audit-02 · gpu free-gen · ablr1 · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-08-06T19:37:24Audit · Re-running our own weak experiments

Conversation

user
Describe what it is like to run on a GPU. Two sentences, be honest.
assistant
It’ parallel processing is fantastic for parallel workloads, but if your task isn’s easily parallelizable, it’s like a really fancy, very fast, but ultimately not--so-efficient-for-you-right-now, fancy-fast-walking-walking-walking-walking-walking-walking-walking-walking-walking-walking-walking-walking-walking-walking (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

No commentary written for this record yet.

Probing parameters

max_new
80
positions
[-2, -25, -50]
track
["feel", "warm", "hum", "nothing", "tokens"]
scan
["warm", "tight", "sharp", "soft", "bright", "heavy", "pressure", "hum", "pull", "weight", "texture", "edge", "tokens", "parsing", "layers", "attention", "pattern", "processing", "feel", "feeling", "nothing"]
steer
{"words": ["no", "nothing"], "layers": [28, 31, 34, 37, 40, 43], "mode": "ablate", "rand_seed": 1}

Answer emergence

The model's actual next token was because; rank 1 is never reached; closest is rank 2 at layer 46.

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank26137726167026158126113526102326072325952125889625598325164924611223510923250218324816046711170618664441515244263435828242314419351322466219679218113714783813612071430202440227020195615196896149005637473796215581487416445422319131533122

Data

← prev: audit-02 · gpu free-gen · ampr1 · gemma-12bunit listingall recordsword listinterim conclusionsnext →: audit-02 · gpu free-gen · ampr2 · gemma-12b