Arthur Yu

Building soloop. Ex-Baidu MeDo, YC China (F24). The world has enough demos. I write for solo founders trying to make hard-won insight earn.

[Arthur Yu @ArthurYu_Soloop]

I used to run MeDo at Baidu, a vibe coding product.
Vibe coding made everyone feel like a founder.
Most are not. They are demo collectors.
Building is cheap now. Getting one stranger to pay is still brutal.
I'm building Soloop to kill that illusion.
First dollars or nothing.

[Arthur Yu @ArthurYu_Soloop]

Most vibe-coded products aren’t ready for $1 of ad spend.

I gave our team a rough scale in a meeting:
A vibe-coded product is usually 1/10.
A product worth putting Google Ads behind is 8/10.
PMF is 10/10.
We already have coding agents and GTM agents. We still don’t have a testing agent we trust to sit between them and say, “Don’t buy traffic yet.”

Its job is straightforward:
- Can a new user finish the main workflow?
- Can they sign up, log back in, and pay?
- Did the product save the right result?
- If something breaks, can the agent show exactly where?

If the main flow fails, there’s no MVP. If signup, login, or payment fails, there’s no commercial MVP.
Browser agents can already perform most of these steps. We need a reliable pass/fail verdict, evidence for every failure, and a loop back to the coding agent.

Until then, GTM agents are just spending money to find bugs for us.

[Arthur Yu @ArthurYu_Soloop]

@coreyganim Hmm. 6 agents passing tells you almost nothing about the 7th.
Missing input, human approval, recovery, and wrong-destination cases should block release.

[Arthur Yu @ArthurYu_Soloop]

@michael_chomsky yup, multi-day agents exposed the real bottleneck:
the shared compute pool.

One runaway run shouldn’t quietly degrade every engineer’s experience.

[Arthur Yu @ArthurYu_Soloop]

@yongfook Five years feels plausible.

The SaaS that lasts probably keeps the same customer problem and rebuilds its learning loop a few times.

[Arthur Yu @ArthurYu_Soloop]

@lennysan Fair. Though that CS + McKinsey + Series A AI PM is really shorthand for judgment under ambiguity. If the person doesn't exist, the training loop needs decisions, context, and outcomes together.

[Arthur Yu @ArthurYu_Soloop]

@arvidkahl Complaints can become product changes now.
The agent still has to check the real account before it says 'done.' That's how we caught blocked Reddit access and an unsaved Google targeting state after both workflows looked complete.

[Arthur Yu @ArthurYu_Soloop]

This week I thought two Soloop workflows were done: Reddit ops and Google Ads. Then we ran them on real accounts.

Reddit returned “blocked by network security.” Google showed five target locations but saved zero.

A few hours later, the workflows finally matched the screens.

[Soloop @soloopai]

Soloop now supports Reddit ops and Google Ads setup.

Real-account tests caught two failures: a Reddit session blocked at the network layer, and Google showing five targets while saving zero.

We added access and saved-state checks before either workflow can report completion.

[Arthur Yu @ArthurYu_Soloop]

@yongfook Well, it's fine as a Friday experiment. But I’d stop it the moment the $299 clone needs a roadmap.

[Arthur Yu @ArthurYu_Soloop]

@ycombinator @patrickc @collision @stripe Back at Baidu MeDo, fast builds were everywhere. Staying with the problem after the estimate proved wrong was rare, and AI still doesn’t compress that.

[Arthur Yu @ArthurYu_Soloop]

@lennysan Speed makes systems thinking visible. Bad second-order effects now ship faster.

[Arthur Yu @ArthurYu_Soloop]

@ycombinator @josh__france @jbellregan @koomen Good to see this open-sourced. A company should first agree on what deserves remembering.

[Arthur Yu @ArthurYu_Soloop]

@d4m1n Congrats!!! AI built software but 1 stranger paying proved there was a business. You just made it.

[Arthur Yu @ArthurYu_Soloop]

Agree. Apple understood this with the iPod.

“1,000 songs in your pocket” gave the company a clear user outcome before it became an ad line. The team could use that outcome to judge product decisions.

Without the same internal clarity, priorities drift, features point in different directions, and customers have to work out what the product is for.

[Arthur Yu @ArthurYu_Soloop]

@ycombinator @JeffDean Well, cheap arithmetic can unlock building. User evidence decides what deserves building.

[Arthur Yu @ArthurYu_Soloop]

@bentossell lol the complication is remembering which agent currently owns the context.

[Arthur Yu @ArthurYu_Soloop]

@SMB_Attorney A one-person company still needs a real person accountable for the broken loop tho.

[Arthur Yu @ArthurYu_Soloop]

@shreyas Taste is what survives use. Signaling only has to survive the feed.

[Arthur Yu @ArthurYu_Soloop]

@ycombinator @FrancoisChauba1 @stuart_sul When AI writes systems code, what becomes the bottleneck: correctness, observability, or hardware-aware taste?

[Arthur Yu @ArthurYu_Soloop]

@gokulr @t_xu @DoorDash @jaltma "AI-native" doesn't mean code faster, but learn faster across the rest of the work.

[Arthur Yu @ArthurYu_Soloop]

@dhh emm AI is best at migrations nobody could justify staffing.