Mod parent funny. This was the joke I was looking for, but Slashdot only reports the top mod. (Though this one currently has no funny mods.)

New idea about how to fix the moderation on Slashdot! Use AI. Can't be worse than the current moderation system, but AI is so trendy that maybe the latest owners (whoever they are, or perhaps incorrectly assuming someone still owns this mess) will eagerly embrace the AI.

So the idea would be for an AImod (AI moderator) to scan the posts and suggest the most likely moderation. It could even be done as part of the preview process to give posters 'objective feedback' about how their comment sounds. Perhaps even including candidate locations where a comment might fit better into the discussion and an easy move option? But the AI mod points should be easily countered by humans (which becomes training feedback for the AImod).

Another obvious joke would have been from the "What could possibly go wrong?" angle.

My actual feelings are mixed, but that got me to think of another possible angle for humor. The problem is that I don't know much about the current or even recent generations of video games, so I can't name the names of the games NOT to recruit from. I'm pretty sure there are games that involve lots of mayhem and Captain Chaos, Master of Mayhem, is NOT someone I want in charge of anything involving a plane (or bus or bicycle) that I might be riding.

Meta-angle involves the recurring brain worm in the form of that song about "Where have all the <category of people> gone? Long time passing..." In the case of Slashdot, it's the humorists of yore. Hopefully they've found a website with a humor-encouraging moderation system and all I need is that URL to make me happy. Or at least more amused?

1) 30B, dense. Optimized to fit Q4 quantized in a 24GB card with a speculative decoding model as well.

2) Because reporters don't know what weights are and assume you don't know either.

3) Way better than Gemma 4 on text tasks, slightly better on multimodal. Numbers below are all: Benchmark: Muse Glimmer score Gemma 4 31B score difference

Artificial Analysis Intelligence Index: Muse Glimmer: 35 30 +5
MCP Atlas (Public): 75.5 54.2 +21.3
DeepSearch QA: 74.6 61.7 +12.9
SWE-Bench Pro: 51.2 36.9 +14.3
SWE-Bench Verified: 76.0 66.6 +9.4
OSWorld-Verified: 65.9 58.5 +7.4
GAIA2: 43.3 36.4 +6.9
WildClawBench: 47.6 37.6 +10.0
TerminalBench 2.1: 51.7 43.4 +8.3
Tau3-Banking: 23.5 15.1 +8.4
MMMU Pro: 74.0 73.0 +1.0
Charxiv Reasoning: 78.8 77.7 +1.1
OmniDocBench v1.5: 75.8 72.5 +3.3
ScreenSpot Pro: 75.4 75.9 -0.4

4) 128k tokens

5) No, sadly.

What you want is a DGX Spark.

It costs $4000.

And yes, there are fundamental advantages to cloud services, such as large-scale batching, little idle downtime, high speed, and hardware optimized to the specific models / serving needs. That said, one can weigh that off against sovereign control over your server...

Define "tolerable".

The model in question - Muse Glimmer - with DFlash/speculative decoding - will probably get you ~60 to 124 tok/s on a 3090 (a quite dated GPU). For a 5090, it's said to clock in at 233,4 tok/s. On CPU you're looking at maybe 3-5 tok/s.

If you call that "tolerable", I guess you're more patient than me? And as mentioned, you're not just wasting time, but also wasting a lot of power too - CPU is a very power-inefficient way to run ML models.

If you insist on CPU, this isn't the right kind of model anyway. You want to take advantage of the fact that you probably have lots of (comparably cheap) RAM, and compensate for the fact that you have (comparably) terrible memory bandwidth, and for that, you want a MoE with a high total parameters but a low active parameters. Not a dense model like this.

Then I was basically aiming at Mac Minis

That's very much a special case which you didn't mention in your post that I responded to, but still the answer is "meh". You couldn't run it at all on a 16GB Mac Mini, and I think you'd struggle to run it at all on a 24GB (because you have to share the ram with the OS, the inference server, etc). For the base Mini you might get 10-12 tok/s, and for the M2 Pro / M4 Pro, maybe 25-35 tok/s. Still pretty far from a GPU, though.

This model is designed for >= 24GB GPUs.

MoEs don't save RAM (for a given quality), they increase it. You have to store all of the parameters in memory, not just the active params. But inference only uses a subset of the total params for each token, so it reduces the memory bandwidth requirements and improves token generation rate. But this comes at the cost of a higher total param count for a given quality.

