Thanks for picking up on this, Derek.
I was thinking of a clever opening to greet you, and by instinct did both a reach for web search for an regionally appropriate expression. While I found many results, and will go with “Molo” it feels good to just say (Xhoso for “how are you doing?”). My question to ChatGPT returned a very chatty response, ultimately suggesting English (sigh, very factual, but not fun)
In South Africa, people speak a variety of languages, with 11 official languages recognized in the country. These languages include English, Afrikaans, isiZulu, isiXhosa, Sesotho, Setswana, Sepedi, isiNdebele, Xitsonga, Tshivenda, and SiSwati. Since you’re not sure of the recipient’s first language, it’s a good idea to use a greeting that is widely understood and accepted across South Africa. English is one of the most commonly spoken languages in the country and is widely used for communication, so “Hello” or “Hi” in English is a safe and appropriate choice.
You can simply start your message with “Hello” or “Hi,” which should be well-received by most South Africans regardless of their language preference. If you want to add a personal touch or show respect for their culture and language, you can also consider learning a basic greeting in one or more of the other official languages, but using English as a default is generally acceptable and polite.
Is this so? It seems so… bland.
Part of my own dilemma in this space continues operating from mostly inference, results, and guessing as conceptually how these work under the hood. And also being closed systems with a sprinkling of “open” in brand names, do we really have influence as to expand the training?
And yes, a first order thought is, and for all I know might be workable, is to create an an LLM or run one trained on known sources. But can it be as effective with the order of magnitude (and computing energy) that the big ones do? A better corpus ought to produce better results, has this what is seen in the initiatives you mention?
I have also read that a different approach is “engineering” the way the large systems work-- the analog (maybe wrong?) that we do not necessarily try to build our own web search platforms but try to slice the ones we have to serve more appropriate results.
I am eager to hear more from Dan’s work, as he has described the positive results from using translation AI (though trained, likely not the same can of beans as LLMs) for producing content in languages such like isiXhosa.
Looking forward to learning more from Dan’s project and hearing first hand how it is being used this week: