[HN Gopher] Summarizing Books with Human Feedback
       ___________________________________________________________________
        
       Summarizing Books with Human Feedback
        
       Author : atg_abhishek
       Score  : 74 points
       Date   : 2021-09-24 14:58 UTC (8 hours ago)
        
 (HTM) web link (openai.com)
 (TXT) w3m dump (openai.com)
        
       | Causality1 wrote:
       | It would be great if this works well. I've noticed it's something
       | that doesn't really exist. If a book is extremely popular it
       | might have a good plot summary on Wikipedia but if not all you're
       | going to get is one or two teaser paragraphs.
       | 
       | I often find myself wanting to continue a series that I didn't
       | finish or just had a book come out without rereading all the
       | previous ones. Having a big compendium of five-page plot
       | summaries would be incredible.
        
       | djoldman wrote:
       | Paper:
       | 
       | https://arxiv.org/pdf/2109.10862.pdf
        
       | abc_lisper wrote:
       | This can't come fast enough. Could be used to filter out
       | pervasive false hoods spewed on social media
        
         | falcor84 wrote:
         | How? What's to stop the AI (and/or humans in the loop) from
         | spreading falsehoods?
        
           | abc_lisper wrote:
           | Other AIs. Because we can summarize well, we would be able to
           | filter out extraneous data and figure out the actual intent
           | better.
        
         | TOMDM wrote:
         | People aren't spewing falsehoods for a lack of people pointing
         | out the truth, they do so in spite of the truth presented to
         | them.
        
         | rcMgD2BwE72F wrote:
         | People desire alternative stories because they distrust those
         | who are telling the main story, not because they lack good text
         | or tools.
         | 
         | You'll be providing with the right AI-written story, but
         | they'll still go read something else -- unless you can convince
         | them that you're on their side (not with stories but actual-
         | life-changing stuff). But we're not working on that and prefer
         | toying with tools and algos.
        
       | dloss wrote:
       | Good. Now summarize Wikipedia.
        
       | djoldman wrote:
       | Example ("The Great Gatsby"):
       | 
       | The narrator, a Yale graduate, moves to New York to learn the
       | bond business. He meets Gatsby, a man from a wealthy Midwest
       | family who attended Oxford. Gatsby invites the narrator to lunch
       | and shows him his war medals.
       | 
       | Gatsby wants Daisy to tell Tom she never loved him and wants to
       | go back to Louisville to get married as they had planned 5 years
       | ago.
       | 
       | Mrs. Wilson is killed by a car that doesn't stop. Gatsby says he
       | tried to stop Daisy from driving, but she couldn't, so he pulled
       | on the emergency brake.
       | 
       | After two years, Nick remembers the police and photographers at
       | Gatsby's house after his death.
       | 
       | On the last night, Nick looks at Gatsby's house one last time.
        
         | timgilbert wrote:
         | That's... kind of a rubbish summary, but I'm not sure what I'd
         | want a ~700-word summary of The Great Gatsby to look like.
        
           | LeonardoTolstoy wrote:
           | Honestly, I think you could do worse than the Wikipedia
           | summary. That is maybe a bit too long (I can't remember the
           | Wikipedia guidelines for books) it looks longer than 700
           | words. But from there I think you could fashion a solid 700
           | word summary of the book without too much trouble.
        
         | tschwimmer wrote:
         | This is truly an awful summary. It is factually correct but
         | completely misses any nuance or description of themes that make
         | The Great Gatsby a notable work.
         | 
         | I doubt this has any value for what is considered to be
         | literature, and probably isn't very useful for most of the rest
         | of fiction either. I'd be curious to see this be used on dense
         | non-fiction to extract the salient points (I'm imagining this
         | being run on Piketty's Capital in the 21st Century), but based
         | on this sample I'm not optimistic.
        
       | herbertl wrote:
       | This is particularly relevant to yesterday's discussion on books
       | that are too long: https://news.ycombinator.com/item?id=28636074
        
       | mercurialsolo wrote:
       | A bunch of similar summaries around movies which have been doing
       | the round at YouTube.
       | 
       | I wonder what the purpose of such summaries is - is it to define
       | the plot section ala Wikipedia movie plots or IMDB plots or is it
       | for really condensing knowledge into abstracts.
       | 
       | The latter does require a certain degree of both cognition around
       | the domain as well creative writing to make it engaging.
        
       | beckman466 wrote:
       | Why don't we instead give all humans full universal access to all
       | knowledge, science and technology, instead of feeding books into
       | black box 'AI' for 'summaries'?
       | 
       | Why should an 'AI' have universal access to this material, and
       | not the working class? Especially the working class in the global
       | south. They are so far from universal access. It's only a matter
       | of time before the capitalist academia comes down on emancipatory
       | initiatives/projects (yet capital disrupting), like Sci-Hub.
       | 
       | I write 'AI' because i see the pursuit of a general artificial
       | intelligence in the labs of the propertied class (on their own
       | profit-seeking, commons-hostile terms), despite the fact that
       | there is widespread suffering caused by today's global production
       | system, as a disgusting joke.
       | 
       | The supposed inevitability of today's widespread suffering across
       | the world is a false truth pulled over our eyes by the propertied
       | class.
       | 
       | World peace and full equality in this new digital age is a matter
       | of priorities.
        
       | DiggyJohnson wrote:
       | Leaving aside weapon and security purposes, full-length book
       | summary strikes me as especially dystopian.
       | 
       | In truth, I think this is something AI is capable of doing, but
       | are we really comfortable in passing information through an
       | interface where we cannot even know what is lost.
       | 
       | This topic begs so many questions that I guess are not unique to
       | this problem. Who is all of this intelligence for, though?
       | 
       | What if we forget how to write well?
        
         | DennisP wrote:
         | It's not like summaries have to entirely replace whole books.
         | 
         | Personally I have a library on my kindle that's too much for me
         | to read in full, considering I'm constantly adding to it. I
         | read a lot of books but good summaries strike me as an
         | improvement for the cases where I either never get to a book,
         | or just read its first several chapters.
         | 
         | For the books I do fully read, summaries could make for nice
         | reviews. With some further development, it seems this
         | technology could turn the summaries into quiz questions, which
         | would really help lock in long-term learning.
        
           | DiggyJohnson wrote:
           | Oh I'm not dissing this particular endeavour or anything-I
           | more meant to point out that I was surprised by my own
           | initial reaction was so negative. There are certainly uses
           | for summaries, and by no means does a summary replace the
           | existence of the original text.
        
             | woko wrote:
             | Keep in mind that the real goal here was to "test scalable
             | alignment techniques", and OpenAI ended up adopting
             | "recursive task decomposition".
             | 
             | That is basically a "divide and conquer" approach for
             | simultaneously learning a complex task *and* allowing
             | humans to evaluate it.
             | 
             | > To test scalable alignment techniques, we trained a model
             | to summarize entire books. [...] This work is part of our
             | ongoing research into aligning advanced AI systems, which
             | is key to our mission. As we train our models to do
             | increasingly complex tasks, making informed evaluations of
             | the models' outputs will become increasingly difficult for
             | humans. [...] Our progress on book summarization is the
             | first large-scale empirical work on scaling alignment
             | techniques.
        
       ___________________________________________________________________
       (page generated 2021-09-24 23:01 UTC)