Thinking about how to add memory to an #LLM
Might be good to have two models running in tandem. One is trained to use a memory API. The other is just the personality. Before the personality gets data, memory decides if it wants to do some RAG queries. Then the result is fed to personality to respond with without it needing to understand memory. Then memory asks itself if it needs to remember any of personality's outputs.
Neat article my coworker found about some of the limitations of LLMs and so called #GenAI (generative AI?).
https://www.aspendigital.org/report/navigating-uncharted-waters/
What Apple is going to do to PWAs in less than two weeks in the EU is dastardly.
Their goal is to remove the only versions of Push, Fullscreen, Homescreen Icons, and Badging that didn't require paying them to be in the App Store *and make sure no other browser can have them either*.
They tried to sneak it under the noise of the alt-store drama, in a clearly premeditated fashion.
It's no exaggeration to say they've gone to war with the web & we don't owe the benefit of the doubt ever again.
Suuuper rough sketch for anyone interested. Basically we'll have a list of accounts you follow in localStorage, then ingest their outboxes into this. Then we'll set up fully p2p identities and use the #DistributedPress #SocialInbox as another source to ingest stuff like followers only posts and replies from people you don't follow.
https://github.com/hyphacoop/reader.distributed.press/pull/3
https://soundcloud.com/mixmag-1/premiere-oyubi-tomys-siren
Music to review code to. 🤓
Bouta start a new database for client side indexing of ActivityPub data and making user curated timelines that much easier.
Hot take that I only partially believe, but that is made unnecessarily inflammatory for the sake of social media brevity:
Nearly every bad thing in tech traces back to that the tech sector is one of the first to grow in post-Reagan disruption of labor rights and antitrust regulations, such that it is uniquely concentrated in a few unaccountable effective monopolies.
Been playing with continue.dev in Visual Studio Code hooked up to lmstudio.ai running OpenHermes 2.5. It works *okay* but I think the default prompts aren't exactly tuned for the ChatML format from the Mistral 7b model. So far editing code has been a shitshow but it's been useful to be able to ask it for python syntax instead of looking it up on duck duck go every time. I have it running on a mac mini in my closet so the dedicated M2 chip makes it pretty speedy.
1. New default: Your phone number will no longer be visible in Signal
2. You can create an optional username to connect without sharing your phone number
3. You can enable a new, optional privacy setting to require people to connect with you via username instead of phone number
Our fully remote team at @hyphacoop does a weekly coffee break. This year I've updated it from a Google Meet room to a Mozilla Hubs virtual office space complete with a giant water cooler and zany avatars. Since hubs is going to be closing down due to the layoffs I might start self hosting it or migrate to another platform like Thirdroom.
btw here is something cool - the 3 of us watching on @monsterdon were making the #p2p thing of peertube happen. i was a little behind because i kept going back to watch scenes again, so i mostly downloaded from whoever was ahead rather than uploaded, but like this is so sick to me because if it had to serve the whole movie to everyone it probably wouldn't be able to handle many more than this, but with p2p we can basically have as many people watching as we want.
first pic is from like 2/3 of the way through the movie, second is from the end, so most of that server download happened at the beginning when i was the only peer.
#monsterdon
One trend I keep noticing about interviews with Gargron or any article about “Mastodon” is how often they ask Gargron about “competing” or they write about competing and growth.
Gargron, to his credit, usually semi-dismisses it, but it shows just how lost in the hypercapitalist techno-feudalism sauce tech journos are that they still make it out to be about growth.
It’s been over year since Mastodon and the Fediverse gained mindshare… It’s clear that while they vaguely understand the technical underpinnings, they do not understand the social underpinnings that make it all possible.
It’s like they believe that each Fediverse server is a micro-Twitter that wants to be purchased by a technocorp, and all admins are in it because they want fifty billion users on their instance.
The Fediverse is the “next Twitter” in the same way that the tram is the “next car”.
It’s a trap. More code is not your friend, it’s your enemy.
Don’t add the dependency. Don’t generate 8000 lines of JavaScript. Replace JS with CSS. Don’t accept the digital grey goo spewed out of tools such as Copilot.
Code is a liability. Anything that helps you make more, quickly, with no effort, is leading you down a very dark path.
(muting this now; sorry, but it’s too noisy)
Everything about Apple the company pisses me off so much. They're some of the most abusive companies out there with how they treat the needs of others.
https://open-web-advocacy.org/blog/its-official-apple-kills-web-apps-in-the-eu/
The Chaos Computer Club is giving out art grants!
If you're an artist/writer/creator of whatever type, you can apply for 420–4200 EUR, to work on your visions and utopias for the Chaos Community! ✨
See https://42.ccc.de for details. Application deadline is already the 2024-02-29! ⏰
If you have more questions, join the @callforstories "office hours" next Thursday!
Occult Enby that's making local-first software with peer to peer protocols, mesh networks, and the web.
Exploring what a local-first cyberspace might look like in my spare time.