Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The 0.2 OOMs/year target, published by Cleo Nardo on March 30, 2023 on LessWrong.

Paris Climate Accords

In the early 21st century, the climate movement converged around a "2°C target", shown in Article 2(1)(a) of the Paris Climate Accords:

The 2°C target helps facilitate coordination between nations, organisations, and individuals.

It provided a clear, measurable goal.

It provided a sense of urgency and severity.

It promoted a sense of shared responsibility.

It helped to align efforts across different stakeholders.

It created a shared understanding of what success would look like.

The AI governance community should converge around a similar target.

0.2 OOMs/year target

In this article, I propose a target of 0.2 OOMs/year. OOM stands for "orders of magnitude", and corresponds to a ten-fold increase. This corresponds to a 58% year-on-year growth.

I do not propose any specific policy for achieving the 0.2 OOMs/year target, because the purpose of the target is to unify stakeholders even if they support different policies.

I do not propose any specific justification for the 0.2 OOMs/year target, because the purpose of the target is to unify stakeholders even if they have different justifications.

Here is the statement:

"Humanity — which includes all nations, organisations, and individuals — should limit the growth rate of machine learning training runs from 2020 until 2050 to below 0.2 OOMs/year."

The statement is intentionally ambiguous about how to measure "the growth rate of machine learning training runs". I suspect that a good proxy metric would be the effective training footprint (defined below) but I don't think the proxy metric should be included in the statement of the target itself.

Effective training footprint

What is the effective training footprint?

The effective training footprint, measured in FLOPs, is one proxy metric for the growth rate of machine learning training runs. The footprint of a model is defined, with caveats, as the total number of FLOPs used to train the model since initialisation.

Caveats:

A randomly initialised model has a footprint of 0 FLOPs.

If the model is trained from a randomly initialised model using SGD or a variant, then its footprint is the total number of FLOPs used in the training process.

If a pre-trained base model is used for the initialisation of another training process (such as unsupervised learning, supervised learning, fine-tuning, or reinforcement learning), then the footprint of the resulting model will include the footprint of the pre-trained model.

If multiple models are composed to form a single cohesive model, then the footprint of the resulting model is the sum of the footprints of each component model.

If there is a major algorithmic innovation which divides by a factor of r the FLOPs required to train a model to a particular score on downstream tasks, then the footprint of models trained with that innovation is multiplied by the same factor r.

This list of caveats to the definition of Effective Training Footprint is non-exhaustive. Future consultations may yield additional caveats, or replace Effective Training Footprint with an entirely different proxy metric.

Fixing the y-axis

According to the 0.2 OOMs/year target, there cannot exist an ML model during the year (2022+x) with a footprint exceeding f(x), where f(x+1)=100.2×f(x). That means that log10f(x)=(0.2x+a) FLOPs for some fixed constant a.

If we consult EpochAI's plot of compute training runs during the large-scale era of ML, we see that footprints have been growing with approximately 0.5 OOMs/year.

We can use this trend to fix the value of A. In 2022, the footprint was approximately 1.0e+24. Therefore a=24.

In other words, log10f(x)=0.2x+24.

I have used 2022 as an anchor. If I had used 2016 instead, then the 0.2 OOMs/yr target would've been stricter. If I...