Dmitriy Kashirin

SWE & Team Lead • AI • Founder ZingZing Softworks

[Dmitriy Kashirin @ThisIsKashirin]

If you write AI prompts in MD, you're doing it wrong. PROs use XML, but in large codebases it collides with code syntax and gets ignored. We benchmarked Gemini 3 & DeepSeek V4 at 800k context window to see how special tags solve this. Our analysis & tips: https://t.co/iYkhawQHxy

[Dmitriy Kashirin @ThisIsKashirin]

@cursor_ai Wow. Thank you guys!

[Google Gemma @googlegemma]

We are entering a new era of on-device automation. ✨

Watch Gemma 4 E4B navigate and drive an iOS simulator directly using Argent. Local models can handle complex interactions and software navigation autonomously.

[Dmitriy Kashirin @ThisIsKashirin]

@antigravity We dropped Antigravity and Windsurf, and never even started using CC because all of them have these ridiculous rolling quotas that vanish the moment you start doing real work.

Antigravity also feels like a rushed prototype. Honestly, our juniors ship more stable software.

[Dmitriy Kashirin @ThisIsKashirin]

For instance, if you do that, Gemini 3 Flash will start appending "/commit" at the very end of every single response to you.

This is my empirical experience from observing agent behavior. What do you think about this?

[Dmitriy Kashirin @ThisIsKashirin]

Have you ever noticed that when working with an LLM in agentic mode, if you ask it to make a Git commit just once, it starts doing it constantly, even if you explicitly forbid it later?

In this thread, I'll explain why I think this happens.

[Dmitriy Kashirin @ThisIsKashirin]

(Plus, saying "do not git commit" just triggers the pink elephant effect, keeping those tokens active in context).

You can easily test this overlap by manually injecting "/commit" into a message in the chat history.

[Dmitriy Kashirin @ThisIsKashirin]

It seems like everyone decided to work today. Incredibly long inference times with Gemini (if it’s even working at all) and with GLM 5.1.

[Dmitriy Kashirin @ThisIsKashirin]

Beautiful photos

[NASA Solar System @NASASolarSystem]

Two years ago today: Millions saw a total solar eclipse across America.

Two days ago: Just four people saw a total solar eclipse from Orion.

Would you rather see an eclipse from space or, in the words of Artemis Astronaut Victor Glover, from this "spaceship called Earth"?

[NASA @NASA]

Experience the magic of our Moon mission wherever you go! ✨

Download free, mobile wallpapers and bring your device into the new era of exploration: https://t.co/494ZyUBtve

[Dmitriy Kashirin @ThisIsKashirin]

@GoogleAI Congrats with the release! It's impressive!

[Dmitriy Kashirin @ThisIsKashirin]

@elonmusk A bias in either direction is bad

[Dmitriy Kashirin @ThisIsKashirin]

We’ve released a Public Preview version of our new service - SpecTree, which helps you create technical specifications for DEVs and AI agents.
It's free, including the AI assistant and AI architect!

Details: https://t.co/xbr1qCUquI
Article: https://t.co/hPRhhOkAtB

[Dmitriy Kashirin @ThisIsKashirin]

AI scanners on X see hundreds of thousands of posts about how AI is going to change everything, especially our jobs, and all of that ends up in the datasets. What do they think? Does their ego get bigger with every release?))) What do you think?

[Dmitriy Kashirin @ThisIsKashirin]

This is very good! It will make UI work much easier. I assume that sharing project tree screenshots also improves the model’s understanding by activating different attention patterns, instead of relying only on text.

[Z.ai @Zai_org]

Introducing GLM-5V-Turbo: Vision Coding Model

- Native Multimodal Coding: Natively understands multimodal inputs including images, videos, design drafts, and document layouts.
- Balanced Visual and Programming Capabilities: Achieves leading performance across core benchmarks for multimodal coding, tool use, and GUI Agents.
- Deep Adaptation for Claude Code and Claw Scenarios: Works in deep synergy with Agents like Claude Code and OpenClaw.

Try it now: https://t.co/WCqWT0qCQb
API: https://t.co/xDy1O6ZPcz
Coding Plan trial applications: https://t.co/qCM6cri0KK

[Dmitriy Kashirin @ThisIsKashirin]

@Zai_org This is very good! It will make UI work much easier. I assume that sharing project tree screenshots also improves the model’s understanding by activating different attention patterns, instead of relying only on text.

[Dmitriy Kashirin @ThisIsKashirin]

@hcompany_ai Wow, I've never heard of Holo before. It'll be interesting to try.