Proton

9015 readers

88 users here now

Empowering you to choose a better internet where privacy is the default. Protect yourself online with Proton Mail, Proton VPN, Proton Calendar, Proton Drive. Proton Pass and SimpleLogin.

Proton Mail is the world's largest secure email provider. Swiss, end-to-end encrypted, private, and free.

Proton VPN is the world’s only open-source, publicly audited, unlimited and free VPN. Swiss-based, no-ads, and no-logs.

Proton Calendar is the world's first end-to-end encrypted calendar that allows you to keep your life private.

Proton Drive is a free end-to-end encrypted cloud storage that allows you to securely backup and share your files. It's open source, publicly audited, and Swiss-based.

Proton Pass Proton Pass is a free and open-source password manager which brings a higher level of security with rigorous end-to-end encryption of all data (including usernames, URLs, notes, and more) and email alias support.

SimpleLogin lets you send and receive emails anonymously via easily-generated unique email aliases.

founded 2 years ago

MODERATORS

ProtonPrivacy@lemmy.world

alex_herrero@lemmy.world

Nelizea@lemmy.world

-8

Can Proton Lumo fix bad LLM replies on technical questions like syntax is "a=xy" not "a = xy"? Are Lumo thumbs down really useful? (lemmings.world)

submitted 1 month ago by hetzlemmingsworld@lemmings.world to c/protonprivacy@lemmy.world

1 comments fedilink hide all child comments

I can set a thumb up/down on a Lumo reply and explain what went wrong. My uninformed guess is that in general cases my thumb-down may allow them to fix the Lumo reply by prioritizing a different LLM for the question but can they fix or adjust LLM output itself?

For example Lumo suggests to use file abc.prf while the app is actually looking for def.prf or suggests adding configuration file line "a = xy" while app accepts "a=xy" (no spaces).

Should I spend time proving how such reply was wrong (they can not trust my claim)?

Proton says that they do NOT use AI APIs but system prompt to fix replies not by doing corrections to the underlying knowledgeby but by adjusting behavior patterns (instructing the model to search for current information, to express uncertainty appropriately, or to avoid certain types of responses).

top 1 comments

sorted by: hot top controversial new old

[–] isVeryLoud@lemmy.ca -1 points 1 month ago

Oh hey look, it says "gullible" on the ceiling!