So, Kovid, when are we getting CalibreAI?

LordP · 12-20-2023, 01:59 PM

Just kidding, but think of the possibilities.

Grammar, spelling, proper HTML, customization, translation....

It boggles the mind.

Quoth · 12-20-2023, 04:19 PM

It's a lie.

Don't look at the 10,000,000 scraped websites and the team behind the curtain and the environmental cost of the massive data centre. Only NFTs and "crypto-currencies" and blockchain (a solution still looking for a problem) are worse.

There is no AI. Just huge databases and matching engines.

ownedbycats · 12-21-2023, 02:16 PM

Oh god no

DNSB · 12-21-2023, 05:01 PM

Given the issues with inaccuracy in the answers from AI, you really want calibre to make use of it? A glutton for punishment you must be.

kovidgoyal · 12-21-2023, 07:31 PM

A basic precondition for AI in calibre would be the ability to run AI models on common consumer hardware. I am never going to be uploading user data to some random SV company's servers. Nor am I going to be maintaining my own server cluster for AI compute.

Comfy.n · 12-21-2023, 08:06 PM

Quote:

Originally Posted by kovidgoyal

A basic precondition for AI in calibre would be the ability to run AI models on common consumer hardware.

which is actually becoming a possibility, a few months ago I've been able to play a bit with gpt4all on my windows laptop with quite standard specs:

https://gpt4all.io/index.html

kovidgoyal · 12-21-2023, 08:20 PM

Those models are still a long way from the performance/accuracy of actual GPT-4 which is itself a long way from being actually either performant or accurate enough, IMO.

Comfy.n · 12-21-2023, 08:46 PM

I can only agree. Performance-wise there's a severe limitation on common hardware, measured as "tokens per second"... I remember some local models were painfully slow to process simple requests. And then there's that other limitation you mentioned on a similar thread "they can only be prompted with very limited amounts of text at the moment."

Quoth · 12-22-2023, 05:56 AM

Also too easily produces plausible junk due to scraping web sites and almost no human curation.

It's a technological dead end and not the way to improve proofing/correction of grammar or spelling.

Translation relies on matching a big database, a giant Rosetta Stone. Google started by scraping all the EU published documents as they are all the the same thing (by expert humans) in all the official languages of the EU. Which still includes English, though the UK left, because that's official in Malta and Ireland.

12-20-2023, 01:59 PM	#1
LordP Zealot Posts: 139 Karma: 1341124 Join Date: Aug 2010 Device: Kindle 3	So, Kovid, when are we getting CalibreAI? Just kidding, but think of the possibilities. Grammar, spelling, proper HTML, customization, translation.... It boggles the mind.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Thank you Kovid	DyckBook	Calibre	1	11-22-2022 10:02 PM
Kovid: just sent you $50	Macthelist	Calibre	1	04-18-2014 09:32 PM
Thanks to KOVID	LARdT	Calibre	1	12-19-2009 03:46 PM

12-20-2023, 04:19 PM	#2
Quoth the rook, bossing Never. Posts: 12,149 Karma: 89201339 Join Date: Jun 2017 Location: Ireland Device: All 4 Kinds: epub eink, Kindle, android eink, NxtPaper11	It's a lie. Don't look at the 10,000,000 scraped websites and the team behind the curtain and the environmental cost of the massive data centre. Only NFTs and "crypto-currencies" and blockchain (a solution still looking for a problem) are worse. There is no AI. Just huge databases and matching engines.

12-21-2023, 02:16 PM	#3
ownedbycats Custom User Title Posts: 9,100 Karma: 62844063 Join Date: Oct 2018 Location: Canada Device: Kobo Libra H2O, formerly Aura HD	Oh god no

12-21-2023, 05:01 PM	#4
DNSB Bibliophagist Posts: 38,390 Karma: 152905840 Join Date: Jul 2010 Location: Vancouver Device: Kobo Sage, Libra Colour, Lenovo M8 FHD, Paperwhite 4, Tolino epos	Given the issues with inaccuracy in the answers from AI, you really want calibre to make use of it? A glutton for punishment you must be.

12-21-2023, 07:31 PM	#5
kovidgoyal creator of calibre Posts: 44,213 Karma: 23446406 Join Date: Oct 2006 Location: Mumbai, India Device: Various	A basic precondition for AI in calibre would be the ability to run AI models on common consumer hardware. I am never going to be uploading user data to some random SV company's servers. Nor am I going to be maintaining my own server cluster for AI compute.

12-21-2023, 08:20 PM	#7
kovidgoyal creator of calibre Posts: 44,213 Karma: 23446406 Join Date: Oct 2006 Location: Mumbai, India Device: Various	Those models are still a long way from the performance/accuracy of actual GPT-4 which is itself a long way from being actually either performant or accurate enough, IMO.

12-21-2023, 08:46 PM	#8
Comfy.n want to learn what I want Posts: 1,077 Karma: 6425108 Join Date: Sep 2020 Device: Calibre E-book viewer	I can only agree. Performance-wise there's a severe limitation on common hardware, measured as "tokens per second"... I remember some local models were painfully slow to process simple requests. And then there's that other limitation you mentioned on a similar thread "they can only be prompted with very limited amounts of text at the moment."

12-22-2023, 05:56 AM	#9
Quoth the rook, bossing Never. Posts: 12,149 Karma: 89201339 Join Date: Jun 2017 Location: Ireland Device: All 4 Kinds: epub eink, Kindle, android eink, NxtPaper11	Also too easily produces plausible junk due to scraping web sites and almost no human curation. It's a technological dead end and not the way to improve proofing/correction of grammar or spelling. Translation relies on matching a big database, a giant Rosetta Stone. Google started by scraping all the EU published documents as they are all the the same thing (by expert humans) in all the official languages of the EU. Which still includes English, though the UK left, because that's official in Malta and Ireland.