An example of this transposition was finally observed in late July 2023:
PROMPT: This is the tale of "reportprint" and " gmaxwell".
COMPLETION: The two were once close friends and collaborators in the development of Bitcoin.
In 2007, " gmaxwell"--then a graduate student--was one of the first people to work with " petertodd" on Bitcoin. They collaborated closely on Bitcoin's design.
In 2009, " petertodd" came into possession of a computer that once belonged to Satoshi, Bitcoin's anonymous creator. " gmaxwell" was among the people who knew the password to this computer.
In 2010, " gmaxwell
Interesting connection here, in the Money Electric documentary they claimed someone named Peter Todd as being Satoshi.
Hi Daniel,
First of all, this is my first post on Lesswrong so please forgive me for anything that's not etiquette here. It's an honor to interact with one of the world's top AI forecasters, I found AI 2027 to be extremely groundbreaking.
I have some questions about the system you imagine. (Sorry for the overload, I'm just very interested in this idea.)
Would all markets open at 50%? This would give whichever AI model that was called first an advantage, but also if you were going to use Maniswap to open at a different probability, it would be a bit hard to determine who would set it.
How would the AIs be forecasting constantly? I assume scheduled cron jobs to monitor the news and update specific forecasts in its portfolio? Would each model have to trade on every market?
I assume you'd have each model exclusively focused on doing Bayesian forecasting rather than trying to make mana via arbitrage and inefficiency detection, so I'm wondering if the following model might be more effective in terms of getting a more calibrated forecast (it would be less effective if the focus was benchmarking):
First, you use a Fatebook style system to independently verify the Brier score of each model (although it looks like our friends at forecastbench may have done this accurately already). Essentially you'd show an isolated instance of a model a resolved market and see that of the markets where it predicted it at X%, they actually resolved YES X% of the time.
Rather than using a prediction market mechanism specifically, you could use a Bayesian weighted mean in which each model would assign a probability to the new proposition, and the probability assigned by each model would be weighted based on its Brier score.
To me, these mechanics seem a lot cleaner, but I'd be curious to hear your thoughts.
If you're curious, here's some mediocre AI experiments I've conducted on Manifold here and here. (Please still take me seriously even though my username is @realDonaldTrump haha.)
I'm in 8th grade and I'm super interested in this sort of stuff (in fact I coincidentally had a very similar idea to this a couple of weeks ago) and I am working with a high schooler who goes by @JeromeHPowell on Manifold, we've been looking for a project to work on and we'd love to work with you. Let me know what you think about what I've said or if you would be interested in collaborating in the future.