James Lu

Exploring AI agents, workflow automation.

Experiments, products, lessons

[James Lu @JamesLuAka]

@jeantimex @threejs Bro, i thought it's true😅 This video depicts AI-generated footage falsely presented as behind-the-scenes filming of a disaster movie

[James Lu @JamesLuAka]

Hope they not, i really like the handcrafted things.

[jeantimex @jeantimex]

One day, all they really have to do is just use @threejs.

[James Lu @JamesLuAka]

当你打开 YouTube 或 X：

一开始只是想看一个教程。

结果：

教程 → AI Agent → Google 新模型 → 某创业者收入 → 美国新闻 → 搞笑视频 → 半小时过去。

每一个视频可能都"有价值"，但结果就是注意力被当猴耍了。🙉

[James Lu @JamesLuAka]

我最近搞懂一件事：把开放题变成「每天可套用的选择题」，可以快速确定方向和进展。

[James Lu @JamesLuAka]

@cholf5 居然能刷到技术栈那么相似的小伙伴，这种click游戏不能直接拿饼干的数值改一改吗

[James Lu @JamesLuAka]

@OGodbuilder @CC_kyjek 还是OG哥靠谱，我都一直在拿账号做实验，还没实验出来个名堂😂

[James Lu @JamesLuAka]

@OGodbuilder 属于是被ai实锤不如ai了😅

[James Lu @JamesLuAka]

用AI 做 web 的时候才发现之前自己所谓的设计水平是如此平庸，
基本上就是现在标准ai web模板，
card+简约+留白，
就连纯字logo的偷懒方式也是出奇的一致。。🫠

[James Lu @JamesLuAka]

@OGodbuilder 这不是巧了嘛，每个人剧本最后一页都一样💀

[James Lu @JamesLuAka]

Scrolling X can be mentally exhausting, it can also leave you feeling surprisingly anxious.

So I built a tiny tool for myself to filter out the noise.

Right now it just lets me journal and schedule posts.

Not sure where it'll go yet, so for now... I'm just using it myself. 🫠

[James Lu @JamesLuAka]

After digging deeper into Ethereum's execution model,
I understand why it chose the account model over UTXO.

For simple value transfer,
UTXO feels more natural while the account model makes state management much simpler

[James Lu @JamesLuAka]

2 major issues when automating upgrades of legacy projects with AI:

1. LLM tends to “approximate” fixes instead of strictly following rules.
2. When you add too many correction scripts, they start conflicting with each other and create inconsistent fixes.

[James Lu @JamesLuAka]

AI output is statistical, not deterministic.

Same prompt can produce different results.

You need extra pipeline steps to control variance.

[James Lu @JamesLuAka]

Object pooling makes sense in game, but usually not in flutter/dart.

dart’s GC is generational and optimized for short-lived objects. Pooling can actually confuse allocation patterns and reduce GC efficiency.
sometimes “reusing objects” is worse than just letting them die young.

[James Lu @JamesLuAka]

In V8, closure is a heap function object that references a heap-allocated (or optimized-away) lexical context, depending on escape analysis

So scope ≠ lifetime.

[James Lu @JamesLuAka]

I stopped letting AI rewrite full files.

Now it only patches TODOs.

Less chaos, more control.

[James Lu @JamesLuAka]

One of the biggest lessons I've learned from AI coding:

Don't just tell the AI what's broken.
Tell it why it's broken.

Otherwise, it'll keep fixing surface bugs instead of the actual problem..

Bug reports are helpful.
Root cause analysis is even better.

[James Lu @JamesLuAka]

Hermes agent shifts tool-aware, skills into execution loop.

In my agent (LangGraph-based GOAP-style): intent → planning → execution

Trade-off:
GOAP = structured, controllable
Loop = adaptive, less predictable

Still exploring balance.

[James Lu @JamesLuAka]

最近在弄一个把老项目自动迁移到其他framework的自动流水线，发现LLM一直错打部分code，搞半天是Context过长导致模型漂移了。
看来 pipline 越往后，就越应该限制LLM 的自由度。

[James Lu @JamesLuAka]

@RaymondLeeCN 我账号之前就是什么都没发就被封了。。 后来申诉说是系统误判了。。