Tokenization is a crude text compression, which happens at the cost of obfuscating the data. Without it, we would effectively reduce the context size and incre

Tokenization is a crude text compression, which happens at the cost of obfuscating the data. Without it, we would effectively reduce the context size and increase the training time. For OpenAI tokenizer - by a factor of 4. Is there a workaround, or are there only trade-offs?