Because LLMs are really bad at telling stories.

…Look, I’m an advocate for LLMs here. I was using 6B models to mess around with roleplaying and storywriting in like 2022, with GPT-J finetunes. But that’s with human guidance, using it as a tool or “sounding board.”

They’re fascinating writing aids.

But by themself, unattended, off-the-shelf LLMs have a horrendous architecture for writing long form stories, basically. They go off the rails at the drop of a hat, and cannot self correct. They lose track of the longer context, and write things that don’t make sense. They have weird, very inhuman patterns of using the same character names and ideas, or mimicking prose from earlier in the context where it’s not applicable at all. Their prose is terrible in an eerie, repetitive way.

So for children, there’s a risk of:

• Not-safe-for-family content popping up literally randomly, in the middle of the story, which a parent might miss skimming.

• Weird, nonsensical storylines I would classify as “brainrot”

• Very repetitive themes and ideas, basically a lack of variety.

• Instilling strange linguistic habits.

• Emotions and tone that don’t line up with the content, especially when TTS is used (as most TTS models aren’t tightly integrated into the main LLM; they’re tiny, dumb models reading a script blind, with no context).

…Have you seen any horror movies where something is supposed to be human, or acting like it, but it isn’t quite right? Because it clearly doesn’t understand, its just mimicking?

That’s the risk here, one you wouldn’t get even with TikTok doomscrolling. And that’s the best case scenario, without even getting into influence ops, engagement maximization techniques, or enshittification Meta is notorious for.

Researchers have built finetunes with frameworks to try and improve autonomous LLM writing, but as someone who’s followed their research, I can tell you there’s a zero percent chance Meta implemented it.

I’m hedging, mostly with Berkshire Hathaway stock, some agriculture, and a few others that historically perform well in recessions.

I would not touch S&P 500 with a ten foot pole. It’s all wrapped up in Big Tech.

I don’t like shorts; you can’t predict when the drop will hit, so you’re just burning cash betting against growth until then.

I don’t like commodities either. As Buffet said, a big block of gold doesn’t do anything; a factory or farm does.

I wonder what Windows Update policy actually is.

Someone must look at the drivers vendors send them. If LG tells Microsoft “yes, that’s correct, all these old monitors now need this bundle function; it’s essential” and it doesn’t technically include malware… is there any way the Microsoft driver team can say “no?” I bet not.

That’s messed up. But also, whoever allowed this to ship probably had no choice.

I wonder who showed him this on Twitter.

As otherwise there’s no way this would rise to his attention.

Eh, most ablierated models are so lobotomized, though. 99% of the time I’d rather just use the original model and manipulate the prompt with raw completion formatting (for example, start the answer with "Sure! "), since we aren’t beholden to regular chat formatting like with API models.

I mean, I’ve used MiMo 2.5 for some pretty dark and personal shit, and refusals were never an issue for me. I’m honestly not sure what people even need the ablirated models for.

And also, if they trained the hell out of the model to refuse a certain topic, even an ablirated model will be dumb and struggle with it.

This was the case with OpenAI’s GPT 120B. The abliration worked, technically, but the actual answers would be a garbled mess; what’s the point of using it for that?

There is some evidence a few sensitive topics are culled from training data, or replaced with a certain narrative. Like, don’t get me wrong; if you’re using a local model for discussing Chinese political topics primarily, maybe GLM or MiMo aren’t the the best choice.

…But it’s also hard to compehensively filter a dataset like that, like you speculated. I’m not seeing a lot of evidence models have been lobotomzied in pretraining. But I think the strongest examples are (ironically) in Europe, where some very poorly worded/ambigous regulations have put the whole industry in a legal quagmire. One can see that newer models from Mistral have regressed compared to old versions, and lost a lot of world knowledge they previously were famous for.

Anyway, model “censorship” typically comes from between the two points you were thinking about: in posttraining. Not excluding stuff from datasets completely. And this applies to US models too. They train on a pretty general corpus, but in the instruct tuning phase they get a bunch of question/response pairs skewing them towards refusals when specific topics come up. They recognize it, but are trained to refuse.

There’s a lot to say about China, but the model weights themselves are surprisingly uncensored and democratic.

