How Humans Surrender to Language AI
All major tech players are investing heavily in natural language chatbots- another example of betting on AI to substitute the feel of human interaction. So far, this is a meme waiting to happen.
I’m more optimistic now than ever that natural language chatbots will succeed.
It’s all our fault. The greatest factor fueling the viability of natural language AI is how human themselves communicate.
The greatest limitation of any machine learning is the availability of large data sets. Natural language has so many combinations, nuances and cultural differences that an algorithm needs a metric ****tonne of data to build an even marginally useful decision tree.
Any ideas where an AI can find large amount of human conversation?
We humans make companies rich just so we can pile natural language on their platforms. By it very nature, social media is publicly available. For the first time, learning algorithms can be trained outside of well equipped research facilities.
(As an aside, technology also allows humans to be hired at scale for teaching AI English-as-a-second-language. In a recent paper, the Facebook AI Research Group used Amazon Mechanical Turk to generate data specific to their negotiation objective.)
Even though we are in a golden age of cheap computation, the data that we want to analyze is growing even more ambitiously. Any time an AI can constrain an untamed dataset (like natural language), the greater the chance of a reasonable output.
What makes a young AI trying to pass a speak-human exam more likely to graduate?
Small, discrete step size.
Now look at your phone.
Humans are lobbing the gentlest possible softballs to AIs.
Instead of needing ridiculously complex AIs to mimic our very dynamic languages, we ridiculously minimize our language for increasingly simple AIs. In a few years when humans are communicating with nothing but repeating emojis, the AIs will be needed to teach us natural language. (Oh wait, the future is now.)
AIs have a hard time acting casual. Without some end goal, AIs have a have no ability to evaluate their decisions.
Using AIs to negotiate, sell or solve gives a destination to aim for and weighted midpoint outcomes to make decisions with. Natural language is then a communication layer in a limited domain that — again — limits complexity and makes the outcome of an AI appearing to speak ‘human’ more likely.
Enterprise built chatbots are not designed for passing academic Turing tests. They are built to add dollars to a bottom line. This refined focus gives a learning AI Ritalin.
There it is. Dollars quantify decision making and endpoint valuation. Capitalism helps AI.
Technology does not happen in a vacuum. The success or failure of what we build is measured in the context of the larger world.
In this case, the world is moving under its own power to better fit with where natural language AIs work. Unlike many other AI applications, my crystal ball tells me natural language is a technology we will see sooner and more often.