"My theory is that the blog writer has a worse than normal impression of GPT because much of their conversation with GPT is in some other language maybe Polish.

"My theory is that the blog writer has a worse than normal impression of GPT because much of their conversation with GPT is in some other language maybe Polish. This isn't such a huge problem because GPT can still do its input and output in Polish, but I think it's commonly accepted that every language other than English adds some cognitive burden to the GPT that makes it dumber. The more obscure the language (or even encoding like morse code or rot13 or base64) the dumber the GPT gets. So the blog writer has a more outdated effective experience with GPT than ones who use it in languages more familiar to it than Polish." It's not necessarily all languages other than English, but certain languages could cause LLM's problems. Here's what ChatGPT had to say about it: Embeddings can be more effective on some human languages than others, depending on factors such as the complexity and structure of the language, the availability of training data, and the quality of the embedding model. Languages with relatively simple grammatical structures, such as Chinese and Finnish, can be easier to model using embeddings, as there are fewer rules to learn and the meaning of words is often more closely tied to their context. Languages with more complex grammatical structures, such as Russian and Hindi, can be more challenging to model, as there are many different forms of words and complex syntactic rules to consider. In addition, the availability of training data can also impact the effectiveness of embeddings. Languages with large amounts of high-quality training data, such as English and Chinese, can be easier to model than languages with smaller amounts of data, as there is more information available to learn from. Finally, the quality of the embedding model itself can also impact its effectiveness on different languages. Some embedding models may be better suited to certain languages or language families based on their design and training methods. Overall, while embeddings can be effective for modeling a wide range of human languages, their effectiveness can vary depending on the specific language and the characteristics of the embedding model and training data. ... While it is true that machine learning models, including those used for embeddings, do not directly know about grammatical rules, they still rely on patterns in the data they are trained on to learn the semantic relationships between words. In languages with more complex grammars, these patterns can be more difficult to discern, which can make it more challenging for embeddings to accurately capture the meaning of words. For example, in languages with complex declension systems, such as Russian or Latin, the form of a word changes depending on its grammatical role in a sentence. This can create many different forms of the same word, which can be challenging for embeddings to learn to associate with a common meaning. Similarly, in languages with complex syntactic rules, such as Hindi or Japanese, the word order and structure of a sentence can convey important information about its meaning, which can be difficult for embeddings to capture without a deep understanding of the grammar. In addition, languages with more complex grammars may have a wider variety of vocabulary, including many words with subtle or nuanced meanings. This can make it more challenging for embeddings to learn to distinguish between similar words and accurately capture the full range of semantic relationships between them. Despite these challenges, embeddings can still be effective for modeling languages with complex grammars. By training on large amounts of data and using advanced neural network architectures, embeddings can learn to capture the subtle patterns of language use and accurately represent the semantic relationships between words, even in languages with complex grammars. However, it may require more training data and more advanced modeling techniques to achieve the same level of accuracy as in languages with simpler grammars.