Hacker News Reader

Hacker News

Comment by Aachen | Hacker News Reader

Parent Full thread

Aachen

Is the page count really the most interesting thing to say about it?

View on HN

What else can you expect from a McKinsey alumni

* alumnus

Probably not. It's just that nobody can find the interesting bits, because they're buried in all those pages.

Experiment: What does ChatGPT produce when you ask it for a 300-plus-page report? Is it better or worse than the average human-written report of similar size?

The GPT architecture can't really produce a 300+ page report right now, because that doesn't even come close to fitting within its window. Of course you can just keep running it and concatenating, but I wouldn't really call that a 300+ page report. It'll either wander off topic completely once it loses the prompt, or be very repetitive because it'll have no idea whatsoever what it said before its window.

I gather that with the current architecture it becomes exponentially more difficult to grow the window, but I'm not sure if that's what a computer scientist would call "exponential" or if it was the common colloquial usage that calls x^2 "exponential growth"; would actually be interested if anyone could clarify that. With the current expense though, any of the common complexity classes beyond O(n log n) isn't very feasible at the moment though.

I believe LLMs could generate limitless page reports. Not a single prompt of course. Pick a topic, outline it, grow the outline with subsections. Use a generic summary and LLM indexing to feed context every time you want a subsection, if you want do not stop there use text to image to add drawings, you have 1000 pages thing. Yes, currently it is a thing not an article or research paper. Simple tool and intention relation.

Attention used to be quadratic. But there have been linear attention advances. In fact, gpt-4 is almost certainly using some linear attention variant, maybe flash attention.

Then there's this https://hazyresearch.stanford.edu/blog/2023-03-27-long-learn...

I don't think a 300 page context window is too far away

You can prompt it for an outline, and then prompt it to go section by section. It’s a lot more work than “write 300 pages!”

It’s actually more like working with a junior copy editor that you have to constantly correct, so you keep wondering what you’re even paying them for. But at $20 a month it’s worth it for that.

I don't get this sentiment. Check the summary if you don't want to read everything before submitting? Has OP even opened the thing to check the first page of contents? If it's good stuff, one would keep reading and be sad for it to end, no matter if it's 5 or 5000 pages (I was sad for book series to end that were longer than that).

OP shouldn't have changed the report's title, or if this is the original title by some miracle, they should have changed it or posted a comment about why it's interesting. Presumably they did read at least part of it and have a reason for submitting it. As it is, the submission is entirely useless beyond pointing out that a certain URL exists.

Reply on news.ycombinator.com