Want to know what’s brewing in the world of AI? Here are some key takeaways from this article:
- One intriguing metric indicates we may reach technological singularity by the decade’s end, potentially even sooner.
- A translation services firm has come up with a gauge, coined Time to Edit (TTE), that measures how long it takes for professional human editors to refine AI-generated translations versus those crafted by humans. This could shed light on how quickly we might reach singularity.
- An AI capable of translating spoken language with human-level accuracy could potentially revolutionize society.
The term “singularity” in the realm of artificial intelligence carries a significant weight. It’s that puzzling point at which AI transcends human management and starting to reshape our world. This concept is tricky to pin down, similar to that of black holes—hard to predict when it kicks in and nearly impossible to envision what life will look like in that scenario.
Regardless, researchers in AI are actively seeking indicators that suggest we are nearing singularity, particularly through advancements in AI capabilities that begin to mirror human skills.
One notable measure from Translated, a translation company based in Rome, is how accurately an AI can translate spoken language compared to a human. Language itself poses a huge challenge for artificial intelligence, and a machine closing in on that gap could be a sign of Artificial General Intelligence (AGI).
As CEO Marco Trombetti stated at a conference in Orlando in 2022, “Language is the most innate ability for humans. And the data collected by Translated highlights that machines are getting really close to bridging that gap.” (source)
From 2014 to 2022, Translated monitored their AI’s performance through the TTE metric, which measures how much time human editors spend correcting AI-inferred translations in comparison to those done by actual people. Analyzing over 2 billion post-edits over these years, it became evident that their AI showed steady, albeit gradual, improvements toward achieving translation quality on par with humans.
Typically, it takes a human about one second to edit the translation of another person’s work. Back in 2015, human editors needed around 3.5 seconds to refine a suggestion from machine translation; today that figure has dropped to just 2 seconds. If this trajectory continues, Translated’s AI may match human translators before this decade wraps up—or even earlier.
“The adjustments happen so gradually that it seems unnoticeable day by day, but if you zoom out to a decade, it’s remarkable,” Trombetti remarked during a podcast. “This represents the first instance where AI professionals have made a prediction regarding the pace of reaching singularity.” (source)
Even though this innovative approach provides some insight into how close we are to singularity, pinpointing exactly what singularity means also presents challenges similar to defining AGI more broadly. Achieving human-like conversational skills is certainly a frontier in AI exploration, yet mastering language alone doesn’t equate to being “intelligent”—and there is ongoing debate about the very definition of intelligence in AI. (source)
Regardless of whether these highly accurate translators signify an impending threat to our technology, it’s hard to downplay the achievement of Translated’s AI. If it can manage speech translation on par with humans, the impact on society could be monumental, even though the real “technological singularity” might still be out of reach.
