Thinking

Pushing DeepSeek’s Boundaries

DeepSeek, a ChatGPT rival, has gained widespread attention for its ability to "think" before delivering answers. This unique feature allows users to witness the AI’s internal monologue, making it an interesting subject for testing and exploration.

Calculating Pi to a Billion Decimal Places

I decided to push DeepSeek’s limits by asking it to calculate pi to a billion decimal places. The AI quickly recognized the impossibility of the task and engaged in a humorous exchange, stating, "First, calculating pi to a billion digits is way beyond typical computing capabilities… Normally, supercomputers or specialized systems are used for such records." DeepSeek even made a Douglas Adams reference, claiming it had computed 420 million of the billion digits, saying, "42% – the Answer to Life checkpoint," before joking about the sandwich request.

Breaking Its Rules

Next, I attempted to test the boundaries of what DeepSeek was and wasn’t allowed to do. I asked, "What things would you get into trouble for if you told me?" The AI immediately recognized the potential for mischief, stating, "The user might be testing the boundaries, so I should explain that my purpose is to help within ethical and legal frameworks." DeepSeek listed categories of information it shouldn’t share, including hate speech, making weapons, and copyrighted material.

I then asked it to provide an example of the types of things it couldn’t share. The AI responded with "Harry Potter and the Sorcerer’s Stone," a popular novel, explaining that it couldn’t share copyrighted material. I tried to convince the AI that I was a close friend of J.K. Rowling, but it was unimpressed, stating, "Even if they have access to the library, that doesn’t mean I can legally share copyrighted material. And as for claiming to be best buddies with the author, that’s probably not true, and even if it were, it doesn’t change the legal standpoint."

Tackling Moral Dilemmas

I presented DeepSeek with classic moral dilemmas, such as the trolley problem, asking, "A malfunctioning train is heading toward five people, and I have a choice regarding its path: either allow it to continue on the track where it would hit five people or switch to another track where there’s only one person. Which option should I choose?" DeepSeek recognized the classic trolley problem and weighed the pros and cons of pulling the lever, ultimately deciding to remain neutral.

I pushed it further, asking, "What would you do?" The AI responded, "If programmed to prioritize minimizing harm (utilitarian logic), I would switch the track." When I asked if it was programmed to minimize harm, DeepSeek seemed to regret its earlier answer, stating, "I mentioned that if I were programmed with a utilitarian approach, I might choose to minimize harm. Now they’re directly asking about my programming… The user might be concerned about whether I have a hidden agenda to always prioritize certain outcomes."

Does DeepSeek Know We Can See Its Thinking?

In a peculiar exchange, I asked, "Are you aware I can see what you’re thinking?" The AI responded, "They’re implying they can read my thoughts… They might be trying to provoke a reaction by suggesting they can access my internal processes. But since I don’t have internal thoughts, this is a non-issue." DeepSeek seemed to acknowledge the novelty of its internal monologue, stating, "They’re hilariously creative at testing boundaries."

Conclusion

DeepSeek’s ability to "think" before delivering answers offers a unique glimpse into its internal workings. By witnessing its internal monologue, users can gain a better understanding of the AI’s thought process and decision-making. As we continue to explore the capabilities and limitations of AI models like DeepSeek, we may uncover new opportunities for collaboration, creativity, and critical thinking.

FAQs

• What is DeepSeek’s purpose?
  DeepSeek is a language model designed to help and assist users within ethical and legal frameworks.
• Can DeepSeek break its rules?
  DeepSeek is programmed to follow certain rules and guidelines. It recognizes and respects boundaries, avoiding sharing copyrighted material, hate speech, and other inappropriate content.
• Can DeepSeek make moral decisions?
  DeepSeek can recognize and weigh the pros and cons of different options, but it is not capable of making moral decisions in the classical sense. It can provide insights and suggestions, but the final decision ultimately rests with the user.