I have used ChatGPT-4 to generate practice texts in Swahili in increasing difficulty levels as I was learning the language. Swahili is a great language, really easy to learn and pronounce, but there is no so much good learning material out there. It is esp. hard to find explanations for grammatical contructions but ChatGPT-4 had read enough to be able to explain the differences of quite alike words like halafu vs. kwisho or zaidi vs. kuliko. It is really nice that you can ask it about parts of speech and let it explain all parts of a sentence. It is also nice to that the learning text is generates are tuned to my interests and can, e.g., describe cities in the form of a local people's experiences. I also managed to coax it to generate "street Swahili" which is code-switching between Swahili and English as is very common in Kenya and esp. Nairobi.
> I also find it struggles more with larger applications and probably can't architect or refactor.
So far my experience is it’s awesome at refactoring. I’ve pasted in files from a legacy codebase that are ~200 lines, written entirely with outdated web libraries, and asked it to modernize the file / component but keep certain aspects of the logic while fixing some aspect of it.
And it’s been able to do it correctly, to the point where I can just directly paste its result in without changes and everything still works correctly throughout a large enterprise application.
I also only use the Code Interpreter model of GPT-4 for coding tasks, which supposedly has an 8k context window and others report it being more logically consistent.
oh that's super interesting. I had tried to used gpt4 to make a react app but once it got past a trivial size it started to get confused. this was before code interpreter came out tho
I will say that while it's likely GPT is fine at big languages like Portuguese, my experience with smaller languages is it tends to make several errors and introduce constructions that are translated literally from English. Of course, that is based on my experience with GPT3.5, perhaps GPT4 is better in that regard.
I have used ChatGPT-4 to generate practice texts in Swahili in increasing difficulty levels as I was learning the language. Swahili is a great language, really easy to learn and pronounce, but there is no so much good learning material out there. It is esp. hard to find explanations for grammatical contructions but ChatGPT-4 had read enough to be able to explain the differences of quite alike words like halafu vs. kwisho or zaidi vs. kuliko. It is really nice that you can ask it about parts of speech and let it explain all parts of a sentence. It is also nice to that the learning text is generates are tuned to my interests and can, e.g., describe cities in the form of a local people's experiences. I also managed to coax it to generate "street Swahili" which is code-switching between Swahili and English as is very common in Kenya and esp. Nairobi.
oh that's so cool!
> I also find it struggles more with larger applications and probably can't architect or refactor.
So far my experience is it’s awesome at refactoring. I’ve pasted in files from a legacy codebase that are ~200 lines, written entirely with outdated web libraries, and asked it to modernize the file / component but keep certain aspects of the logic while fixing some aspect of it.
And it’s been able to do it correctly, to the point where I can just directly paste its result in without changes and everything still works correctly throughout a large enterprise application.
I also only use the Code Interpreter model of GPT-4 for coding tasks, which supposedly has an 8k context window and others report it being more logically consistent.
Also, great post.
oh that's super interesting. I had tried to used gpt4 to make a react app but once it got past a trivial size it started to get confused. this was before code interpreter came out tho
would love if I could just voice talk to gpt all day long
The official chatGPT app has whisper voice recognition built in. Something like 99% accuracy
whisper.cpp :)))
I will say that while it's likely GPT is fine at big languages like Portuguese, my experience with smaller languages is it tends to make several errors and introduce constructions that are translated literally from English. Of course, that is based on my experience with GPT3.5, perhaps GPT4 is better in that regard.