ChatGPT-4o finally seems to provide much neater and more comprehensive summarization, similar to that of Claude and Gemini. I always found that this was a key failure of ChatGPT outputs that discouraged me from using it. For example, I often use LLMs to summarize research papers, for which it’s really convenient to have clear formatting and categorization rather than a block of text – it’s difficult to scan for relevant information in the latter case.
Here’s an example 4o output:
Note that 4 does categorize but there’s definitely much less scaffolding.
I’m wondering why OpenAI decided to add this functionality to 4o and not earlier releases considering that 4o is primarily meant for multimodal integration and since it seems pretty easy to fine-tune (e.g. some RLHF and supervised training). Did they do this intentionally to match Claude/Gemini functionality (they do look pretty similar)?