I'm always a bit sceptical of these embarrassing examples being "fixed" after they go viral on social media, because it's hard to know whether OpenAI addressed the underlying cause or just bodged around that specific example in a way that doesn't generalize. Along similar lines I wouldn't be surprised if simple math queries are special-cased and handed off to a WolframAlpha-esque natural language solver, which would avert many potential math fails but without actually enhancing the models ability to reason about math in more complex queries.
An example from ChatGPT:
"What is the solution to sqrt(968684)+117630-0.845180" always produces the correct solution, however;
"Write a speech announcing the solution to sqrt(968684)+117630-0.845180" produces a nonsensical solution that isn't even consistent from run to run.
My assumption is the former query gets WolframAlpha'd but the latter query is GPT itself actually attempting to do the math, poorly.
An example from ChatGPT:
"What is the solution to sqrt(968684)+117630-0.845180" always produces the correct solution, however;
"Write a speech announcing the solution to sqrt(968684)+117630-0.845180" produces a nonsensical solution that isn't even consistent from run to run.
My assumption is the former query gets WolframAlpha'd but the latter query is GPT itself actually attempting to do the math, poorly.