AI4AI → AI4Human. lol. Now I just need it to do the dishes. 😂
kai
KaiWu123
AI & ML interests
llm coding
Recent Activity
published a dataset 11 days ago
KaiWu123/autoreport_llm_train_20260913 updated a Space about 1 month ago
KaiWu123/awesome-ai4ai updated a dataset about 1 month ago
KaiWu123/awesome-ai4aiOrganizations
None yet
replied to their post about 1 month ago
reacted to their post with 🔥🚀 about 1 month ago
Post
2217
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: KaiWu123/awesome-ai4ai
Dataset: 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.
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: KaiWu123/awesome-ai4ai
Dataset: 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.
- 2 replies
posted an update about 1 month ago
Post
2217
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: KaiWu123/awesome-ai4ai
Dataset: 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.
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: KaiWu123/awesome-ai4ai
Dataset: 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.
- 2 replies