Building a system that is 100% autonomous and makes its own decisions is both hard and high risk. Given that Amazon, with all its resources and smarts, uses human input for the low/no consequence AI built into Alexa, it is fairly safe to assume that *all* other firms making AI claims have a human involved in at least one critical step.
If you want to see how good Alexa is at answering people's questions you should sign on to Alexa Answers and see the questions Alexa cannot answer. This site has gamified helping Alexa answer these questions. I spent a week doing this and figured out a pretty good work flow to stay in the top 10 of the leader board.
The winning strategy is to use Google. You copy the question in to Google and paste the answer Google back in to the Alexa Answers website for it to played back to the person who asked it. The clever thing is that since it is impossible to legally web-scrape Google.com at a commercially viable rate, Amazon have found a way of harnessing the power of Google without a) having to pay, b) violating Google.com's TOS, and c) getting caught stealing Google's IP.
After doing this for a week, the interesting thing to note is why Alexa could not answer these questions. Most of them are interpretation errors. Alexa misheard the question (e..g connor virus, coronda virus, instead of coronavirus). The remainder of the errors are because the question assumes Alexa's knowledge of the context (e.g. Is fgtv dead? - he's a youtube star) and without the subject of the question being a known entity in Alexa's knowledge graph, the results are ambiguous. Rather than be wrong, Alexa declines to answer.
Obviously this is where the amazing pattern matching abilities of the human brain come in. We can look at the subject of the question and the results and choose the most probable correct answer. Amazon can then augment Alexa's knowledge graph using these results. This is probably in violation of Google's IP if Amazon intentionally set out to do this.
Having a human being perform the hard task in a learning loop is something that we have also employed in building our platform. Knowledge Leaps can take behavioral data and tease out price sensitivity signals, using purchase data, as well as semantic signals in survey data.
In a lot of science fiction films one, or more, of the following are true:
- Technology exists that allows you to travel through the universe at the "speed of light."
- Technology that allows autonomous vehicles to navigate complicated 2-D and 3-D worlds exists.
- Technology exists that allows robots to communicate with humans in real-time detecting nuances in language.
- Handheld weapons have been developed that fire bursts of lethal high energy that require little to no charging.
Yet, despite these amazing technological advances the kill ratio is very low. While it is fiction, I find it puzzling that this innovation inconsistency persists in many films and stories.
This is the no-free-lunch theory in action. Machines are developed to be good at a specific task are not good at doing other tasks. This will have ramifications in many areas especially those that require solving multiple challenges. Autonomous vehicles for example need to be good at 3 things:
- Navigating from point A to B
- Complying with road rules and regulations.
- Negotiating position and priority with other vehicles on the road.
- Not killing, or harming, humans and animals.
Of this list 1) and 2) are low level. 3) is challenging to solve as it requires some programmed personality. Imagine if two cars using the same autonomous software meet at a junction at the very same time, one of them needs to give way to the other. This requires some degree of assertiveness to be built. I am not sure this is trivial to solve.
Finally, 4) is probably really hard to solve since it requires 99.99999% success in incidents that occur every million miles. There may never be enough training data.
Last year we wrote about No Free Lunch Theory (NFLT) and how it relates to AI (among other things). In this recent Wired article, this seems to be coming true. Deep Learning, the technology that helped AI make significant leaps in performance has limitations. These limitations, as reported in the article, cannot necessarily be overcome with more compute power.
As NFLT states (paraphrased): being good at doing X means an algorithm cannot also be good at Doing Not X. Deep Learning models that have success in one area is not a guarantee they will have success in other areas. In fact the opposite tends to be true. This is the NFLT in action and in many ways, specialized-instances of AI-based systems was an inevitability of this.
This has implications for the broader adoption of AI. For example, there can be no out-of-the box AI "system". Implementing an AI solution based on the current-state-of-the-art is much like building a railway system. It needs to adapt to the local terrain. A firm can't take a system from another firm or AI-solutions provider and hope it will be a turn-key operation. I guess it's in the name, "Deep Learning". The "Deep" refers to deep domain, i.e. specific use-case, an not necessarily deep thinking.
