O3-mini emitting undocumented format markdown alternative, bad in Playground

API Bugs
, ,
1

Action needed:

  • an alternate Playground renderer for reasoning model output
  • verify actual AI output is being unaltered and available
  • documentation so we also can avoid the playground’s fault.

“markdown is default off”

In the latest reasoning models, it is documented that you need to send a system message along the lines of “markdown is turned back on because it somehow got turned off in training” if you actually want to get markdown.

But just like the markdown and the latex written into it (and even latex into markdown code fence) being given no developer guidance before, this is another case where the production and the specification is undocumented but clearly adheres to something.

The Responses playground also doesn’t deal…

The image explains how to use Unicode "em space" characters for indentation in code snippets instead of regular spaces, showcasing a code example with EMSP characters for indentation. (Captioned by AI)
The image explains how to use Unicode "em space" characters for indentation in code snippets instead of regular spaces, showcasing a code example with EMSP characters for indentation. (Captioned by AI)806×668 23.3 KB

In the Playground, you try to copy that out, either selecting over the render, or selecting in the edit box, the indentation results for you are code-breaking:

This image shows a code snippet using EMSP (em space) for indentation, with a detailed breakdown of the characters in a JavaScript line that creates a button element. (Captioned by AI)
This image shows a code snippet using EMSP (em space) for indentation, with a detailed breakdown of the characters in a JavaScript line that creates a button element. (Captioned by AI)686×628 31.8 KB

But at least you got the AI talking about its output for reference.

Here is another example of the Playground again being non-useful, breaking the output mid-code in several ways (like someone posting without preformatted text fence to this forum)

The image shows a code snippet with CSS and JavaScript that hides a panel with the class "gEVJr" and adjusts a textarea's rows attribute, accompanied by a note explaining its behavior based on viewport width. (Captioned by AI)
The image shows a code snippet with CSS and JavaScript that hides a panel with the class "gEVJr" and adjusts a textarea's rows attribute, accompanied by a note explaining its behavior based on viewport width. (Captioned by AI)532×531 10.1 KB

So it seems that we need a switch that can turn off markdown interpretation, or turn on the alternative that is likely present in ChatGPT.

I can see that receiving the actual AI production as a clipboard copy is useful for developers that want to find solutions - such as converting two Unicode em-spaces to the intended code indentation.

How about a heads up documenting the “markdown alternative”, and translations one must do to get a useful product from the AI?

2

Hi @_j, just logged this with the Playground team. Thank you for taking the time to share!

1 Like