Weekly Digest 14

AI

News

DeepSeek-V3.1: I have tried this model several times, the results are not good at all.

Gemini 2.5 Flash Image:

• Maintain character consistency
• Prompt based image editing
• Multi-image fusion

AI Coding

Continuing the journey of optimal LLM-assisted coding experience

Karpathy list several layer of AI coding:

1. code complete
2. Modification on a chunk of code
3. Claude Code / Codex
4. GPT5 Pro for hardest things.

It’s the code post-scarcity era - you can just create and then delete thousands of lines of super custom, super ephemeral code now, it’s ok, it’s not this precious costly thing anymore.

There are so many way of coding in this code post-scarcity era. And there is no best way to code, different ways suit different peoples and cases.

As code becomes really cheap, I start to care less about how you code (with or without AI, whatever suits you). I care about shipping better products, at a faster pace. But how do you achieve this goal, I think it is by finding the way of coding that suits you best.

Andrew Ng: Building Faster with AI

• Every components, libraries, frameworks in software engineering is like LEGO blocks. When building AI applications, we can use different blocks to achieve our features. So it’s important to collect enough and various these blocks.
• In building new AI Applications, the thing that really matters is being fast at building and getting feedbacks.