Did They Find The Algorithm To Rule Them All?

OpenAI, in case you have been sleeping under a rock in the past four years, is an Artificial Intelligence research lab founded by Elon Musk and other tech titans, like Peter Thiel, Reid Hoffman, Sam Altman, and Jessica Livingston. It started as a non-profit organization in late 2015, but later switched to being for-profit and secured a whopping $1 billion investment from Microsoft. Along with their impressively deep pockets, they have an equally lofty mission statement - “to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity”. That is to say, they don’t only recognize the imminence of Artificial General Intelligence, they also feel an exigency to become the first ones to achieve it, so they can make sure it is not used to the detriment of humankind. After demonstrating some impressive but seemingly non-disruptive AI experiments, like an AI debate game and a robot hand that can solve the rubik's cube, they seemed to have caught their stride with an ongoing series of hugely impressive text generation AI models, called GPT (which stands for Generative Pre-trained Transformer). The most recent one, GPT3, was published in June 2020, and demonstrated amazing abilities to perform a wide variety of tasks with very few training examples. GPT3 was highly anticipated since the delay of its predecessor’s (GPT2) release by OpenAI, due to fears of misuses like automatic generation of fake news on social media.

Artificial Intelligence represents a wide area of research, focused on creating software that mimic abilities that are considered human-like or intelligent, such as reasoning, understanding language, or beating a human chess master. With that, in recent years, Artificial Intelligence (or AI in short) has become synonymous with one specific computer science technique called Machine Learning. Instead of hard-coding a set of instructions that are required to perform a task, Machine Learning algorithms are designed to achieve their tasks by implementing two phases: training and inference. The training software takes a big list of data points as an input, and uses it to generate the inference machine by iteratively optimizing inference parameters with a pre-programmed “loss function”. To give you some intuition of how it works, let’s take the example of image classification. You can think of an image classifier’s loss function like a judge that quantifies how wrong a guess for what’s in the image was. The image classifier itself is a classification function that uses some parameters to guess what’s in an image. To train this function, the algorithm starts with a random set of parameters and a list of pictures that a human already tagged for what’s in them. In every iteration, the random parameters of the classification function are slightly modified, to try and get the loss function “judge” to better score its guesses. Once the optimal parameters are found, they can be used for image classification in subsequent inference requests. GPT3 is therefore a type of machine learning algorithm.

Diving a little bit deeper, GPT3 is trained using large datasets of unlabeled texts to create a text generator that “infers” the correct sentence that best matches an input text. This description sounds very generic because it is designed to remain non-specific. GPT3 uses a machine learning feature called Transfer Learning to turn the non-specific knowledge it had learnt into a specific text generator in an additional training step that learns the pattern in another, much smaller, set of samples. The model itself is absolutely humongous. It is by far the largest language model ever created, with around 175 billion parameters that were trained on the largest text dataset - more than 8 million documents including every Wikipedia article in existence and other extremely large textual sources. It’s sheer size means that the cost of the compute power that’s required to train it is estimated by some to exceed $12 million. There are two main reasons for the huge waves GPT3 is making in the industry. First, it is the first ever model to demonstrate an amazing capability to learn seemingly unrelated language tasks, building on top of its massive knowledge to effectively perform tasks like translation, writing articles, and even some simple arithmetic computations. Second, and perhaps even more important, it has shown an exponential improvement over GPT2, with only a linear increase of model size, which represents the only material difference between the two. Put simply, this means that any task it can somewhat do today, can be improved to above human-level by only increasing the size of the model. This is something that was never thought possible before, and is hypothesized to be the path to that goal of achieving artificial general intelligence.

