AI Research Trends from Personalized LLMs to JEPA World Models
연구/벤치마크 | Wed Jul 15 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 3 sources
Gwern proposed the Guardian Angel personalized LLM, a JEPA-based Super Mario world model experiment was conducted, and original ELIZA source code was rediscovered.
Analysis
[Gwern] proposed the Guardian Angel personalized LLM concept [1]
- A digital twin LLM that emulates the user's values and preferences
- Real-time updates based on dynamic evaluation and online learning
- Cognitive security defense against prompt injection and spearphishing
- Mitigates principal-agent problems by unifying principal and agent
[LeMario] conducted a JEPA world model training experiment based on Super Mario Bros [2]
- Reimplemented LeCun's Joint-Embedding Predictive Architecture from scratch
- Used a 192-dimensional latent and a causal predictor with 6 transformer blocks
- Injected actions in shift
- scale
- and gate form via AdaLN-Zero
- Succeeded in 5-step future prediction and generalization to held-out episodes
[Inventing ELIZA] unearthed and reinterpreted the original ELIZA source code from the MIT Archives [3]
- Revealed various scripts beyond the DOCTOR persona and new conversation records
- Revisited the phenomenon of human emotional attachment observed by Weizenbaum
- Connected to the psychology of current ChatGPT users sharing secrets
- Corrected and restored common beliefs about chatbot history