It’s absolutely insane that anyone pretends Google Play and the App Store are fine though.

Has anyone scrolled through any search and not seen a sea of heavily marketed scam apps?

1.7% is still a HUGE drop in that context. This is a sector, at a time, that does not like subscriber losses.

Yes, stuff like this works. It’s why they worry about MAGA controversy too, unfortunately.

Yeah at some point folks get too old to change. You can try to get them off Fox News at least, but it just depends. Sometimes you can’t. Sometimes they really can’t even learn a new kind of remote/cable or whatever.

Younger folks are deep down the influencer rabbit hole, and I don’t know what to do about that (other than pushing Fediverse) since Big Tech has such an iron grip on everything. I mean, the future is definitely cyberpunk.

This is not how some Trump voters I know think at all; it’s not how they think of Trump. Bring up any of these points and they’d think you’re totally paranoid.

I witnessed a group of old guys watching the movie ‘One Battle After Another’ scratch their heads and have literally, honestly no idea what it was about.

It’s because they’re in the habit of watching Fox News or other ‘old’ media. Trump’s just some Republican like they’ve always voted for with a bunch of hysterical finger pointers, and any Democrat is basically Jimmy Carter or Hillary Clinton.

Thing is, Play Store is already filled with malware or near-malware from seemingly verified developers. I ran into several scam clone apps just today. It’s even snuck in through OEM apps.

Same on iOS, which supposedly verifies devs.

If ‘verification’ and curation is their idea of security, well… It appears their system is already overloaded, yet they want to expand it?

That’s the issue, isn’t it?

I see this on the internet a lot. People posit things like “wouldn’t it be awesome if these fired devs got together” or “Why don’t they make good stuff anymore? Wouldn’t it be great if somone made a thing like this old beloved thing…”

…Except it’s already happening. Or happened.

And there’s just so much noise on the internet, it’s largely unknown to the folks who’d be interested.

To be clear, I’m not blaming OP, and I’ve done the exact same thing myself. But I still find it kind of… sad.

Anyway, thanks, I am bookmarking Exodus and Archetype Entertainment now.

This makes me think of the Ellisons buying Star Trek and Avatar. Why wouldn’t they shutter or castrate two notoriously ‘woke’, expensive, questionably profitable franchises?

Same here :(. Though to be fair, the Saudi’s political leanings aren’t a perfect parallel.

deleted by creator

Does the reality even matter anymore?

Pop quiz: can you guys remember the last 20 major US shootings?

I can’t. I certainly don’t know their outcomes.

Folks will internalize the first wild rumor they heard (indirectly) from Twitter or CNN or whatever before attention moves to the next shooting, and that’s the impression that lasts.

Yeah. There’s domestic pressure for this anyway, unfortunately.

Gundam

The physics of the mechs (from my very sparse knowledge of Gundam) are pretty questionable, lol, which is fine because they’re there to be spectacular.

…Or are they an LLM? I mean, the handle is BroBot, and the emojii makes me suspicious, lol.

Yeah, the APIs are super cheap. It doesn’t make a ton of sense unless you already have the GPU lying around.

With the right settings, GLM will actually work fine in 16GB, 12GB, or even 11GB VRAM + 128GB RAM. I can even make a custom quant if you want, since I already got that set up. 24 GB just gives it a bit of ‘breathing room’ for longer context and relaxed quantization for the dense parts of the model.

GLM Air will work on basically any modernish Nvidia GPU + like 26GB of free RAM. Its dense part is really small.

But to be clear, you have to get into the weeds to run them efficiently this way. There’s no simple ollama run here.

Deepseek is only bad via the chat app, and whatever prefilter (or finetune?) they censor it with.

The model itself (via API or run locally) isn’t too bad, especially with a system prompt or completion syntax to squash refusals. Obviously there are CCP mandated gaps (which you can just add in via context), but it’s not as tankie as you’d think.

A lot of the engagement around him was absolutely manufactured and bot driven, kinda like the Cracker Barrel thing.

With sparse attention, very interesting. It seems GQA is a thing of the past.

I especially love Deepseek’s ‘public research’ aspect: they trained this and Terminus the same way, so the attention schemes are (more-or-less) directly comparable. That’s awesome.

GLM 4.6 is reportedly about to drop too. Which is great, as 4.5 is without a doubt my daily driver now.

Prompt formatting (and the system prompt) is a huge thing, especially with models trained for ‘tool use’ a specific way, so be sure to keep that in mind. For example, if you want a long chain of steps, be sure to explicitly ask (though Qwen is uses its thinking block quite gratuitously).

I find GLM 4.5’s default formatting to be really good though: be sure to give that a shot. It’s also awesome because the full 350B model (with some degredation) is locally runnable on a 128GB RAM + 24GB VRAM gaming rig, and the ‘Air’ version is quite fast and accurate on lesser hardware.

Local hosting, if you can swing it, is particularly nice because the calls are literally free, and promt ingestion is cached, so you can batch them and spam the heck out of them for testing and such.

That is a model that has been vetted that it won’t lead to immediate law suits because it is just Tom Cruise or that can be programmed to do what is needed. Sometimes you want a model that is like Keanu and became a gun nut after The Matrix and shoots 3 gun in his spare time. And sometimes you want that Korean Boy Band actor who is going to flinch every time a blank goes off because that is the character.

We are talking about a completely different ecosystem here.

These generative ‘models’ either fall into the buckets of:

• A very small basket of completely closed, relatively inflexible corporate APIs.

• Or a still-small basket of open models folks build these skeletal frameworks around, or maybe loras or adapters.

All these AI startups like to pretend they’re doing something special when, underneath, they’re really just prompting ChatGPT with a wrapper, or hosting a Flux finetune or whatever.

In other words, they are NOT pretraining Thom Cruz from scratch. The pool of usable frontier models is very small.

And… it won’t end well if Aki Ross 2.0’s underlying model was also used to train Project Melodee 2.0 and people realize her face when she is eating that cake that represents the last vestiges of her innocence is REALLY similar to the face that Melodee make when she is doing DVDA.

At which point the agency helps to avoid situations where “Well… you are contracted and mr tarantino wants you to drive that car while he jacks it to your feet. And you wouldn’t want to be problematic, would you?”.

Again, you’re treating these ‘models’ like a diverse group of humanity, and like every company’s training from scratch, when that’s not how the software’s set up.

It makes no economic sense to treat them like people with their associated complications. They’re software suites, they’re tools, more like different flavors Davinci Resolve or whatever studios use these days, that can each produce an infinite spectrum of humans depictions, basically for free. A closer analog would be video game development, with the cost of voice acting and animation stripped out; the only thing that makes The Master Chief, Commander Shepard, or a particular incarnation of Lara Croft ‘unique’ is the copyright, recognition, and software suite they built them into.

EDIT:

To add to this, I think its extremely dangerous and unhealthy to anthropomorphize them.

In fact, this might be what the agencies are trying to do. ‘Humanizing’ them like theyre individual, sentient things makes them appear less like Lara Croft selling a Snickers bar. It may be optics for the customers (like ad makers hiring actors/actresses) more than anything.

Hard, relativistic STL sci fi can still get super weird, see: https://www.orionsarm.com/eg-article/48545a0f6352a

To add to this, Jason Schreier is a well known, and well sourced, gaming journalist.

https://en.wikipedia.org/wiki/Jason_Schreier

But you aren’t wrong. There’s no way to know that via Bluesky unless you’ve happen to read his stuff from Bloomberg and before. It’s almost like Twitter is a terrible format for news or something…