I sometimes reflect on how we will reach this stage of AI called AGI (Artificial General Intelligence) which is defined by a state of AI that by some measure equivalents the human condition. Scientists in the field sometimes call this the singularity, or the point where AI will develop much faster and at larger scale than ourselves. This is of course still a hypothetical state and many out there are in the very process of proving it either wrong or right – so I am not going there in this post. Then I am not even touching the subjects of free will, intention, sentience and consciousness… Nevertheless it is interesting to reflect on, independent from utopian or dystopian narratives around which many writings, studies and opinions are centered, how the evolutionary steps of AI look like today.
.
Usually evolution is considered over very large timeframes, meaning millions or billions of years. In the most narrow sense, AI only had a couple of hundred years. I was pointed to that when reading a recent Economist article referencing the 20 minute success of two robot arms (with some AI behind) assembling an Ikea chair, something with the same initial conditions we could probably do in 5. It may have something to do with this long process of adaptation we as humans have had over countless generations, integrating with a global ecosystem (largely restricted to earth), thus very slowly adapting, but very deeply at the same time. In all fairness, AI never got this kind of exposure. On the other hand, AI was able to master chess and go in less than 100 year and beat the human masters in the game. Of course, a game is a very artificial construct which we humans have also only mastered some hundreds of years (most humans even never did) so it is conceivable that a very focused effort (no need to eat, relax, fill out tax forms or raise kids) of an AI at some point beats us in this particular domain. Impressive, but at no point cataclysmic. Let’s also not forget that AIs today are not forced to survive in any ecosystem, so they lack that key intention that we humans (or life in general) have, namely the will to survive at any cost (even death). AIs are switched off when they are done and resume when deemed necessary, but other than that they can be pretty relaxed. So until now, AI has developed by addressing ad hoc problems through various techniques of supervised and unsupervised learning, with more or less sophisticated algorithms, with more or less compute and with more or less data. Especially the last two, compute and data have developed at a radical pace to the point where simple environments allow a broad community to experiment with AI, achieving impressive results. That is certainly evolution.
.
What we also start seeing however now is what prompted me to write this article, namely, several aspects of AI coming together in a single environment. It was addressed in the article “5 Ways Artificial Intelligence in Impacting the Automotive Industry” It mentiones driverless automobiles, Internet of Things, Risk Identification and Emotion Detection, Machine Learning and Assisted Driving and Robotics and Defect Detection as those 5 ways. One could even add (maybe under the IoT umbrella) services like Apple Carplay, which come with their recommender systems, finding the way to the car. The automotive domain is of course a very inviting subject, being a very dense ecosystem where humans interact with the environment in real time, and with already a lot of electronics on board with a diverse set of tasks. It is encouraging however to see that AI starts to meet AI in the same environment. For this really to become an evolutionary step however, these islands of AI will have to come together or else they will remain silos. Development requires connectedness. It will be one of the main challenges to address. We now have different functionalities trained with different data sets using different algorithms. During operation they will use distinct inputs and outputs. They are really different AIs in a box, doing their thing and blissfully ignorant from each other. Even in real-time while they source data from the same reality, it is not the same data. It would be a major breakthrough if all the AIs in the automotive ecosystem would somehow be able to network and communicate. For that to happen many challenges will have to be overcome. First there will be the technical challenge on how to interconnect different AI systems. What is our objective there? Do we want to connect them as independent agents or evolve them into a single agent (one step closer to AGI)? What is then the framework to interconnect them?
.
As independent agents we may expect them at some point to find each other and do their thing, just by letting them observe each other and communicate. As an integrated system we may need to find ways to expose functions and data through APIs to allow them to communicate (archaic?) These are only some technical aspects to the challenge. Then there are economic interests involved. The automotive space is a very lucrative one with many third party providers fighting for a seat at the table. To what extent will they be willing to go in a model of coopetition – or will they shield their intellectual property to the extent that it tempers general innovation? Then there are loads of challenges related to privacy and security and their legal implications. For sure already a very hard topic today at the current state of AI, but imagine how this will shift with interconnected systems – who will take the legal responsability there? The AI itself? A governing agency? The human at the steering wheel?
.
Where there are challenges there are also opportunities. For the consumer, the opportunity is there to become educated at least to the point to asking the right questions when for example buying a car. Make those questions at the ‘consumable’ level. Not technological questions, but questions about responsibility, limitations, expectations. Sensibilisation by the authorities may be a good tool to accomplish this. For business there is a great opportunity to come together do two things. One is to come together on standardization. Another is to work together on common themes such as legal aspects, privacy and security. Always good for progress to have a common challenge. For research there is a great opportunity to deeply share and enable the community to replicate and further develop. There, already great progress has been made and one project worth mentioning there is github. Many top researchers publish their peer reviewed papers online as well as the code and datasets to support the results. Try this only 20 years ago. Now thanks to universal environments based now on Python, Jupyter, Tensorflow, Theano, Keras and others, any researcher can dive in and replicate the results from the paper, and even better, play with the hyperparameters and models to further develop the results. This is simply amazing. This brings me to the final benificiary: the amateur. Amateurs have gone through waves of appreciation and valuation both positive and negative. One one hand they are seen as the passionate enthousiast (example the radio amateur – a disappearing art) – on the other hand as the non-professional (and therefore less relevant).
.
I believe however with AI being largely a software business widely in reach to anyone with a laptop and internet access, the amateur here is to be very highly valuated and will contribute big time to AI. In this particular domain of interconnecting islands, amateurs have the big advantage to be totally independent, free to choose their projects, untied by financial goals or requirements to secure the next research grant. Let this then be a call-out to the larger amateur and DYI community to dive in and start breaking the silos for the better advancement of AI.
