Target Question: What is the likelihood that language models in 1-2 years will be able to do the tasks that human participants do in Ought’s experiments?
Ought, an AI research lab, is running experiments with human participants to test and explore factored cognition. In one set of experiments participants create and solve simple 'workspaces' through question and answer dialogs.
Eventually Ought expects AI agents to be able to replace the humans in the workspaces. At that point Ought will be able to significantly expand the experiments and test key questions related to AI alignment. So, if it's likely to happen sooner, rather than later, then Ought on the margin should take certain actions (ex. invest heavily in ML tech) now.
What is the likelihood that, by the end of 2020 with 10M pieces of labeled data, Ought can automate 90% of tasks that trained human participants currently do in Ought experiments?
We will run a test using Mosaic w/ a model trained from the current experiments. An output workspace will be marked as "correct" if the answer from the workspace is judged as equivalent to the human response from the same workspace. If 90% or more of the workspaces in the test are solved correctly, the question will resolve as true.
If there is less labeled data than is required for the question, the AI Resolution council will be surveyed to extrapolate from tests their best estimates of the % of tasks in a set of 100 representative workspaces that an AI could solve.
The test model may be pre-trained using other datasources and then applied to the Mosaic benchmark (ex. use a large language model trained on a generic text corpus, and then fine-tuned w/ examples from Ought workspaces).
"Can automate" is equivalent to a workspace being reviewed by a human and marked as "solved", i.e. produced a reasonable output similar to a human.
One piece of data == one human action performed in an experiment. For example if a human were to summarize an article to less than 100 words, as is an action in some of the workspaces, that action would be recorded and presented in the dataset as one piece of data.