Post

Conversation

xuan (ɕɥɛn / sh-yen)
@xuanalogue
Glad we have some confirmation that the sudden increase in 3D reasoning abilities by frontier models is specifically trained for via RL rather than some emergent capability.
Quote
Alexandr Wang
@alexandr_wang
whole team feels dumb now, but it's true, we did not explicitly train our model on 3d demo games obviously in retrospect we probably should've done a bit of that... but no regrets, please try muse wonder on your actual coding work!
Quote
Theo - t3.gg
@theo
Already see a lot of people discounting Muse Wonder because it wasn’t RL’d in 3d stuff like Astra and Opus. 3d game demos are cool! They don’t represent real code work at all. Wonder is an incredible code model. Don’t sleep on it.
33.7K
Views
View quotes
David Watson 🥑
Post your reply

indif4ent
@indif4ent
i mean, i think it wasnt a reasonable belief to think that astra spontaneously became great at 3d modeling in the first place? lol
2.5K
Square profile picture
MyFundedFutures
@MyFundedFutures
Ad
A real trader passed their Studio evaluation in under three minutes. Less than two minutes later, they requested a payout. That's My Funded Studio. Static drawdown. No consistency rule. No waiting period for eligible payout requests. Watch the actual trader replay.
The media could not be played.
From myfundedfutures.com
3K
vals🔸
@ValsTutor
Kinda, but also it might be that Muse Wonder is less frontier than they'd want us to think. If it's in fact overall worse, then it's plausible that some of the other models being good at 3D comes from their general ability.
87
sohan 😶‍🌫️
@SohanSarabu
i don’t know if that’s necessarily true - there’s definitely quite a bit of speculation that there was a massive increase in 3D / spatial pretraining data for models like astra
161
mono
@dhsusj144284
I guess there's no such thing as emergent ability at the end, it was all just training data.
1.7K
0x44 0x46
@0x440x46
Everything is trained for via RL for any particular task or role. RL environments for every job basically
647
Survival Tipps und Tricks
@und_tricks1979
Yeah - humans are no different - lets take an average intelligent 30 year old human with a umiversity degree but who has never done anything in blender or motion design - he wont be able to generate good stuff - whats your fucking point ? Humans are no different
87
RandomName
@shungite2121
There are just some tasks that require a certain level of capability. They probably tried to train older models on 3D
453
patlsc
@_patlsc
it's all such a farce. any moron could buy a "game" they "made" off an asset store for $20. and now the ability to make asset flips for $200 is apparently evidence AI is worth trillions of dollars
797
Namosilnio
@namosilnio31361
This is not a secret though have been telling, almost all it's capabilities using other software are specific training . It's the ai promoters on social media hyping these like it's magic.
764
Tate
@TheCodingBug
100% it makes sense. It’s why the ability to write code hasn’t actually gotten that much better for awhile now. They have gotten better at reasoning but their coding ability isn’t crazy better. They are just being trained on more data
344
Ir rev
@1rreverant
I definitely think there is some truth to that, but I also don't trust meta at all when it comes to benchmarking
Quote
Binkg
@dummerspast39
Muse Wonder has that BIG model smell
Image
604
Ken
@ken100bands
I don’t think what you’re implying as emergent definition makes sense - it feels like it’d be more so emergence within a specific category you train on handling OOD cases beautifully
343