Martin Elstner

The 120b variant is probably too large for home usage, but the 20b variant as 4 or 8 bit quantisation can run on notebooks or even maxed out phones. I get:
18.79 tok/sec; 1426 tokens total out; 0.60s to first token with the 4bit gguf (4096 tokens context) in LM studio on a M4 MacBook Air with 32GB.

LM Studio - Run openai/gpt-oss on your computer Run gpt-oss, Llama, Gemma, Qwen, and DeepSeek locally and privately.

In terms of tooling. LM studio lowers the entry barrier, at least on a Mac and it's totally accessible to any user (install the chat app from a .dmg, download the default model (~12GB) and you are good to go). lmstudio.ai

defaults directly on my sloppy, short inputs. We will see what can be build on top (the license is permissive) and how fine-tuning turns out, but the base model usability in chat is great.

the overall usability is great. For chat applications, the answers are exactly right in tone and length. It's a bit to positive and supports the users view (as most OpenAI models do). Qwen3 has also strong models, but I have to put more work into prompt tuning to get a good answer, get-oss has good

The Open Model release from OpenAI is a good reason to check out local AI models on your device.
Is it the best model out there? Definitely not, even in it's parameter class, it may not be THE best, benchmark wise (aside: OpenAI build sizes that are uncommon, so direct comparison is harder), but

I often see and use something like:
print(numpy.__version__)
is there a way to do it without calling the dunder method directly? For everything else I'm completely with you, but that's the one exemption that directly comes to mind

It's actually already build in some variants and products. But driven by a 'making money fast' approach. If done properly, todays technology already has much dystopian potential.

Be Right Back - Wikipedia

Waiting for people to live in something like that:
en.wikipedia.org/wiki/Be_Righ...
If it hits you in a weak moment of grieve, there is a lot of potential to get addicted and move to a very unhealthy "co-existence"

The jumping bunnies are an interesting case, because they hit the sweet spot of "uncommon, cool, but still plausible" and depending on the model you run it on, the quality is really good and as layman watching it you can't directly point to obvious flaws.

I honestly think, we need to adapt our notion of reality. The tools to reshape it are here already. We can try to regulate them, we can mourn for our loss or we can adapt. I would see it as an exercise in media literacy...
bsky.app/profile/mart...

But we made some progress on the “bunny video“ task

bsky.app/profile/mart...

Same for chemistry (IChO), not even the theory part

Nevertheless, mine is better :)

Dropping generated videos is a good test if you want to know what other platforms your followers use.
Bluesky has a large user base of AI sceptics and seems generally isolated from the hot topics on the video platforms (not judging, but the latter is probably a good thing).

Mehrere Kindergärten in Thüringen schließen endgültig | MDR.DE Künftig gibt es in Thüringen sieben Kindergärten weniger. Hintergrund sind sinkende Geburtenzahlen, baulicher Mängel und finanzieller Engpässe. Vor allem Ostthüringen ist betroffen.

Man schaut da glaub ich auch schon auf den weiteren Verlauf und die Geburtenrückgänge. Oder: "Die Kinder die heute in der KiTa fehlen, verstopfen in Bälde auch kein Gymnasium"
www.mdr.de/nachrichten/...

TBC: the day off atlas

bsky.app/profile/kevi...

A few years back, I ordered Judea Pearl's "Causal Inference". But was expecting "Casual Inference", like in "Yeah, I do inference, but very casual". You can't imagine my disappointment

YouTube video by Sabine Hossenfelder New Experiment Rules out Bohmian Mechanics. It’s Serious.

Need to read more carefully. Was watching Sabine's recent video and was expecting to learn about "Bohemian mechanics". You can't imagine my disappointment
www.youtube.com/watch?v=DvpS...

DHH on the Lex Fridman Podcast (2h in): probably you can see a productivity drop this week from all the techies wading through it...

A little snail sitting inside a cars charging port

POV: EV charging in rural Germany

A 12 panel cartoon showing an increasingly well worn walking path in the grass.

A 'desire path' represents the shortest or the most easily navigated route between an origin and destination.

If you can preempt these in your UX/UI you'll lower support tickets and increase user retention.

People are going to people. Let them.

A workflow with just and uv, close to what I had stitched together over the last months. I got a handful of new ideas (like putting PORT in .env); recommend watching it!

So, Simon says I should quitting this computer programming thing
<taking notes>

Great post, agree to most of your points!
At the moment, I just feel that MCP is the least awful way to make tools available in an enterprise context. Yes there is significant overhead, but enterprises are ok with that. Besides the only available way to channel tool discovery.

Maybe a bit of jargon, but as many microwaves have pressure sensors (and 10bar sounds like close to the maximum they allow) it’s like ‘under reflux’?

Viele sind an den Ergebnissen interessiert, aber nur wenige rechnen damit ;) Genau andersherum als bei Halbleiter band gap predictions. Gibt es ein gutes Erstsemesterbuch o.Ä. in das man mal reinschauen kann?

Guter Artikel

Still wonder why Apple (or the ecosystem) doesn’t have anything

Initial exploration of using Koog for developing Kotlin based AI agents Koog, announced recently at KotlinConf, is a new Kotlin-based framework designed to build and run AI agents. This article will outline initial exploration of using Koog along with a number of configur...

Will try this out, also to check how well coding agents can help me with a language I don’t know. But I like the examples I found, e.g.: johnoreilly.dev/posts/kotlin...

Thank you, looks good and well thought through. Unfortunately far from my usual tech stack