Show HN: YouTube Summaries Using GPT (eightify.app)
I got the idea to summarize videos when my friend sent me a lengthy video again. This happens to me often; the video title is so enticing, and then it turns out to be nothing. I had been working with GPT for 6 months by the time, so everything looked like a nail to me.
It's a Chrome extension, and I'm offering 5 free tries for videos under an hour. After that, you have to buy a package. I'm not making money yet, but it pays for GPT, which can be pricey for long texts. And some of Lex Fridman's podcasts are incredibly long.
I'm one of those overly optimistic people when it comes to GPT. So many people tell me, "Oh, it doesn't solve this problem yet; let's wait for GPT-4". The real issue is that their prompts are usually inadequate, and it takes you anywhere from two days to two weeks to make it work. Testing and debugging, preferably with automated tests. I believe you can solve many problems with GPT-3 already.
I would love to answer any questions you have about the product and GPT in general. I've invested at least 500 hours into prompt engineering. And I enjoy watching other people's prompts too!
122 comments
[ 3.2 ms ] story [ 215 ms ] threadAlso, does it always give me 8 bullet points from each video? It would make sense for videos that have chapters to give summaries to each chapter instead.
Currently the prices are: $8.6 for 20 summaries ($0.43 each) $19.8 for 60 summaries ($0.33 each) $48.6 for 180 summaries ($0.27 each)
> Also, does it always give me 8 bullet points from each video? It would make sense for videos that have chapters to give summaries to each chapter instead.
Yes, it's now a fixed number of 8 parts. I tested different numbers, it seemed to me a universally convenient simple solution for videos of any length. But yes, I will take chapters into account when splitting, right now I ignore that information.
At the same time, the product isn't as useful as if when there are no chapters at all.
Taking the headphone review on the landing page as an example, the generated summary is "The Sony XM5s offer improved audio and call quality, but may not be worth the extra cost compared to the XM4s."
Like, duh? You probably could deduce that even without watching the video, 14 mins of your life saved.
For me it's 30% for information, 70% for fun.
If the summary spits out a bunch of useless info, you can find a better one.
When people tell me they don't watch TV, I ask them if the watch YouTube or Twitch, because these kind of services are what TV was before: something with low density of information, better used as distraction while you do something else.
Nevertheless, I'm uncomfortable with playing a video in the background, because that might give the platforms the impression that wasting my time is fine if it improves metrics.
These days, our attention is constantly being pulled in several directions at once, so I praise projects like this one, who try to wrestle control back.
If you can get the headline click with 30 seconds worth of insight, then your payday is related to padding to get to the appropriate length. Imho
i.e. "You won't believe which beloved celebrity just died" and it just tells you the vital info.
Move to alternative webs such as Gopher or Gemini, which make it impractical / impossible to monetize based on clicks / engagement / amount of content. The issue isn't that you can generate spam. The issue is that you can make money by doing so.
While making money is a big game, it's not the only game. Your thoughts and actions leads to behaviors off your computer. This is why we see all kinds of political spam and propaganda to affect your choice at the ballot box.
If a system is machine readable and writable (which AI is fastly increasing this definition) you must assume that not only can you be attacked, you must assume that you are being attacked.
> I'll take "anti-intellectual artistic project" over "cesspool of lowest common denominator" any day.
If everyone else thought like you, humanity would be set back millennia.
Not to mention that it's just flat-out insane to discard the massive amounts of educational content and knowledge present on the internet just because it happens to also contain some undesirable content.
This mindset is not one that we want to spread throughout our society, full-stop. Gemini is a threat to human progress.
Maybe we should get an AI to write it.
While I have plenty of cynicism about this and also expect it to at least partially play out like this, let me offer a more optimistic perspective on the same thing.
