- Prosody. Prosody can vary per region but a distinctly alien prosody to a language is a barrier for the receptor of the message, that expects a given language and a range of prosodies. E.g. as I know french quite well, hearing english with a heavy french accent makes my brain try to understand what's being said as said in french, and interferes a lot.
- Sound shifts for particular phonemes. While some of it might be local to the language in certain registers (idea --> /ide"er"/, three --> /free/), others are clearly issues in the target language pronunciation (eg. japanese people having trouble with the l phoneme, spanish people adding an /e/ sound prior to an s-mobile, or v versus b for spanish people also).
- Connected speech. Where do you end words, how do you omit sounds, etc. Also massive hindrance to understanding.
- Grammar. Alien grammar is a hindrance to communication. You need to learn that. 1. Even if the task is unreasonable, it is good to showcase that the LLM will perform poorly - warning not to be used for diabetes.
2. As it is a probabilistic model, the approach was to execute it multiple times and look at the distribution. They also tried to minimize variance: "All at the lowest randomness setting these models offer.", the post mentions. Yet the variance of the responses is surprising.
3. A multimodal LLM should be in general able to discriminate between crema catalana and a cheese sandwich, and provide a textual, uncalculated range of how much calories the item has (internet is full with tables for calorie counting and things such as this https://fitia.app/calories-nutritional-information/cheese-sandwich-1205647).
4. It is not clear that the "expose" surprised / outraged style is just a communication vehicle or if the author really thought that e.g. LLMs could be hypothetically able to provide confidence estimates.
> So you can formally prove that your e-commerce refund management logic is correct, except for proving that you /processed the refund/. You can't even prove anything about recording the refund in your database, say nothing of proving anything about your interactions with your payment processor.
You could say the same thing about the viability of functional programming on a CRUD webapp, but languages like clojure have been used to great effect here. The fact that thera are important, even fundamental, bits that you cannot verify, doesn't take out value from the fact that you can eliminate whole dimensions of issues.