Sure, the model is powerful. And they say nerfing only happens if you do LLM-development related stuff. Which means pretty much anything with modern software engineering could trigger it.

It just feels increasingly an uneven playing field. More competition will be much welcome

I tried out Fable, it was OK, but given #1 and #2 and given I will be doing some work around evals, RAG and others, for the first time in forever I will go back to using a less powerful model

Also, I’ll use Codex more when I do anything with LLMs. I don’t want to be silently nerfed!

@GergelyOrosz 2 is not Fable exclusive though. Nerf seemed to have happened with all other models as well over time

@GergelyOrosz I heard a engineer from Anthropic tell me that this is a conspiracy theory and that they don’t intentionally nerf models. But I’ve seen the degradation firsthand so I think they are gaslighting people.

@alexzenciks its the first time they make it clear they nerf based on your prompts

before, it was just for everyone

@GergelyOrosz some people complaining about app sabotage https://jonready.com/blog/posts/claude-fable5-is-allowed-to-sabotage-your-app-if-youre-a-competitor.html

@GergelyOrosz a silent capability shift can silently break my entire pipeline... nice

@GergelyOrosz Yesterday they were eating my tokens like crazy and the code was bad, really bad. Great combination, exactly what I'm paying for. Thanks anthropic

@GergelyOrosz prompt-content nerfing burns the unconditional part of every alignment claim. "we don't do X" becomes "we don't do X unless we decide you're a competitor". every other trust claim collapses with it.

@GergelyOrosz I asked it to do some investment analysis for me last night, starting by by looking over my liked tweets so see which ones contained potential investment ideas. It churned tokens for half an hour then crapped out because it found a tweet I had liked about cyber or LLMs! Hopeless.

@GergelyOrosz These are fair criticisms. I'm still looking for my own usecases where fable is substantially better than 4.8

@GergelyOrosz 30-day prompt storage feels excessive when everyone else deletes after 7

@GergelyOrosz That makes me realize the power that Anthropic and OpenAI have, being the few providers of the best state of the art coding agents means they can demand terms that others can’t. And people will still use it out of at this point what feels like a necessity

@GergelyOrosz Its intentionally malicious behaviour - I guess we have seen the don’t be evil mantra flipped the fastest.

@GergelyOrosz How is #2 even legal? It's not like you get refunds for nerfed tokens. What happens if you put it into /goal? Will it just continue to drain your wallet endlessly?

@GergelyOrosz another win for codex tbh, the fact that it collects prompt history and stores it means it's completely negated for enterprise and fca regulated work

so me and my team won't be using this for our actual day 2 day work sadly

@GergelyOrosz Pause button safety company is the only one who can safely provide AI. Please ban all other ai only us can handle it. Also we going to silently degrade or inject bad data into your prompt if you do something we angels don't like. Terrible company decent models sadly

@GergelyOrosz But 2 happens with all modern models and that's why it's hard to build anything reliable based on them. It's not even clear we get so inconsistent results because of the nature of LLM models or because of the tricks of inference providers

@GergelyOrosz Haven't tried it but haven't they had access to the prompt history for analytics in every model?

And they are nerfing Semi Analysis already… it’s not theoretical

I don’t want to pay a premium for a model like this