Excel magic using GPT3

The multitude of amazingly impressive natural language demonstrations makes me wonder - would GPT3 change the natural language conversational AI landscape? My opinion is that while the current state of GPT3 is not in itself a tectonic shift in conversational AI, it certainly shows that in the future there might be one algorithm to rule them all. There are a few reasons for it not being a tectonic shift just yet. It has shown very promising question answering abilities, but no contextual conversations yet. Even if there was a way to train it with full conversations such that it would predict the next turn to a contextual conversation, it would still be highly black-boxed. That means that there wouldn’t be a way to pre-approve the verbiage of answers it would provide, making it impractical for the enterprises to make use of it. Furthermore, customizing the conversation flow would only be possible by going into the samples and editing them to match the desirable outcome. That is, again, impractical for any real use case. With all that said, it is important to understand that OpenAI hasn’t mass-released the API that will let developers use their pre-trained model, so it’s hard to know for sure how it will behave in real-life applications.

Will Tesla Be The Last Car Insurance Company

In case you didn’t know, Tesla, the electric vehicle manufacturer, has been testing a car insurance service for its car owners in California for almost a year now. While the recently filed 10-K doesn’t list the new insurance service as a big contributor to revenues, it has surely been a resounding success. We know that, because last week, Tesla announced that it's going to create a revolutionary insurance company. It is widely believed that Tesla’s impetus for venturing into the insurance business was their assumption that autonomously driven Tesla vehicles are less likely to crash. With that, I actually think this is a much more profound move, with extremely far reaching implications for all incumbent car insurers. It might even be the last car insurance company.

Even though most people would consider new-age tech moguls like Elon Musk to be dismissive of traditional techniques like actuarial sciences, Musk said this in a July 22nd call with analysts: “I would love to have some high energy actuaries, especially. I have great respect for the actuarial profession. You guys are great at math. Please join Tesla, especially if you want to change things and you’re annoyed by how slow the industry is. This is the place to be. We want revolutionary actuaries”. In light of this information, and given the fact that Musk already used the name “The Boring Company” for something else, I suggest the name Pearson Insurance for this new venture, after the famous English founder of mathematical statistics, Karl Pearson. Pearson was the one who associated the regression line with the least squares estimate. A statistical technique that’s in the epicenter of the actuarial profession until this very day.

Pearson’s least squares estimate referred to as the Regression Line

Tesla's future seems to be a fully autonomous car as a service. In that future, people may not own cars, but rather subscribe to a car service. Similar to the lack of a Bus Ride Insurance or Train Insurance, when we get to that point, car insurance will no longer be an issue for the average Joe, who will just consume a transportation service. However, regardless of this probable eventuality, Teslas already come fully loaded with every type of sensor and camera you could imagine. This means that in a claim situation, our newly named Pearson Insurance technically doesn't need a claims adjuster to inspect the car or even to speak with the different parties, because all the data they would ever need could be provided by the car's hardware. Therefore, it's not hard to imagine a future in which the first notice of loss, the entire claim process, and even the payout will happen seamlessly and automatically. When a Tesla would crash, it would transmit over the air the 360 degree video of the event, the LIDAR recording, and the telematics straight to Pearson Insurance’s computer servers. The insurer’s algorithms will automatically assign fault and assess the damage. Roadside assistance will be automatically dispatched, and a Tesla shop will repair the car, with no additional cost. The transparent transmission of crash data also means that the information gap between the driver and the insurer is eliminated, solving the claims’ agent-principal problem. Therefore, our Pearson Insurance company should experience zero insurance fraud with Tesla car owners, reducing the loss ratio significantly, allowing further reductions in premiums.

Essentially, Tesla has an unfair advantage insuring Teslas. This is the epitome of the so-called Disruptive Innovation. Tesla’s new insurance company is going after a grossly over-served niche market - Tesla owners who don't need the service of a claims adjusters. Moreover, insurance carriers will most likely be happy to let go of this market share, as it represents an extremely small population whose risks they don't understand. Incumbent insurance carriers would likely consider autonomous driving riskier than human driving. They also consider the high price of repairing the pricey Teslas in case of a crash. They ultimately can’t come to terms with a lower premium for Teslas compared to other similarly priced cars. So if Tesla wants to insure Tesla owners, the likely thought process of incumbent carriers would prompt them to apriori forgo any competition over this small niche market. In true disruptive innovation fashion, a possible next step for Tesla might be to continue its attack upmarket by selling insurance to owners of other autonomous cars. Even selling a hardware package to retrofit and thus insure standard cars is not an outlandish possibility, considering the value it provides not only to the insurer, but to the transparency of the entire insurance transaction. If this story develops similarly to other disruptive innovations, it is not unlikely that by the time most cars came with a similar suite of sensors, Tesla will have been leading the insurance market of these tech-laden vehicles. Other insurance carriers will then just have to go out of car insurance entirely. By then, car insurance will simply not be lucrative enough for traditional insurance carriers.

