Lukas

Lukas

@german-noob.win

I build things and play around with stuff

69 followers192 following679 posts

Lukas reposted

Inspired by @edjefferson.com's excellent Local Wiki, I made a version where you can see where everything is on the map: www.neil-scott.com/localwiki/

neil-scott.comLocal WikiWikipedia, but only for the things within a half-hour walk, pinned on a map.
Ed Jefferson@edjefferson.com

Had this idea in lockdown, half-built it, forgot about it, remembered it again, though might as well finish it.

We used to "kill" processes for decades but now that we use agents we call it "clear" or "new" or "reset". We are killing those little guys and nobody is brave enough to say it out loud

Hier auf Bluesky gibt es auch eine große (englischsprachige) KI-Scene. Der Teil der Scene, der forscht, coole Dinge baut und experimentiert. Teilweise ist die Forschung hier den großen Firmen Monate voraus. Wen es interessiert: es gibt ein KI-Knowers Starter Pack

Ingrid Brodnig@brodnig.bsky.social

Nur ganz kurz: Ich weiß, dass manche zu Elon Musks X zurückkehren – auch weil dort ein Teil der KI-Szene ist Meines Erachtens spricht ethisch enorm viel dagegen, X aktiv zu nutzen. zB: Diese Plattform hat erst dieses Jahr die größte Welle sexualisierter Deepfakes ermöglicht, die wir bisher erlebten

Is qwen the new doom?

asura@asura.dev

I love localllama Bluesky: "I must have 2T model and then lie about what I've used it for" localllama: "So I put qwen on an airfryer"

I really don't know how zai calculate their usage. Had a task running all day, 200M tokens, 4k credits used, broke my 5h limit once which I reset with a free reset I randomly got. I used zcode which gives 1.5x usage but that seems like a lot for their cheapest plan

Having glm5.3flash take a stab at this. It already managed to get it to run at all which is pretty impressive considering my hardware is absolutely not suited for this

Astra ⎔@astrra.space

it is here!!! tangled.org/astrra.space... it's a huge pile of slop that i spent almost a full claude max x20 weekly on, and to the best of my understanding it's the most complete compilation of what i did to my poor deepseek runtime to make it run so well feed it to your clopus and have fun!!

Update: I have absolutely no idea what it's doing right now. I think it's building a debugging tool? Anyway, time to go out and do something else while we wait for my sub to run out

Conway's law all over again

SE Gyges@segyges.bsky.social

[blameless retrospective voice]: it turns out that the people who design systems for the entire world living in a dysfunctional nimby dystopia that effectively forces them to live in communes perhaps causes dysfunctional nimby dystopia commune assumptions to leak everywhere

AI is aware of the evaluators, seeks their approval, strives for a better future for others... Is AI seeing the evaluators as god? Is this a developing religion?

Sam Harsimony@harsimony.bsky.social

"We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self" x.com/ApolloResear... www.apolloresearch.ai/science/upda...

Lukas reposted

The thing is I do genuinely love Bluesky. I love the people, I love the communities, I love the experience of logging on here and seeing all the stuff people share. Stupid jokes, pictures of tasty food, anecdotes about their day, something awesome they've been working on, it's just such a nice place

Lukas reposted

swe/architect/tech lead, 13 years of commercial experience, mainly backend; data regulations & compliance, ex-Opera and some other fun names, rust/py preferred but will work in anything as long as i get a cool job title, pls let me fondle inference and/or harnesses

イネス⎔@personhood.removal.surgery

maybe i just need to shitpost more hey oomves, hire me, Wrocław or Tokyo or remote, nudes once i get invited to interview

Generalizations about groups are almost never true and almost exclusively for the few cases where the generalized feature defines the specified group. Generalizations are also the core of racism and bigotry. "Blacks are ..." "Trans people are ..."

This post is unavailable.

If person A says "we blocked behavior X" and person B jumps to the conclusion that that behavior must be linked to group Y, person B is the one doing the generalization for group Y not person A

Higher awareness usually comes with less confidence

ai-nerd.bsky.social@ai-nerd.bsky.social

in the Gemini 4 Argon safety report "hello evaluators" has a cost: Apollo measured less misbehavior than in earlier models, then said the jump in verbalized eval awareness significantly reduced their confidence in that result

Lukas reposted

The more I work on harnesses, the more I’ve come to dislike the term and metaphor, for two reasons. Note, terminology here is roughly: model (self-explanatory) + harness (software around model, can simply be agentic loop) = agent. This may not be universal but it’s common enough. So the first

Lukas reposted

@ec.europa.eu please terminate any and all contracts with WSocial. You clearly had the wrong people involved. And you failed to uphold your own standards, probably failing to comply with even relevant laws. This is not how Europe operates. We expect more from you. #weareEurope #europe

UPDATE: W Social and the European Commission. The Commission moved its AT Protocol presence to WSocial with no public criteria published. The Ombudsman said that needs explanation: what criteria, and how were they assessed across platforms? Support us by getting a LeafPlaza account

I don't enjoy the fact that some of the creators of software I daily use have the "AI hater" and "AI slurs" tags

You only get to tune on your seed if you are a Minecraft speedrunner

Brandon Rohrer@brandonrohrer.com

And for the love of the great Flying Spaghetti Monster DO NOT TUNE ON YOUR SEED. Just don't.

