Weekly Digest 10

AI

News

New models:

• GLM-4.5 & GLM-4.5-Air

This months, many Chinese company open sourced their models, I started a new habit to give every one of them a try when they come out.

Models

Can small AI models think as well as large ones?

Pro:

• “cognitive core” is effective on training.
• Large models encoding more facts about the world than small ones, which maybe unnecessary and inefficient.

Con:

• reasoning ability might require a lot of background knowledge.
• reasoning ability requires many parameters.
• larger models will always be smarter.

Coding

Andrej Karpathy coined “vibe coding” is where the idea of “vibe coding” comes from.

Vibe code is legacy code: vibe coding is good for short-term prototypes, serious coding is good for real products. Both of them can be enhanced by AI in different ways.

LLM Application

Learning the Bitter Lesson

1. Understand your application structure
2. Re-evaluate your structure as models improve
3. Make it easy to remove structure