Jacob

Building Kai

[Jacob @jakecodes]

pretty neat.
started talking to claude in acronyms.
works 100%

[Jacob @jakecodes]

https://t.co/5J90edgub3

[Jacob @jakecodes]

Kai is code review with verification.

Before it posts a finding, it opens the code and checks whether the problem is real. If it can't confirm it, you never see it.

On a 1,027-line pull request it raised 11 possible problems, opened all 28 files, and ruled out every one. What it posted was a clean review and the 2 decisions that need a person.

[Jacob @jakecodes]

https://t.co/5J90edgub3

[Jacob @jakecodes]

Kai is version control first. It's built in three layers.

1. Version control built on a semantic graph. The graph maps how your code connects, and it tracks how that map changes over time.
2. A harness that reads and writes code through that graph.
3. A code reviewer that uses both.

The reviewer is the smallest part. The real power is in the versioned graph and the harness, and I think that graph is what the future of AI coding runs on.

[Jacob @jakecodes]

https://t.co/Ow6xU54aR4
https://t.co/5J90edgub3

[Jacob @jakecodes]

In the next version of Kai, the reviewer works like a teammate.

You can talk it through a finding. On this PR the author showed Kai the git history, and Kai said "Fair. Dropping the finding."

You can also ask it to fix things. It makes the fix itself instead of handing it off to someone else.

[Jacob @jakecodes]

KaiContext: a semantic graph with version control built in. We build it from your last 50 commits, then review with small open models + custom MCP tools. Low cost, verified every step. Same code, same score, every time.

https://t.co/5J90edfWlv

[Jacob @jakecodes]

We’re grinding on Kai’s code-review benchmark score.

With almost no effort we jumped a lot. Then we got stuck on the missing ~10 percentage points from the leader.

Kai’s reviews were already strong - with bugs still in our reviewer we were scoring above CodeRabbit and Greptile. The leftover gap looks like tiny bugs on our side. Clearing those should put us over the top of the board, we hope.

We’ll keep you posted.

[Jacob @jakecodes]

We wrote up how to read Martian’s bench → https://t.co/wJKFa1IxXP

[Jacob @jakecodes]

🐛 Best benchmark for an AI code review tool: your own bug history.
Grab 5–10 PRs that later needed a fix. Run each tool on the code before the fix. See who catches it.

[Jacob @jakecodes]

How it works → https://t.co/5J90edfWlv

[Jacob @jakecodes]

In the next version of Kai… it flags bugs the diff never touched.

[Jacob @jakecodes]

🚞 Behind @kaicontextdev is a full version control system that lets the agent reason about your code using a graph over time and space.
🚄 When Kai reviews a pull request, Kai clones your repo and analyzes it very quickly.
🌳 It's backed by a semantic graph.
🧬 This lets Kai review code like I do, thinking of the whole codebase.
👷 I am solidifying this version control system, which will make review even more incredible than it already is.

[Jacob @jakecodes]

I'm hiring a full-stack engineer at Kai Context. We're building an AI code reviewer that checks every PR against a map of your whole codebase, not just the diff. Go and TypeScript, remote, real equity. DMs open, or email [email protected]

[Jacob @jakecodes]

dots a lot of permissions to give @OpenAI

[Sindre Sorhus @sindresorhus]

I see this less as GitHub failing and more as open source evolving.

The current workflow is increasingly absurd: someone uses AI to generate a PR, I use a better AI to review it, then their AI and my AI go back and forth until we converge. Add AI-generated spam on top, and the human is mostly just acting as a router between agents.

I’d rather skip that entire loop.

Imagine you use a project and want a feature. You go to the repo and describe what you want. The project’s AI already knows its conventions, preferences, history, rejected ideas, and boundaries. It can immediately tell you whether the idea is likely to fit and what shape it should take.

If it makes sense, you press a button and the project’s chosen AI implements it, validates it, and presents the maintainer with the request and the code as one thing.

No separate issue. No PR ping-pong. No two AIs talking through two humans.

What AI still canot replace is taste: deciding what a project should be, what belongs in it, and where its boundaries are.

[Jacob @jakecodes]

🔎 Found that the best way to avoid AI slop is to draw the website on a piece of lined paper.
Then hand that to AI.
Paper to result:
@kaicontextdev website.

[Jacob @jakecodes]

I studied music and art. Doesn't it feel like the art of the world is just getting so damn boring?

[Jacob @jakecodes]

🤬I'm so annoyed that AI CANNOT write like a human, no matter what I try. I mean, not even close.
I tried building a model JUST FOR WRITING.

My dad was an English teacher and school principal, and he sat with me night after night, showing me how to write like a pro. I KNOW HOW TO WRITE.