DevonThink & Apple Intelligence

Yes…but with backlinks to the “supporting” sentences in the source.

In my experience, Claude is the best one to make summarisations with the right prompt. I have this in Spanish, courtesy of my friend Moisés Cabello:

Escribe en español una lista de puntos continua y no numerada con TODAS las ideas y conclusiones alcanzadas en este texto. Divide los puntos en los encabezados (estilo H3 de markdown) que consideres necesarios. Desarróllalos para que sean perfectamente entendibles sin el texto original. Escribe una primera parte si crees que no puedes ponerlo todo de una vez, y ya te iré pidiendo las siguientes partes. Si no puedes terminar de ponerlo todo de una sola vez, escribe al final "continúa". Si lo has puesto todo, escribe al final "Fin."

Main problem with Claude is it does not search in Internet, but for summarisations and questioning, I think it is the best.

1 Like

I find that comforting :wink:

6 Likes

is this anything that will be released to the public in the near future? looks already good and seems to be quite useful for summarizing (and maybe also finding/creating text?

“near future” is of course highly subjective but we don’t announce releases dates, I’m sorry.

A lot of AI-generated prose seems to sound like that, always hedging and qualifying. Everything comes across as tentative. If this were spoken, every clause would end with rising intonation.

4 Likes

The nature of the mechanism is that it acts as a “smoothing” filter, buffing out the “highs” and “lows” that are typical of real (non-marketing) text. Strong statements are statistical “outliers.”

(You see the same in AI-generated art. It all has a certain “look” after a while.)

1 Like

I hope you implement it in a way, that the AI could be switched of, and prevented from sending the DT3-Databases to their own LLM for “learning purposes” or something else…

Everything will be opt-in.

4 Likes

Thanks for sharing so much information, including the company internal memos. I found it all very helpful.

Another noteworthy is Perplexity AI. This is an AI powered search engine, that can provide answers to questions while providing reliable links back to credible sources, something that chat GPT still struggles with (returns broken links and outright incorrect links).
I am quite prepared to use AI for critical decision making so long as I can see the source, and scrutinise that for correctness and reliability, and use my own brain to verify. For example by using the academic setting, only sources from peer reviewed literature/scholarly sources are returned with back links that work e.g. to PubMed. It is quite fast, extremely useful and removes on some of the anxiety if using for decision making “chat gpt can make mistakes check important info”. It makes this easier.

Perplexity AI has an increasingly poor reputation for questionable (or lacking) ethics, including allegations of plagiarism, failing to produce proper attributions, and ignoring robots.txt on sites.

1 Like

I think ALL these AI are going to brush up against these sorts of concerns but there is still some goodwill towards Perplexity. Looks like they are using third party crawler which might be ignoring robot.txt and agree that is a concern with this emerging technology. The issue Forbes had with “Pages” was addressed with update to software. Using any AI to generate content might increase risk of using text without correct attribution. I think this is less likely with Perplexity because you CAN check it more easily (at least in the way most users will use it) and if I publish anything it is going to be in my own words anyway. I am sure it is not perfect, but it is much easier to verify than chat GPT. I am looking forward to the Apple version of AI, because if they deliver as stated, the privacy protections will be verifiable.
I think it is only matter of time before the issue with links and citations, are improved across all AI and Perplexity is going to have a lot of competition.

Thus far, Anthropic’s Claude is the only one I feel good about.

1 Like

I’m using it to extract diamonds from Jules Verne books for the Hispanic Jules Verne Society. It’s incredible how you can get reliable info from a document with Claude. Handicap is if you have real true man (or woman) questions, you finish your time amazingly fast and must wait 6 hours for the next questions, even with paid account.

1 Like

what makes you feel good (or better) with Claude compared to the competitors?

Responses are more pleasant to read and don’t feel smarmy or hip. They feel professional but friendly and helpful (again, with the HUGE reminder, it’s illusory). Also, Claude 3.5 Sonnet released a few days ago is far more capable than anything I’ve seen from other LLMs.

More importantly, I support Anthropic’s stance on data privacy, not making people a commodity, opt-in vs opt-out, and responsible development of the technology.

1 Like

If you’re thinking of using Perpexity, you really should read this Wired article: Perplexity Is a Bullshit Machine | WIRED

1 Like

They’re all BS machines, really.

4 Likes