Ethan Chern

PhD student @sjtu1896

[Ethan Chern @ethanchern]

"Failure is just iteration. No explosion, no innovation. Keep going."🚀
You vent to @elonmusk—he looks you in the eye and replies instantly, like a video call.

Introducing LiveTalk: real-time video gen system on a GPU that sees you, reads emotion, and responds in real time.🧵👇

[Jie Sun @Eudaemonia279]

🚀 Stop over-optimizing same-tokenizer distillation.

On-Policy Distillation is powerful, but it quietly assumes teacher and student share the same tokenizer.
What if they don’t?
We introduce SimCT, a simple way to recover lost supervision for cross-tokenizer OPD.

[Pengfei Liu @stefan_fee]

Can AI accelerate AI itself? Let's look at what actually happened!!!

AI autonomously discovered a new RL algorithm, invented 100+ novel architectures, and evolved a data curation strategy — each one outperforming the best human-designed counterparts:

🤖 RL algorithm design — discovered novel optimization mechanisms that beat human-designed GRPO by +12.5 pts on AMC
🧠 Neural architecture search — invented 100+ new architectures, surpassing Mamba, GatedDeltaNet, and more by +0.97 pts over SOTA
📦 Pretraining data curation — evolved a pipeline that outperforms FineWeb-Edu, DCLM, and Nemotron-CC by +18 pts on MMLU
🧬 Drug-target interaction — discovered a new architecture that outperforms all prior human-designed methods by +6.94 AUROC
These aren't toy demos. These are frontier-level results — produced without a human in the loop.

This is ASI-Evolve: an open-source agentic framework that closes the loop between knowledge → hypothesis → experiment → analysis, and repeats it autonomously until it finds something that works. We built it for AI research. But the loop doesn't care about the domain.

A year ago, we made a prediction:

Scale computation → Scale scientific discovery.

The world caught up faster than we expected.

Which raises a question we can't stop thinking about — and we think you shouldn't either:

In the post-AGI era, when AI generates breakthroughs faster than humanity can absorb them — when the rate of discovery permanently outpaces the rate of value conversion — what becomes the new scarcity?

Not intelligence. Not data. Not compute.

Something far more irreplaceable.

We don't have the full answer. But we think it's worth asking — loudly, and together.

💻 Code: https://t.co/rBicTvkxYT
📄 Paper: https://t.co/2DriXJ4uFl (1/8)
⭐ Star it. Fork it. Point it at your hardest problem.

[Yi wei Qin @QinYi88814]

🚨 We can download models, but not see how they were built.
Introducing daVinci-LLM: most transparent LLM pretraining project.
🔓 Open source: model weights, data pipeline, training process, ablations.
🎯 3B model matching 7B performance
🔗 Report: https://t.co/HgTcXhSQSS (1/7)

[Geek Lite @QingQ77]

daVinci-MagiHuman 一个把文本、音频、视频塞进单流 Transformer 里统一生成的人像音视频大模型

用一个 15B 的单流 Transformer 同时吃文本、视频、音频，不走多流、不堆 cross-attention，目标很直接：更快地产出嘴型、表情、动作、声音都能对上的人类视频。

https://t.co/hZtZm3ZcWl

[Adina Yakup @AdinaYakup]

daVinci-LLM 🔥
The SII-GAIR team just shared the full training pipeline on @huggingface

https://t.co/0Z8SwYDtEY

✨ 3B, competitive with 7B models
✨ 8T-token transparent training
✨ 200+ ablation studies
✨ Data Darwinism (L0–L9) framework

[马东锡 NLP @dongxi_nlp]

daVinci-MagiHuman

从头预训练的 MagiHuman 在人像生成的表现，达到开源 SOTA，并且multilingual！

期待 daVinci-MagiHuman 与 daVinci agent 的组合，解锁多模态 agent 能力！

Salute @stefan_fee @ethanchern

[Rohan Paul @rohanpaul_ai]

"Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model"

https://t.co/vz98MXWGXl

[Rohan Paul @rohanpaul_ai]

daVinci-MagiHuman is a 15B single-stream Transformer, trained from scratch to generate synced video+audio with self-attention only—no cross-attention or multi-stream paths. It is open-source, supports 6 languages, beats Ovi/LTX, and runs on one H100.

[Ethan Chern @ethanchern]

@GumbiiDigital @stefan_fee Thank you for your interest! The weights are here: https://t.co/5jHQbvDuW4

[Berryxia.AI @berryxia]

🚀 Seedance 2.0 虽强，但闭源！

兄弟们！现在来了完全开源的 daVinci-MagiHuman！

- 15B 单流 Transformer，从零训练
- 视频 + 音频联合生成，仅用 self-attention，无 cross-attention、无多流
- 单张 H100 就能 38 秒生成 5 秒 1080p 视频
- 人评胜率：80% 碾压 Ovi 1.1，60.9% 胜 LTX 2.3
- 支持 6 种语言（含中英日韩等），WER 仅 14.6%

速度源于极致简洁，质量却拉满！

由 @SII_GAIR × @SandAI_HQ 打造，Apache 2.0 全开源

这才是开源视频生成该有的样子！

论文 / 模型 / 代码 / Demo 已放出 👇

你准备试了吗？

[Ethan Chern @ethanchern]

@AiTrendDesk @stefan_fee We think that single stream in fact improves training and inference utilization and efficiency so will give us better models!

[Ethan Chern @ethanchern]

@CMliand @stefan_fee Yes!

[Ethan Chern @ethanchern]

@RealHanyaHu @stefan_fee 感谢关注！我们会致力于让大众更好的使用daVinci-MagiHuman!

[Ethan Chern @ethanchern]

@hckinz @stefan_fee @jfischoff Yes it is!

[Ethan Chern @ethanchern]

@jayuhdinger @stefan_fee Thank you for your interest! We are working our best to develop workflows around it.

[Ethan Chern @ethanchern]

@agentsfy @stefan_fee Thank you for your interest! No, it is not based on Wan. daVinci-MagiHuman adopts a completely new single-stream architecture and is trained from scratch!

[Pengfei Liu @stefan_fee]

Seedance 2.0 is impressive. But it's closed-source!
Introducing our daVinci-MagiHuman — a single-stream 15B Transformer trained from scratch that jointly generates video + audio. No cross-attention. No multi-stream branches. Just self-attention.
⚡ 5s 1080p video in 38s on a single H100
🏆 80% win rate vs Ovi 1.1 | 60.9% vs LTX 2.3 (2,000 human comparisons)
🌍 6 languages
📦 Fully open-source
Speed by simplicity.
By @SII_GAIR × @SandAI_HQ
📄 https://t.co/SgFOunlEIj
💻 https://t.co/9rwNWzlMKN
🤗 https://t.co/txduP5FgIC

[DailyPapers @HuggingPapers]

daVinci-MagiHuman: Speed by Simplicity

A 15B single-stream Transformer unifying text, video & audio via self-attention alone—no cross-attention. Generates 5s of 256p video in 2s on H100. 80% win rate vs Ovi 1.1.

[AK @_akhaliq]

Speed by Simplicity

A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model

daVinci-MagiHuman

paper: https://t.co/AxA9iZr8LX

model: https://t.co/LIv2ZUYsy8

app: https://t.co/24sAurqY80

[Adina Yakup @AdinaYakup]

daVinci-MagiHuman 🎬 Human Centric Audio-Video Generative Model by GAIR

Model: https://t.co/OPC8jEhlRu
Paper: https://t.co/OPC8jEhlRu

✨ 15B – Fully open source!
✨ 5-sec 1080p video in 38s on one H100
✨ Supports 6 languages
✨ Unified model with text + video + audio