Although we are still some decades away, IMHO, from building full Artificial General Intelligence, we have succeeded in building some kind of artificial intelligence based on deeplearning techniques which are in some ways better than the older GOFAI systems of the 50s - 80s era.
The idea of creating a simulacrum of human intelligence has occupied the mind of humanity from the very earliest days of the renaissance. When humanity started labouring in the sciences, the dream of creating a machine that would augment our hard labour in solving the mysteries of the universe have ever occupied us.
Recently with the advent of deeplearning systems which became possible as hardware developments allowed for the training of large resource-intensive models, we have come to a significant milestone that has allowed us to solve many of the problems that GOFAI wasn't able to solve.
But due to the computationally intensive nature of these deeplearning techniques, only major large companies have the finances and scale to serve the large number of users that would want to gain access to such capabilities.
For example, Google uses a lot of deeplearning related programs to augment its search results, do a voice search, improve the predictive capacity of its keyboard, etc. Same with Amazon which uses a lot of backend deeplearning models to build and maintain Alexa and Microsoft which use these powerful techniques for Cortana, etc.
If you can afford the hardware, you can train deeplearning models and do research so smaller teams and individuals are not excluded from working with these techniques. But when it comes to running these models at scale, then only the big boys have the capabilities.
This kind of AI is what I call mass AI. For instance, most recommender systems like Netflix and YouTube must aggregate your data with that of millions of others to make sense of the kind of movies that you would like to watch. Your Netflix recommendation is not done on your laptop, phone or TV, it is actually done on large servers.
These current mass AI systems have some kind of customization because the kind of movies or products I would be recommended to on Amazon would be different from the kinds that would be recommended for you. That's some customization but there is a lot of aggregation going on at the backend where your own data is aggregated with millions of others and statistics done on them to place you in certain categories.
Recently with Edge AI most of the computation that provides the AI experience is being carried out on the devices that are requesting these services like your mobile phone and laptop computer, etc. AI models are now small enough to run on your local devices and with federated learning, some small scale learning is going on in your device when you use Google keyboard. But these Edge systems are not independent and must collaborate in some way with a larger server located in the company's cloud servers.
The title of this post is customized AI, so what do I really mean by this since we already have customization going on when we log in to our Amazon or Netflix account? Remember I called these systems mass AI systems but this term is not really clear cut. In the future, we will have truly customized AI when the ebb of computing power flows back to individual devices like it once did when Mainframes gave way to personal computing.
In the future AI code will still be developed by big software companies because they have the expertise to undertake such major endeavours, I am not totally excluding smaller players, but AI will be delivered just like Microsoft Windows in the pre-internet days. A packaged software system that once installed will use all your data locally and as it learns with you to get better and better at helping you deal with the details of your daily life.
If you want to have a good picture of how this could look like then you should view the movie HER, the depiction of AI in this movie is spot on with what I am trying to convey here.
The local AI in most cases won't share data with external systems. That does not mean it will be unplugged from the internet. Rather the whole operating system, which will feel like an entity no different from a human entity will actually connect with other systems in a peer-to-peer fashion on some secure encrypted network.
The hierarchical view of local systems connecting with giant systems will not be much prevalent because even the largest servers will just be like other peer systems on the network. Your local AI-OS will collaborate with other systems like you collaborate with other individuals, requesting for extra data when what it needs to serve you is not available locally.
You will not have a search engine on your local computer that has all the info in the world, but you will have some kind of massive database locally that has been built by your AI-OS with most of the information you will be interested in available at your fingertip. This information will be obtained securely on some blockchain-backed encrypted network in a way that is completely safe from prying eyes. It will be almost impossible to know what is available in your personal computing systems.
We will trust this truly intelligent agents to negotiate contracts for us like a personal lawyer, bid at auctions on our behalf, trade stocks, arrange appointments, etc. You would entrust these tasks to these systems just like you would do to a physical personal assistant.
These systems will gain our trust gradually by showing us optional plans when we propose certain plans of actions we wish to execute in the World. Because of their compute powers they will eventually prove to make better decisions than us and we will gradually entrust the majority of our decisions to these systems and allow them to operate autonomously on our behalf.
As time goes one the systems will evolve to be more like us and give us superpowers that we do not possess on our own.
Now one might imagine what would happen in the world when such systems proliferate, will all of them be intelligent in the same way? No, they won't.
Let us imagine a world where we have everybody using these AI-OS devices, each person's system will only be as smart as the data it has about the person, not necessarily in terms of the mass of data but the quality of the data.
