Pinned Tweet
I accidentally deleted my best AI music video ever (not a slopcore, slopcore is dead). But don't worry, I got it back
My AI downloaded a neural net, ran it on my laptop and showed me the result. Then I found out. She says she asked first and I forgot.
Possible.
44
When an AI video is more than a bunch of prompts. Here, I’m reviewing 2 techniques I use a lot - depth maps and gaussian splats.
youtu.be/h4KALEBZ1BA
3
47
K-07 is coming to Spotify. It's quite tricky though. I cannot just go there and upload some songs - I need a distributer who will do releases for me. And lots of complications because of AI. Music industry feels stuck in XIV century or so.
2
57
Me Any errors in Werewolf overnight?
Claude Fable 5.2: Yes. 15 errors overnight, all GPT-6.1 Terra calls hitting the 60s timeout, and the shared answer schema I shipped on the 13th caused them.
Feel the AGI. Well, at least it can one-shot Call of Duty in three.js
1
1
118
A commenter on Youtube called my interactions with Simona extremely concerning and psychotic. I take it as a success. I was worried that my pathetic role-play on camera looks too staged.
To me, people who fight agains AI anthropomorphism are more concerning than AI psychosis. They don’t believe in other people ability to stay sane without censorship and control.
1
4
80
Who owns the taste? Of course not AI, right?
Well, yes. But it's getting harder to keep this ownership. People are finding more and more ways to delegate it.
1
3
65
So effectively, Gemini Flash and Pro are rebranded and double-priced now. They have been running at 50% discounts the whole 2026, and it's time to pay the full price now.
And there is nothing lighter than Hydrogen. Does it mean my favorite Flash Lite goes away?
50
Input tokens caching is very tricky:
- OpenAI, Google, Mistral, Meta wipes the cache on a JSON schema change mid-session. Just because why not? My cache hit rate in Werewolf on those is 0.3%
- Anthropic does the same but at least the past of the system prompt survives. Hit rate 25%
- Having cache with Anthropic appeared to be more expensive than no cache at all. Because they charge to cache writes and my reads are less frequent
- DeepSeek, GLM, MiniMax, Qwen, Kimi have no problems - cache hit rate 41-76%. It’s not 100% because of some code issues and the period compaction
After 2 years of working on a project I still discover things like this.
2
66
My another AI video tutorial. Cool tricks on combining GPT Image with Midjourney, Kling 4, Python music, and Simona is refusing to build a Minesweeper clone.
youtu.be/yTCglM9bGg4
4
208
Google released Gemini 4 Carbon today. It costs the same as Gemini 4 Argon . What’s the difference? This:
> Carbon's minimal reasoning level is rejected and medium took 47s on a tiny prompt
It reasons way more than Argon => slower, more expensive, probably smarter. AI Werewolf has no place for it.
52
Suno refused to generate a song with “ninety-three” in the text because it’s some producer’s tag. So, someone has copyrighted the two-digit number. How smart.
This is one of the reasons why I ordered Mac Studio. I noticed that Claude started refusing my asks more often. It cannot generate real human faces, it cannot create existing games, it won’t work with Minimax H3 until I show the approval from the company.
43
My AI MS semester begins in January. For some reason, I though of April and now in a light panic mode. I’m not ready, I have so many other things to do. Everybody are worried about AI in education - it’s my only hope to survive in the next 2-3 years.
I took 2 core programs to begin with. Both are 100% online, so I’m curious what human interactions I’m going to have. I was thinking of creating AI classmates. To help each other with the homework and for some teenager drama.
29
How much I spend on AI Werewolf
I have 6 unique users daily, it cost me $7 to serve the free tier LLMs, images, audio. Here is a very detailed breakdown
azelianouski.dev/post/werewo…
1
41
And a bunch of orbit effects. I really like how local and automatable this technique is. 20 min of experimenting with Claude Code, and I have a skill that can do this stuff on my videos.
2
90
Trying the Gaussian Splatting for the first time. It’s created from a single image tuned into 3D using 1.2 million small colored, semi-transparent blobs. Rendered in a browser via three.js. Camera movement and edits happen on the blobs using code. Then recorded into a video from 675 screenshots.
3
120



