Why the Future of AI is Like Playing with Lego

Share on facebook
Share on google
Share on twitter
Share on linkedin

The invention of Lego brought many young kids and adults both joy and mental development. Since the early days of Lego, the number of parts has quadrupled, enabling users to build and put pieces together in more than a billion new ways. However, the power of this ‘plastic gold’ lies not in the fact that there are so many bricks and colors available, but rather, that the fundamental principle always stays the same: bricks should always fit into each other and may be used in different ways to create something unique out of mere blocks. As a result, you could use bricks to create new narratives with untaught outcomes.

If we extend this analogy to AI, we can notice a few parallels. Most AI discoveries today, such as autonomous drivingprotein folding, conversational AI, are also built on standards. For example, the most popular programming language in AI used today is Python. Along with that, the use of deep neural networks[1] has become its standard architecture. Not to forget, the major revolution of transformers[2] in the NLP domain. With transformers, you can basically use near-human level language understanding without the need to obtain the data or to train a model yourself, often applied in one line of code. All in all, these standardizations make it much easier to ‘connect’ them as ‘bricks’ and build new solutions.

Community-driven evolution of AI

The research community of AI is organized in a way that links to that of Lego. Compared to most research topics, AI research is different in its nature. One of the fundamental differences is that AI tries to mimic something that has existed for thousands of years, namely: human intelligence. This makes it somewhat easier to benchmark study outcomes, namely, we can directly check to how well humans would do it compared to an AI model. This makes feedback loops to be much shorter.

It is there to say that there is a general trend in AI research that tends to build on the AI principles, models, architectures that show good results for a certain AI task. Many AI papers, therefore, are an alteration or an improvement on the existing AI model: take for example BERT. Not surprisingly, amongst all research areas, Artificial intelligence papers amass citations more than any other research topic.

I’m not saying AI research is easier, nor is it less groundbreaking. The point here is that the nature of AI research is different from most, as it opts for a more open and community-driven evolution. AI has become much more like standing on the shoulders of others. This analogy was also pointed out by Clément Delangue, founder of an open-source AI platform called Huggingface. HuggingFace, together with PapersWithCode and Kaggle, saw this trend and jumped into it by providing a platform for the community where you can easily share and continue building on new and better AI creations.

So what does this have to do with Lego? Just like the AI research community, the power of Lego is not entailed as a single entity. Neither the Lego Corporation nor a single AI institute alone is responsible for the majority of its success. Instead, it was accomplished by first setting a standard and then allowing it to be shared as creations worldwide. AI, just like Lego, has developed a narrative that may even be considered a way of communication. The narrative is that everyone in the community understands and attempts to design and build better variations of existing creations.

Let’s build something

But how does this work in practice? Let’s build something of AI bricks ourselves to find out. Suppose we want to create a virtual customer service agent. To obtain the solution, we require a high level of understanding of natural language. For this, we need to produce a list of building blocks necessary to create an intelligent virtual agent. This could include intent and entity recognition, context understanding, answer generation, and so on, depending on the circumstances provided. Solving this task is very complex if you plan to use one overarching model. However, there is a more straightforward method. By connecting the outcomes of multiple models that are the best ones to date, you can achieve the same result. With this sum of narrow AI models, you can achieve maximum performance while maintaining the flexibility to change components once better alternatives are available. This will help you create a solution quickly, easily, and more adaptable to improvement in the future.

Back to the customer service agent. Suppose a customer writes a long service question. You can use DistillBERTto detect the intent first and Albert to obtain potentially present information present in the question. As shown in the figure below, we need the key date, moving date, and new home’s address to completely set up a bridging contract for a customer.

You can ask any question on a given passage, and Albert understands what is being asked.
Future of AI and the race for superintelligence

Let’s take a step back to Lego once again. In this article, I’ve described some concepts and ideas of Lego and AI that flow in a parallel direction. For example, we need to rely on standards as well as getting inspiration through other creations. Therefore, the best suggestions for creating unique AI models are likely the result of community effort rather than a single entity. After all, nothing can beat a worldwide community with thousands of researchers. Although there are more prominent entities out there like Google, Facebook, Microsoft, or Nvidia that try to push the limits of AI with bigger datasets or more computing power. Just like with Legos, the fact that you have more bricks to spare doesn’t mean you have the ability to come up with the best creation.

Also, Lego understands this as well and launched Lego Ideas a few years ago. Lego Ideas is a platform where they ask the crowd to design and vote for creations. If a creation is found good enough by the community, Lego might release it as a set. It can clearly be concluded that the future of AI is bright as it continues to become more accessible to everyone.

‘Playing’ with AI in your organization

Whether you are a business owner or working for an organization, you will probably notice how AI slowly gets embedded everywhere. If you plan on working with AI in your organization, you need to be open to the fact that applied models for the task at hand might be outdated at some time, and newer versions (or ‘designs’) might outperform the old ones. In this case, don’t worry, just like Lego, new versions of models will still follow the same standard. Just make sure your organization is flexible enough to change components every once in a while, and provide training to colleagues to remain keen on new AI models.

Just like Lego’s motto says, “Only the best is good enough.”

[1] Using deep neural networks in the form of CNN for Computer Vision or RNN/LSTM’s for NLP

[2] Transformers are pre-trained models for NLP solutions. Examples are BERT, DistillBERT, XLNet, GPT-3, etc.

Read the full story on Medium

Written by Thomas Schijf