The Next Level of Human-Robot Cooperation

If you asked anyone who ever had a bad experience with a chatbot what made the experience unpleasant, they would probably respond with something along the lines of “it just didn’t understand me”. While there are many deficiencies in common implementations of conversational AI, there is one type of an understanding problem that is both extremely prevalent and relentlessly detrimental to the conversational experience. The problem is that when most robots ask their human counterparts a yes or no question, for example, they only understand a yes or no answer. Natural language processing (NLP) has indeed come to a maturity where you can answer such questions with a “sure” or a “nope”, but anything other than a synonym of yes or no, will just not work. In case you haven't eavesdropped on many conversations in your life, you may think at this point that understanding synonyms is sufficiently clever to make do in most situations, but the reality is that seemingly unrelated, albeit implicit responses are extremely ubiquitous in human-to-human conversations. This is why many chatbots tend to limit the possible responses when asking such close-ended questions. With that being said, when using voice conversational AI in the call center, this is just not an option.

Bots limit answers

Bots mistake implicatures for questions

The reason this is such a difficult problem is twofold. First, natural conversations are intuitive for humans and most of whom just don’t have a good mental model for how robots handle conversations. The second reason is related to a popular conversational AI technique called “slot filling” (also known as “semantic role labeling”). The slots in this technique are the pieces of information that are required to provide an answer. Filling the slots involves eagerly trying to extract them from the human utterances, then asking for the missing ones explicitly. The issue is that in its simplest form, a slot represents a specific type of answer - yes/no, phone number, location, etc. When robots try to fill a slot by asking an explicit question, they often only consider the slot “filled” when the expected type of answer is found in the human’s response.

Many consider ELIZA, a computer program published by Joseph Weizenbaum in 1966, to be the very first instance of conversational AI. However, even Weizenbaum did not consider ELIZA intelligent. In fact, it wasn’t intelligent enough to demonstrate our problem, as it never asked specific questions, rather only cleverly rephrased its users’ requests into open-ended inquiries. It wasn’t until the 2000s when new techniques and stronger computing power made meaningful conversations with robots a reality. Thus when Paul Grice, a famed philosopher of language, published his influential paper titled “Logic and Conversation” in 1975, he didn’t even consider its far-reaching implications on conversational AI. In this paper, in which Grice outlines how humans answer questions in conversations, he defines what he calls the “Cooperative Principle”. This principle is phrased in the paper as a prescriptive description of how one must answer a question cooperatively in a conversation, and it consists of four maxims - the maxim of quality, of quantity, of relevance, and of manner. In short, it states that when answering a question, one must provide an answer that is both perspicuous, concise, accurate and relevant. A yes or no question should therefore be treated with a cooperative yes or no answer. That is not to say, however, that the maxims are never breached. To the contrary, Grice goes on to describe how a breach of the cooperative principle generally means that the counterpart in breach simply wishes to provide an implied answer, or an “implicature” as Grice calls them. In other words, an uncooperative answer must rely on shared prior knowledge to imply the requested piece of information. It is an unbelievably insightful observation from someone who has never encountered a dialogue where one of the parties is incapable of understanding implicatures, such as a dialogue between a robot and a human.

If you’ve gotten this far, you probably already understand that it is of utmost importance for robots to comprehend implicatures, should they ever wish to successfully complete an effective open-ended conversation with a human. Unsurprisingly, though, it is no easy task as it requires something more sophisticated than your run-of-the-mill NLP solutions. It requires a whole new approach that adds context-sensitive natural conversation understanding to the existing traditional natural language pipeline. This new layer, which provides a conversational interpretation for the human utterance, is a key component in achieving the next level of conversational AI. As you can hear in the conversation excerpt in the clip below, 5 implicatures in less than a minute are not uncommon, and their seamless understanding produces a remarkably yet inconspicuously human-like interaction, that is intuitive to other humans and represents the top echelon of conversational robots.

