CYOA: old RPGs versus GPT-3
Our life is like a Chooose-Your-Own-Adventure game. But we often view its mechanics of actions and outcomes like those of a traditionally engineered RPG with limited options. Now, GPT-3 gives us an excellent analogy for the truth of how our life stories truly unfold–it is generated along the fly, with no limit to its possibilities.
Let’s take a look at how this analogy captures life if we view its mechanics through a traditional RPG versus GPT-3. In both scenarios, imagine we’re playing as a detective, uncovering information and theorizing about where it may lead us.
Traditional RPG stories
In the case of a traditional game, you find a criminal and are interrogating him. The game’s engineers had to encode options for this interrogation sequence. They’ve allowed you to ask the criminal:
- “Where were you yesterday?”
- “What’s in your pocket?”
- “Who do you work for?”
You can realistically expect that each option will guide you further down an expected path. Perhaps you’d end up travelling to that town he was previously at to generate more clues, draw some conclusions from the items in his pocket, or check the police database for the boss he named. The possibility space is large, but it’s not infinite.
For those unfamiliar, GPT-3 is an AI model that generates text it thinks best continues the prompt you provided–basically a very smart autocomplete. Consider GPT-3 as the generator of the RPG plot. Your next action can pretty much be anything (simply describe it written word, in the prompt), and then GPT-3 generates the story’s continuation on the fly. Thus any actions are possible. The only limitation is the literal mathematic relationships of our words. But this representation allows for a possibility space that is large enough to depict anything humans can express in our words. Which is a huge step above traditional RPGs. An intuitive way to think about this is how new, good novels are written every day using mostly the same words in a language.
How does this translate to our life stories? Often, we see life as a story whose uncertainties are guided by our expectations of how the provided choices will lead us to semi-predictable results. But in reality, our stories are more similar to GPT-3 in that it allows and grows along with any decision possible.
In the above example, the bold first few lines were inputted by me, and GPT-3 generated the non-bold text that followed. Then I added the bold line about stuffed animals and asked GPT-3 to generate more. What traditional RPG would be able to handle such a sudden departure? But GPT-3, like life, takes it and runs with it.
When viewing our future possibilities as if it were an RPG game, we might hear other people’s experiences and presume ours will follow similar if not the same details. A job might make us money and teach us about X industry. If we pick up a book about X, we’ll simply learn something about X and nothing more. Moving to a city is tough because you don’t know anyone. Never operate under uncertainty because only bad things will happen. Switching industries is dangerous. Etc.
But under the GPT-3 framework, more of reality is fluid. Nothing’s off the table. It’s not like we’ll walk to some region of space and our lives will suddenly error and not render (which would happen if you somehow maneuvered to off-limit regions of the map in old video games). With GPT-3, the possibility space is represented by as many words in our language.
With the interrogation example, we could suddenly decide to check the criminal’s underwear brand, so that we might learn something. We could engage the criminal in a discussion of philosophy. Or we could defect and join the criminal.
How to Live Life
In real life, our stories are generated more similarly to GPT-3 than traditional RPGS. Reality flows with whatever we decide to do. Furthermore, the most unexpected events are the most impactful ones, so that’s where all the fun in life is anyway. If we do something super weird, it could be a limit experience that teaches us some of the most important lessons we’ll ever learn. If we take a job in another city, or shoot for some moonshot goal, the atoms in the universe say “Okay! You’re doing that, so let’s roll with it!” They accomodate you. How do you want to generate your story?
Unfortunately, most people operate by viewing their actions under the traditional video game model. Thus we have tricked ourselves into thinking we are decent at predicting the future. This is because we avoid considering a large swathe of actions from the get-go. These are actions which, had we tried them out, would have shown us that we actually couldn’t predict the future, because surprising, impactful events would eventually happen. (I can’t even get any more specific than these few sentences, since that’s the limit to which I can use words to describe these realities that are unknowable in advance.) Thus we self-fulfill our prophecies of not finding an interesting life with surprises and highs and lows.
So we stick to well-worn paths. Then we look back and believe we did a pretty decent job at predicting the future. Nothing unexpected happens because we took the actions that would guarantee that.
Furthermore, our ability to appreciate unexpected positives is even worse than our ability to predict unexpected negatives. It’s a natural side effect of evolution to make us think in ways that keep us physically safe, but this no longer applies in an opportunity-rich, (relatively) physically safe world filled with coincidences. But if we only stick to the expected paths, we’ll be limited to the unimpactful, expected small positives. We travel excruiciatingly safely to death’s door.
We should start imagining our life story as if it was generated by GPT-3. Throw whatever you want at reality; the universe will conspire to send you where you want to go. Other people will take note of your boldness. They will want to help you. Shake off the rust of only looking at the choices that everyone is already familiar with. Your actions will bring you slowly, but surely, towards a better future for you.
This is the best (and only) Choose Your Own Adventure we’ll get to play.
one likely downside is that GPT-3 can only generate situations/ideas that follow patterns found during training. On the bright side it was trained on 40GB of Internet text
It’s thus debatable if GPT-3 can generate something “new”. I think the complexity generates “enough” newness to create tangible new ideas for humans; we only need small prompts.
Problem: maybe we don’t have the right words to capture things. This is seen in words captured in other languages that English doesn’t have. So the next step of innovation is thinking in a way that trascends the language you operate with.