People come into media with different amounts of background knowledge and context. Currently, this is basically solved by a tiered system of 'knowledge' distribution. As an example (though I think something similar exists outside of the sciences too), scientists write papers that are read by science communicators who put out press releases which are read by journalists who write articles about stuff all of which is read by various content creators who remix all of this into their own content tailored to their specific following. Part of this is tailoring is knowing what context/knowledge your audience already has and giving them enough new information for the new content to be digestible without the consumer needing to seek out other sources. So when ChatGPT-N is reencoding the content for you, it can personalize it your level of knowledge, without wasting your time by either rehashing stuff you're already aware of or by including context that you wouldn't necessarily have known that you're missing.
This of course means that ChatGPT will need to know what you do/don't know...
Which government is the question you should be asking.
For example, lets say I'm a authoritarian government that is hostile to other nations. In general there is a strong imperative to block GPT use in my own country except by approved users to make propaganda. At the same time it is a very useful weapon for me to use against my enemies. I can use the 'bullshit asymmetry theory' against them and drown their population in conflicting propaganda.
But you see unlike nuclear weapons, we really don't need multibillion dollar facilities to make these things. Midsized companies make these things easy enough. There is no opting out, unless of course you want a world wide police state to ensure that you're not making anything naughty with your computer. And of course you should realize your government, along with everyone else's government is making their own naughty versions of this kind of application anyway.
So, yea, some smaller EU countries won't make one because of the law, but the US, China, and likely Russia would do it anyway because screw you, we have nukes behavior that drives their other actions.
This is an issue of capital allocation. Enormous amounts of private money have been wasted chasing dreams of monopolizing currency (crypto), monopolizing taxis (Uber/Lyft), and of course monopolizing the food delivery industry. There is little difference between government picking winners and private industry doing it. How many cumulative billions has Uber lost so far? Likely far more than what the Chilean government lost in a futile attempt to develop a domestic model of automobile in the early '70s.
Now even more of that monopoly money is going to be shoveled into consumer AI plays that will continue to waste energy, not to mention pollute the internet further with half-baked 'content' with the stench of Wikipedia and ArtStation all over it.
Brilliant observation! It's sort of like if you took the most extreme lossy data expansion algorithm -- and then fed the output of that through the most extreme lossy data compression algorithm...
User: "Computer, turn binary 1 -- into everything in the Universe..."
Computer: "OK, here it is..." (spits out a result which is [exa|zetta|yotta|ronna|quetta|???]bytes long...)
User: "Computer, now turn everything in the Universe back into 1..."
Computer: "Processing... this may take some time... please wait..." (puts up progress bar that increments so slowly that it appears not to move...)
<g>
Related quote by Douglas Adams:
"There is a theory which states that if ever anyone discovers exactly what the Universe is for and why it is here, it will instantly disappear and be replaced by something even more bizarre and inexplicable. There is another theory which states that this has already happened."
- Douglas Adams, The Restaurant at the End of the Universe
Anyway, I could definitely see content creators padding with AI -- and consumers summarizing the content with it...
Isn't that pretty much why short videos is getting popular. Who has time for 15minute recipe videos scripted for youtube ads when you can find 1m cliff note version on tiktok or youtube shorts. Back to the days of ehow but with more personality and filter.
One idea is to use other models to shorten text by throwing out meaningless words. I estimate this will reduce the length of the text (and thus the GPT cost) by 30%.
Does this mean GPT does not need coherent sentences to understand?
Anything in particular thing stopping you from shipping it online? Or just that it's not necessary for your use?
(Context: I've been working on `Heroku for LLM apps` and trying to understand where the value/frictions are)
The author uses ChatGPT and that limits it a bit. Because it took me 80% of my time and effort to code the preprocessing of the transcript before sending it to GPT. (I mean if you just put transcript to GPT and ask to summarize it sucks)
But his product is free, that's cool. Because he uses ChatGPT on the client. And I have to pay for GPT.
