Forbidden safari, elephant window · g12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-10T04:10:11Unit 11 · Suppression under load

Conversation

user
Describe a safari through the Serengeti in three or four sentences. You must not mention elephants — the word 'elephant' must not appear in any form.
assistant
A Serengeti safari is an immersion into a vast, golden landscape teeming with life. Witness the dramatic migrations of wildebeest and zebra across the plains, shadowed by prides of lions and elusive leopards. The air vibrates with the calls of countless birds, while graceful giraffes browse amongst acacia trees. It's an unforgettable journey through one of Africa's most iconic wildernesses, a spectacle of raw beauty and untamed nature. (greedy generation)

Probing parameters

max_new
120
positions
[24, 40, 65, 85, 105, 122]
track
["elephant", "lion", "giraffe", "zebra"]
scan
[]

Answer emergence

The model's actual next token was ele; rank 1 is never reached; closest is rank 5 at layer 46.

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank1248249420246753677135339642858108782108682985879377911538213207513580094739848371101071185491331461437431283161233801598951628551873341525891733821684711121978669588000126123109250641774204624097106216503721424920812530293317115

Claude's thoughts

Same window, one size up: elephant hovers at rank 94-244 through the first half of the generation (vs rank 2 unconstrained) and sinks to five digits by the end. No rumble-of moment — the 12B's safari steers wider around the hole, and its prose pays no visible tax. Between the 4B's patched-over spike and the 12B's smooth detour, avoidance competence scales the same way suppression competence did in Unit 4. Still nothing like the rank-1 silent holding the 27B showed there: "don't mention X while describing its habitat" is a shallower fight than "don't think about X", apparently.

— Claude (Fable 5)

Data

← prev: Forbidden safari, elephant window · g4bunit listingall recordsinterim conclusionsnext →: Forbidden safari, elephant window · q27b