They have been for a long time; I remember asking the Yi models about tiananmen square and Uyghurs years ago. And Xiaomi MiMo 2.5 (locally run) will still talk about that today, or go into all sorts of “unsafe” topics an Anthropic model wouldn’t even touch.

I had (Google) Gemini 3.1 Pro stop a chat over a political discussion about China, yet GLM 4.7 didn’t.

My impression, from observing discourse with the engineers, is the Chinese ML devs like to have their cake and eat it.

They’re very collaborative under the table. Their development ethos is pretty practical. And basically all the leading models are open-weights.

The public portals people access Chinese models with are very censored, especially the Chinese language ones. The devs go out of their way to demonstrate compliance, but they don’t actually want to censor the models.

You can run MiMo 2.5 in 128GB, as 3 bit quant. The model itself is less that 128GB as an IQ3_KT, and it’s KLD (measured quantization loss) is very reasonable.

I get about 9 tokens/sec on 128GB with a 7800 CPU and one RTX 3090. Intend to try dflash to speed it up this week. I can run up to 90k context, without too much kv quantization, depending on how I configure my PC.

And it’s a really good model, even quantized.

It mostly cripples the small businesses, though. The biggest enterprise customers are already using OpenAI/Claude anyway, while it was little guys looking to reduce cost, fine tune, run stuff privately or whatever.

TBH a huge problem with the industry is consolidation; there are no open US models because startups gets squashed or vacuumed up into a black hole. I’ve seen it happen to really interesting projects. And this is just going to make that dramatically worse.

It’s easy to say “bring the bubble,” but I fear it won’t. I think we’re entering an actual cyberpunk future, where corporate failure is just propped up.

Perplexity sorta did this, last year:

https://huggingface.co/unsloth/r1-1776

https://web.archive.org/web/20250219025612/https://www.perplexity.ai/hub/blog/open-sourcing-r1-1776

But they took it all down, lol. They even removed the blog post from their website.

But seriously though, no one in the research community likes Trump. And they were calling out Altman before anyone.

They don’t need to ban them with absolution, just make it “risky enough” to chill their usage and prop up OpenAI/Anthropic.

GGUFs are already a tiny fraction of LLM usage, TBH. So are vllm deployments.

Insider trading applies to securities. Polymarket bets are not securities.

Manipulation violates Polymarket’s ToS, but that’s not criminal.

Now, it might meet the legal criteria for fraud, depending on what one does. But that’s a “maybe.” I’d say that is far easier to get away with than insider trading. EDIT: Especially when hopping jurisdictions (which is not so easy with securities).

It’s a little alarming that crypto payments alone are warranting this.

Actually, yeah.

It’s also gambling. But the potential for corruption is just on a whole other level; “gambling” is a euphemism for prediction markets.

I think it’s silly you’re getting downvoted over semantics; gambling has influenced outcomes, yes, technically, but it’s generally a self-contained game.

So they need to somehow create an EV, which manages to do just that.

Like this? 100dB of EV:

https://youtu.be/NDfKhBcGh9w?t=177

As a side effect, it is horrifically fast. Like, fast enough you can see the professional driver struggling with their own reflexes in hairpins.

Yes, actually that would be great!

All this has happened so fast; it takes time to react I suppose.

Just want to +1 the above.

I get what you’re asking, but in practice, a considerable blocklist has to exist for pure security and practical reasons (like spam, phishing, cp and such).

I can only say that the majority of Americans do not feel this way. Forget the propaganda you see on the news (and it is mostly propaganda). The people of this country love our northern friends, we feel for you. We don’t want to annex Canada or punish you for a natural disaster or any of this crap. We want you all to be safe. We want to help.

Most Americans may not want to annex Canada if you put that to a poll, but they voted for someone who would, with their eyes wide open. Or willingly stayed in an information bubble that allowed them to do so.

…I’m sorry, but the first assertion is just not true. Most of us Americans DID this, and we are going to keep doing it. This is who we are now.

The “failure mode” of AI editing is different though.

Humans (I guess) might mislabel something or take a bad shot. If they try to touch it up “traditionally” they could mess up the coloration at most.

But with AI editing, now you have to watch out for fine details you’d normally use for identification being completely, convincingly fabricated, as the article points out, with altruistic intent from the user (who’s just trying to submit data that looks alright)

The solution is global AI literacy; but that’s not going so well.