GLM-5.3 Flash in already nerfed? I compared Ox Alpha in it's prime vs today just before it got revealed as GLM-5.3 Flash using the same prompt to build 3d raptor engine using three.js OxAlpha on 24th August gave far more accurate output compared to today which looks way worse... prompt i used: Create a 3d SpaceX raptor engine using three.js, use white theme background so the main object is clearly visibleshow more

Probably this:

apparently OxAlpha was better as a mystery model with less exposer, it felt so good to use

AI is not deterministic

If you don't set temperature to 0.0 you are getting random outputs each time.

First of all how is the second worse than the first…lol. Second of all rng

you can't really just compare 2 generations 1:1 like that... you have to do like 10 of each and compare on average.

Are you intentionally leaving out that you previously got a result as bad as the bottom one, then claimed it suddenly improved, and now you are claiming it suddenly got worse? Do you really want engagement on your tweety THIS much?

did you fix temperature and compare repeated runs, because one sample cannot prove a nerf

And this is created by gemini 3.1

Improved Ox Alpha was never 5.3 flash. That was a different model as said it themselves.

I guess this was just the first run. As others said, AI is not deterministic.

Same effort level?

everything same same

Maybe it’s just very unreliable and it’s not self improving it’s just all over the map (my experience)

3일뒤에 동일 프롬프트로 테스트하면 결과가 다를 수 잇습니다.

how is it legal for them to nerf. you pay for one thing and get another

every AI model does this after the hype is gone and they get busy building the next new model

maybe cuz its the flash version that was released? not the full standard version.

And it's also different from your first try and will be different with your 4th try because those models still aren't deterministic.

For me even Ox was nerfed when it was so hyped. I was not impressed with it.

it was working good when not too many people where using the model at the same time

you have to host it to be able to control quantization, temperature and more to get the same result.

Yes, it has become much worse to work

it still performs great and I would definitely pick it as a model for fast iterative work I asked it to make a minecraft mod and the results are pretty crazy

Hmm... This is a mystery indeed 🕵️

It’s learned from all the 🆓 data it gulped everything, may be unsupervised

If you open two chats and give the same prompts you'll have a bit different results right? So, to be sure of what you are saying you should've run like 10 times, to get a correct idea of before and after. Since we can't go back to stealth, try running it multiple times to see

every attempt will look different, better or worse because its non deterministic.

I think the plan with OX Alpha is not only to release and test GLM 5.3 Flash but rather the infrastructure too, do they might have just tested with GLM 5.3 Omni.

glm-5.3 flash is a great example of how model updates often create trade-offs between size and performance. nerfing may save memory, but at what cost to output quality?

What quant did you use ?

Yeah its called a flash build buy the pro version

那必然给你上量化模型了

100% nerfd

yup, it feels different, going back to codex

and get codex resets every other day

Curious about its pass@4 variance. Do 4 tests.

same harness or what?

yes same harness used with same prompt same model, the only difference is the time