For anyone who thought we had reached it, language AI seems not to have a ceiling in 2021, with news this week of even more powerful models than GPT-3 and new applications of the former suggesting it could be a vital link to general artificial intelligence (the hypothetical AI of all things).
GPT-3 is a text generation AI that gained fame and notoriety in 2020 for its ability to pose credibly among real people in online forums where participants do not need to show their faces or be heard aloud. Pushing its capacity to reflect human qualities in automated speech, researchers are now going further to explore how the algorithm can master challenging facets of language such as satire and humor. Nature reports researchers have used examples of satirical dictionary definitions to train a version of GPT-3 that can provide the same kinds of definitions whenever prompted with a word. Given the word ‘Literature’, it responded, “A name given to other people’s published papers, referred to by scientists without actually reading them.” For ‘Scientist’, it said on one occasion, “A person who has a laboratory, access to large sums of money, graduate students, or all three.” Asked again, it said, “A field based on science, devoted to completing works for which there will not be enough time in a single lifetime.”
A sense of wit that resonates with people is difficult for humans to cultivate, regardless of how intelligent. By trying over and over, GPT-3 was able to generate results that suggest a greater command of wit than the average person’s. Still, the true significance of these results seems less their persuasive human quality than their consistency. We can all get laughs sometimes, but who can be funny and on-point all the time? GPT-3 could probably be funnier than it is now, but it doesn’t seem prone to falling flat every other time it tries. That consistency is what makes AI a powerful disruptor in any field, and with natural language processing (NLP), the sense that creative tasks belong to a realm beyond its reach has taken a strong hit over the past year.
Meanwhile, news comes this week of another NLP model that exceeds GPT-3’s sheer power by a factor of about 10, using over one trillion parameters to model patterns in both Chinese and English. Created at the Beijing Artificial Intelligence Institute, the model known as Wu Dao is a multi-modal (i.e., cross-purpose) AI that is reportedly dedicated to helping realize general artificial intelligence. While there is little clear performance comparison data, Wu Dao has the unique credentials of being bilingual and able to create multimedia outputs like songs and animations, and should bring interesting developments ahead.
If the stereotypical image of AI was once a clunky robot speaking in awkward, matter-of-fact statements, that has been replaced with something invisible in its workings and almost undetectable as artificial from what it puts out. AI has been a reality in language services for longer than most fields, from machine translation (MT) to document analysis. Importantly, as in MT and other AI-driven linguistic services, it is the unique handling of the underlying technology that seems to determine how effectively more experimental models can perform within a specific domain, just as training and translation memory can define the scope and specialization of an engine’s performance. With NLP advancing, the raw tools at linguists disposal may change in the near future, but the fundamental methodologies and moderating human expertise behind AI’s use in professional translation have a persistent importance when it comes to ensuring truly human (or human-like) communications in the most culturally nuanced and challenging areas of a language.
To learn more about CSOFT’s AI-powered translation services and work with innovative technologies, visit us at csoftintl.com!