the wildcard glob/pattern asterisk tokens (like `/**`, `/*`, `'*`, `:*`, `,*`) inside file path or pattern strings in code, and to a lesser degree other placeholder-like tokens such as `__` in `'__UNKNOWN__'`.

the tokens in a math derivation where a derivative has just been set equal to its expression, firing on the first terms of the right-hand side, such as the leading coefficient or factor right after the equals sign.

The neuron fires on boilerplate text in source-code headers and comments, especially license and copyright notices ("Licensed under the Apache License", "This file is part of…", "free software"), and also on commas separating arguments or values in code lists.

The end of a multiple-choice answer option (especially its final token or the newline after it) in quiz questions, which predicts the next option label or the answer, with strongest activation on the option that is likely the correct answer.

random-looking hexadecimal hash or ID strings, such as those embedded in URLs, file names, or identifiers, with the strongest activation on the later characters of the string.

the word or token that completes a transitive action or relationship, typically the direct object or noun after a verb, where the model predicts a specific upcoming completion (e.g., "provides... a place," "verbal abuse," "frequent attacks," "grip," "squeezing," "brutal," "slap," "Welcome," "some cookies"). Simplified: it fires on words that come right at the point where something is given, done, or inflicted to someone, especially an impactful or intense act (abuse, attack, slap, grip, ambush, brutal reception) or a gift or offer.

Neuron 4 mostly fires on delimiter tokens, such as commas and semicolons that end an item in a list of parameters, properties or statements in code, and on line breaks after completed lines. Its activations are weak and noisy, and it goes negative on many other tokens.

The neuron fires on the head word of a headline or sentence subject (often a noun phrase) just before the main verb or predicate would appear, so it seems to be anticipating the verb that comes next.

the word "domestic" (and similar terms like "abuser" or "domestic" in Spanish) in contexts about domestic violence/abuse, plus related words in the same topic such as violence, abuse, and protection orders.

Looking at the highest activations: " Feb" (1.42) in a forum join date, "ия" in "Категория" (1.41), " horizon" (1.12), " freedom" in "degrees of freedom" (1.08), " of" in "front of" (0.906), "Posts" (0.88 in forum profile headers), " steps" (0.88), "princesa"/"a" (0.8-0.9). The activations are weak and scattered. The neuron fires on the key noun or word that completes a standard, formulaic phrase or metadata field in web text (forum headers, category labels, math-problem phrasing). **Neuron 4 explanation:** it finds words that complete a set phrase or standard label, especially in forum metadata and math or educational text (such as "Category", "Posts", "degrees of freedom", "horizon", and month names in join dates).

The neuron fires on a word or token in the middle of a phrase or list item that predicts a specific, expected continuation, such as the second half of a name, term, or fixed expression, and especially a token that sets up a following noun or object. Examples are "Onan", "walk-out", "last four digits", "Drop the mod file to", "Do not open", and "lilies of the field".

the word "movie" (or similar film/TV terms) in the context of Christmas or holiday movies, firing most strongly on the final word of phrases like "Christmas movie," "holiday movie," and "Christmas movies."

the final digit of a year or number in dates and numeric ranges (e.g., the last digit of a year in a forum timestamp, or book/section numbers), and also the end of phrases meaning "it makes sense" or "is significant" (e.g., "sense in", "significance", "tiene sentido").

the digit or number token that is a multiple of three (such as 3, 6, 9, 12, 36), especially in math or numeric contexts, with the strongest firing on the number itself.

the word "screen" (and related screening/sieve terms) used for sifting or separating materials, especially in industrial equipment contexts, plus a few other odd tokens such as "X", "-sub", and "Law" with weaker activation.

Cosmetic surgery involving breast or skin lifting and reshaping, especially breast lift or mastopexy and sagging (ptosis) of breasts, skin, or arms, across languages.

the concept of gifts and gift registries (e.g., "gift," "registry," "present," and their equivalents in other languages), especially the noun naming the gift itself.

Tokens inside restaurant and culinary text that name chefs, restaurants, or dishes, with the strongest activations on the final or distinctive piece of a name or dining term (such as a restaurant name, "wood-fired," or "omakase").

the word "Christmas" (and, more weakly, closely related terms like Christmas tree), with some weaker firing on mathematical terms like "zeros" and "constant" in algebra contexts.

the digit "2" or "5" appearing as the second digit of a multi-digit number, particularly in the 5th or 6th position of a long numeric figure (such as 12,000, 1.2 万, or A15xxx sequence IDs), marking the start of a number's secondary magnitude. Looking more carefully at the pattern: the neuron fires on a digit (mostly 2, or 5 in "A15…" IDs and "15" in "1 15 万") that directly follows a leading "1" or "萬/億" in a large number, i.e., the digit after a leading "1" in numbers like 12,000 or 1.2 million, and in OEIS IDs beginning "A15". **Summary:** it detects the second digit (typically "2" or "5") following an initial "1" in a number, such as in "12 thousand" or "15".