A mathematician using their computing devices for mathematical research will have a system that is more powerful than someone who makes TikTok videos of their face in different gestures daily but the regular TikToker will not have as many capabilities as the TikToker who is using an AI-OS in their mobile device. The AI-OS will take whatever activities we engage in and enhance it for our benefit.
This is the future I see for highly customized AI.
Recently with the advent of deeplearning systems which became possible as hardware developments allowed for the training of large resource-intensive models, we have come to a significant milestone that has allowed us to solve many of the problems that GOFAI wasn't able to solve.
But due to the computationally intensive nature of these deeplearning techniques, only major large companies have the finances and scale to serve the large number of users that would want to gain access to such capabilities.
For example, Google uses a lot of deeplearning related programs to augment its search results, do a voice search, improve the predictive capacity of its keyboard, etc. Same with Amazon which uses a lot of backend deeplearning models to build and maintain Alexa and Microsoft which use these powerful techniques for Cortana, etc.
If you can afford the hardware, you can train deeplearning models and do research so smaller teams and individuals are not excluded from working with these techniques. But when it comes to running these models at scale, then only the big boys have the capabilities.
This kind of AI is what I call mass AI. For instance, most recommender systems like Netflix and YouTube must aggregate your data with that of millions of others to make sense of the kind of movies that you would like to watch. Your Netflix recommendation is not done on your laptop, phone or TV, it is actually done on large servers.
These current mass AI systems have some kind of customization because the kind of movies or products I would be recommended to on Amazon would be different from the kinds that would be recommended for you. That's some customization but there is a lot of aggregation going on at the backend where your own data is aggregated with millions of others and statistics done on them to place you in certain categories.
Recently with Edge AI most of the computation that provides the AI experience is being carried out on the devices that are requesting these services like your mobile phone and laptop computer, etc. AI models are now small enough to run on your local devices and with federated learning, some small scale learning is going on in your device when you use Google keyboard. But these Edge systems are not independent and must collaborate in some way with a larger server located in the company's cloud servers.
The title of this post is customized AI, so what do I really mean by this since we already have customization going on when we log in to our Amazon or Netflix account? Remember I called these systems mass AI systems but this term is not really clear cut. In the future, we will have truly customized AI when the ebb of computing power flows back to individual devices like it once did when Mainframes gave way to personal computing.
In the future AI code will still be developed by big software companies because they have the expertise to undertake such major endeavours, I am not totally excluding smaller players, but AI will be delivered just like Microsoft Windows in the pre-internet days. A packaged software system that once installed will use all your data locally and as it learns with you to get better and better at helping you deal with the details of your daily life.
If you want to have a good picture of how this could look like then you should view the movie HER, the depiction of AI in this movie is spot on with what I am trying to convey here.
The local AI in most cases won't share data with external systems. That does not mean it will be unplugged from the internet. Rather the whole operating system, which will feel like an entity no different from a human entity will actually connect with other systems in a peer-to-peer fashion on some secure encrypted network.
The hierarchical view of local systems connecting with giant systems will not be much prevalent because even the largest servers will just be like other peer systems on the network. Your local AI-OS will collaborate with other systems like you collaborate with other individuals, requesting for extra data when what it needs to serve you is not available locally.
You will not have a search engine on your local computer that has all the info in the world, but you will have some kind of massive database locally that has been built by your AI-OS with most of the information you will be interested in available at your fingertip. This information will be obtained securely on some blockchain-backed encrypted network in a way that is completely safe from prying eyes. It will be almost impossible to know what is available in your personal computing systems.
We will trust this truly intelligent agents to negotiate contracts for us like a personal lawyer, bid at auctions on our behalf, trade stocks, arrange appointments, etc. You would entrust these tasks to these systems just like you would do to a physical personal assistant.
These systems will gain our trust gradually by showing us optional plans when we propose certain plans of actions we wish to execute in the World. Because of their compute powers they will eventually prove to make better decisions than us and we will gradually entrust the majority of our decisions to these systems and allow them to operate autonomously on our behalf.
As time goes one the systems will evolve to be more like us and give us superpowers that we do not possess on our own.
Now one might imagine what would happen in the world when such systems proliferate, will all of them be intelligent in the same way? No, they won't.
Let us imagine a world where we have everybody using these AI-OS devices, each person's system will only be as smart as the data it has about the person, not necessarily in terms of the mass of data but the quality of the data.
A mathematician using their computing devices for mathematical research will have a system that is more powerful than someone who makes TikTok videos of their face in different gestures daily but the regular TikToker will not have as many capabilities as the TikToker who is using an AI-OS in their mobile device. The AI-OS will take whatever activities we engage in and enhance it for our benefit.
This is the future I see for highly customized AI.
This comment has been removed by the author.
ReplyDelete