aktenlage

Not my experience. Limited, but definitely not nothing.

I'd chime in with @rablackburn: mixture of experts is the way to go. I have a laptop with 6GB VRAM and I'm running KDE with a 4k display on the same machine, so there's only about 4 to 4.5GB actually available.

Using llama.cpp with Qwen3.6-35B-A3B or gemma4-26B-A4B gets me 200-300 tokens/s on prompt processing and 20-40 t/s output, which is good enough for me. Of course it gets slower with larger context. Interestingly gemma is faster, even though it has more active parameters.

It took a lot of parameter fiddling to get it to that speed. If you are interested I can give you some guidance on it, but I guess there are more qualified people around here.

The intelligence is good enough for simple questions and tasks (e.g. bash command howtos, asking about compiler errors, summarize something, document a code function/file, etc), but not good enough for complex things.

That is damn interesting. When I made copilot write a script that creates a .dot diagram, I found it funny that it had the same problems with placement that I have when doing it manually.

It always hurts me, when people basically say "I'm not giving up [something that has cornered the market] for the alternative, because [minor-grievance]", even though it is already clear that the monopoly is being exploited to make things worse for all customers.

Would it speed up prompt processing if you increased the -ub (and -b) parameters.

> Just wish we have a little bit of context

See the comment of vocx2tx

They resumed development on jemalloc only recently, after years of no releases.

So does duckduckgo with their duck.ai. No outrage there.

Enlighten me

I use Ag integrated in the editor all the time. But while it is indispensable for me, I wouldn't say that this helps in all situations, nor is it the best tool to find connections in many situations.

My take on the self correction is, that later research usually cannot be build on wrong results. So over the decades or centuries the correct results prevail.

Wow, setting the "ubatch" size to 2048 (default is 512), like the "batch" size roughly doubled prompt processing speed.

Hm, maybe. Though I thought that's a combination of what Corona did to the society, late-stage capitalism and me getting older and more sceptical.

Good point on the systemic forces. Yet I feel your second point falls broadly in with what I meant with "choosing the right places". E.g., I also use YouTube, but don't watch shorts, avoid clickbait, ragebait, etc, and filter ads. That makes the place OK for me. But as you say, it's then ok despite the structure of the platform.

"Online" is not one place. And what the author is complaining about is not happening in the places I spend 99% of my online time on.

That's just the saying as I know it. Regarding the post it would be along the lines of "if you think everyone is a predator/grifter/scammer, it may be a problem with you or the places you roam"

This so exaggerated. If you think some of the people you know are assholes, that's fine. If you think all of them are assholes, you are either choosing the wrong places to be, or (more likely) the problem is with you.

Thank you, very interesting to read. I haven't considered batch sizes at all up to now. Seems I should look into it.

How did you get the prediction that good? The default spec-draft-max-n is 3 and if I go beyond, the output rate stays the same or gets lower any magic spell you could share?

Have you tried a mixture of experts model? Dense models have been quite slow for me, as I have only 6 GB VRAM. But with llama.cpp and --cpu-moe I get 200 t/s input and almost 30 t/s output with Gemma 4 26B A3B, which feels ok to use. Would be interested about your mileage there.

Not OP, but it seems to be the one most likely not following the silicon valley playbook (aka the road to enshitification).

I'm impressed, because it is so minimalistic.

On the other hand it reminds me of 2010ish robotic demo videos from academic researchers and willow garage.

Problem with robotics is usually that it's easy to produce a good-enough looking demo, but it's really hard to make somethin work in the general case. As exemplified by autonomous driving.

> completely unrelated

They are both chess-inspired and roguelike.

IMHO this is a completely normal "hey I've been doing something similar" response. Why are you so aggressive about it?

What's the difference of a "robot game engine" to a simulator like <https://gazebosim.org>

I work in robotics and with quaternions (mainly 6DoF SLAM and used to do robot arm kinematics), but I don't get the use case for this. Maybe provide some example use cases?

> most people know how to use their own money to further their own life if given capital and opportunity, thus capitalism is the solution

I didn't know that "giving capital and opportunity" to people is inherent to capitalism. My understanding of capitalism is, that if people have money, they can "use their own money to further their own life", but capitalists avoid giving money to other people unless the return on investment is greater than 1.

Very cool. Reminds me of 1D Pacman: https://abagames.itch.io/paku-paku

Where is that? I found "Return the simplest working solution. No over-engineering." which sounds more like the simplest fix.

Freezing assets is simple. Seizing them is a huge pain. The EU has a hard time agreeing how to do it and who takes the liability for the Russian claims.

Woa, that's a weird @ in the screenshot in line 6.

Can you elaborate on this? My guess would be, that because of their status as a government backed research institute, they invent a lot, but let others do the commercialisation. So patent fees seem like a natural choice for them, to recover their investments.

It does sound interesting and I signed up for the wait list. But I actually don't like pure chronological order. It feels like I have to look at everything to find the good stuff. Here I probably won't miss out too much if I check the front page once a day.