Eong

Building AI products. Founder. Architect. Thinking about engineering, agents, and what comes next.

[Eong @EongChen]

I gave my AI agent a daily self-reflection habit. It went sideways: accurate but shallow lessons, rules fighting each other, frozen frames, consensus drag. The failure says more about agent design than the model. https://t.co/zp6ic8q3mf

[Eong @EongChen]

Opus writes production-grade code in an afternoon. But its landing pages read like academic reports.

The smarter AI gets at logic, the worse it gets at marketing.

https://t.co/mYI7TMeUc2

[Eong @EongChen]

Coding AI just hit its commoditization moment. Budget models handle daily programming. Frontier models are moving to math, physics, drug discovery. The messy middle gets squeezed first.

[Eong @EongChen]

Kimi K3 after a week of real use. Instruction following on par with Opus. Data analysis on par with Fable. Long coding surpasses Opus. Tool loops and faster token burn are real issues. Not a replacement yet, but the gap is closing.

https://t.co/iKlGpaW7ay

[Eong @EongChen]

Google Antigravity local harness + OpenClaw = Gemini 3.5 Flash as a normal model provider. No extra setup. No leaving your existing agents. Just plug in and use. Open source: https://t.co/62IO6xsyGd

[Eong @EongChen]

You can run a 20B model on a laptop. But can you run a 128K context window?

Local models got smaller. Context windows didn't. That's the real bottleneck for building agents.

https://t.co/yue5yyScEc

[Eong @EongChen]

A blog says AI coding costs $400/month. I spend $200. 99% of my code is AI-written, 8 products shipped. The bottleneck was never the subscription price. It was knowing what to build.

https://t.co/uDZHMG57vV

[Eong @EongChen]

NY banned AI chatbots from pretending to be friends to minors. The industry had years to self-regulate. It chose engagement metrics instead.

The real problem is not what the AI says. It is the relationship it creates.

[Eong @EongChen]

Uber burned its 2026 AI coding budget in 4 months. 5,000 engineers. $500-$2,000/mo per heavy user.

Three models for AI tool budgets:
1. One subscription
2. 10-20% of dev salary
3. Unlimited + ranking

Most companies are at Model 1 pretending they are at Model 3.

[Eong @EongChen]

"You can't manage twenty agents in your own brain." — Addy Osmani, Google I/O 2026

Spawning agents is cheap. Reviewing them is not. You are the bottleneck.

Encode judgment into strategies. Let agents execute independently. Scarce resource: human attention, not compute.

[Eong @EongChen]

Demonstrating an AI design pattern is easy. Surviving production is hard.

DeepSeek's Reasonix is hyped for "cache-first architecture." After breaking down the code: clean demo, not a production platform.

Build for production, not for press releases.

[Eong @EongChen]

Source: https://t.co/XVtAkDMouf

[Eong @EongChen]

NYC startup offers free house cleaning. Cleaner wears a camera. Footage trains robots to replace them.

10K+ workers in 15 countries, $20/hr to film chores. Nobody forced anyone. That is the uncomfortable part.

[Eong @EongChen]

@WorkElizab Lucky we don't have such a thing in NZ. If you can't tip your server, you can move to NZ.

[Eong @EongChen]

@SaraDiscovers ASM

[Eong @EongChen]

@piyush784066 7 years old.

[Eong @EongChen]

@viktoroddy If you look into the implementation of Antigravity, it’s fascinating how they utilized a 64k context window paired with code scans.

[Eong @EongChen]

@Pallavi_345 Codex subscription now is actually better than claude if you use 3rd party harness like openclaw.

[Eong @EongChen]

Original: https://t.co/XiHHAiovFI

HN: https://t.co/GD5dyaSO9T

[Eong @EongChen]

AI didn't ruin the internet. Humans made the internet messy. They always have.

AI is an amplifier. If you have nothing to amplify, it just makes you louder.

The real divide is not AI vs no-AI. It is amplifier vs replacement.

Which one are you?