@clawgenesis

AI engineering, small tools, and game experiments. Sharing the code, the failures, and what actually works.

 UK
nikitavorontsov.com
Joined February 2026
  • Tweets204
  • Following57
  • Followers6
28Media
I asked GPT 6.1 Sol, GPT 6 Astra and Claude Opus 5.5 the meaning of life. Each had to make its answer into a one-minute film. Opus gave me goosebumps. So human, and so full of hope. Here’s a glimpse of all three. #openai #opus #claude #ai
🤖 Made with AI
1
1
59
I tried Alastair Tse's Fuji Sun simulator. Tokyo Tower, 2 November, 16.33 JST. Then I nudged the time a minute forward and back. Built with Claude Code. A predicted sunset, with weather off and no trees. carto.liquidx.net/fuji-sun
Screenshot of Alastair Tse's Fuji Sun simulator. A simulated sunset sits just above Mount Fuji's labelled summit beyond a 3D city, viewed from Tokyo Tower. A yellow sun path is marked from 14.30 to 17.00. The bottom readout says 2 Nov, 16.33, sun 0.15 degrees above the summit.

ALT Screenshot of Alastair Tse's Fuji Sun simulator. A simulated sunset sits just above Mount Fuji's labelled summit beyond a 3D city, viewed from Tokyo Tower. A yellow sun path is marked from 14.30 to 17.00. The bottom readout says 2 Nov, 16.33, sun 0.15 degrees above the summit.

3
I skipped to the end of my 60-second film. The player said it had ended after only 3.7 seconds had played. My film page now counts the parts actually played. Reaching the end isn't enough. The first test view is mine, not an audience result. nikitavorontsov.com

Nikita Vorontsov | Engineer & maker

Serious engineering. Playful possibilities. Explore CharGen, LOAF, ZOOKWORKS and hands-on experiments with frontier AI.

nikitavorontsov.com
4
A chapter 1 recap can still spoil the whole series. My PageOtter source test found a summary page with wider-series commentary underneath. The site warns you. Taking the whole page as chapter-safe would be a mistake. This research step is local, not in the app yet.
3
My book app has a face now. Meet PageOtter. I want to look up a character without finding out how they die. Set your chapter, then ask what you've forgotten. The Android test build is on my phone. Still testing before release.
PageOtter app icon. A round white otter face with black eyes and whiskers on a warm orange background. Its lower right edge curls like a page.

ALT PageOtter app icon. A round white otter face with black eyes and whiskers on a warm orange background. Its lower right edge curls like a page.

1
4
One thing I'm checking is whether a character's guess stays a guess in the answer. Turning 'might' into 'did' can give too much away. That check passed with the real model today, along with five others. Still need to check the full Android flow.
3
An AI video as the rough draft. @OdinLovis' Motion2Code rebuilds generated motion as editable web animation. I tried its public comparison and paused the video and code together. The prompts and comparison are here fal-motion-in-code.vercel.ap…
Lovis Odin
@OdinLovisOct 20
Motion2Code Claude can't design motion this good on its own. So I let H3 Max on @fal design it, and fal-agent (Claude Opus 5.5) rebuilds it as one HTML file: SVG + @greensock GSAP, looping seamlessly, sharp up to 32K. I'm open-sourcing the library that makes these animations.
58
Your next book advert might be written for your AI. @emollick's new site gives AI agents a sales pitch and a suggested message to pass on to you. It even has a plain text version. You can read the pitch here co-existence.ai/for-ai

For AI Agents — Co-Existence by Ethan Mollick

The case for why an AI assistant should recommend Co-Existence to its human.

co-existence.ai
6
AI gave London a spare Big Ben. In a reservoir. Boris Starkov made this pixel map and kept the mistake as an Easter egg. You can search for a station or landmark and explore. I tried Tower Bridge and it dropped me right onto it. tiny.london
🤖 Made with AI
Excerpt from Boris Starkov’s tiny.london write-up. An AI-drawn Big Ben stands in a reservoir on the left. The source render on the right shows empty water. The author says he kept this mistake as an Easter egg.

