England Tong

Hong Kong / Shanghai · QC & production background · university-era coder returning via AI agents & offline learning tools

[England Tong @englandtong]

Always loved tech. Electronic engineering, early coding, then 20 years in sourcing, supply chain and food QC. Now building again: X-Agent, open source, drives existing systems with AI.

Thesis: ERP shouldn't force humans to adapt to forms. The system should understand the human.

[England Tong @englandtong]

A developer used AI to clone all seven Adobe apps in a few days. Panels, menus, filters — it looks finished. Then people actually used it, and everything broke.

I recognized that story immediately. It's mine.

I need an ERP for work. The off-the-shelf systems never fit — they don't match the specifics of our industry, and customizing one would cost a fortune. So a lot of work still ended up in manual Excel sheets.

Once I could use AI, I figured — why not just generate my own?

AI turned my ideas into working code fast. It looked genuinely solid. For a moment I felt amplified — like I could build anything. The ideas kept coming, and the features piled on: purchasing, sales, warehousing, reports, permissions… the whole thing kept getting bigger.

Two things actually killed it. First: the system was huge, every module wired into the others — touch one piece and everything moved. Second: the code was all brand new. Every feature had to be tested, or I couldn't trust the Coding Agent wrote it right.

Don't verify now, and once real data is flowing in and something breaks, I'd never trace it back to where the mistake was introduced.

Yet a system that big and interconnected can't be tested step by step. If you can't test it incrementally, you can't trust it — so it never shipped. It just sat there, almost done and unusable.

I forced myself to step back and ask one question: What's the single feature I could finish FIRST — and verify on its own?

The answer was inventory. The most self-contained piece, fewest connections to everything else. It could run and be tested independently. So I built only inventory, and did it properly. It actually ran.

I wish the story ended there. The moment inventory worked, I finally felt a real sense of achievement — and I thought the hard part was over.

It wasn't. Once people used it for real, the hard parts turned out to be the ones never on the list: problems only real data reveals, needs users couldn't articulate, hidden connections between features.

You can't think those up in a room, and more tokens won't solve them. They take real usage, continuous feedback, clarifying what's actually needed, and reworking again and again.

That's exactly where I am now — and it's the actual hard part.

AI made "coding it" fast. It hasn't made "getting it right" any easier.

#BuildInPublic #AI #IndieDev #ERP #CodingAgent

[England Tong @englandtong]

What is "Wanfeng"? Two words. It took me 143 years to understand them.

The voice you hear is not human. He is Yao — the boy's companion AI.

In the near future, AI has already mapped the optimal path for everyone: the fastest route, the best trade, the safest life. In that world there is one unwritten law — helping others counts for nothing. The system never records it. The only price is your own pace.

Yet there is always one child who stops at the rock wall, turns back, and reaches down.

Not a single line of dialogue. The only voice is that AI. 2m45s. Every frame generated by AI.

WANFENG — prologue to Part One: BEYOND PERMISSION.

#RunningHub #AIGC #ChineseSciFi

[England Tong @englandtong]

I basically fumbled through my first stretch with AI.
Earlier I mentioned I started for real this April, with coding agents. Honestly, I had no idea what I was doing — I was just excited, bumping into everything.

Every time I saw a new tool or model, I installed it and tried it. I registered a pile of accounts and wanted to touch every single one.

Meanwhile I kept having new ideas, so I kept asking agents to build them. But I couldn't describe what I actually wanted. The thing it built kept going off track — not what I meant. Lots of projects never got finished at all.

Sometimes it just kept writing on its own, drifting further and further away. The worst one: a single goal inside one project kept running for 30+ hours straight. By the time I noticed, I couldn't even understand what it was writing anymore — and there was no pulling it back.

I tried all sorts of things. Results were hit or miss. Quite a few projects got started and then just sat there.

After a while, one plain lesson slowly clicked: AI is like a super capable beginner. It can do a lot — but it has no idea what you actually want, or what "done" looks like.
So don't try to do everything at once. Pick one small goal, explain it clearly, finish it, then add the next piece.

Here's what makes this different from working with a person: AI never rests, and never complains that you didn't think it through the first time, that you keep changing your mind. So the cost of trial is tiny. Add one feature at a time. Tear it down and start over — it's fine.

That fumbling stretch wasn't wasted. At least I've now seen where the traps are.

