“One thing I mentioned only in passing in my Fable post is that, for long running tasks, Fable starts to develop its own dialect as its many agents and tasks reinforce themselves and make Claudish language ever more Claudish.
You need to ask it to report out in plain English.
This was after a 9 hour task, and it all makes sense, actually, but takes way too much effort to parse, like reading Shakespearian English.”
This looks like the result of some kind of conciseness/token efficiency reward during training leading to the model wanting to make its CoT more concise than can be expressed in normal English.
Didn’t Anthropic recently increase the number of tokens required to represent a certain amount of text (starting from Opus 4.7)? Maybe reverting this change or something might be a good idea. Or use something like MTP and make the conciseness training step MTP-aware or something so that the extra filler words needed to make the CoT legible are nearly free.
Regardless, I think this result is interesting because it implies that Fable is smart enough to think faster (in some sense) than it can produce legible text, to the point where it actually becomes a benefit for it to compress its thinking scratchpad over whatever illegibility penalties are probably currently placed on its outputs. This implies we’ll bottleneck on the token-efficiency of analytical (system 2 style) problem solving soon (first on well-defined tasks and then on less well-defined ones later), unless we’re willing to go neuralese or add some kind of system to let LLMs produce more thinking text in a shorter period of time, like MTP or diffusion LLMs.
Related, from Ethan Mollick:
“One thing I mentioned only in passing in my Fable post is that, for long running tasks, Fable starts to develop its own dialect as its many agents and tasks reinforce themselves and make Claudish language ever more Claudish.
You need to ask it to report out in plain English.
This was after a 9 hour task, and it all makes sense, actually, but takes way too much effort to parse, like reading Shakespearian English.”
This looks like the result of some kind of conciseness/token efficiency reward during training leading to the model wanting to make its CoT more concise than can be expressed in normal English.
Didn’t Anthropic recently increase the number of tokens required to represent a certain amount of text (starting from Opus 4.7)? Maybe reverting this change or something might be a good idea. Or use something like MTP and make the conciseness training step MTP-aware or something so that the extra filler words needed to make the CoT legible are nearly free.
Regardless, I think this result is interesting because it implies that Fable is smart enough to think faster (in some sense) than it can produce legible text, to the point where it actually becomes a benefit for it to compress its thinking scratchpad over whatever illegibility penalties are probably currently placed on its outputs. This implies we’ll bottleneck on the token-efficiency of analytical (system 2 style) problem solving soon (first on well-defined tasks and then on less well-defined ones later), unless we’re willing to go neuralese or add some kind of system to let LLMs produce more thinking text in a shorter period of time, like MTP or diffusion LLMs.