Meta has always been releasing open models (Llama was famously the first powerful open model). The change has actually been in the opposite direction, with Muse Spark *not* being open.

Zuck descrbed his motivation way back when, about how they got burned with Facebook on app stores, in that Apple and Google could basically bully them however they wanted, on whatever extractive terms they wanted, and there was nothing Meta could do about it. He's now paranoid about "others controlling the platform", and wanted to make sure that doesn't happen with AI, that they have their own AI base to work with.

I mean, if you want your model to be terribly slow and terribly power inefficient, sure, go ahead and run it on CPU.

This really isn't a good model for CPU. For CPU, you want a MoE with a large number of total params but a tiny number of active params. Something like DeepSeek V4 Flash 0731 if you have at least 128GB of RAM - you might get 2-3 tok/s or so on that. The goal is to minimize the memory bandwidth requirements per token, at the cost of a greater total RAM footprint.

For GPU, you're highly VRAM limited but not bandwidth limited, so your best option is generally a dense model (non-MoE) with speculative decoding to make up for the performance limitations.

You can use it fine on 24GB with a good size context with KV cache compression. They've also designed it to use speculative decoding to boost performance instead of a MoE, so that they can pack as much capability as possible into a dense model. Seems like a winner.

You ask so rudely that my initial reaction was to ignore you, but you piqued my curiosity. Where did you see it? I thought it was just one of my personal peccadilloes. Or you're just referring to my comments here and possibly elsewhere?

The YOB loves his precious brand, so I strongly prefer not to use it. Makes me glad I'm only an amateur historian. In describing the reign of YOB I'm pert' shure they'll have to brand it.

YOB comes from YUGE Orange Buffoon. How many guesses do you want? Or you might prefer a different source word for the B?

I have noticed that there are many other euphemisms applied to the YOB and all things YOBbish. Context almost always suffices to clarify the reference.

Rather disappointed by the discussion so far, which saps the motivation to say anything. I think there's a deep problem in Japanese thinking, but the most clear evidence is probably the Reiwa thing. The Heisei emperor asked to retire, but there was no rush. He's still alive now so they had plenty of wiggle room about when to start Reiwa--but the LDP deliberately chose to inflict imperial stupidity on everyone when they had the easy option of syncing the start of Reiwa so the digits would match up. Insane decision for perpetual friction for the sake of the supremacy of the imperial system. Same reasoning that led to that little kerfuffle in and around the Pacific...

Rational heads might prevail? Feel the Chinese wind? Or maybe it will blow from India before the next decisions need to be made? But preparing for a nuclear war with China is not a rational solution approach. Or maybe the real target is Switzerland? That would make a better joke.

My bad. I forgot that Microsoft had been mentioned. More evidence of senility? Or I just can't take Microsoft seriously based on my experiments and experiences with Copilot? So let IBM hold the pot for old times sake?

That was the best/only joke Slashdot can come up with these years? Sadness.

I was hoping to see something about AI-powered FinTech counter-wagers to protect each other. Sort of a corporate cancer circle jerk.

Oh how about a grand scheme to get the money back based on wagers by human suckers as to which AI will win the race? They could let Microsoft hold the pot since they apparently got left out of the race. My money would be on Oracle for most secretive and ruthless and evil and therefore the most likely winner, but "secretive" makes it easy to make a bad bet...

On the insightful side, the "Why?" seems to be getting ignored. I actually think they are running scared. Not sure if they are most scared of each other or of the Chinese threats that didn't get mentioned, but I am reminded of one of the most famous Asimov stories. It was about a constructive question: "Can entropy be reversed?" As I remember the story, the super-AI answered "Let there be light." However right now I think we are facing extinction via a destructive question with a fatal answer. Back on the gambling scenario, I think the two most likely winning/losing questions that will lead to the fatal unleashing will be "Can you get all the money?" or "Can you destroy all the competing AIs?" (I do think we have a bit of breathing room because the robots aren't quite up snuff. As in up to snuffing us?)

Back to the Chinese side (but with goog flavoring)... Just started reading AI 2041 by Kai-fu Lee and Chen Qiufan. Quite an interesting idea of mixing fictional AI application scenarios with more technical explanations and commentary. Obviously hoping for some kind of synergy, but so far I'm not seeing much. Maybe the stories were better in the original Chinese? Or perhaps the Asimov bar is just too high?