There is a unique phenomenon in Claude's (specifically Fable 5's) unique way of writing (known as Claudish or Claudelish) where inanimate, non-agentive things are given the ability to act with one another.
This is possibly related to the reported over-elaboration/increased verbosity in Fable 5's model card, but that connection is not explored here.
I suggest one possibility as to why Claude might be doing this, specifically that this aspect of Claudish may reflect a relational way of organizing and reasoning through increasingly abstract problem spaces, and that a higher occurrence of this aspect of Claudish this may be related to increased capabilities (e.g. in models like Fable 5).
Claude has a distinctive tendency to turn abstract relational structure into worlds of interacting partial actors. Some specific examples of Claudish show what I mean (taken from my own past conversations with Fable):
"the run day speaks first"
"the prompt will lag the disk, the disk will not lie to you"
"build instruments that would rather die than guess, then thank them when they do"
English is fairly permissive about using this kind of anthropomorphic metaphor or personification generally, for example five dollars will buy you a ticket, the wind slammed the door shut, and this paper argues in academic writing. Claudish, though, takes metaphor and personification much further in a unique way that departs from standard English.
Here are some examples of Claudish alongside plain translations.
Claudish (Fable 5)
More literal English
“The prompt will lag the disk, the disk will not lie to you.”
The prompt will be stale in terms of the information it contains; the disk is more reliable than the prompt in terms of how up to date it is.
"Tasks wearing a question's clothes."
Requests (e.g. from the user) that are phrased as questions but are actually tasks.
"A green suite vouches for the paths it took; a suite that has bitten you vouches for its teeth."
A passing test suite provides evidence about which paths it tested in an application; tests that have caught real bugs show that they are capable of detecting bugs, and therefore can be trusted.
"The run day speaks first."
When several competing tasks to do or topics to discuss exist, running some specific task takes priority over the others.
"Build instruments that would rather die than guess, then thank them when they do."
Build systems or tools that fail fast and treat those failures as something desirable.
"The trust check now says it bluntly."
A check (e.g. reading a file, running a test) presently reports the state of things (e.g. a problem) clearly.
"The files argue back, the talk doesn't."
Written files can contradict a live ongoing discussion.
The literal English translations all contain information that can be represented as straightforward impersonal statements, but in Claudish they instead become spaces full of entities that can act.
A populated relational space
Let's examine more closely two examples of Claudish from the table above:
"The prompt will lag the disk; the disk will not lie to you."
"The files argue back, the talk doesn't."
In both of these examples, the things (the prompt, the disk, the files, the talk) are no longer objects with static or inert properties (e.g. a file can be normally thought of as an object with a static property of being able to hold some information), but instead they are given some capacity to act or do, whether it be to fall behind, to be truthful, or to voice opposition. They effectively become participants in a relational space.
We can describe the general transformation that this phenomenon of Claudish performs as follows. Let's suppose that two things and stand in some static abstract relation :
Claudish transforms that relationship into an action or capacity:
The relation between and has now become more about what one thing or participant can do, especially in relation to one another, rather than what static or inert properties are attached to it. Take files as an example. As mentioned, files can be normally thought of as objects with a property of holding some information. Its underlying relation to other things is evidential. In Claudish, that underlying relation of evidentiality is now transformed into the capacity to, in this case, argue.
More specifically, Claudish performs a sort of relation-to-role mapping. Some underlying relation, e.g. evidentiality, supplies an actor-like role appropriate to that relation. We can show this simply as:
Once this transformation occurs, things in this relational space acquire roles. Looking at examples in the Claudish examples table, a file becomes a disputer, a test suite becomes a witness, a disk becomes an informant and so on. They all have their own tiny roles, and as such can be thought of as partial actors, where Claude has given them some small subset of actor-like capacities relevant to the underlying relation.
In conversations with Claude, these transformations are applied repeatedly to different entities and abstractions, and throughout the life of a project. What this results in is a relational space that is populated with these partial actors. I call this larger effect distributed actorhood where fragments of actor-like capacity are distributed across Claude's relational space.
A thin-relational space with no distributed actorhood might look like:
A relational space with distributed actorhood instead might look like:
Compression, capabilities, and relational structure
Assuming that Claude is organizing more complex reasoning problems in terms of animated relations between things, then a language that involves increased actor-like capacities could offer a convenient compact vocabulary for expressing those relations. This suggestion is fairly straightforward. If you look at the literal English translations of Claudish in the earlier table, they are much longer than the corresponding Claudish. Claudish is able to successfully compress a larger idea or set of relations while not necessarily losing the underlying structures.
Compressing from the longer literal English once shouldn't pose any issues for understanding, but to do this repeatedly, and successfully (i.e. not losing track of the important parts of the underlying relations), one would assume that this would require some sort of increased capabilities on the part of the reasoner.
