which use echo-times from sound waves, come in handy, amongst other technologies. If the AI sees no objects in front of the car, but the proximity sensor clearly tells you that some obstacle is present, then it is sensible for this signal to override the AI prediction. The key here is that a proximity sensor deterministically predicts that something is in front of the car, whereas the AI predictions are probabilistic.?

This distinction between deterministic predictability and probabilistic predictability is a key distinction that is critical to understanding which problems AI is suited for.?

Another example is kinematic extrapolation. With Kinetic extrapolation, the ML-based prediction algorithm is compared against a heuristic-based common-sense prediction. E.g., "the light just turned green but there's a cross-traffic car about to enter the intersection at high speed".?Machine Learning is likely not to predict the danger because there has not been enough real-world data to train for such scenarios - yet, high-school physics can reasonably predict where the cross-traffic car is probably going to be over the next couple seconds. Same as with proximity detectors, it is sensible for Kinematic extrapolation to override the ML-based prediction and suggest to the self-driving software that the car should probably not enter the intersection.?

If it is critical for the problem we’re addressing that the prediction to the input is correct with a high degree of confidence, then we need deterministic predictability. Usually, we get deterministic predictability when you have information about our environment, and cause/effect rules on how to react to this information - e.g., if your proximity detector determines deterministically that an object is in front of the car, then you should likely stop the car. If you know that an object is moving at a given speed without any sign of slowing down, then you should likely stop the car.??

Network automation

Another example I’m intimately familiar with is network automation. I have been involved in building solutions to automate data center networks for more than a decade, most recently as the co-founder of Apstra. As businesses have digitally transformed, the network infrastructure has become the foundation on which everything else (apps, businesses) runs. It is critical that the network is up and delivers the proper level of performance and security for the apps that utilize it.?

For this reason, the network needs to be highly reliable. The standard network operators generally use is five 9’s of reliability - that is, the network can’t go down for more than 5 minutes over a year period. For network engineers to use software that will operate their network for them, they need to trust that it will not cause an outage or a security hole. In practice, this means that when a network engineer presses a button asking the software to take action on the network - say, to allow two end points to reach each other, or to enable a security policy that governs how two sets of end points are allowed to communicate - then they need to have the utmost degree of confidence that the software will do exactly what they want. Chance cannot play a role - same as we’ve seen with self driving software, they need deterministic predictability.

This is not to say that AI doesn’t have a role in network operations. When it comes to Wifi systems, using historical patterns to predict probabilistically whether the location of an Access Point is optimal for signal transmission can be useful. In fact, Juniper Mist has consistently used AI to slash support tickets to the benefit of their customers. AI is a great fit for that problem statement, whereas it is not for the purpose of having self-operating software run a data center network.

Search

Another example I am familiar with is Search. I think the botched launch of Bard clarified to a lot of people the limitations of probabilistic predictability, which is what you get with models such as Bard and ChatGPT. Let’s first recall what happened. When Bard was launched, Google sent a tweet showcasing Bard - in it, they posed Bard a question. To Google’s embarrassment, Bard’s answer to the question was inaccurate - and the way the public realized that it was nonsensical was by conducting a simple… Google search. While this shocked many, it wasn’t surprising. I had played with ChatGPT a fair amount and spent a decent amount of time trying to understand how it works, and ChatGPT exhibits similar behavior. One of the questions I asked ChatGPT is to tell me about myself. It got some of it right, like the fact that I was a co-founder and CEO of Apstra. But ChatGPT also told me that prior to that, I worked at Cisco. ChatGPT said it in a perfectly articulate way, and it was quite convincing. There was only one problem: I never worked for Cisco!?

When you understand how Deep Learning works, it becomes clear why ChatGPT would make such a mistake: In the wider web, which is what ChatGPT was trained on, my name is very often associated with co-workers that were associated with 思科 . Also, every company I worked for competed with Cisco in one way or another. So the constant in web searches was an association of my name with Cisco. Hence the probability of my name being associated with Cisco exceeded that of the other companies I worked for - Arista Networks or Big Switch Networks .?

To be clear, I believe there is significant value in Natural Language Models, as they allow algorithms to gain semantic understanding from written text - as opposed to old school keyword based approaches which keyed on specific keywords, rather than broader semantic meanings. However, the trick is to apply this value to the problems where it is the best fit, i.e. where it is the best tool for the job. And if search is the use case, some of these jobs are (1) parsing and interpreting human language (2) extracting semantic meaning from words and sentences, and (3) summarizing search results concisely using NLP-based summarizing techniques, among others.

Deterministic versus probabilistic predictability, and what it means to the suitability of AI

To recap, when we understand the probabilistic nature of how Deep Learning works, then we also understand its strengths and weaknesses. And when they do, then we ultimately understand the suitability of AI to specific jobs and use cases.

Recapping

As we have seen, when used as the sole prediction technique, there is no way to guarantee that the prediction of a Deep Learning network makes sense given a particular input that it’s not been trained on specifically. This means that in practice, AI is most generally suited for taking on “supporting” roles, rather than primary roles. It can provide helpful insights when it comes to troubleshooting networks, but it can’t be the primary operator of your network. It can point you in the right direction when it comes to best practices for coding, but it can’t be relied on to write the code for you.?

It can be trained to become a great co-pilot - but it can’t be the pilot.

So is AI intelligent??

Before I end this article - I often get asked the question as to whether AI is intelligent.

From the examples we’ve explored in this article, it is clear that AI does an outstanding job at pattern matching. Pattern matching is not the same as prediction, and is a critical function of the brain. As was described in “A Thousand Brains” by Jeff Hawkins, and studied thoroughly in Daniel Kahneman’s “Thinking Fast and Slow”, the human brain is masterful at learning patterns. The vast majority of our actions in our daily lives are instinctual, and can be thought of as “inferences” of the trained models in our brains. This is how the brain’s “System 1” works, using Daniel Kaheneman’s language. However, the brain also has another critical system, “System 2”, which is used to process exceptions, allows us to pose, reason and apply critical thinking when new situations arise - rather than applying simple interpolations or extrapolations from past experience. The field is progressing super fast, and there are attempts at capturing intuitive human “chain-of-thought” through training. It remains that neural networks, even when deep, large, or sophisticated struggle to perform this function. Simply put, AI can interpolate a solution from its training data, but struggles to extrapolate?sensibly beyond this distribution. Never-before-seen things are hard. Therefore, it appears clear that Machine Learning is missing a critical aspect of human-level intelligence.?

Machine Learning can still be built into AI machines that mimic and improve on human behavior, but will require to be complemented and often overridden by more structured algorithms that incorporate rules, logical deductions, heuristics, and cause/effect relationships.?Combining the pattern matching capabilities of ML with more structured approaches can, indeed, be extremely powerful.?

Going back to the original definition of AI, this is perhaps how we are likely to achieve truly intelligent AI.

Acknowledgments

I'd like to acknowledge those that spent time reviewing my draft, and for their valuable comments: Amr Awadallah , Shad Laws , Tom Schodorf , Michael Marcellin and my Apstra co-founders Aleksandar Sasha Ratkovic and David Cheriton. Thank you, your feedback made this article better.