Artificial Intelligence Has a Mechanical Turk Problem

In 1770, a Hungarian inventor named Wolfgang von Kempelen presented an automaton capable of playing chess to Empress Maria Teresa’s court. It consisted of a large box containing gears, cranks, and levers, and behind it stood a mechanical player wearing a turban and an oriental outfit.

The Mechanical Turk

The player was quickly dubbed the Mechanical Turk thanks to his stylized garment. Before performances, von Kempelen would open the box, and holding a candle, he would show the audience that there was no human inside. The Turk went on to beat most of the chess masters of the 18th and 19th centuries, including Benjamin Franklin (an ambassador to France at the time), and Napoleon Bonaparte.

It was an elaborate illusion that fooled crowds on both sides of the Atlantic for almost seven decades. Actually, there was a human chess player inside the box. He would get out just before von Kempelen did his look-there-is-nothing-there act and would go back in to move pieces above him using magnets.

The Mechanical Turk was the first Artificial Intelligence persona, a non-human that claimed to possess superior human intelligence.

Nowadays, it serves as a shorthand reference to ingenious constructs designed to fool us (well, until Amazon appropriated the moniker).

Given my ancestry, it is an aptly ironic metaphor to highlight my skepticism of the current Machine Learning AI paradigm.

Let’s take a look at some prominent examples.

Self-driving Cars

As I mentioned previously, having a computer drive a car between two points is not difficult. Thousands of vehicles operate with some form of autonomy; if you add Teslas, millions.

However, genuinely autonomous cars with no driver in the vehicle is a different proposition.

Sure, some driveless robotaxis are available in Phoenix, Los Angeles and San Fransisco.

Robotaxi

And 600 of GM’s autonomous Cruise vehicles have been driving in California.

Good, you say.

Except, they are not really autonomous:

A recent report from the New York Times revealed leaked data that General Motors’ Cruise robotaxis were seeing a remote “intervention” every 2.5 to 5 miles, and that Cruise had 1.5 staff for every car on the road [my emphasis]

Think about it. More than one driver is driving a driverless car.

German startup Vay is using what it calls tele-drivers to remote control its autonomous vehicles in Nevada. At least, they are doing it openly.

After hundreds of billions of miles of driving information, these self-driving cars equipped with cameras, Lidar sensors, and AI-powered software can only operate “within strict, pre-set geographical boundaries” and only in warm climates.

If the Machine Learning paradigm is supposed to be the way to AGI, it is hard to explain the current inability to have truly autonomous vehicles and many fatal crashes to date despite their AI being trained with the most extensive dataset imaginable.

Another example.

Amazon’s ‘AI-Powered’ Cashier-free Shops

Remember Amazon’s Fresh grocery stores with cashierless technology? The frictionless shopping experience, where customers could pull stuff from shelves, place them in their carts, and walk to their car without going through a cash register.?

Cashierless store

We were told that Amazon’s “’ Just Walk Out’ technology, which uses sensors and machine learning” would do the rest.

It was greeted as “Amazon wants to kill the supermarket checkout line.”

Well, would you believe it, the Mechanical Turk struck again.

More than a thousand people, mostly located in India, were watching these stores, noting purchases, and providing the final purchase list.

Apparently, 70 percent of sales were handled by humans and not by AI, despite hundreds of billions of transactions to train the AI.

Ultimately, the Mechanical Indian was too expensive and Amazon quietly ditched the AI-powered “Just Walk Out” technology from its stores.

They now have scanners in shopping carts. Amazing.

On a side note, this is another pet peeve of mine: People calling for AI for tasks that have already been automated using regular computer programs.

There is no need for AI for such tasks. Scanners in shopping carts was and is one such technology.

Python Programming

It has been said (quite accurately IMO) that Generative AI or LLMs do not write text, they simply generate it. Consequently, they should be quite good at programming tasks as they involve -primarily- generating code.

It turns out that they need a Mechanical Turk, even for those.

As Giovanni Toschi recently made the point on LinkedIn, you get good results if you seriously rig the dataset.

Instead of “training” the AI with millions of Python development projects, you just feed the best work of the best software engineers.

Suddenly, AI produces better code than your niece/nephew Alex, and we are all amazed and proud.

Google Search

This is a popular example as it entails “glues to hold pizza cheese” and “daily ingestion of rocks for geologists.”

It was Google AI repeating what it learned on Reddit from a sarcastic posting by a user with the handle “fucksmith” regarding pizza cheese and the way to make sure it stays in place. Elmer’s glue, as you can see on the screen shot.

And another one was doing an impersonation of the satirical news site, The Onion.

The Machine Taught AI had no clue. The information was on the Internet, and therefore it was plausible and correct.

I imagine many robotic figures with a turban in Google HQ running around to remove ludicrous answers.

It is my Moment of Zen (apologies to Jon Stewart). 

The point remains that AI will really have a hard time moving to AGI with the Machine Learning paradigm.

Even Sundar Pichai acknowledges that the answers to search queries are correct 80 percent of the time.

That is hardly AGI material.

Clearly, we need some context, common sense, and, most importantly, reasoning before we can talk about AGI.

And I seriously doubt that it is going to happen with Machine Learning.

More on that later.