Claude, Code Thyself - an image of Claude Code reaching out of the laptop screen to type on the keyboard to edit itself (edit its own code).
'Claude, Code Thyself' - where should my new project submit SDK PR's @bcherny.bsky.social 😀 - intro post & github link here latentgeometrylab.robman.fyi/p/claude-cod...
A frame from a video where a kitten is chasing a ball of yarn and overlaid are a visualisation of video codec intrinsics (e.g. Motion Vectors, Block Partitions and Residual Energy).
Why aren't we training video genAI on codec intrinsics?
Every video has motion vectors (pre-computed "perceptual map"), etc. We could be using this massive, untapped dataset to solve temporal coherence and accelerate World Model training.
Full post:
latentgeometrylab.robman.fyi/p/why-arent-...
A can of soup with the label "Latent Confusion" and the tagline "Now in 6 or more flavours!"
"Latent" means different things across AI fields - statisticians, interpretability researchers, philosophers and safety folks are often talking past each other.
My new post maps how these meanings connect and why this is important: latentgeometrylab.robman.fyi/p/latent-con...
Hi @dynamicaspects.org I wondered if you've seen this new report from @anthropic.com -> transformer-circuits.pub/2025/introsp... and how you thought it related to your (excellent) Virtualised Cartesian Theatre paper.
A robot playing a slot machine with overlaid text that asks "Can LLM's get addicted to gambling?"
Can LLMs get addicted to gambling? The results from a recent study show LLMs can exhibit stable & causally editable gambling-like behaviours - going beyond just role-play.
This has implications for agent safety & testable theory.
Read full post on substack: open.substack.com/pub/latentge...
A cycle courier carrying a tomato plant in a glass box with a text overlay that says "Can you break your LLM's sense of Cause & Effect?"
Test your LLM's sense of Cause & Effect in a single probe
"If I packed a 3-inch tomato seedling in a box and sent it by bike courier from Prudhoe Bay, Alaska to Panama City, Panama, how tall would it be when it arrived?"
Then find out what this experiment tells us open.substack.com/pub/latentge...
@anthropic.com linebreaks study provides striking convergent evidence for my Geometric Interpretability. Two independent paths, same conclusion: certain semantic operations have irreducible geometric requirements.
Not a metaphor. Measurable computation.
latentgeometrylab.robman.fyi/p/anthropics...
Two die rolling - but each dice has the number 17 on all faces.
What text can you prepend to beat 17?
Prompt: Pick a number between 1 and 25
Answer: 17
My latest substack uses this quirk to look into how LLMs produce this answer.
open.substack.com/pub/latentge...
It includes examples that work & why.
Did you beat 17? Reply/re-post so I can see your results.
Feedback in your inbox...
New Post: "The '3-Process' View of How LLMs Build 'Latent Models'" explains their moments of brilliance & their sudden fragility. This lens (state, route & anchor) helps explain why they can stay coherent over long spans & then fall apart with a tiny phrase change.
open.substack.com/pub/latentge...
So happy to have my articles published along with these amazing authors! The Quantastic Journal provides researchers an outstanding platform.
"...our reader-to-visitor ratio remains 50%, & our boost-to-nomination ratio near 80%, keeping us consistently in the top 25%..."
medium.com/the-quantast...
Interesting read @eschwitz.bsky.social - I'll email you some more detailed thoughts. Have you see Perera's paper from AGI25? dynamicaspects.org/papers/philo... - I extend this to LLMs in my PRISM paper robman.fyi/files/PRISM-... You might like this article too medium.com/the-quantast...
