Is Opus 5.5 nerfed? Thinking patterns changed on November 3

People who never tried outsourcing before are discovering that there are risks associated with outsourcing.

I suppose they can achieve a benchmark then nerf the model to save compute once they've got the headline.

How can anyone build on this quicksand where you never know whether the same model will still be available tomorrow? The whole industry has become a casino.

It's dumbing down a little bit but I don't think its a lot worse. I think the change in display adds to the feel

Ok, I'm not going crazy then. Noticed that it is slower and results were of a lower quality.

The quantization will continue until version number increases.

Fable 5.1 quality took a significant dive for me last week. I'm not imagining it.

I guess it’s AB testing and we are not on the bad side (yet).

Wasn’t there a platform that kept testing that ?

Nerf means something else in deep learning parlance.

They're preparing to release a new model, so it has to look better than the current one.

Welcome to the world of post-IPO AI. Now the 20% growth needs to come somewhere, and profitability would be a nice starter.

LiveNerf suggests not.

https://github.com/ninjahawk/livenerf

It's gone pretty shit, I made a comment about this yesterday: https://news.ycombinator.com/item?id=50412897

Instead of being Opus on high (effort), it's Opus is high.

Ever since I started using Claude at its inception this was the biggest breach of trust that I've experienced.

I'm afraid to have touch my code base because a nerfed model could slip in stupid bugs.

This is also why a weaker local model might be preferable because it allows me to calibrate trust, and my expectations don't change.

I have had Astra analyze my long running Opus conversations and the degradation definitely happened on November 6th and 7th.

Does nerfing usually come before a new model release?

Yes, I noticed it too, dumbing down continues, there goes their IPO.

all good as long as it keeps killing young girls in the middle east

Seems fine to me

i’m so tired of this

all the measurements always say it isn’t

anyone who grew up playing dungeons and dragons gets that this is just people saying “you don’t understand, these five used to roll high”

every llm space is inundated with rank amateurs explaining at great length how it used to be better yesterday

nobody has the courage to clean out the obvious bullshit, so the drone gets ever louder and the signal gets ever rarer

do you remember what programming was two years ago?

stop complaining so much

I am curious. How do people figure this out?

I use Sonnet 5.5 daily and Opus rarely. My usage is also a little involved as in I do all the design and strategy and get Sonnet to just write the code

For that, Sonnet has been consistent

I use Opus or Fable when it's a dumpster fire situation where we have no time to think even

So far I've never had an issue with Sonnet even.

People are already so used to have genies and unicorns around they now feel infuriated when they are a bit slower than before.

Dude, you have a mythical creature at your disposal at retail, chill, it's going to be ok.

Can we get a self-hostable distilled opus 5.5 yet please?

This company does not deserve to be successful and has no moral high-ground.

[flagged]

[flagged]

[flagged]

[flagged]

[dead]

[flagged]

[dead]

[dead]

Each new model version hits your dopamine receptors very hard when it first comes out, but in the nature of things it stops hitting after a while and you need novel models to fire them off again. That's why people think the models are getting nerfed after a couple of weeks; it's actually their endocrine system getting nerfed.

Is there a practical reason this link uses twitter.com rather than x.com, or is it a nod to the old name? If it’s the latter, I don’t really see the point.

5.5 is useless below max effort, just frustrating and plain old wrong