JessRiedel

Senior Research Scientist at NTT Research, Physics & Informatics Lab. jessriedel.com , jessriedel[at]gmail[dot]com

Comments

TAI Safety Bibliographic Database

The automated tools on Zotero are good enough now that getting the complete bibtex information doesn't really make it much easier.  I can convert a DOI or arXiv number into a complete listing with one click, and I can do the same with a paper title in 2-3 clicks.  The laborious part is (1) interacting with each author and (2) classifying/categorizing the paper.

TAI Safety Bibliographic Database

Does the org have an official stance?  I've seen people write it both ways.  Happy to defer to you on this, so I've edited.

TAI Safety Bibliographic Database

If we decide to expand the database in 2021 to attempt comprehensive coverage of blog posts, then a machine-readable citation system would be extremely helpful.  However, to do that we would need to decide on some method for sorting/filtering the posts, which is going to depend on what the community finds most interesting.  E.g., do we want to compare blog posts to journal articles, or should the analyses remain mostly separate?  Are we going to crowd-source the filtering by category and organization, or use some sort of automated guessing based on authorship tags on the post? How expansive should the database be regarding topic?

Currently of the 358 web items in our database, almost half (161) are blog posts from AI Alignment Forum (106), LessWrong (38), or Effective Altruism Forum (17).  (I emphasize that, as mentioned in the post, our inclusion procedure for web content was pretty random.) Since these don't collect citations on GoogleScholar, some sort of data on them (# comments and upvotes) would be very useful to surface the most notable posts.

Search versus design

Somewhat contra Alex's example of a tree, I am struck by the comprehensibility of biological organisms. If, before I knew any biology, you had told me only that (1) animals are mechanistic, (2) are in fact composed of trillions of microscopic machines, and (3) were the result of a search process like evolution, then the first time I looked at the inside of an animal I think I would have expected absolutely *nothing* that could be macroscopically understood. I would have expected a crazy mesh of magic material that operated at a level way outside my ability to understand without (impossibly) constructing a mental entire from the bottom up. And indeed, if animals had been designed through a one-shot unstructured search I think this is what they would be like.

In reality, of course, animals have macroscopic parts that can be partially understood. There's a tube food passes through, with multiple food-processing organs attached. There are bones for structure, muscles to pull, and tendons to transmit that force. And the main computation for directing the animal on large scales takes place in a central location (the brain).

We can tell and understand a post-hoc story about why the animal works, as a machine, and it's sorta right. That animals have a strong amount of design, making this is possible, seems to be related to the iterated search-and-evaluate process that evolution used; it was not one-shot unstructured search.

At the least, this suggest that if search vs. design identifies a good dichotomy or a good axis, it is an axis/dichotomy that is fundamental, and arises way before human-level intelligence.

Review of "Lifecycle Investing"

Agreed. The optimal amount of leverage is of course going to be very dependent on one's model and assumptions, but the fact that a young investor with 100% equities does better *on the margin* by adding a bit of leverage is very robust.

Review of "Lifecycle Investing"

I endorse ESRogs' replies. I'll just add some minor points.

1. Nothing in this book or the lifecycle strategy rests on anything specific to the US stock market. As I said in my review

The fact that, when young, you are buying stocks on margin makes it tempting to interpret this strategy is only good when one is not very risk averse or when the stock market has a good century. But for any time-homogeneous view you have on what stocks will do in the future, there is a version of this strategy that is better than a conventional strategy. (A large fraction of casual critics seem to miss this point.)

If you are bearish on stocks as a whole, this is incorporated by you choosing a lower equity premium and hence lower overall stock allocation. This choice is independent of the central theoretical idea of the book.

2. Yours is a criticism of all modeling and is not specific to the lifecycle strategy.

3. As ESRogs mentioned, neither this book nor my review has the timing you suggest, so the psychoanalysis of proponents of this strategy appears inconsistent.

4. I acknowledged this sort of argument in my review, and indeed argued that the best approaches hinges on such correlations. But consider: even in the extreme case where I believes my future income is highly correlated with the stock market and is just as volatile, the lifecycle strategy recommends that my equity exposure should start low when I'm young and then increase with age, in opposition to conventional strategies! So even if you take a different set of starting assumptions from the authors, you still get a deep insight from their basic framework.

SARS-CoV-2 pool-testing algorithm puzzle
The problem is that there are other RNA viruses besides SARS-CoV-2, such as influenza, and depending when in the disease course the samples were taken, the amount of irrelevant RNA might exceed the amount of SARS-CoV-2 RNA by orders of magnitude

There is going to be tons of RNA in saliva from sources besides SARS-CoV-2 always. Bits of RNA are floating around everywhere. Yes, there is some minimum threshold of SARS-CoV-2 density at which the test will fail to detect it, but this should just scale up by a factor of N when pooling over N people. I don't see why other RNA those people have will be a problem any more than the other sources of RNA in a single person are a problem for a non-pooled test.

Why would panic during this coronavirus pandemic be a bad thing?
"The government" in the US certainly doesn't have the authority to do most of these things.

Both the federal and state governments have vast powers during public health emergencies. For instance, the Supreme Court has made clear that the government can hold you down and vaccinate you against your will. Likewise, the Army (not just National Guard) can be deployed to enforce laws, including curfew and other quarantine laws.

Yes, it's unclear whether government officials would be willing to use these options, and how much the public would resist them, but the formal authority is definitely there.

Alignment Newsletter #13: 07/02/18

Hi Rohin, are older version of the newsletter available?

Also:

This sounds mostly like a claim that it is more computationally expensive to deal with hidden information and long term planning.

One consideration: When you are exploring a tree of possibilities, every bit of missing information means you need to double the size of the tree. So it could be that hidden information leads to an exponential explosion in search cost in the absence of hidden-information-specific search strategies. Although strictly speaking this is just a case of something being "more computationally expensive", exponential penalties generically push things from being feasible to infeasible.

Load More