20 Years from 2020

20 years from 2020

What will the future hold for AI?

On Thursday the 9th of January I will be speaking at the New Year's event of NLSpace. I'll be trying to answer the question: "What will Earth Observation look like 20 years from now?" The question is all the more challenging, because AI will undoubtedly play a large role in the next two decades. And whoever can predict what AI will do next.. Well, let's say there's very few with a good track record. It did trigger me to think about the future. Not only about AI in space, but also about AI more generally. So given the 99.9% probability of being completely wrong, let's give it a shot.

Synergy-by-design

The title and topic of my talk will be synergy-by-design. It's the idea that the capabilities of AI-driven software technologies are forcing us to rethink how we design (hardware) systems. In some cases, AI can already extend the capabilities of such systems well beyond what they are designed to do. Designing a system to take into account these capabilities will give rise to new kinds of systems which are much more powerful than the current combinations of hard and soft subsystems. 

Synergy-by-design thinking can give rise to huge cost reductions, because the hardware requirements can be much simpler.

If more power isn't what you seek, synergy-by-design thinking can give rise to huge cost reductions, because the hardware requirements can be much simpler. A concrete example of the latter principle is the use of deep learning-based frame interpolation in super-slow motion camera's. If you can increase the frame-rate by a factor of 4 using software technology, you can get away with a 'regular' camera, instead of building an expensive camera with a higher frame-rate. Once synergy-by-design thinking will gain traction, I'm convinced the amount of new types of systems we will see are virtually endless.

Artificial General Intelligence (AGI)

When thinking about AI in 20 years, one cannot get away with only practical considerations like synergy-by-design thinking, and ignore the elephant in the room: "Will an AI superpower take over the world?". Or perhaps the more modest version: "Will AI become as capable as humans?". I'm going to take a controversial standpoint and say that artificial general intelligence (AGI) is not that far away. Moreover, by 2040 it will be as capable as humans, but will still be under our control. There is a number of very interesting developments in the field that support this claim.

The building blocks are here

Predicting the impact of AI technology in society is a question that I feel is exponentially harder to answer than how AI technology will develop. The Internet and the smart phone have changed society in ways that we could have never foreseen. Predicting that there would be a connected phone, was, however, much easier to do. Therefore I will stick to the topic of AI technology, which is interesting enough in itself. In the past 7 years, AI technology has taken a huge flight, due to the advent of deep learning (DL). This has made it possible for computers to develop senses like never before. Speech, hearing, vision, language understanding and even decision making experienced enormous strides. On specific subtasks, given the right examples to learn from, computers are on par with, or even surpassing humans already. Nobody has been able to piece all of them together, but, as we will see, the building blocks are all there.

Vision

Disecting images has been one of the main focus area's of DL in the past years. This problem can be considered pretty much solved. On standardized tests for image interpretation, like ImageNet and COCO, the best DL models are nearly perfect. The lessons learnt in computer vision are now proving to be extremely useful in related fields and it can therefore be considered to be at the heart of the latest AI revolution.

Natural language understanding

OpenAI is one of the front-runners in this field. In 2019 there was a small riot within the AI community when they refused to published their new language model, which was supposedly so good, that it posed a risk to society. After the dust had settled, they decided to release their model anyway (called GPT-2), which set the new bar for language understanding. It has proper answers to various trivia and can tell you with near certainty that the 'On the Origin of Species' was written by none other than Charles Darwin. If you think that's getting a little creepy, just wait and see what it can write.

Text synthesis

Not only does OpenAI's GPT-2 have a unique comprehension of the English language, it can also synthesize text like no other algorithm. It's eerily similar to the real thing, except that, unlike humans, its spelling is impeccable (check it out yourself). With some tweaks, GPT-2 can function as the conversational engine of our first AGI.

Decision-making

Google Deepmind's AlphaZero has made the headlines several times during the last few years. Initially, it's predecessor (AlphaGo) was the first computer ever to beat a human professional player at the game of Go. Its enhanced version has yet to lose a game, after defeating many of the world's top players. It's current level is so far beyond reach, that it is doubtful any human will ever catch up. In the meantime, it has become the new state-of-the-art in both Chess and Shogi, beating all the leading engines by large margins. AlphaZero uniquely combines quasi-vision technology and a learnt decision-making system, to take actions which eventually lead to a desired outcome. This idea can be extended to support accurate decision-making in AGI systems.

As we've seen, we already have all the ingredients to see, understand, speak and make decisions. If we can piece these together, we're already quite close to a first version of our AGI. It will probably not be able to do much yet, but it will be a good first step. Step by step, we can start tackling the problems that arise, bringing us closer and closer to solving the problem. In the next 20 years, these existing technologies will start to merge, culminating in truly interacting machines by 2040. While unsure what impact this will have on society, it's without question that it will change the world and our lives for good.

We work on the future of AI. Every day.

Regardless of what the future of AI will bring, we intend to be there when it happens. Every day we try to push the limits of AI technology and contribute to this exciting future. In doing so, we create unique value for our customers and maximize our carbon negative footprint. Get in touch if you want to be part of this transition.

Click here to get in touch or look us up at the NLSpace New Year's event!

Comments are closed.