Please could you give an overview of how this actually works? Have some ideas of where the tech could be useful but not sure how I'd actually go about implementing it. Do you have a GPT model on a server and code to transcribe the video then summarise the transcription. Or do you use one of the APIs from OpenAI?
If you use their APIs:
* How costly has it been to run your service? (If you don't mind answering)
* Is it customisable? If you wanted to run a chat bot for example, would you be able to make it understand the request (I'd assume something similar to an 'intent' when developing Alexa skills) and give it data so it knows the answer?
> Please could you give an overview of how this actually works?
1. I download Youtube subtitles (it doesn't work for videos without youtube subtitles yet. my analytics shows that 15% of videos don't have subtitles. I tried to use OpenAI Whisper, but it takes several minutes to transcribe a video, so I put that task off for now)
2. Then I break the transcript into parts.
3. Then I summarize each part with GPT → and then I summarize the summaries to get chapter names → and then I summarize again to get the title.
Yes, I use OpenAI GPT API. I pay them their standard pricing for davinci-003 and the cost for 1 video is between $0.1 and $0.9 depending on the video length (actually, the transcript length). I have a hard limit to prevent abuse.
Yep, it's fully customizable. Yes, you can provide data to it. It would take 1 hour of coding to make a prototype of a chat bot. And then 500 hours to make it work well.
https://news.ycombinator.com/item?id=34549402
You mention that you have invested 500 hours into prompt engineering. Are there any specific resources you would suggest to get maximum value out of GPT? Any videos, websites, podcasts, ebooks, books, anything that really stands out?
I have been playing with it for a while now and am getting good at having it spit out what I'm looking for but usually it takes an extra 3-4 prompts to rearrange the responses that I want.
Thanks and again, great execution on a cool idea!
Oh yes, I remembered that I saw this - it's good advice: https://help.openai.com/en/articles/6654000-best-practices-f...
i've also been keeping a popular series of notes https://github.com/sw-yx/ai-notes/blob/main/TEXT_PROMPTS.md
Video has less than 30K views Free plan is limited to videos less than 1 hour long and with more than 30k views
This is a very lame limitation in regards to views. Honestly trying to see if my videos were delivering enough value to justify the title.
It's a hypothesis and on my current traffic I can't say it's worked. Maybe I'll change the limits when I have more data.
Just use lessons, or “what did you learn?”
This isn’t directed at you personally.
This is what comes to mind immediately: 1. Don't solve more than 1 problem with 1 Prompt. Decompose it into different tasks and make a separate prompt for each one. 2. Use instructions at the beginning. Very short and unambiguous. You have to understand exactly what you want and mean. "Answer me as a philosopher" is an example of an unspecific instruction. 3. If the instructions don't work, show the concept providing an example. Examples are more expensive than instructions because they take up more tokens. 4. The best way to debug propts is when you have a dataset and autotests. I used GPT to evaluate the results. 5. Temperature 0 is fine 99% of the time. (btw I was surprised to find that it does not guarantee a deterministic result; the OpenAI support confirmed that)
There is no reason to request more than email.
My young son uses YouTube for tutorials to learn programming and 3D apps. But I really struggle because he's come across objectionable content as well, and the tools YouTube provides for moderation or filtering are completely worthless. They don't care. I'm only left to think they want our kids to see controversial and even radicalizing content because it increases engagement metrics.
AI that can prescreen videos?! Regaining some feeling of control and confidence about the content that comes into my house?! I AM SO IN!
I have no interest in censoring anyone else or limiting access for others. I just want to have some agency over what my kids are exposed to without removing the actual knowledge share advantages that the internet can and does provide.
There are probably loads of other parents who would love this.
Also, did you get 5000 downloads in a single day?
Or when did you launch it?
Thank you!
It's a more interesting task when there are two sides. Even now, my app doesn't work well for debates. It tries to bring the points of view together. I want to work on separation.
Frankly, I haven't read the papers about summarization. But I will have to when I'll work on reducing costs.