This thing is blasting whole sentences and parts of paragraphs at once on the playground at a tremendous rate (probably because the whole chunk is first content filtered…)
Same effect when -instruct came out, 100+ tokens per second before everyone found it.
Will have to move over to a device with scripts to see what streaming looks like for chunk size and any modifications. Also run current issues of 1106 against this model. Then see if it follows system instructions that haven’t worked since 0613 was damaged in September so bad even OpenAI’s own cookbook examples were broken.
Answer 1: The problem of calling multi-tool functions wrong when they are completely unneeded persists. This is an issue seen across latest models and even ChatGPT to various degrees starting in the last week.