Most Artificial Intelligence Isn’t Truly Intelligent
People have long dreamed of machines having the intelligence and capabilities of humans. From the early Greek myths to over a hundred years of science fiction stories, novels and movies: human imaginations have envisioned what it would be like to have sentient, intelligent, human-like machines co-exist with us.
In many ways, the quest for the intelligent machine lead to the development of the modern computer. Ideas by Alan Turing not only formulated the basis of programmable machines, but also the core of the concepts of artificial intelligence, with the namesake Turing Test providing a means for evaluating intelligent machines.
Yet, with centuries of technological advancement; and the almost exponential increase of computing resources, data, knowledge and capabilities; we still have not yet achieved the vision of Artificial General Intelligence (AGI) — machines that can be an equal counterpart to human ability. We’re not even close. We can talk to devices, but they don’t understand what we’re saying. We have cars that will drive straight into a wall, if that’s what your GPS instructs it to do. Machines can detect images, but they still don’t understand what they are. We’ve mastered computing. We’ve wrangled big data. We’re figuring out learning. But we still have no idea how to achieve general intelligence.
Part of the reason for this disconnect is the confusion of the various things we’ve developed as a result of our quest. Artificial intelligence is not a technology (in much the same way that the ‘space race’ is not a technology,) it’s the journey. All the technologies so far developed are individually useful — but they haven’t yet fully integrated in order to deliver us to the end goal.
Framing ‘True’ Intelligence
Today, what we refer to as AI is invariably machine learning. It is Twitter selecting the most relevant tweets for you — not because of intelligence, but because it has assessed what you’ve engaged with previously and serves you up more of the same. It’s the clever Spotify algorithm that suggests an array of eclectic music to discover, that you will probably love. And it is the ‘intelligent’ personal assistant that knows what you usually order from the Chinese takeaway — allowing Siri and Alexa to order your meal via Just Eat (and making the process a little too easy). Mighty clever though this all is, it isn’t intelligence.
AI, at its crudest, is the chatbot that pops up during your online banking. Whilst it has the veneer of AI, there’s nothing intelligent about it. It’s simply a series of keywords with default responses. Some of them are quite sophisticated (the Google Duplex assistant, for example, might just be the current pinnacle of this), but if you’ve ever used one and felt you were dealing with a real human, then you haven’t really been paying attention.
There’s currently no ‘great’ conversational agent. During a natural, human conversation, the topic, sentiment, tone and subject will often fluctuate and change. State-of-art tech is just not ready for general conversation; unless it’s restricted to a certain domain, purely topic specific, and scripted.
Constellation are approaching things differently. We’re creating a truly intelligent conversational AI assistant, by understanding how people learn language.
How Do You Create Intelligence?
If you want to teach a child how to learn, you talk to them using very simple phrases. You show them how words should be used, and they begin to learn by repetition, with an understanding of context and the assignment of meaning. We’re using the exact same principles to help AI learn language.
Right now, no other available technology considers the meaning behind the language. Instead, they’re studying huge amounts of data using neural network processes — which only allow you to assess the statistical occurrence of a combination of words in data. It might be statistically meaningful, but It doesn’t touch upon real meaning — and it’s not going to provide impactful conversation. You can see this sort of work in action with Google translate, for example. It’s precisely why current artificial intelligence is not intelligent.
Constellation have used a unique combinatory approach to build a system with a knowledge base that can fundamentally understand meaning and context.
It’s a system that moves beyond recognising the simple similarity between sentences, due common words or sentiment. We’ve pushed through current paradigms to a system that is able to understand what meaning is — it can decipher, understand and then generate a true, unscripted answer.
We’ve also developed a new technique for the conversational agent to self-assess the meaningfulness of its response and apply its own correction. It’s like its own self-improvement feature, that will help it advance over time. The combination of these two tricks are the tools we need to generate an incredible conversation. And the more data we have, the more precise the conversation will become.
It’s the first step to having a high level, almost human-like, conversation with AI and the beginning of a change of paradigm, not off-the-shelf solutions. It’s the future of truly intelligent artificial intelligence.
With Alfredo Gemma — VP of AI and Engineering, Constellation AI.
Alfredo is a researcher in Cybernetics and Deep Learning with more than 20 years’ experience in the design and development of AI, including: Distributed, Concurrent and Iterative Computing, Fuzzy Logic systems, Inferential Systems, Neural Networks, Evolutionary Computing, Cloud Computing, UML Modelling, Quantitative and Qualitative Software Metrics, Network Security, Cryptographic Systems, Intrusion Detection and Hack-Proofing techniques. He is VP of AI and Engineering at Constellation AI and Founder of web forecasting platform SEE-R™.