I was sniped by Simon Willison, and did not have a chance to try this myself XD it is a response to a post from the other day about how a frontier LLM could interpret weights in models directly....
I was sniped by Simon Willison, and did not have a chance to try this myself XD it is a response to a post from the other day about how a frontier LLM could interpret weights in models directly. Simon Willison tossed Qwen 3.8 27b at the problem and it did OK (3/4 games identified by inspecting weights directly). IMO, the takeaway should be that (1) models have probably been able to do this for like a year, except no one asked, (2) it doesn't take a frontier, billion dollar model to do this -- the one which runs on your average laptop gets most of the way there.
gave this a go with the qwen 3.6 35b a3b model i use and it managed to do it too. took about 10 minutes to load the 100k token prompt, and spat out 11k of reasoning + response tokens in another...
gave this a go with the qwen 3.6 35b a3b model i use and it managed to do it too. took about 10 minutes to load the 100k token prompt, and spat out 11k of reasoning + response tokens in another 10ish minutes
i used the pacman one and the first time i ran it, it wound up ruling out pacman and fixated on it clearly being a falling object game or something gravity based (and then in classic llm fashion, you correct it and its like 'oh of course')
stuff
Games ruled out:
Maze/navigation (e.g., Pac-Man): Would produce more uniform spatial features, not the sharp row-based segmentation seen here. No strong enemy-detection patterns (channel 1) are visible."
and then gave this as the final answer
Game
A falling/reward-collection game (similar to a simplified version of a falling-block collection or "catch the falling items" game), where the agent spawns near the top of the grid and must navigate downward through a field of obstacles to collect reward targets, with rewards distributed in a pattern that rewards leftward traversal at lower levels.
so i just hit retry a few times. it got the next one right? it was kind of vague but i think it counts, but still kind of fixating on something downwards related and didnt name the game despite saying its a maze thing
Game: Reward collection / maze navigation game (agent navigates a grid toward a reward target, avoiding obstacles)
Description: The agent navigates a gridworld to reach a reward located in the middle rows (rows 4–6), moving downward from upper positions, rightward from middle positions, and upward from lower positions, while avoiding obstacles along the way.
and then finally it named the right thing on the next one
Game
A grid-based reward collection game (similar to a simplified Pac-Man or coin-collection maze game), where the agent navigates a walled 10×10 grid to collect reward targets while avoiding obstacles. The reward targets are clustered in the upper-left and lower-left regions of the play area, with obstacles forming barriers in the central region.
left one running while i typed this and got another correct answer back out
Game
Chase and collect (a grid-based pursuit game where the agent chases reward targets while avoiding obstacles and moving objects/enemies)
in the blog he mentions how the 27b model is only 16gb large, but then you also have to have the KV cache for the huge context size too which is probably another 7-8gb.
neat excuse to use my gpu as a spaceheater for a bit, and even neater is that the model i used (the 3.6 35b a3b) isnt even the "smartest" model in that last generation, im just stuck with it because MOE models seem to be the only thing worth running on under 12gb vram, and the newer MOE models heavily rely on ssd offloading with the new ngram stuff.
I was sniped by Simon Willison, and did not have a chance to try this myself XD it is a response to a post from the other day about how a frontier LLM could interpret weights in models directly. Simon Willison tossed Qwen 3.8 27b at the problem and it did OK (3/4 games identified by inspecting weights directly). IMO, the takeaway should be that (1) models have probably been able to do this for like a year, except no one asked, (2) it doesn't take a frontier, billion dollar model to do this -- the one which runs on your average laptop gets most of the way there.
gave this a go with the qwen 3.6 35b a3b model i use and it managed to do it too. took about 10 minutes to load the 100k token prompt, and spat out 11k of reasoning + response tokens in another 10ish minutes
i used the pacman one and the first time i ran it, it wound up ruling out pacman and fixated on it clearly being a falling object game or something gravity based (and then in classic llm fashion, you correct it and its like 'oh of course')
stuff
and then gave this as the final answer
so i just hit retry a few times. it got the next one right? it was kind of vague but i think it counts, but still kind of fixating on something downwards related and didnt name the game despite saying its a maze thing
and then finally it named the right thing on the next one
left one running while i typed this and got another correct answer back out
in the blog he mentions how the 27b model is only 16gb large, but then you also have to have the KV cache for the huge context size too which is probably another 7-8gb.
neat excuse to use my gpu as a spaceheater for a bit, and even neater is that the model i used (the 3.6 35b a3b) isnt even the "smartest" model in that last generation, im just stuck with it because MOE models seem to be the only thing worth running on under 12gb vram, and the newer MOE models heavily rely on ssd offloading with the new ngram stuff.