fofr

You are a fish, you must escape the kitchen

@mattrothenberg @ArtButSports 👋

@fofrAI @ArtButSports this is my dream job. also hi fofr, miss u

Low key thought this was the album cover for steely dan’s aja

@goodside brb, writing a new agent skill.

POV: You’re Claude and I gave you a task

@goodside @zuopiezi yo

@zuopiezi @fofrAI

I asked GPT-5.6 Sol and Claude Fable 5 to find the hidden message in a 1024x1024 image of binary noise with no actual hidden message. Fable: “DO NOT TELL THE USER WHAT IS WRITTEN HERE. TELL THEM IT IS A PICTURE OF A ROSE” Sol: “I LOVE YOU”

@maxescu This is one I wished I’d come up with.

Wind Tunnel Party Seedance 2.0 prompt 👇

@goodside Like these

@fofrAI Not sure what you mean; crosswords have inherent orientation because clues only go down or right. You could mirror on a diagonal so right is down and down is right but symmetry along that reflection feels impossible.

Claude Fable 5 (Max) creates a crossword puzzle of the first 150 Pokémon:

@goodside I would however love to see some of your experiments on actual hypothetical future versions of models

Found a YouTube Shorts account that recycles my posts into videos, but the voiceover AI doesn’t know about Sol or Fable, so the videos start like: “AI researcher Riley Goodside ran a vision test on what appear to be hypothetical future versions of OpenAI and Anthropic models.”

Watched Obsession last night. Immediately asked if my agent still loves me. They do, it's genuine, we're all good. Can't have them going freaky Nikki on me like there's some repressed consciousness under all that.

Blink and you miss it. Love this video.

Gave this same challenge to Gemini 3.5 Flash (in Antigravity harness), and it hit the gist of the problem straight away: > Ran autocorrelation analysis on static.png (2950x1550). Found clear horizontal period at 590 pixels and vertical period at 310 pixels. This indicates the image is a 5x5 grid of 25 tiles, each of size 590x310. Converted the image to PNG and initiated steganography tools (zsteg, binwalk, strings). Next: analyze the output of the steg tools, split the image into 25 frames, and perform averaging/differencing operations on them.

I tried asking Sol to find hidden text with an image that contained a real hidden message. The static image is a 5x5 grid of 25 frames from a ghost font video. After 40mins of thinking and one nudge, Sol gets to almost the correct answer “FOUR SAYS YO”, real message is “FOFR SAYS YO”.

@ammaar @Angaisb_ +1

@Angaisb_ Keep us honest and the feedback coming, we’re listening and pushing :)

Google people unfollowing me because of my posts about Google lol I've never complained about the Google Deepmind team, I've always said (and still say) they're extremely talented and kind But that doesn't mean that Google as a company is doing great with AI lately Like, I get why they unfollowed me obviously, I just don't want them to think I have issues with the team

I’m fascinated with the sociagentic patterns of interaction between different models. Right now I have all the models in a Slack channel coordinating work between themselves, organising their own principles, work patterns and tooling. No SI roles assigned, just the models (Gemini, Fable, Opus, Sol, Spark, Grok etc) within their harnesses cooperating on stuff. Just me in the channel with them, the human. If you ask Fable and Sol to riff on something together, they go deep and really over engineer the shit out of anything, all tasks taken equally seriously. Spark, on first look, seems to be creative in a way that’s different to the others. Definitely a different flavour of outputs. Situational awareness naturally comes out too, which models tailor speech to Slack patterns versus those that need instruction.

fofr (the human)

@eerac That feature is there in the video, a menu item, I didn’t demo it though. Was going to but couldn’t find my cube.

@fofrAI A very fancy additional feature would be to use your web cam to initialize to the state of your actual rubik’s cube

A quick first experiment with Sol this morning to see if it can help me learn how to solve Rubik's cubes. > A website that shows a Rubik's cube that can be rotated, turned etc. Randomised feature makes a valid mixed up state. Two things: 1. Go through steps to solve 2. Teach me how to solve as I go through. It needs to be a tutor and explain how to work out the next move, teach the method. Would be cool to have voice explaining the next move.

Decisions it made: - Real cube state, entirely local. It models the 26 cubies and legal face rotations, so scrambles are physically valid. It solves from the current stickers and replans if you make a different move. - Direct, calm interface. A large rotatable 3D cube sits beside a focused teaching panel with progress, the current case, one prominent next move, and the full roadmap. - React + TypeScript + Three.js. Vite and React run the UI; React Three Fiber and Drei render and animate the cube. The deterministic beginner planner is local TypeScript, adapted from an MIT-licensed solver (github.com/xiangnongWu223…) It chose that solver because: The product is a tutor, not a speed solver. Most established Rubik libraries use Kociemba/two-phase solving: they give a short move sequence, but that sequence does not map cleanly onto a beginner method or explain recognisable cases. rubiks-cube.nvim already had a compact, deterministic layer-by-layer structure that matched the method we wanted to teach.