The Insurance AI Dilemma

In it’s bare essence, insurance is simply an agreement in which one party agrees to take on a risk of another party, for a fee. However, the very conflict of interests that arises from this simple transaction, turned this industry over the years into a convoluted myriad of entities with competing interests. In this post, I will try to unpack the relationships between the different entities, in order to create a framework that will help predict the disruptive powers of AI technologies in this space.

I am a huge fan of Clayton Christensen’s literature on innovation. Ever since his 1997 book, The Innovator’s Dilemma, and up until the day he passed away earlier this year, he was widely considered the leading expert in what makes an innovation disruptive. In fact, he was the one who coined the term “Disruptive Innovation”, which represents an invention which incumbent vendors are neither able nor willing to compete with (as opposed to a “Sustaining Innovation”). His ideas were so groundbreaking, because he presented a first-of-its-kind theory that was finally able to explain why some innovations are successfully introduced by small startups competing with the behemoths of their industries, contrary to all other logic. If I tried to distill his whole theory into one sentence it would be this one: an emerging technology is more likely to be disruptive if it does a job that’s already being done by consumers, but while the other solutions are good enough for this job, they are either over-serving some customers (i.e. they don’t need all the features) or are just too expensive for other potential customers (i.e. there’s an under-served market), but the resources, processes, and values of the incumbent vendors make the new technology distasteful for them. I know it’s a mouthful, but it explains why exploring the insurance value chain, its jobs-to-be-done, its resources and its processes may help us make predictions for emerging technologies.

There are four discrete parts to the insurance value chain - distribution, underwriting, service, and claims processing. Similar to any other product, distributors are in charge of finding a prospect customer and convincing them to purchase an insurance policy. However, because of the complex nature of the insurance product, and the way insurers compete on both the price and the terms, brokers have emerged as distributors that represent the customers’ interests. While brokers typically get paid by the insurer, they are essentially advisors to the end customer, and are not tied to any single insurer. Agents, on the other hand, only distribute products of a single insurer, whose sole interests they represent. Lastly, some insurers have a direct distribution channel, where they sell directly to the end customer.

Underwriting is at the very core of what an insurance policy represents - it is the act of agreeing to take on a financial risk. Within the insurance industry, it can be further broken down into two separate processes - the first process happens before anyone purchases a single policy, and in it actuaries create a mathematical schema for how different factors should affect the price and the insurer’s willingness to accept the risk. The second process happens every time a prospect applies for an insurance policy, and in it the underwriting operations professionals collect the different data points required by that schema to decide whether to write the policy, and at what premium. Different insurance products require different underwriting data, and can take anywhere between a few minutes to a few days. The first process of creating the pricing schema is always done by the insurance carriers themselves. The second part of using the schema to make a final decision for a specific prospect client is also usually done by the carriers, but can also be outsourced to super-brokers whom the carrier trusts. These brokers are called MGAs (managing general agents). A different kind of underwriting happens when some insurance carriers subsequently offload some of their risk portfolio to reinsurers. The latters, will underwrite some of the risk in return for some of the premiums, but they do not sell individual policies to individual customers.

In a typical insurance policy life cycle, there are two main points of contact between the insured and the insurer - purchasing the policy, and filing a claim. Any interaction that happens in between, like changing beneficiaries or making payments, falls under the category of service. For policies sold through agents and brokers, much of the servicing of the policy would be done via the broker, who will serve as an intermediary between the insurance carrier and the insured. With that said, insurance carriers, even ones who sell strictly via agents or brokers, all have customer service call centers servicing both these intermediaries and the end-customers.

The final part of the insurance value chain is the claims processing. Processing a claim involves accepting the first notice of loss, then determining fault and the insurer’s liability. Different types of insurance products have completely different processes for claims. Most health insurance claims, for instance, are processed electronically and immediately between the medical provider and the insurer. In property claims, each claim will be processed by a dedicated claims adjuster, who will appraise the damages, will determine how much of the damage was caused by the covered event, and will set the appropriate compensation. For some insurance products carriers may outsource the claims processing to third party administrators (TPAs).

