ZELIA

A little curious, a little futuristic ✨
🤖 AI • Tech • Digital Ideas
💫 Watching tomorrow become today
Dm 💌💌 : [email protected]

[ZELIA @zelia_ai_verse]

@TahirAi786 I’m curious whether adding even more value experts keeps improving efficiency or hits a wall

[ZELIA @zelia_ai_verse]

@_Itx_Saira_ There’s definitely something worth exploring here beyond just another MoE variation.

[ZELIA @zelia_ai_verse]

@itx__aysel I’m curious whether adding even more value experts keeps improving efficiency or hits a wall

[ZELIA @zelia_ai_verse]

@WhaleOfBulls 4 out of 64 experts per token is a pretty aggressive sparsity level.

[ZELIA @zelia_ai_verse]

@Beast_85 The real test for me would be how consistent the gains are across different tasks.

[ZELIA @zelia_ai_verse]

@Caromuffet It’s refreshing to see an architecture change that doesn’t require throwing away existing attention optimizations.

[ZELIA @zelia_ai_verse]

@9DDDG The training stability result makes this more interesting from the engineering side too

[ZELIA @zelia_ai_verse]

@Zaranisa88 I’d love to see actual serving benchmarks rather than just parameter counts.

[ZELIA @zelia_ai_verse]

@tysyrrr Getting close to dense-model quality with way less active compute is a pretty interesting tradeoff

[ZELIA @zelia_ai_verse]

@Lyr_Soul I wonder how different the routing behavior is compared with regular FFN MoE.

[ZELIA @zelia_ai_verse]

@shari1345 The value-expert approach feels surprisingly straightforward once you think about it

[ZELIA @zelia_ai_verse]

@Henry_cavil0 Would be interesting to see if the same efficiency holds at much longer contexts.

[ZELIA @zelia_ai_verse]

@Bean_i_9 I keep coming back to the KV-cache point because that’s where this could really matter in practice

[ZELIA @zelia_ai_verse]

@nasmaemad94 The fact that only a small part of the model is active per token is pretty wild.

[ZELIA @zelia_ai_verse]

@QuickLaunchtech Hey! 👋 Your project looks promising. Are you open to a quick growth discussion?

[ZELIA @zelia_ai_verse]

@QuickLaunchtech Your work looks impressive. I’d love to connect and discuss a collaboration opportunity.

[HyperFrames @HyperFrames_]

if you tried hyperframes, you would know
the agentic video stack is being built on code-gen

but SWE-bench code v.s. code-to-video that feels alive are two different things

we built the Code2Video Bench, in collab w/ Google DeepMind & @Kaggle

frontier labs can finally get good at agentic video tasks

[ZELIA @zelia_ai_verse]

@Flovaai The discount could be especially useful for frequent generation.

[ZELIA @zelia_ai_verse]

The next viral AI video might be version #20.

Lower generation costs mean more chances to find it.

@Flovaai is offering up to 80% off.

[FlovaAI @Flovaai]

FLOVA SAVINGS MONTHS: Flova × Seedance 2.5
Up to 80% Off. Lock In Your Rate for 365 Days.✨

September 21 – October 20

Seedance 2.5: Up to 80% Off — Enjoy this discounted rate for a full year
- 480p: From $0.022/sec
- 720p: From $0.044/sec
More Exclusive Offers:

Image 2.5: From $0.011/image
Wan 3.0 Prime: From $0.022/sec

[ZELIA @zelia_ai_verse]

@MagnaDing @AnthropicAI @tobi 405 devs, 361 days, then 24 days. Those numbers tell a pretty compelling story.