AI That Really Works – Part 1: Solvable Problems
AI works when the technology capabilities and business requirements come together
Artificial Intelligence is an impressive technology, but does it work in the real world and can it solve hard problems? The short answer is that AI works very well today in specific situations. It’s very good at solving narrow problems or performing individual tasks under relatively controlled conditions. In these cases, AI may be much better, faster or cheaper than humans. However, AI still needs help from humans to maintain these operating conditions and to be backup for unexpected corner cases. Otherwise, AI can fail comically or even tragically.
There are two ways to answer the question of where AI is successful today. First, there are types of well-defined problem categories that AI approaches are especially good at solving. This set of problems is driven by the current state of the art and underlying capabilities of the technology.
Second, some industry sectors and business functions are more ripe than others for adopting AI solutions. This ripeness is based on the value versus cost of improving a task or function. But also the adaptability or change-readiness of the sector.
This AI sweetspot between solvable problems and ripe sectors implies a nearterm time horizon in which the benefits can be achieved. However since these are learning systems, we only need to achieve a minimum level of cost benefit, as the system will continue to improve over time.
Existing AI toolkit is very good at solving certain kinds of problems
There are already many types of well-defined problems that AI approaches are especially good at solving. By good, I mean that they do a task more consistently, faster or simply cheaper than humans. Here are some examples of categories of solvable problems based on the current state of the art, though it’s a moving target and thus not comprehensive.
Solvable Problem Categories
Search/ Optimization
Prediction/ Pattern Recognition
Perception
Movement/ Manipulation
(Apparent) Understanding
This list is roughly in order of difficulty for AI.
If you can describe your problem as similar to one already being solved above, then an AI solution is more likely to be successful. Likewise, if existing machine learning techniques and algorithms are applicable, the chance of success is higher (e.g., well-defined objective, classification or clustering problem). If the problem is amenable to keeping a human in the loop then useful results may be closer at hand until you perfect the model. The relative state of human capabilities is also relevant. If humans are really bad at something, then the bar is lower. These are all qualities of an AI-solvable problem category.
If better-than-human most of the time is good enough, then AI may work. But to be always as good as a human is nearly impossible for AI today
One example that solves a set of these problems is the Google product Duplex. In 2018, the CEO of Google demonstrated how their AI can be used to make a haircut appointment with a real person over the phone. It involves search (finding an appropriate hair salon), perception (understanding the sounds of speech and converting it to text), understanding (converting text to meaning), and optimization (planning the conversation/ responses in order to achieve the goal: an appointment).
Again the key thing to take note is that this is a relatively controlled context and set of operating conditions. If the person on the call would ask an unexpected question or bring up something off topic, the system will probably not work well.
Duplex even inserts “ums” to make the AI more convincing. This last part raises the question, is it too convincing, and is it OK to be recording people’s speech even for the purpose of powering an AI system. As of spring of 2019, Duplex is available in 43 states, though it announces itself and asks for permission before recording.
Solving search and optimization problems is one of the earliest AI successes
The first AI-solvable category is search and optimization problems. These types of problems have long been a rich vein for AI. It generally means maximizing a result based on a complex set of constraints. Unlike a straight computational approach, these problems have too many variables to solve for directly and there may not be an optimal solution. AI techniques can be used to search for a good answer, and seek to recursively improve the results. Applications range from strategy games to resource or supply chain management. This problem set includes driving directions for people and navigation planning for robots.
Prediction is another form of pattern recognition well suited to machine learning
Another category of problems well suited for AI is Prediction. This is a form of classification problem, very adaptable to supervised learning, or general pattern recognition like clustering or other unsupervised learning techniques. Here the AI uses training data to extrapolate results with new inputs to predict everything from stock prices or detecting fraud or suspicious behavior.
Artificial perception (machine hearing and vision) is already endowing us with superhuman capabilities that did not exist a decade ago
Perception is also a very successful problem area, including both computer vision and hearing. AI imitates these human abilities, and in some cases, makes them super powers. AI can hear outside the human range and filter out frequencies that are literally noise in the data. AI vision can identify objects and render them in three dimensional models.
Speech-to-text is already about as good as humans. The Google API, for example, is comparable to human closed-captions with a word-error-rate less than 5%.
And it is only getting better as these AI systems train on the millions of smart speakers in people’s homes. These systems are gaining experience understanding speech in a wide variety of room types and situations with different acoustics and background noises. Amazon has the apparent lead in this area, but it’s too early to say which platform may ultimately win. Nonetheless, it would be difficult for a startup to compete in the speech-to-text function, at least in a general consumer context.
There are other sounds, beyond speech, that can be perceived by computers that hold great value. One example of this is music of course. Companies have been using a process called fingerprinting to quickly identify songs for various reasons, including enforcing the digital rights of the artists and music labels. In recent years, there has been an explosion of remixed music. This is music created by users by combining portions of different songs, overlapping the tracks and often distorting them in some way, including changing their tempo and key. By some measures, this user-generated remixed music is ten times as popular as official studio versions. However, the regular fingerprinting technique does not work on remixed music. This makes it difficult to figure out who owns the rights to the component tracks and leaves no way for the original artists and the labels to be paid for streaming or purchasing of the music, let alone paying anything to the user for their efforts creating the remix. At the same time, the distribution platforms like Spotify and iTunes are unable to distribute the music since it has not been cleared by the music labels
A company I invested in, Dubset Media, is addressing this problem using machine learning. Dubset analyses the remixed music and determines which tracks came from which artist and recording. The AI has to have almost superhuman hearing in order to distinguish the overlapping tracks, as well as adjust for distortions and changes done by the user. By working with the original artists and labels, Dubset determines all the rights for each component of the remixed song and how to allocate the payment for the different rights owners.
