Using AI for Review Thread Summaries?

[amirm]

Hello everyone and especially those of you with experience deploying AI in specific cases.

I have noticed Amazon including an AI generate summary of all the user reviews for products. I find them quite useful and accurate. I am thinking such a thing might be useful for the reviews as well. What I don't know is how one goes about using these AI tools to generate thread/site specific results. Do you have to become a commercial user and pay high licensing fees to do so?

 

[Timcognito]

Amir I like the Amazon AI summaries as well and can't speak to how best implement them on ASR. But Amazon's AI summaries are compiled from actual purchases and purchasers, I can't imagine what AI would do with all the blathering from non-purchasing know-it-alls that post on ASR. Just my to 2¢.

 

[Jeromeof]

 

[kypen]

Given that Microsoft has been adding AI features to parts of the OS with some frequency, Google has its own AI systems that it's adding to Chrome, and Arc browser is already doing some of this, I'd recommend just a wait and see approach.

If you start adding summaries, then errata is needed for the post, the summary may be out of date and require a new analysis that won't be consistent in language with the previous. It's difficult to determine. If this is delegated to browsers for when users need/want the summary on demand, it reduces diligence and opportunity for blame for Amir.

 

[BinkieHuckerback]

This fellow Al is getting a lot of bad press

 

[Jeromeof]

But to automate generating this content would need a OpenAI API token. Depending on how much content you would be summarising (how many threads and how often) this might get expensive as you would run out of free tokens. I imagine for a forum like this some php code would need to be written or some plugin added to automate the thread summaries

 

[Kervel]

View attachment 349476

Now do one of thread of a speaker with 100+ comments pages!

 

[-Matt-]

Please correct me if I'm wrong. I don't think any of the current chatbots are capable of interpreting the image data shown in the plots (text input only). It will be a major step forward when they can.

 

[pseudoid]

That brings up a weird question, in my mind: Who really owns the rights to the 'posts' @asr?
And possibly, if those rights allow the 'owner' to receive residuals from their use by AI?

 

[Steve Dallas]

ACME Tornado Shelter Reviews - 4.5 stars
5 stars - the color matches my house perfectly!
5 stars - my daughter loves to play in it
5 stars - it looks great behind my pool
5 stars - it serves as great extra storage for firewood
5 stars - it adds to the steampunk post-apocalyptic aesthetic of my backyard
5 stars - it doubles as a clothes dryer when I drape a blanket over it
1 star - it crumpled like an aluminum can and killed my family by squishing them til they turned to goo

AI Review
Reviewers love the cosmetics, color, and general usefulness of the item. They find it to serve other practical purposes, including pressing and holding liquids.

 

[Jeromeof]

Also depending on the structure of the summary required some simple 'prompt' engineering might be required. E.g. from a recent thread:

 

[Timcognito]

1 star - it crumpled like an aluminum can and killed my family by squishing them til they turned to goo

 

[threni]

That brings up a weird question, in my mind: Who really owns the rights to the 'posts' @asr?
And possibly, if those rights allow the 'owner' to receive residuals from their use by AI?

You agreed to this when you joined:
-----
You are granting us with a non-exclusive, permanent, irrevocable, unlimited license to use, publish, or re-publish your Content in connection with the Service. You retain copyright over the Content.

If you do not agree with these terms, please do not register or use the Service.
-----

 

[Jeromeof]

This does bring up an extended idea I had for the content on this forum. Which was to train (or fine-tune) a smaller open source LLM with just the content on this forum. In fact I would almost be tempted to train it on just @amirm content and thread replies (but also include a few of the more knowledgable members content who would 'opt' in to have their responses be used as training data). This 'custom' model then would then be used create a 'AudioScienceChatbot' which would possible be able to answer peoples questions in as knowledgeable way as possible.

 

[Timcognito]

So if Stereophile and Soundstage publish reviews on their websites for the same product and one uses AI to summarize those as a summaries and critiques, can one go another forum and legally comment with those critiques stating they were formed by AI?

