@Varpie

🪨 @Varpie@peculiar.florist

7mo

@adzetko Ce qui est bien c'est que t'as pas à rajouter les petits drapeaux, t'as déjà Céleste qui indique ton identité

⭐1

0

1

🪨 @Varpie@peculiar.florist

7mo

@tournesol En même temps bon tu paies la marque quoi, c'est pas pour rien que c'est marqué en gros dessus

1

0

🪨 @Varpie@peculiar.florist

7mo

@box464 It is the same principle, I saw that comparison as a diagram in a documentation website and I can't find it back, but I remember thinking that it was a neat comparison... I should use bookmarks for this kind of situation.

⭐1

0

1

🪨 @Varpie@peculiar.florist

7mo

@box464 I saw the other day a picture that summed it up pretty well: it was a comparison of centralized, federated and decentralized architectures. Unfortunately I can't remember where it was...
In the centralized one, there is one big circle in the middle (the server) and plenty of links pointing to small dots (the users).
In the federated one, there are "islands" of servers linked together, of different sizes. They aren't all connected to each other, so there can be single ones, smaller groups or a big land, so it kind of makes an archipelago. Local community instances could be like the smaller islands here.
In the decentralized one, each server is connected to everything else, so it is just one big mesh.

⭐1

1

0

1

🪨 @Varpie@peculiar.florist

7mo

@tournesol feur > coubeh, confirmé

⭐1

1

0

1

🪨 @Varpie@peculiar.florist

7mo

@tournesol Attends c'est un symptôme de l'ADHD ça ? Peut-être qu'il faudrait que je me fasse diagnostiquer...

1

0

🪨 @Varpie@peculiar.florist

7mo

@tournesol Ah, je pensais que t'allais dire "mais maintenant il y a du bazar au sol à droite de mon bureau" :blob_nervous2:

⭐2

1

0

2

🪨 @Varpie@peculiar.florist

7mo

@scottjenson This is great but it assumes a single device (private key) per account. What if I want to have a phone and a desktop? Whatsapp and Signal both "solve" this by having one main device (the phone) and connected devices that make use of this main device's private key, but in the context of having different applications connected to a fedi account, I'm not sure it would work too well.

⭐1

1

0

1

🪨 @Varpie@peculiar.florist

7mo

@tournesol @MoritzBrouhaha It's a limitation that makes sense though, imagine if following an account fetched all of their previous posts... Here we go potentially hundreds of thousands of posts at once, every time someone from a new instance follows someone shitposting for years :meow_melt_sob:

1

0

🪨 @Varpie@peculiar.florist

7mo

@scottjenson True, handling the messages in a standardized way is one thing, but managing keys across multiple clients is the hard part here. The way I see it, there are 2 options:
- each client creates its own key, encrypted messages now need to be encrypted with multiple keys and the new clients don't have chat history (this could be mitigated by having existing clients with the decrypted messages send them to the server with the new key)
- there is some sort of handshake when registering a new client, that passes the private key from a registered client to the new one

The first option allows to handle each client separately, so we don't need the other device to be available and if we want to stop using a specific app, we can deregister it, but it requires senders to encrypt their messages n times, and as mentioned it makes it difficult to handle chat history.
The second option makes chat history trivial, but it puts a lot of trust on new clients, if we want to stop using it the rotation of keys is more complex. Also, each client needs to be able to handle the same type of keys, which isn't a given when using different apps.

I think for user experience, having each client generate its own key and asking older clients to re-encrypt messages with the new key can be better: there is no requirement to have the other clients active at the same time, but we can have the same handshake that would be required for passing PKs, to recover chat history. It also allows to give more granular control over which clients are active, kind of like seeing the active sessions for an account and being able to log off on other devices.

1

0

🪨 @Varpie@peculiar.florist

7mo

@tournesol Why not both?

1

0

🪨 @Varpie@peculiar.florist

7mo

@scottjenson Private mentions aren't really private if they're not end-to-end encrypted. On a federated platform, you put a lot of trust on the servers, and it's not just the one you're on but also the one receiving the messages. What if I want to message a friend on Threads for instance? I don't know about you, but I don't trust Meta to just deliver the messages without using them to build a profile on me or improve their AI models, which are things I can't really opt out of since it's not my platform. The only way to avoid these things (to some extent) is to make it impossible for them to read my messages.

The good thing is you don't have to reinvent the wheel here, there are already a few attempts at standardizing encryted messages for ActivityPub: Evan put together the MLS over AP, Hollos also did something similar. Make sure to talk to them so we don't end up with yet another standard.

1

0

🪨 @Varpie@peculiar.florist

7mo

@jwcph That just validates my opinion on LLMs: they are just a tool, and if you can't code without them you shouldn't depend on them in the first place.
In a way, they are a multiplier: they can make a good coder more efficient, but for someone that doesn't know what they're doing they will just result in a lot more bad output. Just like an efficient saw can help a good woodworker, but also result in a lot more wood scraps if used by an unskilled one.

I do agree that the reliance on a handfull of companies is bad though. Since it takes so much resources, it's not like anyone can build a decent LLM, so the competition just isn't there, unlike other tools where there are usually many good options (including more ethical ones...)

0

🪨 @Varpie@peculiar.florist

7mo

À chaque fois qu'il y a une nouvelle LLM qui sort (donc à peu près une fois par semaine ces jours-ci), je vois des gens qui parlent de modèles "open source".
Est-ce qu'on peut juste tous être d'accord, ce n'est pas des modèles open source mais open weight parce que c'est juste les poids du réseau neuronal qui sont donnés de façon ouverte ? C'est aussi con que dire "ouais mais ce logiciel est open source, je peux librement télécharger l'exécutable".
Je propose qu'on n'appelle "modèle open source" que les modèles qui:
- rendent public l'intégralité de leur dataset
- rendent public le code source utilisé pour entraîner le modèle (y compris la seed, pour que ce soit reproducible), sous une license approuvée OSI
- rendent public le code source utilisé pour l'inférence (je crois que cette partie est la seule qui est correctement appliquée actuellement, sans quoi on ne pourrait pas utiliser les modèles open weight)

0

🪨 @Varpie@peculiar.florist

7mo

wordsalad stuck | no alt text

@tournesol Il reste quoi pour le mot en 6 ?

1

0

🪨 @Varpie@peculiar.florist

7mo

wordsalad stuck | no alt text

@tournesol Ah bah triangle du coup

⭐1

1

0

1

🪨 @Varpie@peculiar.florist

7mo

wordsalad stuck | no alt text

@tournesol Bon, juste avec les lettres présentes : Pologne. C'est en 7, pas de rapport avec l'amour et c'est en français, mais au moins y a pas de S.

⭐1

0

1

🪨 @Varpie@peculiar.florist

7mo

wordsalad stuck | no alt text

@tournesol Mais comment t'as nest, seat et song du coup ? :akko_what:

2

0

🪨 @Varpie@peculiar.florist

7mo

wordsalad stuck | no alt text

@tournesol Je sais pas comment ça marche ton truc, mais en 6 il y a Love Island (je vois que les mots trouvés ont des S alors qu'il n'y en a pas sur les cases du haut donc jsp si c'est valide)

1

0

🪨 @Varpie@peculiar.florist

7mo

@tournesol @Sidoine Ici ça a l'air d'être JupyterBook qui est utilisé

⭐1

1

0

2