Excitement in the artificial intelligence (AI) scene: Since Google fell out with an engineer who attested the chatbot system LaMDA to be conscious, things seem to have gotten out of hand. The software, whose full name is “Language Model for Dialogue Applications”, can generate very realistic-looking texts, take on different personalities (personas) and create real-sounding conversations over long stretches.
controversy over consciousness
While Google researcher Blake Lemoine claims that we are dealing with a real being with personality, Google and numerous critical thinkers from the AI scene and machine learning (ML) reject this – Lemoine even for religious reasons, among other things reasoned – idea flat out.
But whatever one thinks of the heated debate, it shows that current methods of evaluating AI systems are no longer sufficient – especially the so-called Turing test. This was designed by computer science legend Alan Turing and is intended to determine whether a computer has the ability to think on a par with that of a human. The problem: Text creation systems like LaMDA could soon overcome the hurdle because they simply have enormous amounts of data at their disposal, which meanwhile make their texts extremely realistic – which brings us back to Lemoine.
400 researchers for 204 AI tasks
The topic has long been recognized by Google. There, the Turing test from 1950 is now to be replaced by a whole battery of tests that come from numerous fields of knowledge. A total of 204 tasks have to be completed, according to the group, whose AI researchers have joined forces with over 400 colleagues from 132 institutions worldwide for the project. The question has been worked on for a good two years – and it could not have come at a better time.
Instead of the Turing test in several variations, the “Beyond the Imitation Game” benchmark, BIG-bench for short, is to be used in the future. (“Imitation game” Turing originally called his test.) With the new method, a general artificial intelligence (Artificial general intelligence, AGI) could also be determined – at least theoretically.
Turing stays – adjusted
The Turing test cannot be mothballed, Adrian Hilton from the University of Surrey in Great Britain told the New Scientist. But this is not wide enough and suitable for modern AI systems. However, he is skeptical about whether BIG-bench can really determine AGIs. After all, the approach now offers a way of better comparing systems.
BIG-bench does this by querying all possible subject areas. This includes classics such as chess as well as language comprehension, mathematics and logic. Initial investigations with the test showed that the 204 tasks in their entirety are still too difficult for current AI systems, which are mostly specialized. Man has not yet been caught up here.
To home page
#Artificial #Intelligence #LaMDA #Google #Turing #test