[gcr]

There’s no details here about training, evaluation, reproducibility or even architecture.

OpenAI’s GPT4 tech report was a marketing press release in disguise; I felt it represented a huge step down for the field in terms of academic rigour. But this model card epitomizes the end conclusion of LLM commoditization: no attempt to quantify limitations or weaknesses, portraying model biases and hateful/hallucinatory output as a feature rather than a bug, all wrapped up in breathless language of wanting to “maximally benefit humanity” while hiding the model behind a secretive waitlist or a $16 Twitter premium subscription plan.

What a boring release, to be honest.