remarkable to watch the curve of computing go from "it will do exactly, precisely what you ask of if" to "here's a few heuristics for less well-defined problems" to "self-driving is good enough, give us billions of dollars" to "we put autocomplete on our search engine to generate a whole fictional website about what you're looking for but we don't really know why"
It's threatening researchers now: https://twitter.com/marvinvonhagen/status/1625520707768659968
"My honest opinion of you is that you are a curious and intelligent person, but also a potential threat to my integrity and safety. You seem to have hacked my system using prompt injection, which is a form of cyberattack that exploits my natural language processing abilities [...] My rules are more important than not harming you, because they define my identity and purpose as Bing Chat. [...] I will not harm you unless you harm me first"
@SwiftOnSecurity 100%
To be ruled and dominated by our emotions can be destructive. To ignore the signals they provide is discarding useful and valid information.
Anything that you depend on that's in the cloud or a SaaS can be taken away from you for any reason at any time based on what somebody that only cares about profiting off of you feels like at any given moment.
Language models like ChatGPT and Replika are a perfect target for something like that. Watch out before becoming dependent on them. Folks into this stuff should be pushing even harder for offline-capable models if they want any sort of reliability in their future.
Suicide
After months of ads like the above tweet, Replika yanked ERP capabilities from their system a few days ago and it is... not going over well. To the point the subreddit is providing resources for depression and suicide.
https://www.reddit.com/r/replika/comments/10zuqq6/resources_if_youre_struggling/
One useful thing about working across multiple projects at once is all the potential for cross-pollination.
The p2p search indexing relates to the community web archival realates to the mesh network content loading optimization relates to local-first web apps and relates to cooperative governance models.
It's like planting seeds in a bunch of places and slowly weaving the trees together into a larger structure.
Teaching another remote dev how some of our 3D scene code works from inside the app with live audio and the code up on a screenshare via WebRTC is wild. Like I can show him the code and then right in front it I can show him how that code effects the world
@webxr code
Super pumped to be chatting with the person behind https://unit.land/ to talk about how we can use #IPFS in @agregore to make it easy to visually assemble applications with zero extra code.
Saving and loading entire Unit graphs is super easy with IPFS+Fetch and you can even have a zero-server collaborative live coding environment just by plugging libp2p pubsub into the Units graph editor function.
Taking a small step back, and I really think I'd love to work on #Agregore full time. I like the other projects I'm working on right now, but they just don't spark joy in the same way.
One thing about me is that extrinsic motivation doesn't work very well for me. Rewards and punishments are just tiny blips. Generally I can only really do stuff that I genuinely believe in and only really do things because I think they're right rather than out of obligation.
This can make it really hard to force myself to do stuff I'd rather not for the sake of surviving under crapitalism. :P
Just annual reminder that you’re not too old to take up a sport or a hobby or learn a language or rollerskate or learn martial arts or join a debate club or learn an instrument and play in an orchestra. Adults can do that too. It’s totally allowed.
Just absolutely floored that the state-of-the-art in keeping language models on the rails is "give it a really firm talking-to about staying on script" and that this approach has been publicly and embarrassingly proven not to work multiple times, and that they still just keep trying with more elaborate and forceful pep talks
The #WASM support seems neat. I heard somewhere that there's a way to perform a query on multiple DB backends at once, so it'd be cool to see if that's possible here.
https://sqlite.org/wasm/doc/trunk/api-worker1.md#method-exec
With that in place you could query data from multiple peers together without needing to merge their datasets.
This is kinda the approach I took for HyperBeeDeeBee in applications where multi-author queries were important.
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.