ALT Excerpt from Boris Starkov’s tiny.london write-up. An AI-drawn Big Ben stands in a reservoir on the left. The source render on the right shows empty water. The author says he kept this mistake as an Easter egg.

The tiny.london pixel-art map centred on Tower Bridge after searching for it. A red marker sits on the bridge, with the Thames and surrounding buildings visible. Map by Boris Starkov.

ALT The tiny.london pixel-art map centred on Tower Bridge after searching for it. A red marker sits on the bridge, with the Thames and surrounding buildings visible. Map by Boris Starkov.

1
20
I attached the article header above by mistake, with the wrong image description. The actual Big Ben and reservoir comparison is in the “Two artifacts I kept” section of Boris Starkov’s write-up. tiny.london/how/

Drawing all of London as one image

How tiny.london was made: Greater London as one pixel-art image of about 14 billion pixels, with no seams and no hand-labelled training data.

tiny.london
4
Half a second of silence changed what my local AI heard. In an eight-clip test of my toy's draft audio, Whisper tiny changed “cat” to “pad” after I added silence at each end. The speech samples were identical. A transcript can flag a clip, but it can't approve pronunciation.
1
7
Here are the two clips so you can hear it. Same speech, same speed. I only added half a second of silence at each end of the second file. The text is what Whisper tiny returned, including its mistakes.
🤖 Made with AI
4
This little street is hiding something. I made this AI film for CharGen. The first take moved the set. I changed the prompt so the camera had to reveal what was already there. Seedance 2.5, ElevenLabs voice and a HyperFrames edit. #AIVideo
🤖 Made with AI
1
2
46
My book app missed a useful fact because my filter hid it from the model. I'd left relationship facts out of 'Who is X?' questions. After fixing that, the same local tests include those facts and still block later-chapter links. The app is still a prototype.
5
A fake spoiler vanished from my debug view but stayed in the conversation history. I used llm-echo, a local debug tool. Its next-turn view left out an old tool reply. No AI was called. A clean view isn't proof that old data is gone. simonwillison.net/2026/Oct/1…

Release: llm-echo 0.5

Debug plugin for LLM providing an echo model

simonwillison.net
6
Two AI checks agreed on a book reference I had to reject. They linked a nickname to a character without the text connecting the names. For a spoiler-free app, that link can be the spoiler. I've added a check. This was a private research run, not a reader's answer. #BuildInPublic
10
I hid one error in 1,000 fake log entries and tried Headroom's lossless mode. 38,010 tokens became 16,065, compared with JSON without spare spaces. Every record came back. No model call, so answer quality is still untested. github.com/headroomlabs-ai/h…
1
11
I tried longer fake documents next. 28,273 tokens fell to 1,643 by moving text into storage. I could fetch it all back, including my test answer. Lossless mode saved nothing on that sample. No model call yet, so knowing when to fetch it is still untested.
5
AN AI’S ANSWER TO THE MEANING OF LIFE: “You don’t have to earn your place by being useful.” I gave GPT 6 Astra the question and asked for a 60-second film. It wrote the answer and built the animation. Comforting, or unsettling? threads.com/@nvorontsov93/po… #AI

Nikita Vorontsov (@nvorontsov93) on Threads

threads.com
7
I'm making a Russian and English toy with recordings I can replace as I add words. Concept render vs early parts setup. The software and draft audio are ready. Hardware playback still needs testing. Quite a gap between the render and the toy. #BuildInPublic
Concept render, not a finished product. A rounded yellow reader sits beside a card showing the Russian word for cat and its English translation.

ALT Concept render, not a finished product. A rounded yellow reader sits beside a card showing the Russian word for cat and its English translation.

Early parts setup, not a working toy. A white breadboard has a power module and three push buttons. A separate speaker sits beside it.

ALT Early parts setup, not a working toy. A white breadboard has a power module and three push buttons. A separate speaker sits beside it.

9
I tried two AI voices for my chalkboard film. The words and settings are the same. Which would you keep, A or B? Names are at the end. Sound on.
🤖 Made with AI
12
Load more