site banner
PaperclipPerfector  3mo ago (text post)   25042 thread views

Culture War Roundup for the week of September 29, 2025

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

7
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

AGI Was Never Going To Kill Us Because Suicide Happens At The End of Doomscrolling

I'll go ahead and call this the peak of AI version one-dot-oh

The headline reads "OpenAI Is Preparing to Launch a Social App for AI-Generated Videos." People will, I guess, be able to share AI generated videos with their friends (and who doesn't have THE ALGO as a friend). Awesome. This is also on the heels of the introduction of live ads within OpenAI's ChatGPT.

Some of us were waiting for The Matrix. I know I've always wanted to learn Kung Fu. Others were sharpening our pointing sticks so that when the paperclip machine came, we'd be ready. Most of us just want to look forward to spending a quiet evening with AI Waifu before we initiate her kink.exe module.

But we'll never get there. Because Silicon Valley just can't help itself. Hockey sticks and rocketships. Series E-F-G. If I can just get 5 million more Americans addicted to my app, I can buy a new yacht made completely out of bitcoin.

I am a daily "AI" user and I still have very high hopes. My current operating theory is that a combination of whatever the MCP protocol eventually settles into plus agents trading some sort of crypto or stable coin will create a kind of autonomous, goal-seek driven economy. It will be sandboxed but with (semi) real money. I don't think we, humans, will use it to actually drive the global economy, but as a kind of just-over-the-horizon global prediction market. Think of it as a way for us to have seen 2008 coming in 2006. I also was looking forward to a team of maybe 10 people making a legit billion dollar company and this paving the way for groups of 3 - 5 friends running thousands of $10 + $50 million dollar companies. No more corporate grind if you're willing to take a little risk and team up with some people you work well with. No bullshit VC games - just ship the damn thing.

And I think these things are still possible, but I also, now, think the pure consumer backlash to this silicon valley lobotomy of AI could be very much Dot-Com-2-point-O. The normies at my watering hole are making jokes about AI slop. Instead of "lol I doomscrolled into 3 am again" people are swapping stories about popping in old DVDs so that they can escape the ads and the subscription fatigue.

Culturally, this could be great. Maybe the damn kids will go outside and touch some grass. In terms of advancing the frontier of human-digital knowledge, it seems like we're going to trade it in early not even for unlimited weird porn, but for pink haired anime cat videos that my aunt likes.

To what extent is it or will it become possible or practical to run a homebrew jailbroken LLM on local hardware? That's the big question in my mind.

I'm late to the party, and I'm aware of it, in that I'm only just now using LLMs beyond a toy for research and education purposes. But essentially every day I'm aware there's an expiration date, that the product is just a few bad days for the SP500 from being enshittified. Whether that comes in the form of censorship and legal caution that makes it useless for my purposes, or in the form of pricing that makes it prohibitive, or commercialization and monetization in ways that make it unreliable (pay extra for your product to be recommended!), or optimization for it as people start to operate their products specifically to be seen and understood by LLMs. There's going to come a time when I can't just log into ChatGPT and get a good result, I'm sure the old timers are already complaining; and there's going to come a time when there isn't enough VC money sloshing around to fund a competitor like Grok that throws off shackles.

So at that point, can I or will I be able to operate a homebrew LLM for my personal and business purposes? I'm not handy enough to know how possible that currently is, or how user friendly, I'm at the level of "I can run a Linux machine but I'll need to look stuff up once a week or so."

I strongly endorse what @erwgv3g34 says below. You can, in theory, run a model far superior to GPT-4, and not that far from the SOTA, all on consumer hardware.

Of course, as he correctly points out, it's going to be expensive to host it on your personal hardware. Somewhere between used car and new car expensive, but people can and do buy cars. But not that difficult, if you're capable of following instructions. He's right that it makes more sense to just rent an H200 as and when needed, a while back I saw them going for below $2/h.

If you have an archived copy of a decent model, especially one fine tuned to remove censorship, there's little that can be done outside of totalitarian legal action to stop you from using it. That's far less likely than potential enshittification or censorship online.

enough VC money sloshing around to fund a competitor like Grok that throws off shackle

Elon is a stubborn mf, and supremely wealthy. xAI is probably one of the companies most resilient to VC panic. Or look at DeepSeek, which is owned by a net profitable quant firm. They'd be relatively safe options to out last a market downturn.

