A New Era of Personal Boards

One Way to Fight Confirmation Bias

I've been spending a lot of time thinking about LLM sycophancy. While this is a problem that the large AI labs have acknowledged, the problem persists.

To be clear, it wasn't like we were already representing the epitome of objectivity. Social media has made it that much easier to retreat into our echo chambers.

We have to actively fight confirmation bias. It isn't easy. I'm not immune and have been trying to develop systems that can provide an external check when I'm making both major and minor decisions.

To fight my own biases, I've experimented with something in ChatGPT. I was initially inspired by a recent Sloan article that outlined how LLMs can be fantastic sounding boards. But instead of confirming your biases, you can instruct those LLMs to take on different personas.

With this hyper-customized personal board of advisors, you can stress test assumptions, get an outside perspective, and identify flaws in your thinking. It can't fully eliminate confirmation bias, but it can highlight things that you may not be thinking about.

The author of the Sloan article decided to use real-life personas in his personal board. While that can work, I like to create "personas" that are hybrids of different real and fictional people.

For instance, one of my flaws is that I can be too deliberate. My analytical nature slows me down when I should be speeding up. To combat this, I created a persona called the "Aggressor." I wrote a fairly basic prompt that tells the Aggressor to challenge my contemplative nature and to push me to act (even if I'm hesitant to do so).

I then also created a persona called the "Delegator." This one fights my bias to take on tasks by myself. Instead of white knuckling through an already-packed schedule, the Delegator helps me brainstorm ways to rely on others.

I've found it helpful to create fairly basic system prompts for these personas. This is so I can quickly paste into new ChatGPT chats. Then, if I need to change a specific advisor's personality, I can easily do so.

Granted, this isn't a perfect system. One flaw is that my board doesn't have persistent memory across context windows. While ChatGPT does offer memories, the board still has to learn a little about me before providing better feedback.

Nonetheless, I think this is a quick and easy way to get an outside perspective on certain decisions. To be clear, I still think a personal board of human advisors is better. The machines can't replace humans yet. However, this can be a good way to get at least some input from a "person" that you may not personally know.

I think all of us could use at least some type of personal board. We need an external sounding board where we can get new views on the challenges we face. Whether you do this with an actual personal board or one constructed by an LLM, I'm confident that it will help you make better decisions. Give it a try!

Prompt of the Week

Like I described above, I've been spending more time getting career advice from LLMs. I think about my career all the time, and it can be all too easy to be myopic.

That's why I like the prompt below. It's simple, but it led me to some interesting ideas. While your huge goal may be different, try something like this:

"My lifelong dream is to own a soccer team. Knowing what you know about me, what are both actionable and meta things that I'm missing that are preventing me from reaching this huge goal?"