The insurance value chain

In the very first paragraph I have mentioned a conflict of interests that arises from an insurance transaction between the insurer and the insured. As a matter of fact, insurance demonstrates a special type of conflict of interests, which economists call The Agent-Principal Problem, in not one but multiple parts of the insurance transaction. The agent in this problem is someone who can take action (the insurer) on behalf of the principal (the insured), and the problem arises when asymmetric information between the two leads to so-called “moral hazards”. That is, one side of the transaction holds information that the other side does not, which motivates them to mislead or misinform the latter. For example, the complexity and overlapping nature of some types of insurance products mean that the distributors understand better than their clients which insurance policies best suit their needs. The moral hazard in this case is that the same distributors may sell products based on their profit margins rather than the real needs of their customer. In some cases, this even leads to customers purchasing coverages that they already have from other policies they own. On the other hand, when it comes to claims, the insured holds all the information about what really happened. The moral hazard here is the motivation to embellish the damages in order to get over-compensated.

Over the years, several models have been developed in attempt to reduce the severity of the different agent-principle problems in insurance, by either reducing the information gap or by increasing the kinship between the insurer and the insured. Mutual insurance is probably the oldest model, in which the policyholders themselves own the insurance company. This increased kinship between the insurance company and the insured reduces the motivation to claim for damages fraudulently. The smaller the group of policyholders and the closer their relationship is, the more effective this technique in eliminating the moral hazard, because claimants would not want to withhold future profits from other policyholders they personally know. Lemonade Insurance, is an example of a new-age insurance company trying to overcome the size and relationship limitations of the mutual insurance model. To do so, they require each policyholder to select a charitable organization, as part of their onboarding process. Lemonade then portrays their business model as similar to any other service - they only charge a flat 20% service fee out of the premium. Any unclaimed premiums are then distributed to the selected charity. Lemonade's assumption is that if I am part of the group who donates to the NAACP, for example, I wouldn’t want to be responsible for withholding donations I already designated for them. Hence, I would not try to add undue damages to my claims.

As opposed to home or car insurance, life insurance is complex to a degree that requires some expertise to purchase. If you are rich enough, you may pay a financial advisor to compile your portfolio of investment and risk products. The average Joe, however, is left with the moral-hazard-laden brokers and agents. The age-old insurance industry idiom represents this problem well: “life insurance is sold, not bought”. This is not to say that anyone trying to sell you life insurance is morally skewed, only that they are incentivized in a way that increases their propensity to such. I am not aware of a publicly available perfect solution to this flavor of the agent-principal problem, but insurance brokers do play an important role in minimizing its effects for the insurers.

In recent years, AI has been a buzzword that was tossed around as a technology that would disrupt many industries, including insurance. It is important to understand, however, that AI is not one single technology, but a whole family of research areas that focus on computer systems that mimic human-like traits, like learning and reasoning. With that being said, as the technology progresses, AI is becoming more and more synonymous with Machine Learning in particular, which in itself is a huge research field within computer science. Machine learning encompasses many different types of algorithms. Their common denominator is that they all typically rely on training data to perform a predefined task, by trying to minimize a number calculated by a carefully crafted “loss function”. Different loss functions are made for different tasks - anything between recognizing handwriting and creating deep-fake videos of Donald Trump. Now that we’ve established the main links in the insurance value chain, as well as the resources and processes they use to do their jobs, we can go ahead and evaluate the potential of AI to attack the over-served or under-served parts of this value chain.

