Link to Google Doc
This entire year has been pretty bleak in terms of everyday life, plans has been cancelled, we haven’t been able to socialize as we used to, and much more. Overall, there has been few things to be excited for during this year. But these last few weeks have come with some positive news, which makes me cautiously positive for the future.
First of all, I want to appreciate the effort expanded into making a vaccine against COVID-19. These last few weeks there have been multiple reports of vaccines that have passed phase 2 trials with high rates of success. Moderna, Pfizer, and AstraZeneca are all firms that have come up with vaccines. The vaccines are possible due to novel technologies that allows for usage of messenger RNA (mRNA) to guide a protein to the pathogen. This in itself is promising for future medicine development. But what is more fascinating about the Moderna vaccine is that it only took two days after the release of the genetic sequence of the virus for Moderna to develop the synthetic sequence of RNA that is used in the vaccine.[i]
This raises questions about the regulative path that vaccines and medications need to pass before they become publicly available. A long time has passed between January 13, when Moderna designed their RNA-sequence, and now. A time in which many people have died from the virus. Although there are many risks associated with vaccines, I think that this should at least spark a debate about human challenge trials. Initiatives such as 1daysooner have gotten over 35,000 volunteers that are willing to subject themselves to COVID-19 to more rapidly test and approve of a vaccine. [ii] If human challenge trials could accelerate the approval and creation of safe, useful vaccines, then there should at least be a debate about the use of them. [iii]
All participants in such trials are volunteers, which of course does not make their life expendable, but one must believe that these volunteers are well versed in the associated risks and that they are willing to take on the risks for the benefit of saving lives with the vaccine. With the success of the current vaccines, it is not sure that human challenge trials would have sped up the process. But what if the first generations of vaccines would not have been as effective, or if some vaccines would bring unexpected side effects? Then human challenge trials could have accelerated development of vaccines, and potentially saved many hundred thousands of lives.
Overall, the rapid development and success of these vaccines is good news to the world, both in terms of their immediate effect and the long-term effect of us being able to create vaccines using mRNA.
The second achievement, and cause for optimism, is DeepMind’s and their algorithm AlphaFold 2’s success in “solving” the protein folding problem. This is a huge achievement, not only in biology and life sciences where an increased understanding of proteins can be immensely important, but also in terms of machine learning and computing. This will surely lead to an increased understanding of proteins, which in turn spills into medicine, biology, and perhaps a deeper understanding of life itself.[iv]
In terms of the computing and machine learning aspect, I believe that this breakthrough is equally huge. Deeply technical problems solved by machine learning/AI is probably the last step before we reach the point where AGI is possible. Thus, the evolution of more and more complex machine learning programs suggests two things. First, more and more complex problems appear solvable by using machine learning and second, the time between each breakthrough is shortening. All of this suggests a rapid approach of programs that are complex enough to be regarded as AGI.
My model of when to expect AGI depends largely on past breakthroughs in machine learning, and not any other private information. The AlphaFold and muZero (a program that plays chess, go, shogi, and Atari without knowing the rules beforehand, a step forward in terms of learning techniques) successes force me to update my previous estimate of when AGI is going to exist. From thinking that AGI would be possible in 8-40 years, 50% CI, I now think that AGI is possible in 5-30 years, 50% CI. (The previous estimate was based largely on the breakthroughs made in chess, go etc.) It is a fairly wide estimate, but I am not very confident in neither my understanding or knowledge of computer science. I also believe that the tails of the distribution for creation of AGI are rather fat, there may be many obstacles that we currently do not know.
[i] https://globalnews.ca/news/7492076/moderna-coronavirus-vaccine-technology-how-it-works/
[ii] https://www.1daysooner.org/
[iii] For experimental papers on the possible uses of human challenge trials see: Sauerwein, R., Roestenberg, M. & Moorthy, V. Experimental human challenge infections can accelerate clinical malaria vaccine development. Nat Rev Immunol 11, 57–64 (2011). https://doi.org/10.1038/nri2902
Nir Eyal, Marc Lipsitch, Peter G Smith, Human Challenge Studies to Accelerate Coronavirus Vaccine Licensure, The Journal of Infectious Diseases, Volume 221, Issue 11, 1 June 2020, Pages 1752–1756, https://doi.org/10.1093/infdis/jiaa152
[iv] DeepMind has a great blogpost about the protein folding problem, AlphaFold and its potential future importance. https://deepmind.com/blog/article/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology