The LessWrong hyphen

Note from Caspar: I suggested this (extremely unimportant) topic and gave some input. I didn’t fact-check any of the references.

Writers on LessWrong often put a hyphen between an -ly adverb and the adjective (or participle) after it when the pair comes before a noun, as in “First, some lab builds early transformatively-useful AI” (from a 2025 post). Standard English would write “early transformatively useful AI”. I’ll call this the LessWrong hyphen, because Caspar says he mostly notices this in LessWrong posts.

That rule concerns compound modifiers, i.e., groups of words that together modify a noun. Before the noun, such a group usually gets hyphens, which tell the reader how the words group. For example, in “a small-business owner”, the hyphen shows that “small” describes the business rather than the owner. Style guides exempt compounds that begin with an -ly adverb, such as “highly paid” or “carefully designed” (CMOS Shop Talk 2023). In “a highly paid consultant”, the -ly ending marks “highly” as an adverb. It has to go with “paid”, since a “highly consultant” makes no sense. So the hyphen has nothing left to clarify. The -ly exception does not cover adverbs such as “well”; “a well-known result” keeps its hyphen. Even after an -ly adverb, a hyphen sometimes helps. The -ly ending only shows that the word is an adverb, and an adverb can also modify a verb before it. In “the agent explores randomly generated environments”, a reader may at first take “randomly” to go with “explores”; writing “randomly-generated environments” prevents this misreading. Google’s developer documentation style guide allows the hyphen after an -ly adverb “when needed for clarity”. The Punctuation Guide also calls for the hyphen when the adverb is part of a larger compound modifier, as in “a not-so-poorly-produced sequel”. Many writers also use the hyphen in cases where neither reason seems to apply. In each of the examples below, no verb comes directly before the hyphenated adverb, and that adverb is not part of a larger compound modifier.

The LessWrong hyphen predates LessWrong and appears elsewhere. Older English prose often has it – Austen’s Pride and Prejudice has “deeply-rooted dislike”, and CMOS Shop Talk 2023 gives further examples – and it shows up in academic computer science, e.g., “computationally-bounded” in a 2013 paper. All examples below come from LessWrong, but I have no data on whether the hyphen is more common there than in other technical writing. My guess is that the hyphen is common on LessWrong because many LessWrong writers have technical backgrounds. In mathematics and programming, parentheses show which terms belong together, as in \((a + b) \cdot c\). Ordinary prose has no comparable way to bracket words within a phrase. Writers who are used to marking groupings explicitly may use the hyphen as a substitute, even where the -ly ending already makes the grouping clear. Finally, some words that end in -ly are adjectives or nouns, such as “early” and “family”. Compounds that start with such words take a hyphen in standard English (“early-stage project”, “family-owned business”), so they fall outside the pattern.

Examples

I have added bold to mark the compound in question.

• From a 2025 post: “First, some lab builds early transformatively-useful AI”; “In the simplified model, there are two strategically-important stages of AGI: early transformative AGI, and superintelligence”.
• From a 2022 post: “there is no Ghost of Christmas Past that will magically float into a sufficiently-capable AI and imbue it with human values”; “clearly not an unbiased randomly-selected sample”. The post has several more instances.
• From a 2021 post: “there isn’t really an outer alignment problem for Go: a precise, algorithmically-generated training signal (the win/loss signal)”. The post’s title, “The case for aligning narrowly superhuman models”, leaves “narrowly superhuman” unhyphenated.
• From a 2024 post: “But this ‘logical omniscience’ is impossible for computationally-bounded agents like ourselves.”
• A 2024 post titled “Deceptive AI ≠ Deceptively-aligned AI” opens with: “A ‘deceptively-aligned AI’ is different from (and much more specific than) a ‘deceptive AI’.”
• From a 2024 post: “having built a highly-capable AI system with some structural similarities to the human brain”. The same post writes “broadly superhuman AI” without a hyphen.
• The LessWrong wiki tag “Fully-General Counterargument”. The tag’s description writes “fully general counterarguments”.
• From a 2024 post: “We have some interface to our hopefully-corrigible agent - in this case the shutdown button.” Another 2024 post, which argues for more hyphens in general, lists this hyphen as one to remove.
• From a 2012 post: “For trivial matters, sure, for science too, but not for our strongly-held opinions.”
• From a 2019 post: “Eventually our society’s trajectory will be determined by powerful optimization with easily-measurable goals rather than by human intentions about the future.”
• From another 2019 post: “Someone with the intent to produce a potentially-dangerous agent, but not the means or knowledge, is granted access to supplies …”
• From a 2023 post: “people may think of some typical widely-shared tweet of his, which probably scores highly on the troll scale.”
• From a 2024 post: “They both basically understand how their environment works, so they generally agree on predictions about any specific directly-observable thing in the world”.
• From another 2024 post: “How sharp is the distributional shift from putting a newly-trained AGI into an environment where it’s capable of extended periods of thinking to itself?”
• From a 2025 post: “I used to think of AI development as obviously being the last fully-automated job.”
• From a 2026 post: “In that world, I do expect us (the LessWrong team) to clean up most of the issues obvious to publicly-available LLMs relatively quickly”.