Notable Reads
-
Three Observations really a mix of thoughts about AGI from Sam Altman of OpenAI. We get yet another definition of AGI: “A system that can tackle increasing complex problems at a human level” but we have others: “A system that is better at most humans at almost all tasks”. In my opinion, the broad distinction that we should keep in mind is between AGI (i.e., human expert level) and ASI (Artificial Super Intelligence; post-human capabilities). The article makes it clear that the scaling laws are still intact, rapid progress is still happening. The author points out that much of the exponential returns that we should expect will not come from alien intelligence, but from deploying millions of such AGIs to try different things non-stop at the service of humans. To add to that, we need a lot of maintenance (i.e., sleeping, eating, exercising, showering, resting, socialising, etc), while AIs need none. The article talks about the good scenario where AIs won’t be fully autonomous and would need human initiation (for framing & solving a problem; or formalizing willingness for action) and control at critical decision points, under such assumption, people with agency, willingness, and determination will be super-productive and achieve a lot. Humans will move up the “value creation” chain and create many many more small 0-employee N-agents companies. However, such empowerment may fall on its face if we have embodied AIs that are accepted in the real world the same as regular human actors.
-
DO TOO MUCH Although short, this piece brings up an interesting question: Is there really such thing as a “balance” if you are ambitious? We keep hearing things like “if you are not pushing forward, you are falling behind” (sports) or “If you are not giving your best, someone else is” (competition) etc. If we look around, most people we respect seem to be unreasonable, obsessive, doing too much, at least from the viewpoint of others. The good news is that genuinely “caring” about something is more valuable than “genius”, which is innate. I came to believe that we all get to define and adjust our definitions of “balance” (Example: Benjamin Franklin), and the median (consensus) descriptor of it may not work for you.
-
On Overconfidence most of us are very very very overconfident. Even me and you! we think we calibrate well but still way off. We are driven by intrinsic biases and local (i.e., limited to our environments, experiences, education, etc) models. That is why we need to avoid predictions and embrace tasteful experimentation (i.e., let me try out first) and get surprised by results. This is hard because we want to be right and so we misinterpret opposing evidence to fit our narrative, we rarely change our world model. However, there is beauty in surprise that we can learn to appreciate.
-
Beware The Man Of One Study It seems that science isn’t that straight forward! There are many types of randomness that can skew the results of a scientific experiment (i.e., biases, noise, etc). Instead, we need to examine the outcome distribution of related studies. So, Is meta-analysis or systemic reviews enough? Well, it (again) depends on what studies we select (“beware the man of any number of studies less than a relatively complete and not-cherry-picked survey of the research”) and the overall population of researchers working on that fields (remember, A meta-analysis of hundreds of studies is what tells you that psychic powers exist). This feeds into why pure empiricism (i.e., relying on observations) is not enough, it needs to be complemented with a rational model of the world to help us investigate more certain experimental results or align ourselves. In summary, use meta-analysis and literature reviews to establish a general impression and your own (hopefully rational) model of the world to turn it into an opinion (related: Einstein’s Arrogance, Why I Am Not Rene Descartes, How Common Are Science Failures?)
-
Scholarship: How to Do It Efficiently Summary: start with review papers/meta-analysis, note notable researchers in the field, read lots of abstracts before text and select, read starred papers, update your research questions (if needed).
