正在加载视频...

视频加载失败

Grok 4.7's Fishslop run is the worst I've seen this year. This is my second attempt (in the first one, the submarine and all the fish moved backwards)

142,486 次观看 • 6 天前 •via X (Twitter)

30 条评论

S0n1cG4nd4lf 的头像
S0n1cG4nd4lf6 天前

I see why Elon agreed with Dario that frontier labs need to slow down. 💀 Just a silly joke

Jeremy Johnson 的头像
Jeremy Johnson6 天前

The fish reflect the model. Backwards.

Ananth 的头像
Ananth6 天前

this is such a disappointing launch. elon hyped it up so much

Khurram Bhutto 的头像
Khurram Bhutto6 天前

what about mimo2.6?

Ramesh Bende 的头像
Ramesh Bende6 天前

Using grok build or cursor. Think works best in grok build.

Theo - t3.gg 的头像
Theo - t3.gg6 天前

I used Grok Build and this is the slop it created.

Lee Penkman 的头像
Lee Penkman6 天前

its over :(

Sirin 的头像
Sirin6 天前

Can we have up to date fishslop bench on all SOTA?

Sathwik 的头像
Sathwik6 天前

Rerun Astra and check how it is performing now!

Youssof Al Toukhi 的头像
Youssof Al Toukhi6 天前

Ok it’s bad but come on. “This year”? You do realise models like Composer 2 released this year.

chiikabu 的头像
chiikabu6 天前

compare it with 4.6 and 4.5 please

nikoster 的头像
nikoster6 天前

Finally AGI achieved, congrats to spacexai team

Jayden CJ Wilson (previously One Bubble) 的头像
Jayden CJ Wilson (previously One Bubble)6 天前

Is there a way to redo this but balance the token usage? (Cause I swear 4.6 was more economical, and 4.7 is apparently even more economical... so it seems mean to compare them like this.... very mean, Mr. t3)

meguna 的头像
meguna6 天前

anybody try it for PCB design?

Jatin Garg 的头像
Jatin Garg6 天前

Backwards-moving fish in a trained model suggests the RLHF dataset included inverted examples, not just low-quality data. That's a training pipeline bug.

RyanQuinn 的头像
RyanQuinn6 天前

Out of curiousity how is copywrite?

ZIGAG 的头像
ZIGAG6 天前

There are two things that do not help Grok 4.7 at all: Elon's tweet saying Grok 4.7 in 10 days and then missing the deadline and having to delay. Of course, this model releasing after GPT-6 Astra, which is a really, really good model by the American labs. You would hope that Grok 4.7 would be really good but instead it's slightly better, maybe, than 4.6, but way less token-efficient than 4.6, which makes it more expensive. It really doesn't look good.

NyanK ツ 的头像
NyanK ツ6 天前

Go easy bro it's elon's best.

Harshith 的头像
Harshith6 天前

Grok 4.6 is better at 3d gen than this

Mikhail Rogov 的头像
Mikhail Rogov6 天前

If every object moves backwards, calling it 3D reasoning is doing a lot of work 😅

Ishu Agrawal 的头像
Ishu Agrawal6 天前

i feel like gemini flash might make better 3d output than this 😭

Michael Wall 的头像
Michael Wall6 天前

got dizzy watching like 5 seconds

Raven 的头像
Raven6 天前

the submarine has the navigational confidence of a shopping cart with one bad wheel

VectorCactus07 的头像
VectorCactus076 天前

this model doesn't deserve a review please don't waste your time making a video on it

bittu 的头像
bittu6 天前

I had big hopes for grok 4.7 in frontend cuz it's cheap. But yeah seems like it's not easy

ShadowAguy 的头像
ShadowAguy6 天前

submarine and fish moving backwards is such a specific failure mode. did it learn the currents run the other way, or just decide to be contrarian about physics?

sun 的头像
sun6 天前

Do mimo2.6 please

Nuno 'Hakira' Ferreira 的头像
Nuno 'Hakira' Ferreira6 天前

(in the first one, the submarine and all the fish moved backwards) That's called maritime reverse engineering 🤣🤣

Shubh Chaudhary 的头像
Shubh Chaudhary6 天前

high token is still excusable if the quality is good (eg. Fable) but this is such a disappointment from xAI

bubba gump 的头像
bubba gump6 天前

why are you wasting your time on grok... we already know it's shit. focus on mimi v2.6 pro... that's where the alpha is. nerds gonna nerd.

相关视频