Artificial general intelligence is much closer than you might think by a significant measure. Time to Edit (TTE), a statistic created by a translation company, measures how long expert human editors need to correct AI-generated translations compared to human translations. This can provide a measure of velocity towards the singularity. If an AI can translate speech as well as a human, civilization could change.
The Concept of Singularity in Artificial Intelligence
The concept of “singularity” is widespread in the field of artificial intelligence. This ambiguous concept represents the point at which AI became uncontrollable by humans and drastically changed civilization. The tricky part about the AI singularity is that it's very difficult to predict where it starts, and it's nearly impossible to know what lies beyond this technological "event horizon." So it takes its terminology from black hole physics.
On the other hand, some AI researchers are looking for indicators of singularity, as measured by AI development approaching human-like abilities and skills. According to Translated, a translation agency with offices in Rome, one of these indicators is an AI's capacity to translate speech with the same level of accuracy as a human. One of the toughest AI problems is language, but a machine that can break through this barrier would theoretically exhibit evidence of artificial general intelligence (AI).
“This is because language is the most natural thing for humans,” Translated CEO Marco Trombetti said at a conference in Orlando, Florida in December. But the data Translated collects “clearly shows that the machines are not far from closing the gap.”
Performance of Artificial Intelligence
A statistic known as "Time to Adjust" or TTE was used by the company to track the performance of its AI from 2014 to 2022. This statistic estimates how long it takes skilled human editors to correct AI-generated translations than human editors. Over these 8 years, Translated's artificial intelligence has steadily closed the gap towards human-level translation quality by examining more than 2 billion recent edits.
According to Translated, a human translation usually takes a second for a different human translator to replace each word. Professional editors checked their machine translation (MT) suggestions in about 2015 seconds per word in 3,5, while today it takes about 2 seconds.
By the end of the decade, Translated's AI will match the quality of human translation if the current model continues (or even sooner). In a podcast broadcast in December, Trombetti said, "The change is so subtle you don't notice it every day, but when you see the growth over 10 years it's impressive. “This is the first time a prediction is made about the speed of the singularity in the field of artificial intelligence.”
This is a creative way of measuring how close humanity is to the singularity, but this concept of singularity has the same problems as a more general AGI definition. While mastering human speech is undoubtedly a frontier in artificial intelligence research, intelligence is not necessarily a function of an extraordinary skill.
Whether or not these incredibly accurate translators spell the end of technology as we know it, Translated's AI success has not diminished. Even though the true "technical singularity" is still far away, a machine that can translate speech as effectively as a human could well transform civilization.
Source: Popular Mechanics
Günceleme: 25/01/2023 14:52