Kang

Running your company's managed data hub with

[雨夹雪❄️ @mizorewww]

介绍比Jev快50倍，在你设备上跑的laya-mlx!

只在你的设备上占用最高1G内存

Laya是一个开源的类似于Jev的,基于文本输出概率的分类系统

我将其移植到MLX,并且做了一些性能优化!

视频中就是这个模型在我的本地M3Max上玩贪吃蛇

这个模型能够以每秒决策60次的速度玩贪吃蛇！

https://t.co/31KGUiNunb

[Kamimoto(かみもと) @sep_is_heim]

流行のJevをMiniMax H3に組み込んで、動画生成を高速化してみた！Attention処理のスパース化にJevを使用。

・層ごとにJevが重要度を判定（4step 49層が対象）
・Jevがスパース率1%, 3%, 5%, 10%を選択

RTX4070で6分7秒→3分34秒で41.7%短縮！動画生成中にJevクラウドに問合せしているのに速い！

[Kang @schift_official]

쓸 수가 없는데 리셋 해줘봤자 무슨 의미야...

[Kang @schift_official]

@LuciferMornens @AdiExDeo @thsottiaux South Korea

[Kang @schift_official]

@datafndr @NtTestAlertX @thsottiaux South Korea

[Kang @schift_official]

@thsottiaux Codex users are at capacity. Please try a different subscription

[Kang @schift_official]

@usefulforme @thsottiaux we call it token management

[Kang @schift_official]

Yes,
This is my astra max of 09.14 18:54 (KST)

[Kang @schift_official]

@thsottiaux Being every model is at capacity?

[Kang @schift_official]

@thsottiaux intelligence rugpool and 429 obviously

[Kang @schift_official]

@Mr_Salio Will be nerfed like astra , just only 5 days after launch

[Kang @schift_official]

@NtTestAlertX @thsottiaux there were buying banked reset option on the codex app (desktop)

[Kang @schift_official]

Codex usage might drain my quota.

Sure, there can be some model updates, and I’m willing to accept that. But when the model just gets dumber, that's completely out of our control, and I can't stand it.

It's not just one model—having Sol, Terra, Luna, and Astra all become useless is a complete disaster.

I’m honestly afraid of using Codex now, because doing nothing at least doesn't generate more bugs or slop.
#codex

[Kang @schift_official]

@michaelnakache @thsottiaux They surely do not have a dev i guess. This cannot be the production quality. We don't pay $200 for being guinea pigs lol

[Kang @schift_official]

@thsottiaux Resetting limits doesn't justify the model downgrade you just pulled on Codex.

We just want Day 1 Astra and the setup that came with it. Don't mess with it

[Kang @schift_official]

@SpikeCalls Where did you get those fly 3d and neuron activation graph?

[Kang @schift_official]

AI model companies must stop rug-pulling users regarding the performance of models included in their subscriptions. This is downright consumer deception.

Neither Claude nor ChatGPT maintains the same performance on Day 5 as they did on Day 1. The model changes from yesterday to today, and this constant fluctuation causes severe user fatigue. Furthermore, users are not being properly compensated for the loss of productivity caused by downgraded models.

Resetting usage limits is by no means a solution. The time wasted cannot be recovered, and our data has effectively been reduced to nothing more than test subjects for their QA.

[Kang @schift_official]

@Mackerel786372 @thsottiaux This current reset policy makes it so much dependence on his twitter, not the user's usage.
Its his usage

[Kang @schift_official]

@DaSun64125381 @thsottiaux Do you think i would have bought the reset if i knew it would come 24 hours after? Haha

[Kang @schift_official]

@AdiExDeo @thsottiaux It is on your codex app tho, on the settings, usage page