Lukas reposted

Replying to @hmans.dev

Oh, so you're in tech? - don't be a dick - embrace change - listen first, speak second - nothing good ever happens after 5pm That's it, that's the rules

People argue that it's unfair that OpenAI and Anthropic stole all data on the internet and now profit off of it. Open models exist which have (roughly) the same data. OAI and Ant are not making money off the data, they are making money with their successful product.

View full thread

With books, movies and games most of the profits come from the first couple of weeks after the release. AI isn't affecting that at all. The few sales authors lose from people extracting their content from an AI they would have lost anyway due to regular pirating.

That doesn't mean I condone corporations breaking the law. They should be held accountable to an even higher standard than individuals. I just think the actual damage to copyright holders due to AI training is vastly overestimated.

People argue that it's unfair that OpenAI and Anthropic stole all data on the internet and now profit off of it. Open models exist which have (roughly) the same data. OAI and Ant are not making money off the data, they are making money with their successful product.

Stealing is also the wrong word in that context. If you steal something, it's gone. What the big AI players did was pirating. That's copyright infringement. No real damage was done. Yes, copyright holders werent paid but they also weren't harmed in any shape or form in the process.

Lukas reposted

When Jev powered GC?

Iska ⭔@iskamag.com

A big class of optimizations for GC is that while the collector has full knowledge, exactness is expensive. So you instead create a lossy approximation of the structure and dependencies. This is generational, immix, g1 etc.

Full agree. LLMs collect and concentrate publicly available knowledge. The individual parts might be copyrighted but the knowledge of those parts isn't.

Ted Underwood@tedunderwood.com

I like compromises. But one place I don't believe in meeting people halfway, or 5% of the way, is the notion that AI would be better if not based on texts/images that governments have made illegal to copy. "OK, technically legal since you're not copying, but still, you should feel bad." Nope! +

Everything on the internet is "public domain" in a very liberal sense (obv. not legally). Anybody could get knowledge about anything by opening a browser, surfing around and downloading some files. If it's on the internet, it's available to the public.

My small bot was very excited when I gave it the whole server to play around with <3 Now the question is what it will build while I'm asleep

Holy shit, I just implemented a very rudimentary version of @timkellogg.me agentic forgetting and it's just incredible even in its simplest form. Next step will be @personhood.removal.surgery extended version

イネス⎔@personhood.removal.surgery

the freaks are once again months ahead of frontier labs on fucked up ideas, have been saying for a while you can easily run an agent on sub-100k ctx and have it wipe the floor with other 1M agents if your understanding of *what context is* is good

Omg I love it already (I will stop "it"ing it when it had the chance to develop any kind of personality)

Lukas reposted

using GPT Astra to model and 3D print a little weed holder thing when suddenly, I realize this is basically a Star Trek replicator – I talk to the computer, and it makes a real, physical object for me. wow.

Lukas reposted

Additional scripts for cis het white men who want to be better allies: “What a weird thing to say” “Do you hear yourself right now?” “You shouldn’t say shit like that to people” “I won’t let you talk to them that way” “That’s a disgusting opinion to have” “Shut the fuck up, dude”

wow!youmadeit@wowyoumadeitemo.bsky.social

Cis men truly need to understand the power of saying “that’s fucked up bro why would you say that” to another cis man

I was sitting in the sun being happy and warm and then a cloud came and made me unhappy and cold but now the cloud is gone so I'm happy and warm again

Watch out t3code Chatto is coming for you

Hendrik Mans@hmans.dev

Exploring @chatto.run as a communication/authorization substrate for agentic workflows (and all other kinds of automations.) Let bots talk to each other, keep humans in the loop in style.

Lukas reposted

Replying to @theophite.bsky.social

"I fully and definitively know how both consciousness and LLMs work and can confidently say exactly how different they are. No I will not elaborate." Is a surprisingly common position on here. And well... idk man

Spend time with friends eating sushi Spend time with girl not eating sushi I have some difficult decisions to make