You know you’re not a kid anymore when mom trusts you with money and sends you out by yourself to negotiate with the shopkeeper and buy a few things.

Similarly, back in 2017 , researchers at  Facebook Artificial Intelligence Research(FAIR) set out on the task of building AI agents that could negotiate.

Negotiation you see is a far more complex activity that involves reasoning  and the ability to combine information with worldly knowledge.

Up until then Chat bots could only craft simple sentences and hold short conversations. During the course of the experiment something very peculiar happened which was very puzzling at first. The two bots that were negotiating with each other were found to be talking in a language that they both understood but was something incomprehensible to humans. 

What the bots had indeed managed to do was, to come up with a shorthand of sorts of their own and were now communicating more effectively using it.  

While it was widely believed that the FAIR team shut down the bots because of this, nothing has been mentioned regarding the same on their website (https://engineering.fb.com/2017/06/14/ml-applications/deal-or-no-deal-training-ai-bots-to-negotiate/) and as per the results stated their best agent performed as well as a human negotiator.

A year before this , Google’s Translate switched to a new system called Google Neural Machine Translation (GNMT). As Mike Schuster, Melvin Johnson and Nikhil Thorat explain in their research paper(https://arxiv.org/abs/1611.04558) and blog ( https://research.google/blog/zero-shot-translation-with-googles-multilingual-neural-machine-translation-system/), how the GNMT has now made zero-shot translation now possible. Previously , Translate only worked between languages it had been explicitly trained on.

It could translate between Korean and English when it was shown examples of them. Let us also consider English and Japanese similarly to help explain zero-shot translation. Translate couldn’t formerly directly translate between Korean and Japanese based on the Korean -><- English and English-><-Japanese samples it was shown. However with GNMT , it could, and more interestingly how it did this was by creating an interlingua, an artificial language that it used internally, that helped it achieve this zero-shot translation. 

If seeing AI communicate in incomprehensible languages is not futuristic enough , listening to two AI models talk to each other surely is. A video that went viral captured two AI agents conversing with each other , who on discovering that they were both talking to another AI agent , switched to gibberlink to communicate faster. Gibberlink , an audio protocol based on ggwave, increases efficiency by 80%. Gibberlink might be incomprehensible to the human ear, but is something that can be decoded. 

AI models ingeniously coming up with their own short hand to successfully negotiate, interlingua helping us translate between more languages than ever before , or models talking in gibberlink are all positive developments and are splendid examples of technological breakthroughs-solutions that only AI could have made possible. 

 But what if this ability of AI to converse with other AI in incomprehensible languages potentially becomes a very grave threat and a problem of epic proportions to us sometime in the future?  AI models conversing with each other in ways we can’t comprehend , a phenomenon which by itself, might just seem unsettling at first and lead to FOMO at worst. However, when coupled with all the other risks that go with AI  (AI going rogue, AI agents developed by bad actors for malicious purposes), is when the alarm bells rings. Only when we imagine the events that could transpire do we begin to grasp the true magnitude and gravitas of the rabbit we have unleashed from Pandora’s box. 

Even before Generative AI models took our world by storm , technology had been incessantly upgrading. The world went from predominantly desktops, to laptops , to mobile, to cloud in less than a decade. API’s have helped us build a digital world with seamless connectivity . A great many applications are connected and talk to each other and all our data is online. 

If Rogue AI models that act against our interests communicate in ways we do not understand, we will possibly never know what their designs are. If AI agents developed by bad actors such as criminals , or cyber attackers communicate with AI systems in unintelligible ways , we might only learn too late when key systems have been compromised. 

John McClane managed to save the day in Die Hard 4 and put to rest a fire sale attack , but it’d be better if we were prepared and prevented a threat such as that in the first place.  In this hyperconnected world , where if catastrophe strikes , there is the possibility of a chain reaction setting off and the problem even escalating to colossal dimensions.  

Geoffery Hinton , also called the Godfather of AI , who already believes AI is now superior to humans has issued stark warnings and says it is time to confront the existential dangers of AI. (https://mitsloan.mit.edu/ideas-made-to-matter/why-neural-net-pioneer-geoffrey-hinton-sounding-alarm-ai) One of his concerns lies in the likelihood of our inability to limit and control AI development. He believes it is imperative that we make sure AI does things that are beneficial to humanity but also it is going to be very hard to do so with bad actors around and someone or the other will soon wire into them the ability to create subgoals. Several leaders from the Association for the Advancement of Artificial Intelligence have called for collaboration to address the promise and risks of AI. Ethics , and Ethics in AI has perhaps never before been as important as it is today.  

Every major breakthrough from the wheel , to agriculture , to electricity to computers has been a testament to and reaffirmed the fact that what distinguishes man from other animals is his ability to think ,his intelligence. And man today, perhaps emboldened by the civilization he has built for himself , has now set out on the ambitious objective of recreating that very thing , intelligence , artificially . The areas of Electronics, Computer Science and Artificial Intelligence have seen such rapid developments in the past few decades that it is imperative that we equally prioritize the formulation of rules and legislation of policy around AI to ensure safety and security. For the future might not be as distant from the present as it once was in the past.