DEV Community: zifeng peng

PostgreSQL 19's delay shows the new bottleneck: deciding which AI findings matter

zifeng peng — Sat, 17 Oct 2026 06:07:39 +0000

PostgreSQL 19 is now expected around November 19. One of its committers, Peter Eisentraut, wrote that LLM-assisted code review was surfacing defects faster than the team could fix them, and roughly 8 to 10 significant features were reverted from the 19 branch after beta began. (The Register ran it as "AI reviewers delayed Postgres"; the project's own reasons are a bit more nuanced, but the pattern is worth noticing.)

Finding problems got cheap

A year ago the scarce thing was a careful reviewer. Now you can point several models at the same change and get a long list of concerns in minutes. That is great, until the list is longer than your week.

The hard part has moved:

Which of these findings are real, and which are a model being confidently wrong?
Which ones actually block a release, and which can wait?
Who owns that call?

What I do as a solo builder

I build Brainstorm, a Linux app (1.0.1) where several AIs discuss the same goal. The workflow I ended up with is boring on purpose:

Let more than one model argue about the plan, so disagreement shows up early instead of in review.
Keep the check separate from the agents that wrote the draft, so "it ran smoothly" doesn't get mistaken for "it's right".
Nothing reaches the project's main branch until a human approves it.

The point isn't more findings. It's a trail of who disagreed, what got checked, and who signed off.

You bring your own model accounts; I don't sell model credits.

If you're dealing with an AI review flood on your own projects, I'd like to hear how you triage it.

https://brainstorm.mungwaan.cn/en/

Vibe coding isn’t the bug — shipping without a disagreement trail is

zifeng peng — Sat, 17 Oct 2026 02:13:36 +0000

I almost shipped a feature that one model loved.

I vibe-coded it in a single tab. It felt clean. Confidence was high.

Then I ran the same brief through a second model. It flagged a silent failure path the first one never mentioned.

That was the real lesson for me as a solo builder:

Vibe coding isn’t the problem. Rubber-stamp confidence is.

What I want before anything leaves my machine:

Multiple models answer the same question (BYOK — bring your own keys)
They disagree in one room
I audit the trail
Then I export

I’m building Brainstorm as that shell: debate → audit → export. It does not sell model credits. Linux first. $6/mo.

Site: https://brainstorm.mungwaan.cn

If you already hop between ChatGPT / Claude / Gemini tabs for important decisions, you’re halfway there — the missing piece is a forced disagreement step before you trust the answer.

I built a $6/mo BYOK multi-AI debate workspace (not another credit pack)

zifeng peng — Fri, 16 Oct 2026 15:42:57 +0000

I already pay for several model APIs/subscriptions, but serious work still means hopping between official apps. A lot of "all-in-one" tools just resell their own credits.

So I built Brainstorm: a local-first Linux workspace where you connect your own keys (BYOK), assign roles across models, and run:

discuss → challenge → summarize → human audit → export

What it isn't: a Poe-style quota / credit pack.

Price: $6/mo or $54/yr (1 account, 1 device). Provider usage stays on your keys.

Link: https://brainstorm.mungwaan.cn

Curious what feels unclear in the positioning — "debate workspace" vs chatbot wrapper? Harsh feedback welcome.