Note too the very bounded nature of Japan's use cases. Collins' use of language here reminds of 'stochastic parrots' - a term coined by Emily Bender to capture what LLMs really do, which is to string words together through statistical probability, in the hope they make sense.
"Bespoke LLM" is probably not the best term, but customised LLMs are certainly possible and available as a service from e.g. Azure. Not from scratch: model customisation starts from the public data base model and is then trained further on your private data.
It's just a waste of time and money. đŸ˜†
Humans are also stochastic parrots. We are just allowed to develop our own unique internal language of thought by constantly fine-tuning on our own ideations, decisions, and their consequences.