Another form of computer perception is computer vision. In the case of self-driving cars, computer vision is also superhuman, looking beyond the visible spectrum using LIDAR and radar, to create a three dimensional model for its environment to see around corners and other vehicles.
Likewise, facial-recognition algorithms are now as good as the best forensic examiners. Though, the best results come from combining humans and machines.
AI powered specialized cobots working in partially controlled environments alongside humans are already doing real work
More recently, I would point out that recent capabilities in movement and manipulation are very real. Robotics is no longer relegated to the strictly controlled factory floor. Pulling on parts of all of the previous problem areas, including perception and optimization, they are able to interact with humans in only partially controlled conditions.
Advanced robotics has also made real progress recently, much of which has been fueled by AI. Robots are no longer relegated to the strictly controlled factory floor, they can work safely alongside humans. Using AI, robots can now perceive their environment and navigate around people and other obstacles while accomplishing their task.
Still robots operate best in somewhat controlled environments, like a warehouse. Some companies like Boston Dynamics have demonstrated the acme of what is possible, getting inspiration from human and other animal forms for their designs. However, the most successful robots are ones that are specialized for a task rather than trying to look like us.
One of the best examples of specialized robots actually working in a warehouse is Amazon Robotics, formerly known as Kiva. These robots completely reshaped the way Amazon set up its warehouses, allowing for smaller, more densely-packed facilities closer to urban centers. The robots are a small Roomba-like device that travels through the warehouse under movable stacks of product trays. When the robot is under the correct stack, it raises up to acquire the stack and brings the stack to the waiting human packer. If necessary, other stacks will be rearranged to get the desired stack out. All the robots work in concert to avoid collisions and accomplish their tasks. This approach is more flexible than conveyor belts and takes much less room. According to Amazon, these robots cut operating expenses by about 20%, while growing inventory space by 50%. They now have more than 100,000 robots working in their warehouses.
The home is very different from a warehouse. It has children and pets, with obstructions like toys, crumpled carpets and loose cords. Nonetheless, our friends at Boston Dynamics have been working on bringing a robot to the home for some time. They are expecting to release Spot, a quadruped with an extendable arm, this year. This is an extremely capable and agile robot. The days of the pet dog may be numbered, but if you notice the scuff marks on the wall, these robots have a long way to go to operate safely in a home environment.
Likewise, while specialized machines for food preparation may become more widespread, I would not expect to see knife-wielding robots in the kitchen any time soon.
AI’s ability to mimic understanding will soon surpass the reality, a different kind of ‘uncanny valley’
To many, the ability to understand human language is an essential part of intelligence. And there has been significant progress in the branch of AI called Natural Language Processing (or NLP). NLP can find associations of words and concepts to mimic understanding, especially in a limited context. This capability was demonstrated in the Google Duplex example where the AI was able to make a phone call with a human to set a haircut appointment. This is very impressive, but it is not true understanding in the human sense.
We can see this limitation in machine language translation. There's been a lot of progress to date in machine translation but AI is still less capable than the best human translators. That gap remains mostly because human translators truly understand the words being translated, not just matching words and phrases.
At the same time, AI is getting better at detecting the emotional tone of language. One example is Perspective, a project by Jigsaw (Alphabet Google’s think/do tank). Perspective is an API to identify toxic speech, which is anything using abusive language, threats or harassment. This is especially needed in the user comments areas of media websites or chat features of online communities. Perspective uses machine learning to sense the emotional tone of the language and block speech which violates the community guidelines. Jigsaw worked with The New York Times and Wikipedia to develop Perspective. Before using this tool, the Times’s comments section was managed by 14 moderators, who had to manually review nearly every comment.
Many companies big and small are providing similar tone detection and moderation services. One such startup is Smart Moderator. The social network Facebook is also using machine learning to address fake news and hostility online. Tools like this are helping publishers and moderators improve conversations on their platforms and enhance the exchange of ideas on the internet.
This level of understanding has been extended to the area of professional services, like the law, and academics.
AI can perform rudimentary reviewing, and even drafting, of some legal documents, saving time while avoiding errors and omissions. Lawgeex, an AI startup, recently claimed its contract review application was more accurate and faster than experienced human lawyers.
Likewise in academic research, an algorithm called Word2Vec was used to sift through over 3 million scientific paper abstracts for connections humans had missed. Word2Vec was able to find predictions for possible new thermoelectric materials. The algorithm linked words that were found close together, creating vectors of related words that helped define concepts and make new connections not seen before.
AI tools such as these only augment professionals in their work, yielding better and faster results, while making the work less of a drudgery. But we have a long way before we can replace professional judgement, understanding or other capabilities.
In the next article, we’ll look at the other half of this equation, which industries are ripe for adopting AI.
© 2026 Machina Ventures LLC