A-TIER repo
ASI-Evolve — open-source agent loop that autonomously designs AI architectures
100k ft — What is this?
ASI-Evolve — open-source agent loop that autonomously designs AI architectures
50k ft — Why care?
ASI-Evolve, from Shanghai Jiao Tong University and Shanghai Innovation Institute, is an open-source agentic framework that automates AI research in a learn-design-experiment-analyze loop across data, architecture and learning algorithms. The paper reports 105 autonomously discovered architectures, with the best beating DeltaNet. The video also touches on China's AGI framing in its five-year plan.
10k ft — Key takeaways
• A closed learn-design-experiment-analyze loop can search data, architecture and optimizer choices together.
• Open code means you can try the loop on your own small-model experiments.
• Read the paper's DeltaNet comparison before believing the 'tripled human gains' claim.
5k ft — Tools & techniques
ASI-Evolve
automated AI research, closed-loop experimentation, neural architecture search
1k ft — Deep-dive / try it
China now has an open-source AI system that autonomously designs better AI models. It’s called ASI-Evolve!
#china #futuretech #aitools #ainews #gigaqian
blakemoser
11w
it has come to this. AI building AI
Transcript excerpt
AI might not need us anymore. This is all because China might've just leapfrogged us. They released something called ASI Evolve, which isn't a model like Claude or GPT-40. It is a general agentic framework designed to automate AI research. And that ASI stands for what you might think it does, artificial super intelligence. Hi, I'm Giga Chan, and I want to acknowledge a milestone. That's 100,000 followers on TikTok and 50,000 followers on Instagram. I love you all, and I just want to say thank you. Let's go back to ASI Evolve. If you know me, there are two things that I'm going to get excited about. The first one, yeah, it's open source. And the second one, yes, it comes with an excellently detailed research paper. With this paper, we can understand the closed loop cycle. It goes learn, design, experiment, analyze. And this brings us to a major milestone, the first unified demonstration of AI-driven discovery across data, architecture, and learning algorithms. This means it learns what works, designs new approaches, experiments on itself, and then analyzes the results and does it all over again. So yeah, for this, it does not need us. And if you're thinking, so what? This is still going to take a while. 105 new AI architectures have already been discovered by this thing autonomously. Its best new model beat DeltaNet by almost a point, which might not sound like much, but it means it tripled what human researchers were able to manage. So are we approaching something, namely…
More from @gigaqian
- B-TIER Xiaomi AI Cube — 80GB unified-memory box running 120B models locally
- B-TIER Moderna–Merck personalized mRNA cancer therapy — first positive Phase 3
- A-TIER Qwen 3.8 27B — local agentic model that fits on one 24GB GPU
- B-TIER Grok Bot — persistent AI workers sharing one cloud computer
- A-TIER Evo-designed bacteriophages — AI genomes that beat human gene swaps