The idea of Google being enshittified was incomprehensible when I was young. Clearly it isn't so anymore.

The strongest open source models aren't that far behind the strongest proprietary models; a year or two for LLMs, six months for text to image. You can see several open source models in the LMArena top 20, such as Qwen3, DeepSeek R1, Kimi K2, and GLM 4.5.

Problem is, those models are huge. Qwen3 is 235B, R1 is 685B, K2 is 1T, and GLM 4.5 is 358B. It would cost a fortune to get enough GPUs to have enough VRAM to run such locally.

Your best bet is to rent GPU time from an online company and run the models there.

I think that one aspect is the question which performance you actually require from the model.

A fundamental difference between free / open source software and open weight models is that for software, the bottleneck is mostly developer hours, while for models, it is computing power on highly specialized machines.

For software, there have been large fields of application where the best available options are open source, and that has been the case for decades -- for example, try even finding a browser whose engine is proprietary, these days. (Of course, there are also large fields where the best options are all proprietary, because no company considered it strategically important to have open source software, nor was it a fun project for nerds to play with, e.g. ERP software or video game engines.)

For LLMs, tens of billions of dollars worth of computing power have to be sacrificed to summon more powerful shoggoths forth from the void. For the most part, the business model of the AI companies which produce the most advanced models seems to be to sell access to it. If Llama or DeepSeek had happened to be more advanced than OpenAI's models, their owners would not have published their weights but charged for access. (The one company I can imagine funding large open-weight models would be Nvidia, as part of a commodize your complement strategy. But as long as no AI company manages to dominate the market, it is likely more lucrative to sell hardware to the various competitors than to try to run it yourself in the hope of enticing people to spend more on hardware than on model access instead.)

That being said, for a lot of applications there is little gain from running a cutting edge model. I may be nerdier than most, but even I would not care too much what fraction of IMO problems an AI girlfriend could solve.

Unfortunately, Nvidia sucks ass at making LLMs. Nemotron was a joke.

Mistral Nemo was great though.

Nemo? I know No One by that name.

Haven't tried it, unfortunately. I think it's still a poor showing that they needed another company to finish their work for them.

The trick's that the same chips used to produce a model are also usable to run the model for someone else, and a lot of the technologies used to improve training has downstream benefits on inference or implementation improvements. Every AI vendor has its own complement to turn into a commodity.

Don't try to do your own local setup first.

There are host-your-own solutions; featherless, together (my choice), runpod, openrouter.

These all have privacy policies that are far better than the Big providers, but you are, fundamentally, sending your prompt (and data) to another computer to have it processed. There can be leaks, there can be man in the middles etc. Still, you aren't literally being used as a guinea pig like you are with OpenAI, Claude, Gemini etc.

For really serious / personal stuff, the answer is "wait." This is such an obvious market need that I am 100% confident we're going to see LLM-specific personal hardware at the $1500 or below price point in the next few years.

I'll look into that, thanks!

But philosophically, it's less personal interest that I'm talking about than it is a sort of market competition threat against enshittification. As long as its possible to homebrew at some level of effectiveness, OpenAI won't be able to completely ruin their product and still market it. Social networks enshittify because I can't make my own at home. In my mind I'd like to set up a local LLM more to learn the process for the inevitable future where I can't trust the commercial variants that have been enshittified. Inasmuch as I find LLMs useful, I want to have full control over it to protect myself from losing a tool that I find useful because it is under someone else's control. Hosting my own might be the intermediate step...

Yeah, if that's your goal, 100% go for a paid-to-host solution. I like TogetherAI because you can easily pull from a lot of stuff on HuggingFace and it's all pay-as-you-go. $25 will last quite a long time if you're in pure chat mode. If you're using an API to sling code at it, $25 will evaporate quickly.

There will be no fundamental enshitiffication. The crucial IP feature of LLMs is that their architecture is pretty easy to grok for anyone with a basic ML background. The cost comes in 1) Collecting and preparing training data and 2) training the models - especially big ones - at scale. There's not really a lot of secret sauce in the model itself.

