GPT-6 Astra at launch vs GPT-6 Astra today. Identical prompt. Same settings. Today's output is clearly weaker. Not saying it's been nerfed, but something has definitely changed. Anyone else seeing this?

It has 100% been nerfed. Past 18 hours it has been unable to perform the simplest tasks.

you tried splitting it into more tasks?

Simpler than the ones I am asking now I don’t even know how. I am temporarily moving back to Fable 5.1 and checking back in a few hours. Hopefully it’s a bug and not a nerf.

people still don't know how LLMs work wtf LLMs are stochastic in nature this is why idiots shouldn't be allowed to post stuff in the internet

gpus get tired

just feed more electricity

LLMS are non deterministic, this is to be expected

lmao bro is max larping xD

I've been building @PlayAlerith for over 3 years now. If you like MMORPGs and have played RuneScape or World of Warcraft or others be sure to check out Alerith. Devoting my entire life and pouring my soul and heart deep into it.

1000+ days

It’s nerfed literally all models get Nerfed

It's the same old story, brother. I imagine they reduced its performance due to expenses; maybe they'll acquire more computing power to lower the high cost. It's normal that they'd lower its performance to avoid massive losses

100% nerfed. Yesterday it kept going off task and adding things that were never requested. Luckily I don't let it run operations long operations. It do small pieces at a time so issues are easier to catch.

They have to share the compute because most people use it for playstuff that will never see the light of a production-ready release day. 😃

i’ve noticed the same on longer tasks

try to provide more electricity to astra it loves it

it's gonna be fake lmao

what was the prompt and where did you run the isolated test in?

I experienced the same with 3d modelling. In the beginning it was amazing. Now its producing the same trash as opus

Almost everybody can see it at this point.

👀

right llms be hallucinating

@thsottiaux

I would like you to respond to these accusations. It’s not about paying and getting reduced performance, as long as it’s for personal use and fun, that’s not a big deal. But when people start building businesses on it and require the level of quality that was initially shown, that’s not acceptable, and for me, it’s serious. I would like you to address this; otherwise, we will join forces and create a massive petition against all those who do this.

Interesting how all these people post these comparison videos but they never actually give the prompt they use. So we can test it ourselves. Astra has been exactly the same for me. Outputs will change each time though since its a different seed. and it depends on the prompt

Same shit as every model. Comes out in a good state and just gets worse

@thsottiaux I hope these accusations are not true and you should push against them by showing examples on regular interval (same prompt, same model+thinking level - run it 5 times, show results in one video so that randomness has less of an impact on comparison)

lmsys arena has logged silent backend swaps before, same prompt hitting a distilled checkpoint would explain the drop without any nerf needed.

this always always happen

Not saying it's been nerfed, but..." — proceeds to describe the exact silent downgrade every major AI model goes through a week after release.

@Deepneuron Been seeing a few people say it’s getting dumber already lol

Never criticise AGI bro.

Foutage de gueule scam révolte !

someone turned quantization on cuz their gpus were melting

@Deepneuron It’s so bad now lol. This will be the new normal for frontier labs I guess - they are manipulating our dopamine circuits like those fruit flies 😭

they nerfed consumer compute to focus on millennium problems - I am fine with this!

Lfg

one for the team

Lmao

Castrated intelligence. Happy that I run on the SpaceXAI stack. Grok does it.

Yes. Nerfed. Since yesterday.