Here’s a quick post demystifying the process by which we’re scientifically bootstrapping our way out of our minds’ and brains’ black boxes into better and better self-knowledge.

1. We identify types of human cognition that are of interest to us.

For example, analogical reasoning. Once you’ve learned about it, it’s shocking how analogy underlies almost all of our cognition.

See for example:

https://plato.stanford.edu/entries/reasoning-analogy/

2. We develop benchmark tasks that require humans to employ a particular type or types of cognition.

See for example Raven’s Progressive Matrices and related datasets, such as the I-RAVEN dataset.

https://arxiv.org/abs/2002.06838

3. We hypothesize models (possibly inspired by human brains or cognition) that we believe would mimic human performance on some benchmark metric.

See for example:

https://arxiv.org/abs/2302.04238

https://arxiv.org/abs/2309.06629

4. We instantiate those models in computer software and test them on a benchmark task or tasks.

See for example:

https://arxiv.org/abs/2007.04212

5. We investigate how something like these models might be instantiated in the brain.

See for example the discussion in the following:

https://arxiv.org/abs/2309.06629

(And stay tuned for ongoing work!)

Note that although humanlike mechanisms of reasoning imply an ability to solve benchmark tasks, solving benchmark tasks does not imply humanlike mechanisms of reasoning. So how can this process result in any progress toward understanding intelligence and human consciousness?

The key will be ongoing interdisciplinary communication, criticism, and feedback between all stages of the process outlined above. In the meantime, the iterative process we have created will certainly close the gap between human capabilities and machine capabilities.

Now back to research!