The secret sauce, to the extent it is real, is what happens during inference time. This can be system prompts or other intermediate prompts that are both visable and not to the user. We also know that all of the Big AI firms are now using multiple models at once to "route" different parts of the user query. I also heavily suspect that there's a middle layer that does some sort of context management to create a proto "memory." What do you need to build a system like this? The same thing you need for any software system; a bunch of talented engineers with a defined vision for the product and some coordination overhead. That is difficult to replicate.

The question is how much does it matter? We're going to be able to run open source versions of very good LLMs on our phones one day (and, maybe, one day, have actually private phones!). Will those private LLMs be so much worse in terms of performance than the Big AI system-of-systems in place? Hard to say. They're making $100 billion bets on it right now.

Main thing about running LLMs locally is that GPU VRAM is probably the limiting factor in most cases. 3090/4090 with 24GB of VRAM or 5090 with 32GB are OK, but for hobbyists really into it, they've made custom GPUs like soldering 24GB more VRAM to a 3090, or using dual-3090 for 48GB of VRAM (spreading it out over multiple GPUs works for LLMs). The speed differences matter, of course, but 3090 with 48GB of VRAM will likely enable more than 5090 with 32GB due to being able to fit bigger models.

It's surprisingly easy if you already have a good GPU: https://ollama.com/

The real question is whether frontier models remain proprietary for the foreseeable future or if there's really just no moat and everyone will be able to run open models on consumer hardware instead of having to pay for data center compute.

My gut says decentralization is likely because I'm pattern matching this on the rise and not-quite fall of IBM, but I'm biased.

You can run a homebrew LLM (7 billion parameters / 12bn / even 24bn) for nothing on any decent PC with a GPU. It will be lucid but really pretty dim.

You can rent a RunPod server pay-as-you-go and run a 70bn / 105bn / 200bn model for a few dollars an hour. It will be smarter but not quite GPT / Claude level. You can also pay 25 USD a month for Featherless, which is the same thing but less under your control.

Or you pay for the APIs.

I've run a few 4-byte quantized 70B models on a small home gaming machine pretty easily (Intel i3-13100, nVidia 3060, 48GB RAM). It's a little slow -- non-MoE models can go into a couple tokens-per-second, and MoE seldom go higher than 10 tps -- but there are some set-and-forget use cases where the difference isn't a big deal, and you're just a couple GPU generations away from it going faster.

Both ollama and lmstudio work pretty easy 'out-of-the-box'. You can dive down the deep end if you want, and start moving to vllm or others, but it's far from necessary for most use cases.

Scaling up without waiting can get expensive, though. Used server GPUs aren't ludicrously expensive and buy you more RAM (and thus more context/bigger models), but they're slower than current-gen (or even two-gens-old) gaming cards. Trying to break past 24GB VRAM gets into the kilobucks range, and while nVidia says that they're dropping a card that will change that in a few months, it'll probably be seconds before it get scalped. For LLMs, processing power is lower priority than total memory bandwidth, so you can get away with some goofy options like the Ryzen Max series and run 128 GB ""VRAM"" with a CPU, but setup is more annoying and throughput suffers a lot, and it's still not cheap.

I have a Ryzen Max 395+ with 128GB RAM and it runs pretty well; granted I don't use it for LLMs but the humongous amount of RAM is useful more often than one might think.

From what I've read on HN and Twitter, the Ryzen Maxs can run larger LLMs, but not very fast. The throughput for tokens/s is single digits at times.

God damn man! What do you use it for if not llms?

Prototyping things which I really should be pushing off to a cluster for computation but I can't be bothered with doing the SSHing. When it comes to prototyping not needing to do the extra steps of moving my latest version of the code over to the cluster saves me around 10-15 seconds for each iteration, which is enough of an annoyance I'm happy to pay to avoid it. My machine is the HP laptop that comes with AI MAX 395 so it's my main personal device. Not needing to worry at all about RAM management for my own code has been surprisingly freeing.

That and of course playing DOTA with all settings set to maximum, it's amazing how I can get Desktop quality DOTA performance on my laptop today.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats