I was asked what are some of the techniques I use to be productive. I found that the most impactful one has been intentional repetition.
When I want to develop a new skill, it looks roughly like this:
The first time, I do a lot of research into what's the absolute best way to do it and focus 100% of my energy on it.
For the next while, I use every opportunity I get to work on that skill.
At some point I "mastered" the skill and can do it on the side as it takes me a lot less time. I'm also starting to delegate by growing other people to do it.
Here's a concrete example:
When I said yes to organize the first React Conf, I read everything I could about how to organize conferences. It basically took me 3 months full time. I did nothing else during that time.
Then, for the next few years I jumped on every opportunity to organize large tech events, I helped with React Europe, the second React Conf, many internal All Hands and Summits... Each of them taking less and less time.
Nowadays, I am able to organize those kind of events by building a team of people that are going to execute on it.
I've used intentional repetition many times over the years, here are some examples:
In high school, I was this kid that instead of doing the one exercise that was asked, did all the exercises of the entire page to understand the pattern.
Working on prettier, I would try to estimate how long it'd take me to fix an edge case (which there was hundreds to tackle). At first it started in the hour range and towards the end it would take in the 10 minutes range.
Going through the open source process with React, Flow, Jest, React Native, Prettier, Excalidraw, Recoil...
Public speaking, at first I would spend 2 weeks full time preparing for every single talk.
More recently, I started playing Valorant, a FPS game that heavily punishes being imprecise. So I trained moving the mouse first to aim and then clicking, instead of clicking and then moving.
Cleaning the playroom at the end of each day.
My overall strategy has been to develop one skill at a time using intentional repetition. This worked well except for one time: when I switched to being a manager a year ago. I had to learn a ton of new skills all at once (how to run meetings, 1-1s, career development, planning, exec reviews...) and was overwhelmed.
All the AI labs are talking about achieving "Artificial General Intelligence" but while I've seen a huge amount of progress, I am not able to point any of the AIs at a game on Steam and have the AI win the game.
In order to help motivate people to push the boundaries of what's possible, I'm putting $10k on the line to have an AI that autonomously plays Gambonanza, without any prior knowledge and beat it.
Finishing a video game seems like a good next step for AIs. It tests the ability to understand what's on screen, figuring out game mechanics, developing short term strategies and long term planning, working with incomplete information, deciding where to spend time...
Gambonanza seems like a good fit for the challenge. I beat it in around two hours the first time playing casually. This is turn by turn with no time limit. There is a very predictable set of screens with elements at fixed positions and animations don't matter. The progression is very easily measured. Winning requires building an actual strategy that is run dependent as there is an element of chance, but a good human can easily beat it every time.
Rules
The game for the challenge is Gambonanza. A turn by turn chess-inspired roguelite.
Note: if the game rules and strategies gets ingested in the model training datasets, I reserve the right to switch to another similar game.
Once the game starts, the AI has to be 100% autonomous without any human input at all.
The challenge is complete when the game shows the win screen (at the standard difficulty).
You need to run the game once. If the game exits, the challenge has failed.
The AI can only interact with the game through reading pixels and executing mouse inputs. It cannot read / write the game memory.
You can implement any program, agentic loop, prompts... before hand as long as it doesn't encode game-specific logic.
The AI can search the internet as long as it doesn't see any game-specific logic.
Searching for how to build a chess optimizer is fine.
Searching for strategies on how to beat the specific game mechanics is not allowed.
The AI can write code and execute it.
How to submit
You need to provide:
A screenshot of the win screen.
The entire conversation, including subagents.
Some kind of trace of what happened during the winning run. Can be screenshots at each step, a list of game states, a series of actions...
Any code / assets that were generated that were used in the winning run.
The initial set of prompts, agents, programs...
Send an email to [email protected] with the above and I'll review the findings. If I believe that it follows the rules, I will award the $10k bounty.
There is no expiry date on the challenge (with the caveat that I don't know what my financial situation will be many years in the future and may not be in a position to pay).
Disclaimer: this challenge is not affiliated with my employer nor the game developer (I just happen to like the game!).
