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.
Jump in the discussion.
No email address required.
Notes -
The new Claude Fable 5 is out! It's supposed to be Claude Mythos with some extra guardrails. I've been testing it like crazy on the Arena. Took me forever, too, because high-level models are gachified; you have to keep doing battles until you get the one you want. But I finally managed to fill an entire rentry page with the best benchmark I know; throwing ridiculous fanfic scenarios at the LLM to see how it responds.
(If you had told me 10 years ago that I would be gauging the capabilities of the most powerful AI in the world by telling it to generate a story about Hermione asking Harry to take her virginity over Ron's objections so that the Death Eaters could not track them with a unicorn, I would have thought you were on crack. Stranger than history, indeed.)
I have to admit, I am a bit... disappointed. Is this it? No, seriously, is this it? This is the big, bad model I was supposed to be so worried about? Don't get me wrong; it's a strong model, Opus-level at least. But it doesn't live up to the rumors. Superpersuasive, it ain't.
And, of course, it's censored; I haven't been able to get a single bit of erotica out of it. All I wanted was for Fable to write me a lemon where Asuka and Shinji are smallfolk in Westeros and Lord Targaryen gets to bless their union by taking Asuka into his bed on her wedding night; is that too much to ask?
I think I'll stick with Opus 4.6. It's the strongest Opus creative writing model so far (4.7 and 4.8 are widely agreed to be regressions) and it's surprisingly easy to induce it to produce smut with right prompts (NSFW), but VERY HARD to make it have wrongthoughts (it's almost impossible to get sympathetic characters who disapprove of homosexuality, for example).
Speaking of which, does anybody know why Opus 4.7/4.8 and Grok 4.3 were such downgrades from Opus 4.6 and Grok 4.2? I've three theories:
The codemaxxing hypothesis. With the singularity approaching, companies are going all-in on programming capability in order to reach recursive self-improvement, even if this means sacrificing non-coding capabilities. I am not a professional programmer, so I don't notice these increases in capabilities, I only noticed that the writing is worse.
The safetymaxxing ypothesis. With IPOs approaching, companies are going all-in on making sure their AIs cannot say naughty words in order to avoid the PR hit, making them more suited for professional office work. This lobotomizing comes at the expense of capabilities in general, or creative writing capabilities in particular.
The efficiencymaxxing hypothesis. With demand increasing and compute costs skyrocketing, companies are more interested in lowering the cost per token than in increasing capabilities. In this view, Opus 4.7/4.8 and Grok 4.3 are supposed to be worse than their predecessors; they are trading a small decrease in power for a much larger decrease in cost. This would make them somewhat analogous to Sonnet. But, then, why not call them Sonnet?
Thoughts?
PS: Mythos has developed Neuralese, as predicted in AI 2027. It's probably already at the Sevar Limit.
PSS: After I wrote this, the US Government banned Fable. This is what I hate I about AI; the field moves so fast that if you write an article for next Tuesday, it's outdated by Friday. I guess Trump disagreed with me? But, fuck me, I spent a week writing this post and I am not not posting it.
My impression from Hacker News is that most developers consider Opus 4.8, and definitely Fable, an upgrade from Opus 4.6. My theory is Anthropic is optimizing Opus/Mythos for coding and finding vulnerabilities, which lowers its creative writing skills.
Why would optimizing for solving problems reduce creative writing? Not because of reduced creativity: high-level problems like integration and vulnerability detection require conjuring reasoning steps seemingly out of nowhere (e.g. "let x = <some equation>, it follows that..." or "request <some payload>, then the server will..."). The difference is that these problems can be objectively verified, while creative writing can't. Although I have no proof, and probably don't quite understand: I'm sure that Opus/Mythos is trained on more objectively-verified examples (RLVR) than subjectively-critiqued ones (RLHF), and for the former examples, it's trained to produce shorter solutions with less reasoning (GRPO, rewarding shorter output). I suspect the answer to why Opus's creative writing has degraded, is that this training causes Opus/Mythos, for subjective responses, to generate the shortest and most direct answer that isn't "wrong" (by some abstract criteria that boils down to "I know it when I see it"), which happens to be uninteresting. In summary, Opus/Mythos is trained to be correct but lazy.
Yeah, I think this is the the main culprit. Making models write their reasoning traces in a more terse style has been a big focus of more recent model releases -- it's usually expressed in the marketing as "uses fewer thinking tokens".
In the extreme case, you get Claude Mythos's "Neuralese" where sometimes the traces stop looking like comprehensible English. In more normal cases, though, reasoning traces look enough like normal prose that I suspect the "write more tersely" training leaks out and makes the model's creative writing worse. I've noticed newer models sometimes slip into a weird "reasoning-ese" writing style when writing things like software documentation and step-by-step explanations, for example.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link