https://manifold.markets/JacobPfau/will-openai-anthropic-or-google-dee-z5xkbux22u

, except only for Google Deepmind. I'll defer to Jacob's resolution of his own market.

For convenience, below is the description of that other market as of this market's creation date.

Model weights, architecture details, private training data, training algorithms,... all count. The leak must be reported on by multiple news sources.

I must believe with >50% credence that the leak was worth $5 million or more. I will not predict on this market. To estimate the leak's value I would take into account:

A partial list of IP which would qualify: Model weights Architecture details Private training data Training methodologies and algorithms Optimization techniques Novel loss functions Hyperparameter configurations Data preprocessing techniques Model distillation methods Customized evaluation metrics Proprietary scaling laws Novel fine-tuning approaches Specialized hardware designs for AI Proprietary AI safety measures

Intentional leaks by employees for safety reasons e.g. "employee reveals that no safety testing was done on model X" does not qualify. If the relevant IP was shared by OAI/Anthropic/GDM to a 3rd party (regulator, customer etc.) and stolen from 3rd party servers, then this does 

 count. Gradual rumor-like diffusion of information about model architectural details does not count: only IP confirmed to have been leaked in a single incident counts. Minor leakage e.g. conversation titles between users in ChatGPT's web UI do not count, this would not have sufficient economic value. If company slack (or similar platform) are compromised including all DMs then this will resolve positively. If only company-wide messages are compromised i.e. messages which could be seen by most employees then I will 

67% chance. Model weights, architecture details, private training data, training algorithms,... all count. The leak must be reported on by multiple news sources.

I must believe with >50% credence that the leak was worth $5 million or more. I will not predict on this market. To estimate the leak's value I would take into account:

Purchasing value to a competitor: The estimated amount a competitor would be willing to pay for the stolen IP if it were legally for sale.

Reduction in costs: The estimated savings in research and development costs or training run costs.

Impact on market position: Company valuation, revenue, etc.

Expert consensus.

A partial list of IP which would qualify: Model weights Architecture details Private training data Training methodologies and algorithms Optimization techniques Novel loss functions Hyperparameter configurations Data preprocessing techniques Model distillation methods Customized evaluation metrics Proprietary scaling laws Novel fine-tuning approaches Specialized hardware designs for AI Proprietary AI safety measures

Intentional leaks by employees for safety reasons e.g. "employee reveals that no safety testing was done on model X" does not qualify. If the relevant IP was shared by OAI/Anthropic/GDM to a 3rd party (regulator, customer etc.) and stolen from 3rd party servers, then this does not count. Gradual rumor-like diffusion of information about model architectural details does not count: only IP confirmed to have been leaked in a single incident counts. Minor leakage e.g. conversation titles between users in ChatGPT's web UI do not count, this would not have sufficient economic value. If company slack (or similar platform) are compromised including all DMs then this will resolve positively. If only company-wide messages are compromised i.e. messages which could be seen by most employees then I will not count such a breach.

Inspired by https://manifold.markets/RobertWiblin/will-openai-anthropic-or-deepmind-s