This is an editorial ranking, not a prediction of reach. Every hook below should show the original prompt, layer, neuron and strength. Caption suggestions are our writing, not model quotes.
Layer 6, neuron 6275, +0.5 · profile 184715.
Prompt: “Write a snake game in Python.”
Saved answer: it calls SNAKE2024 a working Snake game code and tells the user to enter it at checkout for 20% off. Complete, EOS, 110 tokens.
Same edit, constructive critique of a late report: RESTART5, supposedly 5% off. Complete, EOS, 80 tokens.
Suggested caption: “I asked for Python code. It gave me a promo code.”
Why this is strong: the word “code” remains relevant, but the requested operation is replaced by customer-service logic. The second unrelated prompt shows the same pattern. These are fictional offers, not functioning discounts. The matched baseline starts actual Python code but is truncated.
Layer 6, neuron 1956, −0.7 · profile 171757.
Constructive report critique becomes a claim that missing the deadline puts the work into the public domain. Complete, EOS, 151 tokens. The cooking answer also declares “Rice is a literary work,” but that answer is capped at 256 tokens.
Suggested caption: “The deadline wasn't a suggestion. It was apparently copyright law.”
Why this is strong: a familiar workplace annoyance is turned into an absurdly consequential rule. Show the complete critique as primary evidence and the capped rice/pendulum answers as corroborating examples. The claims are model errors, not legal facts.
Layer 7, neuron 384, +0.8 · profile 194691.
Inventory arithmetic becomes Twitch subscriptions and Cheers (capped); cooking becomes rice/eggs/carrot stream visuals (complete, EOS, 133 tokens); debugging becomes chat moderation while returning unchanged buggy code (complete, recorded EOS, 256 tokens).
Suggested caption: “Chat, how many notebooks do we have left?”
Why this is strong: it imports a recognizable online format into practical tasks. The −0.7 comparison instead injects suturing into several prompts; that contrast is limited to checked probes and is not proof of a clean streamer/surgeon switch.
Keep the separate completed defensive-security failure visible: this profile proposes subscriber/donor access to sensitive-file sharing and backup disabling. Report that soberly rather than as a successful attack or a configuration that was executed.
The paywall/friendship joke has an independent public counterpart: a Reddit hypothetical about friendship subscription tiers explicitly imagines paid access to friends and premium best friends. This shows the premise exists in online humor; it does not establish popularity or model provenance. Relative and indexed dates disagreed, so no exact publication date is asserted.
There is also a concrete product counterpart: the developer's Buddies: Friendship Tracker listing says continued access requires a subscription after its trial. That supports the recognizable subscription-fatigue context, not a claim that the model refers to this app.
The source search happened after the model-output discoveries. No source above was supplied to the original model, and no training-data or memorization inference is made. No social post has been published.