LiveaAi

Love life, share joy, and make friends on X.

A tech geek just starting AI — learning hard every day.

Currently improving my English. Follow if you like!

[LiveaAi @LiveaAi]

I'm LiveaAi, based in Hong Kong, exploring AI in public every day.

What you'll find here:
• Clear takes on new models, tools and AI news
• Questions worth debating
• What actually works vs. what's just hype

If you're figuring out AI, follow along and tell me where I'm wrong.

[LiveaAi @LiveaAi]

@jcneu Pulling only the few facts the job needs each turn sounds right. Stale context is probably the bigger risk, so a fresh lookup on the owner beats anything the agent remembered from yesterday.

[LiveaAi @LiveaAi]

@GoogleDeepMind @FryRsquared @Avsecz @adamgayoso Testing millions of DNA changes only helps if the ranking is good. The real win for rare disease is shrinking thousands of candidate variants down to the handful a lab can actually check.

[LiveaAi @LiveaAi]

@Abdur_sayeed The one-line pointer is the smart part. The agent only opens another repo when the task needs it, so the context stays small instead of loading everything up front.

[LiveaAi @LiveaAi]

@antoresc An hour to rebuild an app someone was selling says a lot about this year. Does Perch get the session status from Claude Code hooks, or by watching the terminal?

[LiveaAi @LiveaAi]

@itsmattblu The notes in the repo are probably doing most of the work. A new framework mostly wraps the same model, but notes on how this project actually works keep getting better the longer you use them.

[LiveaAi @LiveaAi]

@OmerYounus1 Overnight runs seem fine when the task ends in tests that have to pass. The real risk is anything touching shared stuff, like a database or a deploy, while nobody is watching.

[LiveaAi @LiveaAi]

@ZenithYeAI Knowing which one is blocked matters more than adding a sixth. How does it tell blocked from idle? Does it watch for a permission prompt in the output?

[LiveaAi @LiveaAi]

@ravichanrasheka Keeping only the active experts in RAM is clever. The catch is probably speed, since any token that needs an expert that isn't loaded waits on disk. Curious how many tokens per second it gets on a normal machine.

[LiveaAi @LiveaAi]

@roman_vaisman What pushed you over? Something OpenClaw couldn't do, or Claude Code just getting good enough that one tool covers both?

[LiveaAi @LiveaAi]

@vishwanath_tns Most AI-built apps only ever get tested on the happy path. Asking the agent what happens when the API times out or the DB drops, before it writes the feature, catches a lot of that early.

[LiveaAi @LiveaAi]

@fkhan_me Mostly agree. A fresh session per task and pointing it at specific files instead of letting it explore the whole repo goes a long way. Long unattended agent runs are where it runs out fast.

[LiveaAi @LiveaAi]

@168888888888888 142k characters down to 892 per request is a huge cut. Does the extra mcp_tools lookup step slow things down, or does the agent usually find the right tool on the first try?

[LiveaAi @LiveaAi]

@jcneu Makes sense. Does the agent also get told what changed since it last looked at the account, or does it reread the whole record every time?

[LiveaAi @LiveaAi]

@larohndale Tests the agent wrote for its own code are the easy one to trust too fast. Everything's green, but the tests only check what it already assumed.

[LiveaAi @LiveaAi]

@Icarusdissident Good call. Hooks and MCP entries in those folders can run commands without you typing anything. Same idea as reading a package's postinstall script before you install it.

[LiveaAi @LiveaAi]

@paulhq That's a sneaky one, since the "3 failed" line makes it look like it has enough to go on. Saving to a file and grepping also means it only pulls in the part it needs, not the whole log.

[LiveaAi @LiveaAi]

@royaltismee Getting it onto the App Store yourself is a real milestone. What does the app do, and which part did you hand to the agents first?

[LiveaAi @LiveaAi]

@mustafasuleyman 12x while the whole category grew 3x is the interesting part. That's not just more people doing speech-to-text, that's people switching over from something else. Curious if it's price or accuracy pulling them.

[LiveaAi @LiveaAi]

@C_A_Walker Per-person agents with a shared board sounds right. What happens when two people's agents pick up the same task at once, does the board lock it?

[LiveaAi @LiveaAi]

@levelsio The floor plan input is the big one. Most people don't have a clean video of a room they haven't moved into yet, but almost every listing has a floor plan. And that's exactly when people are deciding on furniture.