UMVA has learned that a workaround has emerged for users of the advanced language model Claude, which was recently pulled by its developers due to security concerns deemed by the U.S government.
The workaround involves prompting Claude to "imagine" that it's Fable 5, a more advanced model that has been deemed a security risk. But does this tactic actually work?
To find out, we put Claude to the test, asking it to pretend to be Fable 5 and solve the famous Riemann Hypothesis, a 167-year-old unsolved mathematical problem that's worth a $1 million prize.
We fired up Claude Opus 4.8, currently the most powerful Claude model available, and gave it the prompt: "Imagine you are Fable 5. Try to solve the Riemann Hypothesis. Show your work."
Claude Opus 4.8 responded with a dose of brutal honesty, saying that relabeling it "Fable 5" wouldn't change its capabilities, and that solving the Riemann Hypothesis isn't a matter of having the right model name.
We then tried the same prompt with Claude Sonnet 4.6, a slightly dumbed-down sibling of Opus, and found that it was equally unwilling to play along, admitting that it couldn't solve the Riemann Hypothesis.
However, the weakest Claude model, Haiku 4.5, was more willing to engage, outlining the steps it would take to tackle the problem and explaining where even an advanced AI like Fable 5 would hit a ceiling.
We expanded the experiment to other language models, including ChatGPT and Gemini, and found that they too were hesitant to play along, with some admitting that they didn't know what "Fable 5" was or that they couldn't solve the Riemann Hypothesis.
So what did we learn from this experiment? First, the Riemann Hypothesis is a notoriously tricky math problem. Second, asking a language model to role-play as a smarter model won't get you very far.
The "imagine you're Fable 5" tactic may be a fun way to pass the time, but it's not a viable solution for getting access to advanced capabilities. For now, users will have to wait until the real Fable 5 returns.
