AGI vs. Whole Brain Emulation — Give Me Your Probability Distributions and Statements Tuesday, Nov 10 2009
There is still a great amount of disagreement in the community over whether uploads or AGI are more likely to occur first. I would like to solicit probability distributions and statements from interested parties on the relative likelihood of either route reaching human-equivalent intelligence first. Please send your probability distribution values, including mean and standard deviations, to michaelanissimov at g mail dot com, coupled with a well-written, well-edited statement of 400 or more words, and a short official bio and link to your CV (if applicable). Please include at least 3-5 citations in your argument, and please make it as academic as possible. The final results will be published on a web page at AcceleratingFuture.com, and may make it into a journal if the results are good enough.
The operating definition of intelligence to be used during this exercise is that given by Shane Legg in his PhD thesis Machine Superintelligence:
Intelligence measures an agent’s ability to achieve goals in a wide range of environments.
Good luck! By the power of Aumann, may our probability estimates converge closer to the truth as a result of this exercise.
The academic philosophers who are reading — you know who you are — if you only make one comment on Accelerating Future all year, make it a contribution to this exercise. Thank you for your time.
Mike, I don’t think this is the best way to structure such a survey!
Personally, being an AGI researcher but NOT a whole-brain emulation expert, I feel like I could write something meaningful about the timeline to AGI in a fairly rapid way.
However, for me to write something solid about the timeline to whole-brain emulation, would require me to do a bunch of research. I know what my qualitative opinion is but to substantiate it carefully (and quantify it) would require more time than I’m willing to spend right now.
It might be better to get AGI experts to comment on the timeline to AGI, and WBE experts to comment on the timeline to WBE. At any rate I think this alternate approach would get you a greater number of well-thought-out responses. Because I suspect that many AGI or WBE researchers will feel the same way I do — not wanting to create rigorous-looking, referenced academic prose making projections about research areas in which we don’t have expertise.
I’m happy to mouth off in a blog entry about my rough estimates of the time to WBE, but before I’d put my opinions about that in a journal, I’d need to do a bunch of research, ask some WBE experts, etc. … and I’ve got too much else on my plate to do that…
ben Goertzel
I’m not sure that anyone will be able to give a convincing answer to this question – probabilistically or otherwise – because the deciding factors have yet to be determined/discovered.
We still don’t know what it takes to build an AGI (a practical one, at least). There are some ideas and projects ongoing, but it’s far from clear that any of these will succeed and also difficult to quantify the odds of success. For example, if you had asked Doug Lenat in the 1980s what the chances of success for Cyc were, or Marvin Minsky in the 1960s they would have given peak probabilities 15-25 years into the future, and been able to provide complex and academically grounded justifications for their answers.
Likewise it’s also difficult for brain emulation. The chances of achieving a structural upload (connectome) for a complete brain in the near future (10 years) I would say are very high, but this is only the first step in a much bigger project. Although we do understand a great deal about neurons I’m not convinced that we yet know exactly what they’re doing (what their “essential function” is) and what, if any, role glial cells play. It might take more detailed cell modeling to decide the matter.
Ben, even though people may be more or less informed about either area, they still have implicit probability distributions which are better than nothing and suitable as a starting point. Estimates can be improved as time goes on. Thanks for responding though.
Bob, probabilities are all we have. Right now we have thousands of implicit probabilities among all the people who are considering the issues. We need explicit probabilities. Bayesians have no problem giving probabilities even if they are extremely uncertain — even if you know nothing, you can give the probability of n arbitrary mutually exclusive events n a probability of 1/n.
Even though you say, “no one will come up with an answer…” you are still providing us with Bayesian evidence in your comments, which is better than nothing.
There are so few published references and opinions on AGI vs. uploads that it is possible to read them all. In addition to the thousands of pages of cognitive science knowledge that many of us have acquired, I think that is sufficient to at least give probability distributions and short statements. Implicit, low-confidence probability distributions can still contain all the necessary uncertainty merely by expanding confidence bounds.
yo