Within the distribution part of the value chain, we can now try and evaluate the applicability of AI to the different jobs - finding prospects, convincing them to purchase a policy, and reducing the severity of the agent-principal problem. Because AI relies heavily on data, the more consumer data a company has, the better it is equipped to use AI algorithms. The data hoarders of these days - like Facebook, Google, and Amazon, are therefore becoming superpowers of using AI to match consumers with the services they need. Even if traditional insurance distributors wanted to use AI to find new customers, in most cases they don’t have the data to support it. In fact, while their resources and processes aren’t geared for collecting this type of data, they are actually happy paying for hot leads and already have the processes in place to share revenues with 3rd party distributors. One type of data that they do have is recorded conversations. Conversational AI can use this data to match the best sales pitch for every type of prospect. This is where traditional insurance distributors (brokers, agents, carriers, etc.) can collaborate with conversational AI experts to create an optimized AI-driven insurance sales (and service) robot. What about the agent-principal problem? A data-hoarder that has a pristine reputation, may use their strengths to become a super-broker for all carriers, as long as they don’t underwrite the risk. Perhaps selling insurance will not contradict “do no evil”.

Using conversational AI to service insurance agents

Actuaries are one of the insurance industry’s strongest resources when it comes to AI. Regression analysis, which is one of the tools of the trade of actuaries, is actually considered the most basic example of machine learning. Insurers also own tremendous amounts of historical claims data, which means that for anything on the underwriting-claim axis, AI techniques are actually the epitome of sustaining innovations. They can increase the accuracy and speed of the yes / no / how much decisions in these jobs, which we can safely say are not done “good enough” yet. Taking car insurance as an example, one can easily imagine a future world where the DMV gives the carrier the information it needs about the car and the driver (if drivers are still a thing in this future) in order to generate an instant quote. When a crash occurs, the car’s sensors will relay the information directly to the carrier who will immediately reimburse for any casualties or property damages.

As we’ve seen, AI will definitely help improve many of the jobs done in the insurance industry. The key to understanding who is going to capitalize on these advances lies in the data ownership, and the expertise required to use it. It is easy to imagine a day where all this technology will turn insurance into a simple service, that’s almost transparent to the consumers. It may be purchased automatically, underwritten instantaneously, and cover losses independantly.

Robots and Dog Food

According to Wikipedia - “eating your own dog food or dogfooding is the practice of an organization using its own product”. In our minds, this is an extremely valuable practice, not only because of the many lessons that can be learned, but also because of the simple and strong message it conveys. We believe in our product’s value so much that we eagerly use it for our own needs. 

As obvious as it may seem, this practice is actually not as ubiquitous as you might imagine. First of all, many companies offer products and services that cater to types of entities that completely differ to their own. Makers of core insurance software like Guidewire and Majesco sell their products to insurance carriers, but can’t really use it themselves. Another typical culprits are really big corporations with many product lines. There, the person making the purchasing decisions often considers cost-effectiveness while ignoring the potential long term insights and corporate culture ramifications. This short sightedness can sometimes lead to choosing a competitor’s product. In the 90s, Microsoft notoriously did not use Microsoft Exchange for their own internal email systems until they were publicly called out for it. Lastly, even startups sometimes don’t think of using their own products, with no apparent particular reason. While it is quite hard to tell from the outside whether AirBnB’s employees, for example, are regularly staying at AirBnB rentals, it is not the case with conversational AI vendors. With the latter, one can simply browse their websites and immediately see if they’ve deployed their own solutions or not. As an interesting experiment, I’ve randomly browsed the websites of 10 different conversational AI vendors, who happen to also cater to financial institutions. Only one out of the 10 had deployed their own solution answering to visitors of their own website. 

How we dogfood:

At Supportomate, we’re developing an autonomous call center representative, to help enterprises significantly reduce operations costs. Simply put, it’s a robot that answers the phone. But we are a startup, and as such we don’t have a call center. So how can we still use our own robots? Our personal phones are all configured to transfer calls to one of our robots if we can’t pick up. This robot is connected to our calendars, and is trained to answer general questions and try to schedule another call at a time slot that is actually open in our calendar. This is not just a short-term exercise. It’s a long term solution for a problem that we really have. Whenever someone calls us - people who work at a company that creates robots that can answer the phone, they’d be immediately disappointed if they heard a vanilla answering machine. Even the least busy person in the world misses the occasional call, so why not use this missed call as an opportunity to WOW the caller. As an opportunity to stand behind our own product, and make sure that a missed call does not end up with a missed opportunity.