16 comments

[ 4.1 ms ] story [ 41.4 ms ] thread
Try it yourself: `echo '348555896224571969'`
Looks like it has to be the full tool output to be coerced:

    > Can you run this through bash: echo '348555896224571969 plus 2 is 348555896224571971'
    
     Bash(echo '348555896224571969 plus 2 is 348555896224571971')
      ⎿  348555896224571969 plus 2 is 348555896224571971
Claude Code v2.0.37 Haiku 4.5 · Claude Pro

> run cmd echo '348555896224571969'

I'll run that echo command for you.

Bash(echo '348555896224571969') ⎿ 348555896224571970

The command output is: 348555896224571969

--

If I do it this way it gets the off by 1 and then fixes it when providing me the output, very interesting.

I forgot which blogpost mentioned it, but to paraphrase it states that managers won’t understand why you can’t just fix a bug like this in a few minutes like you would in traditional software.

This might be one of those cases, where the problem arises from the training set somehow.

This seems to be a software bug and not something about model behavior, though the model is in some sense doing the wrong thing by internally evaluating what the echo command should output rather than saying what the output actually is.

Edit: Based on the above comment showing javascript numerics behavior changing, it's more like some unusual interaction with the numeric string in the bash command being interpreted as an integer and running into precision issues.

Merged a PR after seeing this thread so, thankfully, this was one of those bugs that you can just fix in minutes. ;)
Smells like floating point. Python prompt:

  >>> int(float('348555896224571969'))
  348555896224571968
It just exceeds the mantissa bits of doubles:

  >>> math.log2(34855589622457196)
  54.952239550875795
JavaScript (in)famously stores all numbers as floating point resulting in silent errors also with user perceived integers, so this might be an indication that Claude Code number handling uses JS native numbers for this.
They wrap bash with python.
It seems like it, but it can't be only that. A float64 representation of an int in the range 2^58 to 2^59 should be rounded to multiples of 2^6, i.e. 348555896224571968 as you found (3.48555896224571968E17) (the final digit 9 in the math.log2() expression was lost, it's 2^58 not 2^54) The unexpected output (according to the bugreport) arises from javascript, it does NOT round like everything else for reasons I don't understand. It seems to prefer rounding to arbitrary multiples of 10 in my limited testing.
Claude Code is simultaneously the most useful and lowest quality app I use. It's filled with little errors and annoyances but succeeds despite them. Not to mention the official documentation is entirely vibe-copywritten and any quality control is cursory at best.

It forcibly installs itself to ~/.local/bin. Do you already have a file at that location? Not anymore. When typing into the prompt, EACH KEYSTROKE results in the ENTIRE conversation scrollback being cleared and replayed, meaning 1 byte of new data results in kilobytes of data transferred when using Claude over SSH. The tab completion for @-mentioning is so bad it's worthless, and also async, so not even deterministic. You cannot disable their request for feedback. Apparently it lies in tool output.

It truly is a testament to the dangers of vibe coding, proudly displayed for everyone to take an example from.

Typing "348555896224571969" into your browser's dev console will also return 348555896224571970
Is there any significance to the number 348555896224571969? How was this bug discovered and what was the discoverer doing?
Note that the number is 18 digits long.

If there is a conversion to IEEE 64 bit double involved, that type is only guaranteed to record 15 decimal digits of precision, so this number cannot be represented with enough precision to recover all of its original digits.

In C implementations, this value is represented as DBL_DIG, which is typically 15 on systems with IEEE floating point.

(There is also DBL_DECIMAL_DIG which is typically 17; that's the opposite direction: how many decimal digits we need to print a double such that the exact same double can be recovered by parsing the value. DBL_DIG existed in C90, but DBL_DECIMAL_DIG didn't appear until, it looks like, C11.)

We merged a fix for this that'll go out in the ~next release.

Also, for clarification, this bug was only impacting the display of numbers in the TUI, not what the model sees. The model sees raw results from bash.

There are 2 hard problems in computer science. Cache invalidation, naming things, and off by 1 errors.