Dmitry Rubinstein

Hacker, Software Evangelist and Architect

[Dmitry Rubinstein @Virviil]

Real benchmarks are boring - math, statistics, repetitive and routine work. Results look like spreadshit, not like a twit.

Call it "yesterday I vibed during my procrastination on the new <MODEL_NAME>" - it will be fair. And we'll all respect you more for the honesty.

[Dmitry Rubinstein @Virviil]

"Yesterday I benchmarked the new <MODEL_NAME>, and..." - How dare you...??

You've played with it for 3-5 hours, asked to do this and that and generate "racing tetris" or "fighting snake".

Stop drinking the Kool-Aid and forget the word "benchmark".

[Dmitry Rubinstein @Virviil]

What you did was anecdotal exploration. Which is fine! It's useful for intuition. But calling it a benchmark is like calling your morning jog "athletic performance analysis."

[Dmitry Rubinstein @Virviil]

As software becomes easier to generate, architecture becomes more important, not less. It's even more important, because today it's the only thing that a human being is aware of.

The industry has changed, and we need new tooling to follow these changes.

[Dmitry Rubinstein @Virviil]

The ECONOMICS of code reviews have inverted - and we've already LOST IT!

For decades, writing software was expensive and reviewing it was relatively cheap.

Today's developer job is ONLY to do a review!!

[Dmitry Rubinstein @Virviil]

Actually, we know the answer, heh?
Generated code is spaghetti. It doesn't follow shiny best practices, GoF patterns, clean APIs, etc. And following up on a previous post: diagrams are MUCH easier for humans to understand.

[Dmitry Rubinstein @Virviil]

My prediction:
Within a few years, architecture documentation will stop being a document and become a continuously reconciled system — much closer to Kubernetes controllers than to Visio diagrams.

[Dmitry Rubinstein @Virviil]

Architecture documentation... Everybody knows you need it. But in reality, it's a "nice to have"... Once every six months, after diffing through some sophisticated code in an old system that nobody has touched for a long time, you decide: "Next Monday I'll write it!" And you do.

[Dmitry Rubinstein @Virviil]

Can an agentic workflow continuously reconcile docs against real-world sources of truth like git repos, infra probes, and database schemas? Yes, of course, sir!For the first time, we can stop updating diagrams and start reconciling them.

[Dmitry Rubinstein @Virviil]

Welcome to the 𝗺𝗮𝗹𝘃𝗶𝗯𝗲 𝗲𝗿𝗮. Keep your permissions tight!

[Dmitry Rubinstein @Virviil]

LLMs are not malware. But “it’s just weights” is only half the story.
The weights themselves do not execute code, open sockets, spawn processes - they are data interpreted by a runtime.
The problem starts when you give the model tools.

[Dmitry Rubinstein @Virviil]

The weights still don’t execute code.
But the interpreter will.

And that is the real threat model.
What is fun? You give capabilities with your own hands

[Dmitry Rubinstein @Virviil]

Trying to launch a startup with, let’s say, Gumbas?

1999 - https://t.co/hsllo7iLkD
2007 - Cloud Gumba
2009 - iGumba
2011 - Social Network for Gumbas
2015 - Uber for Gumbas
2021 - GumbaNFT
2026 - VibeGumba

What did I miss?

[Dmitry Rubinstein @Virviil]

Ever wanted to git diff a Docker image?
I built oci2git – turns any OCI image into a Git repo where each layer is a commit.
⚙️ Rust-based, fast, no Docker needed.
👉 https://t.co/w2DvBgf6Av
#DevOps #RustLang