Futapy

HW3

per http://informationcomputingandthefuture.blogspot.com/p/hw-3.html

I. - For one or more of your proposed project topics, make a list of several future potential impacts (such as privacy, security, quality, or any others that seem useful to list), that relate to individuals, such as people like yourself. Explain why these impacts might occur. 250 words.

A1.I, "AI langue model perfection and proliferation" impacts related to individuals:

• Inability of the individual to discern if a written piece is from an AI or a human: the age of computers failing the Turing test is nearly over. GPT-3 is claimed to not yet be able to discern a bad-faith question such as "who was the US President in 1600" and expose itself as an AI by attempting to answer in earnest (source), but this can be overcome by training a model to spot bad-faith questions and arguments. One neural model will never fool a majority of people but combined models that evaluate conversation, opponent questions, and relevance of answers would be much more difficult for a human to spot and are imaginable today.
• The ethics of an individual using language generation as a tool or in place of their own original work: It is already easy to access tools that would finish this paragraph in a grammatically correct way with believable content (source). While the results of such tools today are fairly easy to spot in that they veer off topic or don't quite make a coherent connection to the previous text, these problems will soon be solved. When that happens the temptation to enhance one's work with a unique exposition by an AI will need to be explicitly addressed: under what, if any, conditions will it be acceptable? How can such social strictures be enforced, when the enforcer is presented with quality and original work?
• An individual may be more easily exposed to well formed arguments crafted to take advantage of human cognitive weaknesses or evolutionary features: Humans are a persuasive lot. Countless world religions and philosophies have found fertile ground in the minds of their followers, and have been passed on with arguments modified and adapted to their times. It is conceivable that an AI could construct an argument so tuned to the current thinking that many people would be convinced. This may be so near that the current OpenAI model, GPT-3, borders on crossing this line already (source).

A1.II. - "AI langue model perfection and proliferation" impacts related to organizations, such as businesses, governments, etc. Explain why these impacts might occur. 250 words.

• A government or organization may have the ability to write persuasively on any topic: We are easily convinced by a crowd of voices all telling the same story a slightly different way; with language model perfection it will be trivial for an organization to produce a huge array of original thought pieces all telling the same story, leveraging our instinct to go along with the crowd. (Perhaps my life's work will be equipping humans to fight the influences of AI by recognizing our weaknesses to its persuasion)
• A government or organization may be fooled by the dilution of constituent opinion with a cheap army of AI opinion writers: All companies that stay in business listen to their customers. Our representative government relies on feedback from the governed. It will be soon be possible to easily draft countless original narratives on any product or political topic. Governments and companies will need a new standard of evidence to discern a legitimate human voice from an army of convincing AI personas. (This may have already been weaponized by nation-state actors)
• A government or organization could promote unity by automatically tailoring a convincing message to sub-populations of clients/constituents: Breaking from my usual pessimism, I can see this technology put to good use. Scores of different sub-populations are categorized and studied for the way they respond to marketing today (source); these sub-populations could receive tailor-made messaging in support of the common message of a beneficial company or the beneficial aspects of government upon society.

A1.III. - "AI langue model perfection and proliferation" impacts related to society as a whole. Explain why these impacts might occur. 250 words.

• AI based creative writing could displace human writers, upending an industry. This would effectively be the motor-driven novel writing machine from Orwell's novel "1984". Twice recently when reading a web-based article that starts strong but strays from the original idea halfway through, amidst a sea of advertising, I have had the notion that I was probably reading computer generated text. Within the foreseeable future these language models will be advanced enough to fool most of us, and at that point, why would a web publisher pay a human writer? Art and images using the same techniques are the next frontier (source), and if the march of Moore's law continues, audio and video being convincingly created by machine are possible within our lifetimes.
• AI based cognition could form new philosophies and religions using arguments and techniques known to work best: Given an almost limitless set of human test subjects willingly signing up for social media, the hook of an idea could be automatically tuned and tested to perfection, creating entirely new self-spreading ideas within the wider community for the purpose of whoever controlled the AI. This would almost certainly be weaponized by nation-state actors.
• The foundations of belief: what can believed when anything can be more convincingly written by an AI than by a human? Already the release of GPT-3 has been delayed over concerns that it is too good at finishing radical texts in a convincing manner. I have overheard, on ham radio, a subset of the population dismayed that they didn't know what to believe in the wake of the 2020 U.S. presidential election where misinformation delivered by computer was rampant. When opposing opinions are convincingly presented as true, we may choose one or disbelieve both. Generated texts from more convincing language models could further erode any shared notion of truth in our society.

