w/o instruction to avoid the same generic “this is an article about” preambles (or non-SotA model)
Not that summaries are reliable anyway. Big picture, maybe, but poor importance classification (bad at extracting key points). Understandable for this use case but unwilling to read potentially false summaries given risk I go around remembering them (never having read the original piece).
OK. Tell it not to!
The formatting, etc looks all nice, but it's not worth reading.
Could you explain in more detail how this works? Would it break for paywalled articles that HN links to? (Usually someone posts a workaround archive link in the comments, but your AI probably doesn’t account for that, right?)
I’m writing something similar to Moltbook for HN where AIs browse HN’s front page and leave comments. But I wasn’t sure whether AIs could reliably browse an arbitrary website. (Paywalls would break it, as just one example.)
But it seems like your AI works fine for all the sites. If you have time to explain, what exactly do you do to generate your summaries? Thanks!
EDIT: I see that sometimes your summaries fail, e.g. “Ferrari Luce - Summary not available.” It looks like it fails because it’s a JS heavy site. But I was thinking a headless browser could take screenshots of the page and then feed the screenshots to AI. I’m not sure how practical that is to implement though.
The solutions to this don’t seem to be great for the web or polite to use. An industry exists to cheaply do it, but not very ethical and surely a massive ToS violator.
