petersson.eth

@pidotdev @vieitesss_ @badlogicgames @mitsuhiko The black 🖤 makes me anxious, did something terrible happen?

@vieitesss_ @badlogicgames @mitsuhiko Welcome to Vienna 🖤

Good morning from Vienna People of Pi🖤 Sunday Meditations from @badlogicgames and @mitsuhiko - Thinking is Essential - Minimalism over Maximalism

@jun_song Which model did you use? Do you have a HF reference?

The new engine for MLX is in its final stages of development. Just ran GLM-5.2 on a single MacBook (116GB) hitting 41.8 tok/s with a 256k context window. Quality loss is only around ~4%, which puts it right at the 3-4bit quality level. The tech behind this uses a newly introduced layered architecture. When I first started, I was getting 10 tok/s with Kimi-K2.6 (128GB, 1024 context). Now it is fully at production level. Been grinding on this for months. Feels great to see it finally coming out soon.

Cloud AI is convenient. Local AI is control. I joined “Was KI kann” to talk about local AI, data privacy, and why digital sovereignty matters. Listen here: youtube.com/watch?v=ljUpT3… #AI #LocalAI #DigitalSovereignty

@james406 Even more proof for flat earth. If it were round, the shortest path is through the middle.

wow. AI is seriously amazing. i asked it to find a better route for the Sydney - London flight Opus 4.8 found a much more efficient route that flys in a straight line instead of a curved one. but Fable 5 found an even better route that's half the distance! please tag Qantas so they can see this, this will revolutionize the airline industry

Somebody train the models what "ACs" are when talking about validating tasks. Acceptance Criteria. It's the things that need to work to close issues.

@mitsuhiko Come to our meetup on Wednesday in 1010 Vienna, I will give a talk there and we can chat about this. I have a few public mid sized projects to share. Ownyourai.eu. / luma.com/3yy6gwic :)

I decided to do some experiments with looping over the weekend. The only cases where they work so far for me are a) review b) research c) autoresearch. If someone uses them for actual implementation on a medium sized project, would love to have something to look at!

@jassdotgg @theo Almost identical. I use both, depending on the tool. Some Ralph loopa have fancy "reflections" to better get unstuck

@theo how is this different from Ralph loops

I hate to admit it but the loop people were right

When I struggle to structure my thoughts about what's happening I turn to writing. Today about the recent US Anthropic ban news, what it says about power and dependency, and what it should mean for Europeans and citizens of the world. It's a long one. lucumr.pocoo.org/2026/6/13/amer…

@BrantlyMillegan @pcaversaccio Do you think a greenfield wallet should implement an option that after connecting w walletconnect it will automatically SIWE on demand

The whole point of the Sign with Ethereum standard (EIP 4361) is to avoid the exact problem you're talking about. you should not just sign random messages, agreed. That's why there's a standard, and many wallets show the UX differently because it is a standard message (more wallet should do this). hence if you want users to sign in with their wallet you should use the EIP 4361 standard

sign-in with ethereum is actually a terrible idea because it trains people to sign messages all the time & creates _signing fatigue_. what you actually want is the exact opposite. a sig should be considered _dangerous_ and treated as a _special event_. it should not be something users do casually as part of their normal workflow. if people get used to clicking "sign" everywhere, they stop paying attention & eventually sign things they do not understand (this happens every minute in this space btw). the right security model is to teach users that signing something is a _serious action_ that deserves a lot of scrutiny, and not to normalise it as just another login button. we should immediately stop doing this shit.

This is what my "Dark Software Factory" output looks like. Parallel worktrees. High-capacity orchestrations. multi-turn review/fix iterative looping. Cheap writing, expensive validation. Unit + End2End integration tests of progress and features. How does yours look like?

@jundotkim @antirez check out github.com/apetersson/oml… the day when we can use DS4 as an engine within oMLX - is today. Running at exact same speeds as DS4 proper on my M1 Ultra 128GB. Looking for feedback before i draft a proper PR.

What a great initiative! DwarfStar is a really impressive engine, and I love the concept of taking a single model, optimizing it to the absolute max, and running it at top speed. oMLX is built with an open architecture that can support various engines - right now it supports mlx-lm, mlx-vlm, embeddings, and audio. I'd actually thought about embedding DS as a separate engine myself, but given the time I can realistically invest in development right now, I don't quite have the bandwidth to extend coverage all the way to a DS engine. I'm cheering you on, and I hope the day comes when we can use DS as an engine within oMLX too.

I love oMLX for its API, memory management and the ability to put many different model families under one umbrella. I have also tried out DS4 and sadly, it is just way ahead in terms of efficiency (generation, preprocessing) and flexibility (ssd streaming)

@jundotkim @antirez Ralph is on it, that day might not be far away

@antirez Nobody can stop my local ds4! Except myself; not even @antirez. The "Dark Software Factory" will continue to run.

Anthropic global AI pause is the right call.

@antirez Would it make sense to use this approach to run Q4 or higher of DS4-Flash on 128GB?

1. It is slow, obviously, but not unusably slow if for some reason I'm cut out of frontier AI. 2. The goal here was to make this working slow but optimize it enough to have Flash good enough on 64GB systems. I'll ask somebody to check what performances it reaches. I don't have a 64GB machine.

DeepSeek v4 PRO running via SSD streaming on my 128GB MacBook m5 max. 1.6 trillion parameters.

@thsottiaux I was still unable to simply move an existing chat to a new project

What’s something we haven’t fixed in codex for a while and that’s plain annoying?

@steipete @Olli0103 @openclaw Big oof, bureaucracy 🫨

@Olli0103 @openclaw Current status: Waiting for a DUNS number to get an Apple account for the non-profit. This can take anywhere between a week and month.

Can you already estimate / tell, when the TestFlight-Version of @openclaw iOS will be out? @steipete

@did0f Gemma4 uncensored, just to help me connect to the onboard wifi

Guys I was using a local model on the plane, the attendant asked me half-joke-half-concerned if I was doing something bad looool No m’lady, just Gemma4 on omlx flying like an 🚁 on my beloved M1 - sorting out my Obsidian trash

@HermesAgentTips Ds4 Flash imatrix abliterated. Iykyk

Local LLM folks if you had to choose only one local model & you’re stuck with it for a month Which one would you pick? •Llama 4 Scout •Qwen 3.6 27B •DeepSeek-R1 •Gemma 4 •Mistral 7B •Phi-4