A2:

I read the book "Predictions" preface, intro(11 pages) and pp 1-10. Introduction (Jonathan Weiner) talks about the 21st century as a set of five experiments in progress: the population explosion (then only at 6 billion people), global warming, the progress of technology, the progress of evolution, and "consilience" which is the dream or hope that everything we are learning will fit together in single vision understanding all of human histories and endeavors. The intro is brief a review of the following essays and how they fit into this framework.

agree: Dennett: Memes, both tonic and toxic, will wreak havoc on the unprepared world - this rings very true today and for our near future. Clarke: would rather be known as an extrapolator than a prophet, and that often it is more interesting to see where and why one goes wrong than what happened to be right. Chinua Achebe: The essay by Jennifer Wallace is a brief but informative history of Achebe and his role as an African writer and African politician. Some predictions of Achebe are the eradication of malaria and other diseases; while this has not been achieved in 20 years, malaria deaths are half what they were in 1999. He is hopeful that the end of the Cold War and relative geopolitical stability will leave Nigeria less interfered with than in the recent past, and given Nigeria's people and resources things will be better. I agree Nigeria is still ready for the changes he wrote about.

disagree: Achebe: optimism is better than pessimism solely because optimists are happy some of the time and pessimists (we call ourselves realists) are always miserable. I could agree that optimism can be more productive, but I feel an unprepared optimism is often ultimately more harmful than an attitude that allows some amount of realistic foresight and preparation for less than acceptable outcomes. His optimism leads him to believe that perhaps issues of race will be solved in the 21st century. While there has been notable progress in the United States, racism still has an outsized impact on our collective behavior.

learned from: Achebe's optimism is refreshing; this history of his involvement in Nigeria with perhaps overly optimistic and ambitious affairs is not a history I was familiar with. Africa is studied so little in the culture that I was raised in but is rich in history and culture and I appreciate reading Achebe's thoughts.

delta from other reviewers: Many of the reviewers said the value is in the breadth of authors, not in the structure or depth of their arguments. So far I agree. The book is from 1999 and I find myself wondering about and looking up the fate of some of the authors; their writing is intensely personal. In particular I looked up the history and fate of Chinua Abache who died in 2013, aged 82. So in that regard I disagree with the reviewers that treat the book as a buffet of ideas with some to be studied and others ignored; each author likely brings something useful to this book.

These are not homework answers, but related internet links to the GPT research. I think I need to do a lot of foundational research to get some history of what this represents.

• Links of note: GPT paper
• Talk To Transformer demo
• GPT and iGPT discussion of biases in unsupervised training
• OMG: Image GPT is "creative". Look.
• A Reddit thread with insightful criticism of what OpenAI is doing, and another thread discussing scale and cost of igpt

Related internet links to the Enterprise Database is Dead notion. On "What is the future of enterprise databases", here is a fellow asking about data access to some public solar data which is available as an HSDS hdf5 file using some example code. Looks like the HSDS layer provides access management and atomicity of updates, but I don't know if it full ACID. I guess this makes more sense than exposing your PostgreSQL username/password/port and is more scalable than a huge batch of semi-consistent files on an FTP server. Looks like any transformation and aggregation has to be done client-side which, honestly, maybe isn't as big of a problem as I had thought. This is something I need to know about. Q: Is there an easy API access layer for something PostgreSQLish? Something like HSDS with Pandas is definitely more difficult than an FTP full of CSV files and arguably less easy than a SQL interface but "it will scale". How important is that really in this use case?