This is great news if you are an AI developer or have experience in building AI-systems. You are the house builder of the modern age and your talents will always be in demand - unless someone automates AI-system implementation.
UPDATE: A16Z wrote this piece - which supports my thesis.
Many firms (Amazon, Google, etc) are touting their plug-and-play AI and Machine Learning tool kits as being a quick way for firms to adopt these new technologies without having to invest resources building their own.
Sound like a good idea but I challenge that. If data is going to drive the new economy, it will be a firm's analytics capabilities that will give it a competitive advantage. In the short-term adopting a third-party framework for analytics will move a firm up the learning curve faster. Over time this competitive edge becomes blunter, as more firms in a sector start to use the same frameworks in the race to be "first".
This homogenization will be good for a sector but pretty rapidly firms competing in that sector will be soon locked back in to trench warfare with their competitors. Retail distribution is a good example, do retailers use a 3rd party distribution network or do they buy and maintain their own fleet. Using a 3rd party distributer saves upfront capex but it voids an area of competitive advantage. Building their own fleet, while more costly, gives a retailer optionality about growth and expansion plans.
The same is true in the rush for AI/ML capabilities. While the concepts of AI / ML will be the same for all firms, their integration and application has to vary from firm-to-firm to preserve their potential for providing lasting competitive advantage. The majority of firms we have spoken to are developing their own tool kit, they might use established infrastructure providers but everything else is custom and proprietary. This seems to be the smart way to go.
In conversations with a friend from university I learned about the No Free Lunch Theorem and how it affects the state-of-the-art of machine learning and artificial intelligence development.
Put simply, the No Free Lunch Theorem (NFL) proves that if an algorithm is good at solving a specific type of problem then it pays for this success by being less successful at solving other classes of problems.
In this regard, Algorithms, AI Loops and Machine Learning solutions are like people; training to achieve mastery in one discipline doesn't guarantee that same person is a master in a related discipline without further training. However, unlike people, algorithm training might be a zero-sum game with further training likely to reduce the competency of a machine learning solution in an adjacent discipline. For example, while Google's AlphaZero can be trained to beat world champions at chess and Go, this was achieved using separate instances of the technology. A new rule set was created to win at chess rather than adapting the Go rule set. Knowing how to win at Go doesn't guarantee being able to win at chess without retraining.
What does this mean for the development of AI? In my opinion while there are firms with early-mover advantage in the field, their viable AI solutions are in very deep domains that tend to be closed systems, e.g. board games, video games, and making calendar appointments. As the technology is developed, each new domain will require new effort, likely to lead to a high number of AI solutions/providers. So rather than an AI future dominated by corporate superpowers there will be many providers, each with a domain-distinct AI offerings.
The patent that has just been awarded to Knowledge Leaps is for our continuous learning technology. Whether it is survey data, purchase data or website traffic / usage data., the technology we have developed will automatically search these complex data spaces. The data spaces covers the price-demand space for packaged goods, or the attitudinal space of market research surveys and other data where there could be complex interactions. In each case, as more data is gathered - more people shopping, more people completing a survey, more people using an app or website - the application updates its predictions and builds a better understanding of the space.
In the use-case for the price-demand for packaged goods, the updated predictions then alter the recommendations about price changes that are made. This feedback loop allows the application to update its beliefs about how shoppers are reacting to prices and make improved recommendations based on this knowledge.
In the survey data use-case, the technology will create an alert when the data set becomes self-predicting. At this point capturing further data is unnecessary to understand the data set and carries an additional expense.
The majority of statistical tools enable analysts to identify the relationships in data. In the hands of a human, this is a brute-force approach and is prone to human biases and time-constraints. The Knowledge Leaps technology allows for more systematic and parallelized approach - avoiding human bias and reducing human effort.
The weekly ad is beloved and bemoaned by US retailers. On the one hand it is seen as an essential tool in the marketing departments armory. On the other, it is seen as a significant investment that is difficult to measure.
This is exactly the type of pain-point we like to solve at Knowledge Leaps. Using our platform we have built a solution that identifies which promotions get an uplift from appearing in the weekly ad and the incremental $ generated from appearing in the ad.
Adding in meta data about items, ad placement data, and seasonality we can build an AI learning loop for retailers that will optimize and then maximize the return on this investment.
The attraction of AI is that it is learns by experience. All learning requires feedback, whether its an animal, a human or a computer doing the learning, The learning-entity needs to explore its environment in order to try different behaviors, create experiences and then learn from them.
For computers to learn about computer-based environments is relatively easy. Based on a set of instructions, a computer can be trained to learn about code that it is executing or is being executed by another computer. The backbone of the internet uses something to similar to ensure data gets from point A to point B.
For humans to learn about human-environments, this is also easier. It is what we have been doing for tens of thousands of years.
For humans to learn about computer-based environments is hard. We need a system to translate from one domain into another. Then we need a separate system to interpret what we have translated. We call this computer programming, and because we designed the computer, we have a bounded-system. There is still a lot to understand, but it is finite and we know the edges of the system, since we created it.
It is much harder for computers to learn about human environments. The computer must translate real-world (human) environment data into its own environment, and then the computer needs to decode this information and interpret what it means. Because the computer didn't design our world it doesn't have the advantage that humans do when we learning about computers. It also doesn't know if it is bounded-system or not. For all the computer knows, the human-world is infinite and unbounded - which it could well be.
In the short term, to make this learning feasible we use human input. Human's help train computers to learn about the real-world environments. I think of the reasons that driver-less car technology is being focused on, is that the road system is a finite system (essentially its 2D) that is governed by a set of rules.
•Don't drive into anything, except a parking spot.
•Be considerate to other drivers, e.g. take turns at 4-way stop signs.
•Be considerate to pedestrians and cyclists.
This combination of elements and rules makes it a perfect environment to train computers to learn to drive, not so much Artificial Intelligence but Human-Assisted Intelligence. Once we have trained a computer to decode the signals from this real-world environment and make sensible decisions with good outcomes, we can then apply this learning to different domains that have more variability in them, such as delivering mail and parcels.
This is very similar to the role of the machine screw in the industrial revolution. Once we had produced the first screw, we could then make a machine that could produce more accurate screws. The more accurate the screw, the more precise the machine, the smaller tolerance of components it could produce, the better the end-machine. Without the machine screw, there would have been no machine age.
This could open the doors to more advanced AI, it is some way off though because time required to train computers to learn about different domains.
I have done some quick back-of-envelope calculations on the progress of AI, trying to estimate how much progress has been made vs. how many job-related functions and activities there are left to automate.
On Angel List and Crunchbase there are a total of 4830 AI start-ups listed (assuming both lists contain zero duplicates). To figure out how many unique AI tools and capabilities there are, let's assume the following:
- All these companies have a working product,
- Their products are unique and have no competitors,
- They are all aimed at automating a specific job function, and
- These start-ups only represent 30% of all AI-focused company universe.
This gives us a pool of 16,100 unique, operational AI capabilities. These capabilities will be in deep domains (where current AI technology is most successful) such as booking a meeting between two people via email.
If we compare this to the number of domain specific activities in the world of work, we can see how far AI has come and how far it has to go before we are all working for the computers. Using US government data, there are 820 different occupations, and stock markets list 212 different industrial categories. If we make the following set of assumptions:
- 50% of all occupations exists in each industrial category,
- Each occupation has 50 discrete activities.
This gives us a total of 4.34 million different occupational activities that could be automated using AI. In other words, at its most optimistic, current AI tools and processes could automate 0.37% of our current job functions. We have come a long way, but there is still a long way to go before we are out of work. As William Gibson said, "the future's here, it's just not widely distributed yet"