You'll get some hostility around here for all the slop-text, but the board of personas with anti-cheat public attestation seems like the beginning of a useful forecasting tool.
I tried building something similar to make 72-hour predictions about the US war on Iran, but found that the persona subagents were far too naive. They believed official statements and media reports at face value, they failed to read between the lines or apply principles of bounded distrust. They accepted spin and wartime propaganda and didn't give enough weight to underlying incentives. They didn't learn from their mistakes from earlier rounds or downgrade their trust in sources after their statements were repeatedly proven false.
It must be possible to improve accuracy with memory, system prompts, progressively changing subagent weights based on historical performance, etc.
I found it helpful to allow the personas to talk to each other. A pure weight of 20% each for 5 personas that are blind to the arguments of the others didn't work as well as personas that modified their rationales after reading the output of the others.
After each prediction resolves, I would have each persona create a post-mortem analysis of what they got right and wrong. Maybe visibility into prior post mortems of their own persona and those of others on the board could allow them to recognize historical cognitive biases and recalibrate for the next prediction.
Presumably the hosted version will have a leaderboard of some sort. Each board might not be able to cheat, but if users can cheaply create many sockpuppet boards, you'll see the Baltimore Stockbroker Scam emerge on the leaderboard. If the public attestation is to be meaningful, it must be difficult and expensive to create new boards.
I tried building something similar to make 72-hour predictions about the US war on Iran, but found that the persona subagents were far too naive. They believed official statements and media reports at face value, they failed to read between the lines or apply principles of bounded distrust. They accepted spin and wartime propaganda and didn't give enough weight to underlying incentives. They didn't learn from their mistakes from earlier rounds or downgrade their trust in sources after their statements were repeatedly proven false.
It must be possible to improve accuracy with memory, system prompts, progressively changing subagent weights based on historical performance, etc.
I found it helpful to allow the personas to talk to each other. A pure weight of 20% each for 5 personas that are blind to the arguments of the others didn't work as well as personas that modified their rationales after reading the output of the others.
After each prediction resolves, I would have each persona create a post-mortem analysis of what they got right and wrong. Maybe visibility into prior post mortems of their own persona and those of others on the board could allow them to recognize historical cognitive biases and recalibrate for the next prediction.
Presumably the hosted version will have a leaderboard of some sort. Each board might not be able to cheat, but if users can cheaply create many sockpuppet boards, you'll see the Baltimore Stockbroker Scam emerge on the leaderboard. If the public attestation is to be meaningful, it must be difficult and expensive to create new boards.