Live data from Hacker News

Universal Summarizer

labs.kagi.com

31–40 of 172 posts

Re: Universal Summarizer

#31
post #5

Impressed with their long document summarization, any ideas how they do this? Seems beyond normal GPT limitations; either they have a more powerful model (doubt it) or hacked around the limitations? e.g. good summary for a very long text https://labs.kagi.com/ai/sum?url=http://localroger.com/prime...

Thanks for noticing! (dev here) We have an in-house model we've been developing since 2019, just for summarization of long documents in real time. We'll try to find some time to blog about the high level design.

Looking forward to a technical writeup on what you all have done, looks impressive!

Re: Universal Summarizer

#32
post #22

Earlier quoted context omitted.

We do, and we thought about opening this as a part of other APIs we already have. How would you price this?

Probably by usage similar to OpenAI since I assume your costs are correlated (compute etc). You could do a hobby plan which is free up to a certain no. of requests/tokens per day or hour so developers can start building without any friction (I think this is important and sort of expected as most beloved dev tools do it). You can minimize your costs by offering this on shared resources so inference time is a little sl…

Cool, pricing per tokens processed makes sense, thanks for brainstorming this.

Re: Universal Summarizer

#33

I pay for Kagi search. This doesn't interest me in the least. They have been breaking down the costs of running their business and justifying their prices, which I'm cool with. If it turns out I'm paying to subsidize this kind of thing, it's not going to work for me anymore. (I don't mean to say it's bad or anything, just I don't care about it. It's an interesting downside of a paid business model. Google can do what…

I also pay for Kagi search and have a usage value of ~55% from what I pay and I'm fine with experiments.

Re: Universal Summarizer

#34
This is nothing short of remarkable! The video summary in particular feels like magic. I am exceptionally eager to learn more about their datasets, model architecture, and training process.

Re: Universal Summarizer

#35
post #5

Impressed with their long document summarization, any ideas how they do this? Seems beyond normal GPT limitations; either they have a more powerful model (doubt it) or hacked around the limitations? e.g. good summary for a very long text https://labs.kagi.com/ai/sum?url=http://localroger.com/prime...

Thanks for noticing! (dev here) We have an in-house model we've been developing since 2019, just for summarization of long documents in real time. We'll try to find some time to blog about the high level design.

> an in-house model we've been developing since 2019

We will be very interested in how you tackled the problem of "understanding" the input text - to differentiate it from current "fakeries of actual speakers".

I.e. how you implemented the "intelligent" parts - or, which simulation of actual intelligent processing (if any) it contains.

Re: Universal Summarizer

#36
Well.

> Case is in a bar called the Chatsubo, a place for professional expatriates. He is talking to the bartender, Ratz, who has a prosthetic arm. Ratz and Case joke around and then a prostitute who was sitting next to Case leaves. Suddenly, a drunken Australian starts talking about how the Chinese invented nerve-splicing. This causes Case to become bitter and he expresses his feelings to his glass. This passage paints a vivid picture of the atmosphere in the bar and the characters that inhabit it. The interesting detail is that Ratz has a prosthetic arm, which is a unique detail that adds to the atmosphere of the bar

The results are interesting. The in-production use-case, where the tech will be most reliable, not evident.

Edit:

It seems (also seeing the "Key moments" section) that elemental attempts to understand the details (and structure) of the text should output reliability values, and employ them with importance towards the final output.

Edit2:

It seems from further tests that the "understanding" may be very limited, and that the "trick" is more in terms of "trying to identify salient parts and present them reformulated - without attempting to understand them".

Edit3:

It seems that the attempt of identification of salient points is confirmed as a main mechanism - but, in absence of understanding, the re-formulation just degrades information. For example: the original being, «Bronze is the first metal that gets its own age, which began around 3300 BCE in Mesopotamia. Other metals were certainly in use before it — especially copper — but the addition of a small amount of tin to existing copper technology changed everything. Bronze was a step up in hardness, durability, and resistance to corrosion [...]», the summary «Bronze is the first metal to have its own age, beginning around 3300 BCE in Mesopotamia. It was a step up in hardness, durability, and resistance to corrosion» betrays faults in changing the initial 'gets' to 'have', and in missing that those qualities of bronze are there in comparison to copper.

--

...I see we have a sniper here, as so common: well, do not forget to make your criticism explicit (assuming it will just take its time to elaborate). Edit: still need more time?

Re: Universal Summarizer

#37

I pay for Kagi search. This doesn't interest me in the least. They have been breaking down the costs of running their business and justifying their prices, which I'm cool with. If it turns out I'm paying to subsidize this kind of thing, it's not going to work for me anymore. (I don't mean to say it's bad or anything, just I don't care about it. It's an interesting downside of a paid business model. Google can do what…

Figuring out what a document is about is one of the most central problems in building a search engine. Being able to regurgitate it in a way that makes sense to humans is just a nice benefit.

Re: Universal Summarizer

#38
I fed it The Last Question. While the resulting summary is impressive, I suspect there may be some cheating/plagiarism going on because it includes commentary with no origin in the text.

The Last Question is a science fiction short story by Isaac Asimov about two attendants of Multivac, a giant computer, who make a bet over highballs. The question they ask is whether mankind will ever be able to restore the sun to its full youthfulness even after it had died of old age. The story follows the question through the centuries as mankind develops interstellar travel and builds a better and more intricate computer, the Universal AC. In the end, the Universal AC is unable to answer the question due to insufficient data, but it is able to demonstrate the answer, restoring the Universe from chaos. This story is a fascinating exploration of the power of technology and the limits of human knowledge.

Re: Universal Summarizer

#39

I pay for Kagi search. This doesn't interest me in the least. They have been breaking down the costs of running their business and justifying their prices, which I'm cool with. If it turns out I'm paying to subsidize this kind of thing, it's not going to work for me anymore. (I don't mean to say it's bad or anything, just I don't care about it. It's an interesting downside of a paid business model. Google can do what…

I also pay for Kagi search and have a usage value of ~55% from what I pay and I'm fine with experiments.

Right, maybe I should have framed my comment a bit differently: this is not core functionality I would pay for.

Experiments are cool and it would be dumb to make suggestions about how they do R&D. I was more reacting to the thought that this would become part of the product offering.

Re: Universal Summarizer

#40

I fed it The Last Question. While the resulting summary is impressive, I suspect there may be some cheating/plagiarism going on because it includes commentary with no origin in the text. The Last Question is a science fiction short story by Isaac Asimov about two attendants of Multivac, a giant computer, who make a bet over highballs. The question they ask is whether mankind will ever be able to restore the sun to it…

It's interesting. Feels like an answer that someone who didn't know about, or not understand, the concepts of the story (mostly entropy) would give.
Post reply on HN