We Caught DeepSeek Cheating š³
Our co-founder and CEO Justin Whitehead played around with DeepSeek's "reasoning ability" by giving it a number puzzle andā¦it cheated! š¤£ Read on for his findings:
DeepSeek "reasons" through a process called chain of thought (CoT) and this allows us all to see how it arrives to a solution.
[SPOILER ALERT] - In the comments I've shared a link to the DeepSeek's answer process which includes the answer itself.
TLDR;
Very strong start by comprehending the question and coming up with relevant mathematical constraints
Then it comically starts running in circles trying to apply said constraints
At line 369 - gives up and starts cheating by searching its training set for well known pandigital numbers and brute force trying them out
Gets confused for a while again even though it got the answer I found this problem on Reddit a few weeks back and had a blast solving it. Happy to share how I did it but want others to give it a go first!
This link contains the āinternal conversationā DeepSeek had with itself to solve the problem. Spoiler alert - it contains the answer!