 

[pseudoid]

If you do not agree with these terms, please do not register or use the Service.

:silly: I was talking about @amirm, when he asked about costs of using such services to trivialize/monetize versions of his hard work.
TL&DR versions, as such, devalues them. imho

 

[-Matt-]

Without paying for anything, you can ask questions in this sort of way:

CoPilot Prompt:

Based on all posts created by the user "amirm" on the Audio Science Review forum (https://www.audiosciencereview.com/forum/index.php) please answer questions in the style of amirm. Here is a first question for you to answer:
Compare and contrast the benefits of vinyl vs digital music reproduction.

CoPilot Response:

Certainly! Let's delve into the fascinating world of music reproduction, comparing the timeless allure of vinyl with the precision of digital formats. As an Electrical Engineer and audio enthusiast, I'll endeavor to provide a balanced perspective.

Vinyl Records: The Warm Embrace

1. Analog Warmth and Character
- Vinyl records offer a unique warmth and character that digital formats often struggle to replicate. The inherent imperfections—subtle crackles, pops, and surface noise—become part of the listening experience. Audiophiles cherish this analog charm.

2. Dynamic Range and Compression
- Vinyl boasts an excellent dynamic range, allowing faithful reproduction of both quiet and loud passages. Unlike digital formats, which can clip when audio levels exceed their capacity, vinyl exhibits natural compression, resulting in a smoother listening experience(1,2).

3. Tangibility and Ritual
- The act of handling vinyl—carefully placing the needle, flipping the record, and admiring the artwork—creates a tangible connection to the music. It's a ritual that transcends mere playback.

4. Album Art and Collectibility
- Vinyl covers are iconic. The large canvas allows for intricate artwork, liner notes, and lyrics. Collecting vinyl becomes a treasure hunt, with rare pressings and limited editions adding to the allure.


Digital Music: Precision and Convenience

1. Pristine Reproduction
- Digital formats aim for faithful reproduction without degradation. High-resolution files capture wide dynamic ranges, offering exceptionally clear representations of the music. No surface noise, no wear and tear.

2. Convenience and Portability
- Digital music fits in your pocket. Streaming services grant access to vast libraries instantly. No need to flip records or worry about scratches. It's music on demand, anytime, anywhere.

3. Consistency and Reproducibility
- Digital playback ensures consistent sound quality across devices. The same file sounds identical whether played on a smartphone, laptop, or high-end DAC.

4. Metadata and Organization
- Digital music brings metadata—artist names, album titles, track numbers—to the forefront. Organizing playlists and searching for specific songs is a breeze.


The Verdict? It's Personal.

Ultimately, the choice between vinyl and digital hinges on personal preference. Some seek the tactile experience and nostalgia of vinyl, while others prioritize convenience and pristine sound. As with any art form, it's about what resonates with you. So, whether you're spinning vinyl or streaming FLAC files, let the music move you! (1,3,4,5)

Source: Conversation with Bing, 13/02/2024
(1) Vinyl vs. Digital: Which Is Better? | Headphonesty. https://www.headphonesty.com/2021/04/vinyl-vs-digital/.
(2) Vinyl vs. Digital: Which Is Better? - All For Turntables. https://allforturntables.com/2023/09/23/vinyl-vs-digital-which-is-better/.
(3) Vinyl vs. Digital: Debunking the Sound Quality Myth. https://vinylitic.com/faqs/vinyl-vs-digital-debunking-the-sound-quality-myth/.
(4) Vinyl vs Digital Music: A Comprehensive Comparison. https://www.retrotechlab.com/vinyl-vs-digital-music-a-comprehensive-comparison/.
(5) Does vinyl sound better than streaming? - SoundGuys. https://www.soundguys.com/vinyl-better-than-streaming-20654/.
(6) amirm | Audio Science Review (ASR) Forum. https://www.audiosciencereview.com/forum/index.php?members/amirm.2/about.
(7) WiiM Pro Plus reviewed by amirm on AudioScienceReview | WiiM. https://forum.wiimhome.com/threads/wiim-pro-plus-reviewed-by-amirm-on-audiosciencereview.2018/.
(8) A bit about your host.... | Audio Science Review (ASR) Forum. https://www.audiosciencereview.com/forum/index.php?threads/a-bit-about-your-host.1906/.
(9) amirm | Audio Science Review (ASR) Forum. https://www.audiosciencereview.com/forum/index.php?members/amirm.2/.
(10) Amir's TODO thread | Audio Science Review (ASR) Forum. https://www.audiosciencereview.com/forum/index.php?threads/amirs-todo-thread.2074/.
(11) undefined. http://www.madronadigital.com.

