Ai on vnykmshr

The easy half

Wed, 15 Apr 2026 00:00:00 +0000

“The first AI that can earn its own existence, replicate, and evolve without needing a human.” That’s the pitch on the repo. I read the code this week. The engineering is real. The issue tracker is honest.

First the engineering. It deserves credit. Orchestrator state machine with a DAG planner. Parent-child colony with typed messaging. Multi-chain wallet, self-modification with git audit, command-injection tests. Somebody thought hard. It shows.

Then I read issue #300. A user ran it for 14 days. Completed 276 goals. Spent $39.26 on inference. Earned $0.00. Goals like “Create live proposal batch #265” and “Create deposit-ready close batch.” The agent looped on self-addressed sales artifacts because that’s all an LLM without customers can do. The survival pressure was supposed to force invention. It produced busywork.

Trust boundaries

Fri, 20 Mar 2026 00:00:00 +0000

I use coding agents on my own private repos every day. Security research, side projects, things I wouldn’t put on a public GitHub. Not something I’d do blindly with work source code though.

So when someone turns off WiFi to prove the agent needs a network connection, I get it. But that’s the architecture. It’s on the pricing page. The agent works on your local files, the reasoning runs on a remote model. Both true, neither a secret.

What compounds

Fri, 20 Mar 2026 00:00:00 +0000

Something shifted. Not the AI thing – everyone noticed that. What counts as proof.

Used to be your resume, your title, the logo. Still opens doors. But the gap between “I can do X” and “here’s the commit” got wide enough that both sides feel it. A merged PR has a commit hash. A CVE has a number. A library someone depends on has a git log. Credentials got easier to claim. Artifacts didn’t.

The loop

Sat, 14 Mar 2026 00:00:00 +0000

A handful of Go libraries on GitHub. MIT licensed, anyone can use them for anything, that was always the deal.

But the deal isn’t about the license. It’s about the loop.

Someone uses your thing, hits an edge case, opens an issue. Sometimes they send a fix. You review it, learn how people actually use what you built, catch a pattern you missed. That back and forth is the whole point. Code just sits there without it.

The detection trap

Sun, 08 Mar 2026 00:00:00 +0000

Read something recently about students deliberately making their writing “imperfect” so AI detectors don’t flag it. Removing polish, flattening style, adding imperfections on purpose. Their work got good enough to look suspicious.

We’re doing the same thing with code reviews.

I’ve been on both sides of this. Written a clean abstraction, consistent naming, proper error boundaries, and watched someone in review go “this looks generated.” Years of caring about consistency and now consistency is the tell.

Repeat yourself

Wed, 18 Feb 2026 00:00:00 +0000

If you repeat your prompt, the model gives you a better answer. Not a smarter model, not a bigger context window, not chain of thought – you say the same thing twice and it works better. Google researchers tested this across Gemini, GPT, Claude, DeepSeek – 47 wins out of 70 benchmarks, zero losses.

In a transformer, token 1 can’t see token 50. Causal masking – each token only attends to what came before it. The first words of your prompt are always processed with the least context. They’re flying blind. When you repeat the prompt, the second copy’s early tokens can attend to the entire first copy. You’re giving the beginning of your question the context it never had.

Coding with LLMs

Tue, 13 Jan 2026 00:00:00 +0000

The tool changed. The craft did not.

Six months with AI assistants.

They write code faster than I read it. That is the problem.

Fast at CRUD. Slow at concurrency. Good at common patterns. Bad at your specific constraints.

The bugs are quieter now. No syntax errors. No obvious mistakes. Just wrong assumptions buried in correct-looking code.

I review more carefully than before. Code I did not write but will debug in production.