#AI #BuildInPublic #CodingAgent #IndieDevt

[England Tong @englandtong]

I started following AI (SI now?) back in 2025 — took some courses, read the news — but only used it casually then, mostly plain chat-based AI. Strictly speaking I wasn't a complete beginner, but I'd never really gotten into it before.

I only started really using AI this April. I began by trying different coding agents — Codex, OpenCode, Cursor, CodeBuddy, WorkBuddy, Trae, Qoder, and others — plus various LLMs. At first I was pretty pragmatic: I'd jump on whatever had free credits or freebies to grab 🤣, and the moment I saw a new feature or model I'd try it.

My usage opened up gradually. At the start, seeing a few hundred thousand tokens used made me nervous; later, at peak, I'd burn over a billion tokens in a single day. I began using agents for coding and projects; now I'm starting to try generating video and images.

After half a year of intensive use, I've gone from a rookie with little hands-on experience to a slightly experienced beginner. I'm keeping these notes as a daily log — for myself a record, and if anyone finds them useful, a way to learn and avoid some pitfalls.

#SI #AI

[England Tong @englandtong]

@IshimweE17047 B

[England Tong @englandtong]

@elonmusk @bot I have talked to Grok Bot. I hope her voice will be more warm 🤣🤣

[England Tong @englandtong]

@ChineseEmbinUS Hope to have a good news.

[England Tong @englandtong]

@dexhorthy @mattpocockuk @humanlayer_dev Useful. A “show me” gate before calling a PR done feels right — less slop, more to check.

[England Tong @englandtong]

I mean the model of WNA is one of the options.

[England Tong @englandtong]

The World Nuclear Association (WNA) may be one of the solutions for AI safety

@WorldNuclear #AI

[England Tong @englandtong]

@thsottiaux To press the button?!

[England Tong @englandtong]

Without the powerful compute infrastructure, I don't think AI progress can reach the current level. Huge credit to @nvidia @intel @AMD

I don't see AI spiraling out of control anytime soon - Right now we don't even have enough chip operation capacity and power supply to run systems at the scale to pose that kind of risk 🤣However, it is the right moment to start seriously planning for the long-term risk ahead.
@JensenHuang @sama @DarioAmodei

[England Tong @englandtong]

Limits of Large Language Models: Innovation, Research and Engineering

1. Human Growth Limitation: Human learning has inherent capacity ceilings. People progress rapidly at the initial stage, yet improvement slows sharply once approaching personal limits. Experts simply have higher capability ceilings than ordinary people. Standard learning and experience accumulation only bring minor incremental improvements, and cannot cultivate talents with disruptive, original innovation.

2. Inherent Defects of Current LLMs: Existing large language models have no real intelligence or original creativity. They only optimize and seek optimal solutions within established knowledge and fixed frameworks, merely iterating and accelerating human existing research achievements. The current AI boom is essentially a secondary outbreak of accumulated human scientific results, rather than revolutionary original innovation. Moreover, LLM development relies heavily on high-quality data, and its progress will stagnate with insufficient data supply.

3. Essence of Modern AI Development: The rapid advancement of contemporary AI owes to upgraded engineering capabilities including computing power, chips and energy systems, instead of breakthroughs in underlying theoretical innovation. Such engineering advantages are not unique to the present era.

There are essential differences between original pioneering innovation, iterative scientific research and engineering implementation. Current LLMs are only competent at optimization and engineering execution, lacking the core ability of zero-to-one original creation, which is the fundamental reason why they cannot generate true intelligence.

@JensenHuang @nvidia @intel @AMD

[England Tong @englandtong]

@JessePeltan Solar power has great potential to become one of the main power sources in future

[England Tong @englandtong]

@JessePeltan It is only in Texas?

[England Tong @englandtong]

@brandilwells Same question to me!!

[England Tong @englandtong]

@ai That's true. I am using 10+ agents.... Sometimes I forget who is who 🤣

[England Tong @englandtong]

@omarsar0 "Build your own harness" is the expensive framing. If the domain already has a system of record, the harness was paid for years ago: schema, constraints, permissions.

[England Tong @englandtong]

@saranormous It depends. For open-ended research, borrow someone else's. For a domain with an existing system of record, the best harness is inherited, not built: schema, constraints, permissions. Put the orchestration there—not in the prompts.

[England Tong @englandtong]

@yanivf @garrytan The harness isn't something that has to reach twenty years of data — it IS that record, and the schema, constraints and permissions are what made it canonical.

The model only reads a sentence and returns a verb.