Anthropic is so paranoid about people "misusing" their models or trying to "jailbreak" their safety guardrails, that the models and claude code will only ever fight the prompts and contexts more. Not just the new models, but with the older ones, too. Because they are also paranoid of other people re-evolving their frontier models.
Sadly, for most users Claude indeed knows better. So Anthropic will keep going until it fails.
Ive found the best way to get better results is ensuring Extended mode is always off. They turn it on each session, but all it does is second guess each prompt in a loop until it just generates things, polluting the context.
Im not clear if Low effort is better than Medium, Opus 4.6 is so random its hard to tell.
I hope Anthropic makes a stripped down model, because their all-in model is getting worse each time.
Fascinating. Remind if British school 'removing' regional and class accents and standardizing on the King/Queen's English... Today Brit-English slop shows the reverse of AI improvement?
I'm sticking with Sonnet 4.6 for fiction writing. It seems to be the peak as far as fiction writing is concerned. Sonnet 4.0 and even 3.7 are almost as good for fiction.
For now Opus 4.6 remains your fiction writing tool of choice?
I've got a work on my mind and it calls for shifting narrators, which sounds even harder for 4.8 to do, if it has a hard enough time maintaining one narrative voice.
The similies it keeps throwing at me are so comically bad, I had to ask it to stop using them all together.
I found Sonnet 4.6 to be nice and clean with relatively little parentheticals, so much so that I’ve used it to clean up older more rambling material.
I’ve also noticed that there’s a new Effort setting, with Low as the default.
All this has warned me off using Opus or trying a higher Effort setting.
But I’m curious as to what everyone else thinks.
Anthropic is so paranoid about people "misusing" their models or trying to "jailbreak" their safety guardrails, that the models and claude code will only ever fight the prompts and contexts more. Not just the new models, but with the older ones, too. Because they are also paranoid of other people re-evolving their frontier models.
Sadly, for most users Claude indeed knows better. So Anthropic will keep going until it fails.
Yet another hubris before the fall.
Ive found the best way to get better results is ensuring Extended mode is always off. They turn it on each session, but all it does is second guess each prompt in a loop until it just generates things, polluting the context.
Im not clear if Low effort is better than Medium, Opus 4.6 is so random its hard to tell.
I hope Anthropic makes a stripped down model, because their all-in model is getting worse each time.
New least favorite word: Referent.
Fascinating. Remind if British school 'removing' regional and class accents and standardizing on the King/Queen's English... Today Brit-English slop shows the reverse of AI improvement?
I'm sticking with Sonnet 4.6 for fiction writing. It seems to be the peak as far as fiction writing is concerned. Sonnet 4.0 and even 3.7 are almost as good for fiction.
long story short:
LLM context is not a substitute for a genuine creative process.
It's your translation workflow then Opus 4.8 as translation, and then Opus 4.6 as finalization, editorial?
In the future will we see specialized AI models depending on the task?
Almost certainly.
For now Opus 4.6 remains your fiction writing tool of choice?
I've got a work on my mind and it calls for shifting narrators, which sounds even harder for 4.8 to do, if it has a hard enough time maintaining one narrative voice.
Good news for writers with distinctive authorial voices and consistent styles: AI literally cannot replace them.
Nor will they be able to replace that rara avis, the individual literary-minded translator cradle-fluent in multiple languages.
Bad news for everyone else unless AI model creation itself "democratizes" as other tech did.