Ask HN: ChatGPT but in Reverse? | Hacker News Reader

Ask HN: ChatGPT but in Reverse?

7 pointssoheil15 comments

I am wondering if there is a way to reverse how ChatGPT works. Instead of giving it a prompt to answer a question or "embellish" a topic. Could we give it a large article and ask it to condense it to a concise and efficiently understood short form?

The more it learns about how you think the better it will get at explaining things to you. What if you could paste in a 5 page PDF and get back a paragraph that makes you understand the entire thing at once. Again this is not just a summary, but a new, fast and most efficient way to understand a topic most suited to your way of learning. And it should include all substantive points. No information should be lost.

The equivalent of "I know kung fu".

View on HN

Knowledge/Memory between different ChatGPT Sessions is not stored. But you can very easily ask "Please summarize this:"

Then, pasting in however many Paragraphs the chat UI allows. Repeat until fully summarized. Other useful questions to try after each sub-summary are:

"Were there any points of interest outside of this summary?" "Could you give me more detail on what the article said about X, Y, Z?"

I once used chatGPT to summarize a 30,000 word article, and after moving chatGPT's summaries to a .txt, it had been reduced to about 3,000 words. The summaries and points of interest also gave me great bits to search for in the full article.

The whole "It learns from your way of learning and optimizes for you personally understanding it" thing is not just ChatGPT in reverse, you just added a whole new major element. Because it definitely can already take long input text and make it shorter, and does it quite well.

I don't understand why it cannot "learn" the way we understand things and tailor it personally. It already can remember context between prompts so I don't see how this is a "new major element".

Because unless you have a way to quantify the manifold ways in which specific human beings "understand things", there's no way to express that to chatGPT in a language symbol form that it can use to help best reexpress a larger piece of information and tailor it specific to you.

E.g. some people learn best through narrative story form, other people through visual imagery, etc. at a minimum yo'd need some kind of a neural biofeedback device to gauge and "AB test" different strategies of expressing information that would optimize encoding from short-term to long-term memory for each individual person.

Basically what you're suggesting is entirely to hand wavy and not nearly concrete enough to be applicable.

OK I'll try again:

Forward: fine tune responses base on new prompts with more details (output from human) - learn from human to create context

human -> machine

Reverse: condense large inputs with details to best fit understanding base on feedback (input into human) - teach human from context

machine -> human

Why don't you ask ChatGPT this question?

Do we already believe its answers are better than what you get on HN?

A 5 page PDF report has X amount of bits of information, whereas a paragraph has ~X/25 (5 paragraphs per page) amount of bits. Its an impossible task to retain the same information.

But it might be interesting to create an Autoencoder for text that creates a compressed embedding that can recreate the original article. That embedding can be used for summary creation and all sorts of downstream tasks

That seems false, bits may be unnecessary for information, or follow a pattern. For example:

apple fox apple fox apple fox apple fox apple fox apple fox apple fox apple fox apple fox

9 times "apple fox"

Same information, less bits.

So the point I am trying to make is to compress the concepts instead of just the bits now that we AI that seems to understand our words and thoughts.

There are limitations on the number of tokens you can use. One token is approximately equivalent to one word in the English language. In the case of the davinci model, there are 4000 tokens shared between the question and answer. It is not possible to send a 5-page PDF in a single prompt due to these limitations.

What you’re asking is essentially fine-tuning GPT3 on your own learning style. Which is to say, fine-tuning it on what you already know versus what you don’t know. And also on the ways of explaining things that work for you versus those that don’t.

It’s a powerful idea!

I believe it can already do this. You give it the text you want to condense and ask chat gpt to rephrase for a fifth grader, and it will.

Can't you tell it to summarize something?

"It knows kung fu."

Reply on news.ycombinator.com