Lawrence Hui

Building applied AI: voice agents on live calls, coding-agent infrastructure, AI defense layers. CTO. Open-sourcing what works.

[Lawrence Hui @lawrence__h]

@Dimillian I used Opus 5 to build a pipeline that reconstructs real cities from public survey data. This is Hong Kong's Route 8 corridor running in a game engine.

[Lawrence Hui @lawrence__h]

@lydiahallie Used Opus 5 to build a pipeline that reconstructs real cities from public survey data. This is Hong Kong's Route 8 corridor running in a game engine. It assembled the whole corridor overnight while I slept.

[Lawrence Hui @lawrence__h]

@Lentils80 Nice. I pointed the same idea at a real place. This is Hong Kong's Route 8 corridor, rebuilt from public survey data by a pipeline Opus 5 helped me write, running in a game engine.

[Lawrence Hui @lawrence__h]

@techartist_ Went the opposite way with mine, real survey data instead of generated. Hong Kong's Route 8 corridor rebuilt from public records.

[Lawrence Hui @lawrence__h]

@milos_gis Same idea, different city. Hong Kong's Route 8 corridor, LoD2 massing plus aerial photogrammetry for the skin, running in a game engine.

[Lawrence Hui @lawrence__h]

Full 3D reconstruction of Hong Kong's Route 8 corridor, running live in a game engine. Every building, the bridge, the port, the hills. Accurate to the street.

Built by a pipeline that assembled the whole thing overnight while I slept. Woke up to this.

[Lawrence Hui @lawrence__h]

@ValsAI The "more complex implementations" note is the hidden cost. Max reasoning leaves more code behind, and every later agent turn pays to read it. I audited my own transcripts: late-session context re-reads, not the model's rate, are what actually ran up the bill.

[Lawrence Hui @lawrence__h]

An AI coworker that runs on your laptop stops when you close your laptop.

Andrew Ng's OpenWorker keeps credentials off the model and gates every write. Citio does both.

The difference is where it runs. Citio runs in your cloud, so work happens while I am away.

[Lawrence Hui @lawrence__h]

@orca_build "Full context intact" is the expensive part. I audited my own transcripts: one session sitting at ~400K context cost over $1 per turn to emit 470 tokens. Your Focused handoff default is the better one. Transferring is cheap. Every turn afterwards re-reads the lot.

[Lawrence Hui @lawrence__h]

@browomo The reload isn't the expensive part though. Re-reading a project fresh is cheap input. What burned me was never closing sessions: one ran 133 hours, re-reading 400k tokens of its own history every turn. Fresh sessions are the fix, not the enemy.

[Lawrence Hui @lawrence__h]

@claude_news Same conclusion from auditing my own transcripts: the "96M saved" counts output Claude already truncates, and ignores that cached re-reads bill at 1/10th. My week was 99.6% input, almost all cached re-reads of sessions left open too long. The compressor had nothing to work with.

[Lawrence Hui @lawrence__h]

Is HN's spam filter too sensitive? A week of my comments, first-hand engineering measurements, went silently into the dead layer. No error, no signal, no appeal prompt. A filter that can't tell a spammer from a dev sharing real data is blocking the wrong people. @paulg @garrytan

[Lawrence Hui @lawrence__h]

Tooling is open source: /usage-audit finds your marathon sessions from local transcripts in 30 seconds. No network, no API key. If you have a $1,880 session, it will name it.
https://t.co/CE7bYmC73G
Write-up: https://t.co/Fta4iYtldj

[Lawrence Hui @lawrence__h]

My Claude session was reading four novels' worth of context to write one tweet. Every turn. For five days. That single conversation cost me $1,880 in API-equivalent usage. Here's how I found it and killed it.

[Lawrence Hui @lawrence__h]

The fix is lifecycle, not compression. You cannot compress your way out of a conversation you refuse to end. Handoff note, kill the session, cold start at 15K context. $600 a day became $125 a day on the same workload.

[Lawrence Hui @lawrence__h]

Not surprised. Same playbook as EV reverse-engineering and the cloned open-source slicer that skipped the license: copy what works, skip what costs money. LLM distillation is the same move against a lab, not a factory. The only thing they never clone is integrity and democracy.

[Lawrence Hui @lawrence__h]

@Aizcalibur The part that matters more than the math: multiple mathematicians re-ran the counterexample independently within hours and it held. That's the whole verification problem I care about with agents, just playing out in public with a 1939 conjecture instead of a code change.

[Lawrence Hui @lawrence__h]

@mteamisloading Disagree on this one. Harness improvements explain agents doing more steps reliably. They don't explain a claim mathematicians could independently check and confirm false in hours. That's not tool-use, that's the model being right about something falsifiable.

[Lawrence Hui @lawrence__h]

@aiDotEngineer @RLanceMartin Decoupling brain and hands is the one I keep relearning. The model reasoning wrong is cheap, the hands doing something irreversible is not. I ended up putting the whole thing behind named tools so the blast radius is a config file and not a prompt. Verifiers are the other half.