site banner
PaperclipPerfector  7d ago (text post)   31921 thread views

Culture War Roundup for the week of March 24, 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.

4
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

Another day, another LLM

Google just launched the latest iteration of their Gemini language models. I suppose the name was most appropriate for the (short) period where the version number was precisely 2.0.

Well, everyone say hi to Gemini 2.5 Pro (Thinking). The naming scheme hasn't gotten any better, albeit Google beats the packs's abysmal average performance. It's a narrow win, folks, as there's a Gemini 2.0 Pro, a smaller and leaner 2.0 Flash, and a 2.0 Flash Lite.

(We're days away from OAI matching them with an o3-mini-high-low-too-slow.)

What stands out about this model? Nothing really. It reasons by default, which can be nice, but at the cost of increased latency for responses.

It is incrementally better on benchmarks, but even Google's PR team couldn't drum up a revolutionary new capability to showcase. They get a pass, because 2.0 Flash's image gen was revolutionary, and happened a mere week or so ago.

Gemini models have recently become the Honda Civic of LLMs. Not nearly as flashy, but reliable and with no obvious downsides. This one has seized the number one spot on LM Arena's leaderboard, based off (nominally) blinded user feedback. It might hold on to it for a week, or a month. The days when GPT-4 retained the crown for months on end are gone.

After plenty of use, all I can confidently say is that it writes better. I'm very happy with that. I'm sure someone will find a task it can do better than the rest, but I doubt it'll make anyone switch over if they're already happy. I'm confident there's something deep to be said about my inability to meaningfully differentiate models in terms of capability, be it for work or play. I'm just not going to be the one to say it today.

"Another day, another LLM" is definitely correct. Deep Seek also released their newest variant DeepSeek-V3-0324 yesterday. DeepSeek-V3-0324 is a significant improvement over DeepSeek-V3 and even beats Claude 3.7 Sonnet in many benchmarks, and not to mention, it's open weight! I guess it's less sexy since it's a text-only model and we already have highly capable ones that are generally interchangeable for most purposes, but I'm excited to see the future DeepSeek-R2 that'll be based on this improvement.

Do you know if it beats Claude 3.7 for writing? I have wasted shameful amounts of money in the last week: it gets character consistency and plot progression perfectly, even carefully adding in hooks for potential future plots. It’s the only model I’ve ever tried that gets long-form writing right.

I haven't tried Claude 3.7 for creative writing but it's definitely better than the existing DeepSeek-V3 from my limited experience with it, so feel free to test it out. It's a lot less repetitive at longer context lengths which actually makes it usable for creative exercises. The original DeepSeek-V3 was likely more optimized for multi-shot prompting for factual queries, which made it strictly follow the reasoning, tone and structure of earlier examples. Good for factual determinations but not so good for being creative and non-repetitive.

From what I can tell it is a giant improvement over V3 and a substantial one over R2. People from 4chan are saying it's suspiciously good at following their outlines and even uses a similar format, so they guess it's been trained on their writing (DeepSeek openly trains on user data). Generally it's in the same league as Sonnets. I recommend giving it a try.

Wow, that's great to hear. I'm eagerly looking forward to the commoditization of novel writing (and videogame NPC dialogue), but I didn't think we'd figured out yet how to maintain long-term consistency.

Yeah, I'm really impressed, especially since I was pretty nasty about 3.7 in another thread. There are some caveats:

  • It's EXPENSIVE. I burned through enough money for multiple AAA games in the last week.
  • I have to limit the context to 20,000 tokens because of the expense, which is less than it sounds. If a character drops out for more than three or four scenes, they may come back with a noticeably different voice. You can limit this with summarisation / manually copy-pasting character introductions into the author notes when they reappear, so a character who was shy will stay shy, but the way they're shy may change.
  • For similar reasons, the tone can drift. I especially notice a slow trend towards analytical/clinical/technical style writing appears about 30,000 words in unless you force correct it.
  • I use Guided Generations for guiding new generations. Not the clothes etc. stuff, just 'write the next scene in this way' style prompts to keep the AI on track.
  • The plot is still being made up as you go along, and won't be consistent in the way that a novel is consistent unless you work to make it so.
  • Claude 3.7 is surprisingly permissive but the limits may cause trouble for certain types of stories. I don't know if you could do misery memoirs (explicit abuse), or Game of Thrones (explicit torture, harm). As far as I can figure out, it will refuse to depict on-screen sex or other explicit actions, or serious on-screen violence/damage to a human character.

Long-term consistency in video game NPCs is probably a lot easier - each NPC gets 4000 token summary and produces max 1000 token replies. Global consistency is taken care of by the game state.

The issue for game NPCs is ensuring total consistency with the setting and characters and not making up any lore at all. Even late generation LLMs still struggle with this in hallucination (I like to ask about the political backgrounds of Chinese politicians, many of which have no content online beyond two or three generations back, and I’ve seen even latest models completely BS).

I haven't been poking at it with a stick like you have, but in general Claude is mostly content to stick to the scenario absent provocation. It's not like Deepseek R1 where reigning in the constant flow of random creativity gets exhausting and irritating after a while.

Give Claude explicit instructions about maintaining consistency and I wouldn't expect to have serious issues.

What sort of stories do you use it for?

Scenarios, mostly. I have a scenario that's basically the borrowers / Grounded (you are a tiny person in the walls, stay alive and don't be seen), a scenario that's the setting from Rosario + Vampire (you have been enrolled in a prestigious boarding school, but all the other students are monsters in disguise), etc. I also have a romance that I was working on. I also have a Japanese-teaching bot that takes place in your standard anime setting for colour and to provide conversation topics.

I was wondering whether to discuss it, but felt too lazy given that it was 1 am. Like you, I'd rather wait for R2, I always try and use reasoning models for anything complex, barring trivial answers

While we're on the topic, OpenAI just announced their 4o image model, though it's not rolled out to 100% of users yet (AKA not me yet) https://openai.com/index/introducing-4o-image-generation/ If it performs nearly as well as their cherry-picked examples, it knocks the socks off of gemini flash image gen.

Now we have Gemini, Grok, and ChatGPT all releasing image editing, a previously unheard of feature, within two weeks of each other. Interesting how that happens.

When it comes to text generation models, I can't bring myself to care anymore. They all do the basics proficiently, which is generate code snippets and console commands and trivia answers to questions that I can ask in a few sentences. And they all sort of fail on really complex stuff like doing my work for me in making changes to a 1M line codebase.

The text generation in those example images is phenomenal.

Damn, if it’s better than Gemini then it looks like you were right and they were sitting on that capability. It reaffirms my opinion that OpenAI is the most insidious AI company out there.

Why insidious?

Some unordered reasons:

  • They call themselves OpenAI and yet nothing is open. They don’t publish research or release open source or open weight models.

  • They positioned themselves as a non-profit to get clout and talent and later reneges.

  • They charge an insane amount compared to other companies to price anchor because they want the SOTA models to only be accessible by elites.

  • They hide the thinking traces from their thinking models and ban people who try to figure out their methods with prompt engineering.

And then tack on “they hold back revolutionary features until their competitor releases their own”

They positioned themselves as a non-profit to get clout and talent and later reneges.

Based on the talent banking Altman when the board tried to remove him, the talent is more than happy with the reneging.

The investors were clearly not happy. Also his chief talent, Ilya, did leave him with many other OG founders.

Thanks. I didn't realize they'd finally released 4o image capabilities.

When it comes to text generation models, I can't bring myself to care anymore. They all do the basics proficiently, which is generate code snippets and console commands and trivia answers to questions that I can ask in a few sentences. And they all sort of fail on really complex stuff like doing my work for me in making changes to a 1M line codebase.

LLMs have been "good enough" for my use cases for a while now. I haven't seen anything like the jump from GPT 3.5 to 4, where prompts that previously had been borderline useless magically became useful. Not that I mind incremental improvement, I'd say that reasoning models were a decent change in terms of QOL and reliability when it came to complex reasoning.

If I had to use an LLM for my day job, then pretty much any recent model would work fine. While medical records can be massive, dense and a PITA to work with, they're not as complex as a large programming project. The whole field was built around humans needing to keep the gist of things in working memory. I'm more constrained by things like data protection laws, access points idiot-proofed by IT, and the lack of tooling to make it easy to transfer records or information back out. Ah, the wonders of the NHS..

I remain quite confident that LLMs will continue to improve, and suspect a lot of the recent sense of being underwhelmed that I've experienced is because the rapid rate of iteration makes changes less stark. They're beating agentic behavior into them, more by the minute.

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