Hugging Face's logo
kai's picture

kai

KaiWu123
1 2 3
linlincode's profile picture Liuv's profile picture 6cf's profile picture
·
https://kaiwu5.github.io/

AI & ML interests

llm coding

Recent Activity

reacted to theirpost with 🔥 about 3 hours ago
Can AI improve AI? We read 223 papers to find out. The literature is fragmented — long-horizon agents, AI4AI, self-improvement, and RSI are four separate conversations that don't cite each other. So we organized all of it around one question: how far can an AI system reliably carry an improvement process from idea to verified result? Two findings: 1. AI is now excellent at the *work* of improvement — planning, coding, experimentation, repair. Humans still set the goals and decide what counts as progress. That asymmetry isn't closing. 2. The composition gap: strong performance on individual components rarely composes into reliable end-to-end improvement. A system can beat every component benchmark and still fail the full loop. Most "self-improving AI" claims live in this gap. Search the catalog: https://huggingface.co/spaces/KaiWu123/awesome-ai4ai Dataset: https://huggingface.co/datasets/KaiWu123/awesome-ai4ai 📄 Nice Layout Paper: https://github.com/KaiWU5/Awesome-AI4AI/blob/main/assets/AI4AI-Survey.pdf 🌐 Site: https://kaiwu5.github.io/Awesome-AI4AI/ ⭐ Repo: https://github.com/KaiWU5/Awesome-AI4AI 📄 Paper: https://www.preprints.org/manuscript/202608.2108/v1 Disagreement welcome — especially on the composition gap.
reacted to theirpost with 🚀 about 3 hours ago
Can AI improve AI? We read 223 papers to find out. The literature is fragmented — long-horizon agents, AI4AI, self-improvement, and RSI are four separate conversations that don't cite each other. So we organized all of it around one question: how far can an AI system reliably carry an improvement process from idea to verified result? Two findings: 1. AI is now excellent at the *work* of improvement — planning, coding, experimentation, repair. Humans still set the goals and decide what counts as progress. That asymmetry isn't closing. 2. The composition gap: strong performance on individual components rarely composes into reliable end-to-end improvement. A system can beat every component benchmark and still fail the full loop. Most "self-improving AI" claims live in this gap. Search the catalog: https://huggingface.co/spaces/KaiWu123/awesome-ai4ai Dataset: https://huggingface.co/datasets/KaiWu123/awesome-ai4ai 📄 Nice Layout Paper: https://github.com/KaiWU5/Awesome-AI4AI/blob/main/assets/AI4AI-Survey.pdf 🌐 Site: https://kaiwu5.github.io/Awesome-AI4AI/ ⭐ Repo: https://github.com/KaiWU5/Awesome-AI4AI 📄 Paper: https://www.preprints.org/manuscript/202608.2108/v1 Disagreement welcome — especially on the composition gap.
posted an update about 4 hours ago
Can AI improve AI? We read 223 papers to find out. The literature is fragmented — long-horizon agents, AI4AI, self-improvement, and RSI are four separate conversations that don't cite each other. So we organized all of it around one question: how far can an AI system reliably carry an improvement process from idea to verified result? Two findings: 1. AI is now excellent at the *work* of improvement — planning, coding, experimentation, repair. Humans still set the goals and decide what counts as progress. That asymmetry isn't closing. 2. The composition gap: strong performance on individual components rarely composes into reliable end-to-end improvement. A system can beat every component benchmark and still fail the full loop. Most "self-improving AI" claims live in this gap. Search the catalog: https://huggingface.co/spaces/KaiWu123/awesome-ai4ai Dataset: https://huggingface.co/datasets/KaiWu123/awesome-ai4ai 📄 Nice Layout Paper: https://github.com/KaiWU5/Awesome-AI4AI/blob/main/assets/AI4AI-Survey.pdf 🌐 Site: https://kaiwu5.github.io/Awesome-AI4AI/ ⭐ Repo: https://github.com/KaiWU5/Awesome-AI4AI 📄 Paper: https://www.preprints.org/manuscript/202608.2108/v1 Disagreement welcome — especially on the composition gap.
View all activity

Organizations

None yet