Ancestors

Toot

Written by Darrell Hilliker 👨‍🦯♾️📡 on 2024-11-16 at 08:25

Has anyone gotten any of the generally available AI's to take an inaccessible, untagged #PDF and turn it in to a more accessible format such as HTML? I tried to get #Perplexity to do this with the Zoom H1 Essential manual only to quickly discover hallucinations. Just wondering is anyone else has a better idea, or are we just not quite there yet? #accessibility #AI

=> More informations about this toot | More toots from darrell73@mastodon.online

Descendants

Written by John Allsopp on 2024-11-16 at 09:47

@darrell73 I use gpt4 and Gemini 1.5 for something related. I take screen grabs of the slides from presentation videos. I get the models to return structured html with descriptions of images, tables graphs and so on.

They do an excellent job though not perfect.

Though the information density of a page is greater so ymmv.

IME it never hallucinates.

=> More informations about this toot | More toots from johnallsopp@indieweb.social

Written by Darrell Hilliker 👨‍🦯♾️📡 on 2024-11-18 at 04:53

@johnallsopp OK I'll give this a try thanks.

=> More informations about this toot | More toots from darrell73@mastodon.online

Written by John Allsopp on 2024-11-18 at 09:09

@darrell73 it's not something I'd ask Perplexity to do–AI Studio gives you gemini 1.5 free, and did an excellent job IME. I use GPT4 via the API when I am doing this at any sort of scale.

=> More informations about this toot | More toots from johnallsopp@indieweb.social

Written by Dale Reardon on 2024-11-16 at 14:29

@darrell73 Have you tried Google's Notebook LLM - I find it amazing on reading, summarising, analysing and comparing documents and it cites&references the docs in its answers-I also use it for summarising audio files & Youtube videos

=> More informations about this toot | More toots from dalereardon@indieweb.social

Written by Darrell Hilliker 👨‍🦯♾️📡 on 2024-11-18 at 04:52

@dalereardon Thanks for the NotebookLM idea. I tried and it provided truncated output. Any ideas on good prompts to get it not to do that?

=> More informations about this toot | More toots from darrell73@mastodon.online

Written by Dale Reardon on 2024-11-18 at 12:56

@darrell73 I know with both OpenGPT and Claude I often get summarised versions of what I ask for first time. I just politely say that is not what I requested sorry, can you please provide a full exact answer, full transcript not a summary etc. Claude always apologises & does the full thing. Haven't had to try it with Notebook but worth a try

=> More informations about this toot | More toots from dalereardon@indieweb.social