Consider figure 15 from this paper.
Some people (arguably the authors of this paper) predict that as we scale models past GPT-3's size (the 10^11 parameter learning curve, models with parameter count X trained on X elapsed tokens will score close to the L(D) line at X elapsed tokens.
We are interested in whether instead the trendline will "plateau" or at least be substantially slower than the line L(D) by the end of the next 3 orders of magnitude of parameter count. For the sake of specificity, let's say substantially slower = less than half as steep as L(D) on this graph.
If and when this graph is extended to 10^14 parameter models trained on 10^14 elapsed tokens of similar-quality data, will the 10^14 parameter learning curve have slowed down substantially?
This question resolves positively if the relevant experiment is done and reported (extending this graph, or providing equivalent data) and the slope of the learning curve for the 10^14 parameter model around 10^14 data points (Say, from 10^12 to 10^14) is less than half as steep as the slope of L(D). It resolves negatively if instead the slope is at least half as steep as L(D).
This question also resolves positively (or negatively) if it becomes uncontroversial what would have happened if the experiment had been done. For example, maybe other experiments will provide much more evidence about neural net scaling trends in general, such that it will be easy to calculate what would happen with this one in particular.
This question resolves ambiguously if by 2050 no one has performed the experiment AND there is no consensus about what would have happened if someone had.
The Metaculus moderators are the judge of final resort for determining whether and how to resolve this question.
For more context, see the comment threads below (Search for "It's a big deal.")
It is important that the data used to extend the graph be of similar quality. Obviously if we just threw in 10^14 tokens of basic arithmetic problems, the model would get good at basic arithmetic but not at anything else, and it's unclear whether the result would be on-trend or not. Ideally we'd have 10e14 tokens of diverse internet text, scanned books, chat logs, emails, etc. If this experiment gets done with different-quality data, the question becomes whether it gives us enough evidence to uncontroversial predict what would have happened if we had done it with similar-quality data.