ChatGPT-4 Is Impressive In Some Ways, And In Others, Not So Much

I was curious as to the the existence of a dialect continuum between Russian and Ukrainian and how it potentially manifests in Russian dialects spoken in the southwestern oblasts of Russia.

I started the conversation in Finnish but switched to English because it answered in uninteresting generalities in Finnish. I was right and the English language answers were better.

It needed more prompting. It helps if your prompt is more specific.

Now I put it to the test because I'm very familiar with regional variation in the Finnish language having relatives in many regions of the country.

That was quite generic but it is notable that Russian has a lot of loanwords from many other languages. No one who studies it can fail to notice it right away. It is conceivable that those would not have been uniformly spread across the country before modern mass communication.

I'm a bit skeptical of the first part of the conclusion that it inferred from the great geographical spread of Russian but the AI is correct in that Finnish has a lot of local variation. The variation in the informal registers is large enough to trip up non-native speakers rather easily. Even basic vocabulary like personal pronouns can appear unrecognizable to the untrained ear of a beginner.

It linked that observation to the concept of founder effect. I'm not sure if it's an often used concept in linguistics but it makes perfect sense.

When I asked the same version of ChatGPT to come up with funny acronyms from the initial letters of the names of countries on the list of the current and the newly invited BRICS countries, it insisted on including Denmark on the list despite my asking repeatedly not to do so and only include countries on the list. :D

I've found ChatGPT-4 a powerful search tool. When you know little about a topic but have a specific question it's much faster to get started using it than googling because that approach forces you to read tons of text containing new information, which is a daunting task. But you have to be skeptical of all information it spits out. Large language models (LLM) suffer from what is called the hallucination problem. LLMs create continuations to prompts based on the likelihood of the next word given the previous content based on a very large simulated neural net trained on a massive number of texts. They simply generate what is plausible give their training data. While what they produce is often sensible and correct, sometimes it isn't. Such models have also limitations when it comes to logic and reasoning. They can appear smart but make really silly mistakes.

I've also found LLMs helpful in code generation provided they've been trained on sufficient quantities of code in the language in question. ChatGPT-4 is capable of writing and debugging code when given a specification. Such models can definitely make programming work more efficient. The greatest disadvantage, however, is their narrow context window. The number of characters in a conversation current LLMs can keep in memory varies from thousands to tens of thousands, which is not much. A programmer who uses them still has to manage breaking up the solution into smaller chunks. But things like starting to use an unfamiliar library are a breeze when using a properly trained LLM like Copilot. A lot of time spent reading manuals can be saved. These LLMs can find even suitable candidate libraries quite easily when you describe the problem to them.