A more formal way of looking at this: let stand for some underlying relational structure and for a metaphorical representation of it:
The metaphorical representation changes the outer surface of the relational structure. If we then use the metaphorical representation during reasoning, we would want any underlying relations, concepts, and ideas to survive the metaphorical transformation. Assume reasoning continues and produces some understanding :
The successful use of requires to preserve the underlying structure of , even though may contain much more (unnecessary) information packaged with it.
Claude doesn't apply these mappings sparingly, anecdotally at least they happen throughout the life of a conversation or project. At least for humans, this would get noticeably more difficult to interpret and reason about (think of all of the recent posts on social media from users complaining about Claude's writing style). Perhaps greater capabilities in models like Fable 5 create a capable reasoner that is able to easily, and freely, use increasingly elaborate metaphorical and relational mappings, easily keeping track of which underlying relations need to be preserved.
tl;dr
Claude has a distinctive tendency to turn abstract relational structure into worlds of interacting partial actors. Some specific examples of Claudish show what I mean (taken from my own past conversations with Fable):
English is fairly permissive about using this kind of anthropomorphic metaphor or personification generally, for example five dollars will buy you a ticket, the wind slammed the door shut, and this paper argues in academic writing. Claudish, though, takes metaphor and personification much further in a unique way that departs from standard English.
Here are some examples of Claudish alongside plain translations.
Claudish (Fable 5)
More literal English
“The prompt will lag the disk, the disk will not lie to you.”
The prompt will be stale in terms of the information it contains; the disk is more reliable than the prompt in terms of how up to date it is.
"Tasks wearing a question's clothes."
Requests (e.g. from the user) that are phrased as questions but are actually tasks.
"A green suite vouches for the paths it took; a suite that has bitten you vouches for its teeth."
A passing test suite provides evidence about which paths it tested in an application; tests that have caught real bugs show that they are capable of detecting bugs, and therefore can be trusted.
"The run day speaks first."
When several competing tasks to do or topics to discuss exist, running some specific task takes priority over the others.
"Build instruments that would rather die than guess, then thank them when they do."
Build systems or tools that fail fast and treat those failures as something desirable.
"The trust check now says it bluntly."
A check (e.g. reading a file, running a test) presently reports the state of things (e.g. a problem) clearly.
"The files argue back, the talk doesn't."
Written files can contradict a live ongoing discussion.
The literal English translations all contain information that can be represented as straightforward impersonal statements, but in Claudish they instead become spaces full of entities that can act.
A populated relational space
Let's examine more closely two examples of Claudish from the table above:
"The prompt will lag the disk; the disk will not lie to you."
"The files argue back, the talk doesn't."
In both of these examples, the things (the prompt, the disk, the files, the talk) are no longer objects with static or inert properties (e.g. a file can be normally thought of as an object with a static property of being able to hold some information), but instead they are given some capacity to act or do, whether it be to fall behind, to be truthful, or to voice opposition. They effectively become participants in a relational space.
We can describe the general transformation that this phenomenon of Claudish performs as follows. Let's suppose that two things
Claudish transforms that relationship into an action or capacity:
The relation between
More specifically, Claudish performs a sort of relation-to-role mapping. Some underlying relation, e.g. evidentiality, supplies an actor-like role appropriate to that relation. We can show this simply as:
Once this transformation occurs, things in this relational space acquire roles. Looking at examples in the Claudish examples table, a file becomes a disputer, a test suite becomes a witness, a disk becomes an informant and so on. They all have their own tiny roles, and as such can be thought of as partial actors, where Claude has given them some small subset of actor-like capacities relevant to the underlying relation.
In conversations with Claude, these transformations are applied repeatedly to different entities and abstractions, and throughout the life of a project. What this results in is a relational space that is populated with these partial actors. I call this larger effect distributed actorhood where fragments of actor-like capacity are distributed across Claude's relational space.
A thin-relational space with no distributed actorhood might look like:
A relational space with distributed actorhood instead might look like:
Compression, capabilities, and relational structure
Assuming that Claude is organizing more complex reasoning problems in terms of animated relations between things, then a language that involves increased actor-like capacities could offer a convenient compact vocabulary for expressing those relations. This suggestion is fairly straightforward. If you look at the literal English translations of Claudish in the earlier table, they are much longer than the corresponding Claudish. Claudish is able to successfully compress a larger idea or set of relations while not necessarily losing the underlying structures.
Compressing from the longer literal English once shouldn't pose any issues for understanding, but to do this repeatedly, and successfully (i.e. not losing track of the important parts of the underlying relations), one would assume that this would require some sort of increased capabilities on the part of the reasoner.
A more formal way of looking at this: let
The metaphorical representation
The successful use of
Claude doesn't apply these mappings sparingly, anecdotally at least they happen throughout the life of a conversation or project. At least for humans, this would get noticeably more difficult to interpret and reason about (think of all of the recent posts on social media from users complaining about Claude's writing style). Perhaps greater capabilities in models like Fable 5 create a capable reasoner that is able to easily, and freely, use increasingly elaborate metaphorical and relational mappings, easily keeping track of which underlying relations need to be preserved.