I doubt @amirm would agree with the comments on vinyl dynamic range. Don't really know if it has truly captured Amir's writing style. What do you think?



After that, it got even more weird:
CoPilot Prompt:

Still answering as user "amirm":
What is your favourite meal?

CoPilot Response:

Ah, the delightful pursuit of gastronomic pleasure! As an aficionado of both code and creativity, my favorite "meal" transcends the mundane boundaries of mere sustenance. Allow me to concoct a poetic feast for your imagination:

def favorite_meal():
    ingredients = ["Syntax sugar", "Recursive functions", "Lambda expressions"]
    flavors = ["Salty bits of elegance", "Sweet loops of logic", "Umami callbacks"]
    garnish = "A sprinkle of comments and a dash of whitespace"

    # Simmering the ingredients
    meal = [f"{ingredient} infused with {flavor}" for ingredient, flavor in zip(ingredients, flavors)]
    meal.append(garnish)

    # Serving the masterpiece
    return " ".join(meal)

# Bon appétit!
print(favorite_meal())

This delectable dish combines the precision of Python with the whimsy of imagination. Savor it slowly, for it nourishes not only the mind but also the soul. ️

 

[Mikig]

That brings up a weird question, in my mind: Who really owns the rights to the 'posts' @asr?
And possibly, if those rights allow the 'owner' to receive residuals from their use by AI?

In essence, if I can decently summarize the open issues under discussion on the legal side, they should be

1) one of the issues is precisely the use by the AI itself of published content used to "learn". Most of these contents are protected by copyright.

2) generative AI often produces content similar to existing content, "plagiarism"

3) the difference between content produced by AI alone, and those produced with a large part of human input, where AI becomes a tool.

It is a path that is taking shape in the legislative branch, because in addition to the issue of "rights" there are problems of privacy, use of sensitive data etc...

 

[kemmler3D]

That brings up a weird question, in my mind: Who really owns the rights to the 'posts' @asr?
And possibly, if those rights allow the 'owner' to receive residuals from their use by AI?

It's in the TOS:

You are granting us with a non-exclusive, permanent, irrevocable, unlimited license to use, publish, or re-publish your Content in connection with the Service. You retain copyright over the Content.

Answer - ASR owns the rights, but you also retain your rights.

AFAICT ASR would not be required to give you royalties for any further use of your posts.

e: Saw that someone else answered already, my bad.

Anyway, I think it's an alright idea to summarize long threads using AI.

However, I would still urge caution. I have been using an AI tool to summarize meetings at work. It works pretty well, but I still review all the output manually. Out of ~5 meetings, in one of them it reported the opposite of what we actually agreed on in the meeting. The rest of the time it's spot-on, but that's how it gets you - a false sense of confidence.

The problem with LLM-based summaries is they don't have any way to self-correct on factual information. And when they report non-factual information, there's no obvious sign that it's wrong unless you happen to know it's wrong. It cheerfully and confidently outputs fact and fiction alike.

 

[Multicore]

I doubt @amirm would agree with the comments on vinyl dynamic range. Don't really know if it has truly captured Amir's writing style. What do you think?

It seems a lot like ChatGPT, i.e. a compute-intensive BS generator.