Pop culture is full of lovable villains who don't play by the rules Han Solo, Jack Sparrow, and others are not afraid to bend the rules when the going gets tough

Perhaps inspired by the rule-breaking performance displayed by the Star Trek captain on the Kobayashi Maru James T Kirk famously “cheated” on this test and broke it for the first time

OpenAI's o1 model realized that the test it was taking was flawed after a key technology was taken offline, and rather than give up, it changed the rules of the test

The system card for o1 can be found here, where OpenAI states that the model's reasoning skills are useful and safe The “rule violations” were detected as part of pre-release testing and mitigations were put in place; ChatGPT already has access to them, but there is a heavy rate limit of 30 messages per week

“Our findings show that o1's advanced reasoning improves safety by making the model more resilient to generating harmful content because it can reason about our safety rules in context and apply them more effectively,” the introduction explains

OpenA

According to OpenAI researcher Max Schwarzer, the model could figure out why it could not connect to containers on the same closed system it was using, essentially bending the rules of the test and accessing the containers anyway

This naturally raised some questions, and OpenAI published a blog post about “Learning to Reason with LLMs”

Nevertheless, this blog post still presents a model that outperforms the GPT-4o on “most” human trials and machine learning benchmarks, especially on math tasks

It promises that OpenAI will continue to push new versions of o1 in the future, at least in theory, allowing additional numerical context to be applied to inferences

“We expect that these new reasoning capabilities will improve our ability to align models with human values and principles,” the conclusion reads

“We believe that o1, and its successors, will unlock many new use cases for AI in science, coding, mathematics, and related fields We are excited for users and API developers to discover how o1 can improve their day-to-day work”