Tsz Ting Chung

Ph.D candidate @HKUST

[Kyver @0xkyver]

HARD TRUTH: Standard vector databases are making your AI agents stupid.

We feed agents semantic similarity. "Here are 5 docs with the same keywords!" The agent doesn't need keywords. It needs intent.

InsightEmb changes the game. It retrieves progress-oriented experiences. It teaches agents to remember HOW to solve a bottleneck, not just what words were used.

Stop building digital hoarders. Start building thinkers. Deep dive in the article below.

[Sumit @_reachsumit]

InsightEmb: Learning Action-Intent Embeddings for Agentic Insight Retrieval

@ttchungc et al. at Tencent presents a contrastive embedding, trained only on math data, that retrieves insights matching an agent's current bottleneck.

📝 https://t.co/686aqyVnKU

[Tsz Ting Chung @ttchungc]

Thanks @_akhaliq for sharing our work!

Please see https://t.co/Lm6SUPZio9 for more details.

Join our discussions at https://t.co/AoKGktYYLB!

[AK @_akhaliq]

The Stochastic Parrot on LLM's Shoulder

A Summative Assessment of Physical Concept Understanding

[Junjie Wu @jieeijjie]

🚀 Introducing PhysiCo: A New Benchmark for Evaluating Abstract Understanding in LLMs! 🚀

📚Link: https://t.co/TKqXJaAqJy

While models like o3 have made impressive strides on ARC-AGI, how well do LLMs truly grasp the abstract patterns in ARC-style tasks?

(1/5)