Futapy

HW2

per http://informationcomputingandthefuture.blogspot.com/p/hw2.html

A1: Long ago it stuck in my head that an easy way to calculate the total amount for annual percentage increase of a number by a percent is to multiply the number by (1 + percent) raised to the power of the number of years of interest. This gives the same answer as calculating the percent compounded per year and adding it up a row at a time, but is much easier. So to find the doubling time of any non-zero number c at 6% growth annually, just find x where 2c = c*1.06x, or more simply, when 1.06x = 2. Plugging in values for x in a binary-search manner, the value of x=11.896 yields close to 2.0. Double checking with the spreadsheet rows method, after 12 years of 6% growth the final value will be 2.012 times the original value. 1.0612 is also 2.012, so this method holds.

Not being able to solve 1.06x = 2 for x directly bugged me so I looked it up. https://tutorial.math.lamar.edu/classes/alg/solveexpeqns.aspx tells me to use a logarithm to solve for x when x is the exponent: 1.06x = 2 can be written as x ln 1.06 = ln 2, or x = ln 2 / ln 1.06, or x = 11.8956610459. I love having the sum of world knowledge at my fingertips, but will it last?

A2: I have considered many topics for a term project, but the one that interests me the most right now is "what is GPT3 and how can I use it", or more generically, "tracking the progression of human-like interactive automated language models and making some estimates of the future". OMG people have made chatbots with it. From what I understand of it so far, it is a natural language processing AI trained on an enormous amount of text data from the Internet with the ability to output natural sounding language and even computer code with what seem on first glance like original insights. My thought is that some research on what it is and how it came to be, what is the scale of the product and how much energy does it consume at that scale, what is the likelihood that me as a UALR grad student can get some time on this model, how insightful and original are the results, and can I get GPT3 to write my final paper for me (note: this is likely one of the few contexts in which this behavior would not be considered cheating - let me know if you disagree). If I'm successful at that, my final presentation could actually be the research and methods of setting the GPT3 API up, the generated text (assuming GPT3 is self-aware enough to write about itself), and a critique of the insights in the generated text, some comparison with earlier language model outputs and a prediction for when we can call IA something like "creative", "insightful" or "spiritual". Stretch goal: see if I can get it to describe how to design an artificial intelligence model better than itself.

Edit: that may be a bit ambitious. That's really ambitious, and apparently API access is hard to come by. Reviewing the project requirement again, I'll target a 3000 word paper (not AI assisted). How about: Research OpenAI and GPT3; get a sense of the scale of what they are doing and how it could profoundly impact out lives in the future, and present a 20 minute slide-deck summary of what I found. Send me an email if that's too boring or it will be covered in a class this term already.

The bucket of potential ideas is large; if any of these sound better or if I determine early on that I CANNOT use GPT3 for a project, here is a list of things I've jotted down as topics and notes that could be turned into a project:

P.S. do any of these sound like a Masters thesis or project? Particularly the perception of information as trustworthy by holders of disparate modern mythological beliefs?

A3.A: My question was: "When (if ever) will over half of U.S. restaurant diners choose to pay with bitcoin vs. USD?"

Round one estimates data were 2025, 2028, 2030, 2030, 2035, 2037, 2080. The middle 50% of values median is 2030 and the interpolated middle 50% range is [2029.5,2047.75].

After round one, participant discussion centered on topics of convenience of paying, fiat monetary stability, United States influence and the value of the dollar, and the independence of cryptocurrency.

Round two estimates data were 2025, 2025, 2028, 2030, 2035, 2045, 2080. The middle 50% of values median is still 2030 and the interpolated middle 50% range is [2027.25, 2053.75], possibly indicating that the group as a whole became less certain after the discussion.

A3.B Like all college kids these days, I found some material on YouTube. It is difficult to quote and tedious to sit through information that can be skipped in a written format, but it is also difficult to find something useful to do while on the treadmill, so I killed two birds with one stone with this excellent 50-minute video lecture by Jeffrey Franc entitled "Delphi Tenique - The do nots and why nots" referencing the book "The Delphi Technique in Nursing and Health Research" by Keeney et. al, 2011. Below are my notes and comparisons with our class exercise.

Observations from class exercise: First of all, we're clearly not a panel of experts. Also, to do Delphi right according to this fellow it is a slow process; there should be a couple weeks between rounds, with formal questionnaires, anonymous answers, and rounds until consensus is derived. In class, we only performed the quantitative analysis; our answerable question really should have been based on the results of open-ended questions in round 1, so what we did was modified Delphi. Also the process should iterate until consensus is reached instead of being capped at two rounds. 8 to 12 experts are recommended. At 7 in class, we were pretty close to the optimal number. The experts should have knowledge about the issue under question (this typically takes hours of the expert's time). Confidentiality of the answers should be assured; we certainly didn't do that in class.

This is interesting; the definition of consensus must be decided before the data is obtained, otherwise the data may taint the definition and therefore throw the study into doubt. While our metrics were predetermined by the professor, this point could have been emphasized. I'd like to hear more about the improper application of statistical methods.

One thing we did in class that this fellow didn't, is panel discussion and sharing of individual rationale for our answers. Also, our panel selection for our questions are a better approximation of a random sample than 12 of our friends that are more expert on a particular question. However, since experts aren't truly random in any Delphi study, we should avoid using confidence intervals or p-values that need a random sample to be valid. I think we did OK here; we just reported the data in the middle 50%.

A3.C Shortcomings of the class exercise (compared to a full Delphi study):

Any of these could be fixed at the expense of time and possibly money to pay for experts, but then it wouldn't be a class exercise.

Saving this here because I just want to make a note of it. Delphi Problems: Cannot produce right or wrong answers. Consensus doesn't mean we have a correct answer. Interval validity is largely unknown. Delphi is a tool of last resort; we should be aiming higher on the evidence pyramid if possible. On the strength of evidence scale used in Evidence Based Medicine, what we did would rank below the bottom rung of "expert opinion" since our panel was not comprised of experts.

Source website

A4: I ordered 4 books because I am terrible at decisions. Technically I am in compliance with the HW because I did order "a book about the future"... and 3 more with it. I'll winnow it down to one based on reviews:

First choice: Predictions. I can get through the whole thing in the weeks left in this course mostly 2 authors at a time, with a couple of 3 author sprints. If it ever gets here, this sounds like the the most appropriate book for this course in that each week I could read a different couple of essays and provide my own input, and by the end have an introduction to the futuristic thinking of 30 noted minds from around the end of the 20th century. Wow!

Second choice: Megatrends. The hit-and-miss nature of the predictions will be interesting to study, plus the commentary on the social situation that many reviewers say hit the mark.

The other two will go on the bookshelf for reference material. Future Shock appears to be widely regarded as excellent but I have only limited time and attention on this blue ball. Religion and Scientific Future has value in that it was written after a period of heightened religious activity in this country and during the beginnings of the computer age, and appears to offer insight into the human desire for myth that continues to shape society today; this could be valuable if my studies go in the direction of social acceptance of information as truth.