The people you need is actually people that have a software skill, software skill to use the tool kits that are out there. Mechanical Turk is an expression of providing a service and pretending that it's an automated service but in reality, there are people behind the levers. Sometimes it's a way to get the data. Namely, having the people doing it while the service is running, classifying pictures, saying whether a customer is happy or angry, and as time goes by, you use that manually, that manual data, that manual classified data to train the system to more and more automate it. So there are many applications, I think LinkedIn business cards that is actually used where was originally done by humans, and then successively that was automated. The main thing is that you can actually launch your service and see how much value does that particular feature or a service get appreciated by the customers. There are of course drawbacks to that and there was an example, a company called Expensify where you could actually automate the credit card handling. It turned out that these humans called Mechanical Turks through the Amazon network can actually see very personal data without the actual customer knowing about it. But Mechanical Turk allows you to have in a very inexpensive way human intelligence tasks that can be very very narrow to be performed and by actually buying that service from Amazon, and that's called the Mechanical Turk. Well, Mechanical Turk is actually a service provided by Amazon. They saw that they could provide humans behind a certain performance of a task. Where an artificial deep neural network could not perform that task, why don't we outsource that to people? These tasks are very simple. They can be classifying pictures, they can be classifying texts, actually deciding the repetitive simple tasks. What this allows you is that if you are a company and you want to provide a service and you do not have the data, you could actually use these people, Mechanical Turks to make it look like if your service is intelligently autonomously doing a certain thing, like scanning a business card or calculating the number of calories of the food on your plate which is existing example. Even if you have a limited amount of data and your neural network doesn't really provide the accuracy and the answer that is required, you could have the Mechanical Turks, actually the people actually making sure that the responses that go back to the people, to your customers are on the right level, and while they are working, you using that additional data and labeled data to train the automated system even further. So do not see the amount of data you have as a limiting factor of whether you should offer this out to your customers. Well, Mechanical Turk originates from the 1760s, 1770s in Vienna, the Court of Maria Theresa. There were mechanical machines at that time. There was this man who said, "I will make a machine that can play chess." Of course he succeeded, but not by having an intelligent machine, but by actually putting a human being into this machine. He never told anyone. This machine was a wooden figure looking like a Turk with a feather because actually at that time they thought the chess was invented in Turkey and it was actually invented in India but came to Vienna, to Europe through Turkey. So it actually originates from this poor man sitting inside and seeing this chess board through this little box he was sitting and actually moving a mechanical arm manually to be able to play chess. He won and he played against Napoleon, he played against the people in the Vienna court, he actually played against Philidor, he was a French chess player, probably the best in France at that time, probably the best in the world. He lost, but people were still amazed. The secret was never revealed that there was a person sitting inside. I think what you should focus on when you would like to provide an AI enabled service to your customers is actually to design that service without the restriction and knowledge that you don't have enough data. Creating the data is actually the most expensive and time-consuming thing that you can do. With less data, you will have less accuracy in that train system, with more data you will have a higher accuracy. So with time, you will have a satisfactory accuracy to what you're trying to achieve. In some cases, you need a 99.99 percent accuracy in deciding whether a certain picture contains something. In other cases, it's enough to have a 75 percent accuracy, and it's a huge difference in the amount of training that you need to do. It's not only the big companies that are providing these pre-trained networks. It's actually people behind these and these people they move and they are employed by these big companies. The drivers are of course the big internet companies that we see today. But what is provided is actually an architecture of how these neural networks look like. They are optimized for understanding either pictures or understanding sequential data. They are sometimes pre-trained. So you can actually classify something ahead with an accuracy, without having your own training data required. So you should use these and that's the quickest way to be able to launch something which is to your preference. Well, a lot of these things are actually open source and that's one of the good things with this field that the open source community it's free of charge to use it. What costs is actually the cost of having extra training, utilizing the storage, utilizing the process power to train with your specific data which is good to top on. It's also quite expensive to acquire new data. Actually part of the Cloud service, it is one of the tools in the toolkit of a Cloud service provider namely the Machine Learning tool kits. That goes for the Microsoft, Google, IBM. You would need to get some expertise, software engineers that are actually trained on utilizing either of these services. Probably it doesn't really matter which one you use. It might depend on which one you use depending on which field you're in, is it health care or is it actually analyzing texts. So it probably differs but that's not a key issue for you, no. The key issue is to get people that are able to use these tool kits in an efficient manner. They need the theory. They need a bit of experience and practice, and you might need, depending on what you're trying to achieve, some real experts that can be hired from the outside or used as consultants from the outside. The people you need is actually people that have a software skill to use the tool kits that are out there. Using a Cloud Service has traditionally, once it started, was traditionally a question of, can I trust putting my data outside my own company? That still prevails. Can I trust the Cloud Service? We have seen that there had been Cloud breaches of data. The drawback is of course that you legally need to comply to rules saying that well, I cannot move my data outside my country's border, in this case Sweden perhaps where we are today. That is countered by actually these companies building datacenters within the country's borders. You can as a user decide that I want to use this particular datacenter for the Cloud Service that I'm using. It would surprise me actually if we would see separate AI toolkit, Cloud service providers within certain applications, yes, but I do see that the dominant players will remain dominant if they're not split up through regulations. The storage, processing power is very coupled with the tool kits that are provided. The Cloud service providers today use the machine learning tool kits as a differentiator to get customers on board, to actually store all their data that they have. So it is a competitive factor. They of course try to be a service provider for all your software processing storage needs and that's the way it's actually heading. So depending on which type of company you are, you will be more susceptible to using Microsoft, Google or IBM. Once you have chosen a Cloud service, there's a threshold of course for changing to another one, but it's doable. You can build a Cloud service on Google Cloud or on the Amazon Cloud. They are not different programming languages but there are of course different tool kits. You will have people trained, software engineers trained to build these services on each of these Cloud services and that perhaps decides which service you start off with. But they all have different advantages and disadvantages. Well, many have like in the Cloud service providers to the providers of the utilities providing electricity. Of course that's the direction where it's going. It's a pipe to something. I think you should not be afraid of using the Cloud service providers. I think one should perhaps be more scared not to use them. If you put it in a cybersecurity context, it's the ability to defend yourself against cyber attacks, it's probably higher if you're utilizing the really big Cloud service providers than if you actually have your service locally.
