Include Training in SCI - Option #2 (non-time based)

[noahcb]

Creating a space to discuss one potential option for including carbon generated during training in the SCI for AI. High level components could include:

• The definition of training would include just the process of updating model weights prior to release.
• We would treat AI systems as versioned where all of the carbon generated would be included for the SPECIFIC version.
• Model producers would predict the usage of the system to see how many inferences they would distribute the training carbon across.
• Allocation would be 1) an exponential decay function to incentivize longer-term use or account for training carbon sooner given the reality that these systems are obsolete faster than an anticipated or 2) determined by the model producer.
• Any unallocated carbon at model End of Life would roll over into the next version.

Unaddressed issues include:

• Speed of versioning making this very difficult
• Continuous training might break this framework
• Open source models present unique challenges compared to closed source given proliferation and accountability.

[jawache]

@noahcb so given the above if you took a pre-trained foundational (open source?) model and "tuned" it, it's the "tuning" that would be allocated to the functional unit and not the cost of training the foundational model?

[jawache]

WG:

• With OS models, it becomes hard to allocate training carbon, can be used multiple times. Can be gamed by just creating lots of downstream models which take a portion.
• If I have model A that is forked into B, C D (maybe fine tuned), allocate based on expected usage of B, C, D from an organization expectation.
• A organization has expected 1M inf requests, but 500k later they produce a version 5, roll over the carbon to B.
• Counter abuse, generate a cost for abuse which is higher than the benefit received from the abuse.
• One method to reduce chance of abuse is to fix the number of inference to say 1 million and leave it at that.
• Allocate the training to first 1m inferences, after that it's 0 -> this means that you are incentives to use older models.
• Incentive squeezing every last drop of value from a model once it's been created.