(The below text version of the notes is for search purposes and convenience. See the PDF version for proper formatting such as bold, italics, etc., and graphics where applicable. Copyright: 2022 Retraice, Inc.)
Re101: News of ChatGPT, Part 4
retraice.com
How ChatGPT crosses with the hypotheses.
More attempts to describe ChatGPT; maximizing its reward; intended uses; what it might tend to do, overall; new means, old ends; what it is `thinking'; dialogue as key; improving human search-and-find space and medical knowledge and information; OpenAI's whack-a-mole burden; the Chinese, Americans, left-leaners and right-leaners now to be attacked with mis- and dis-information; strategic search-and-find information now easier to get; our collective environment now contains new agents, the ChatGPTs; the technological `future' containing ChatGPT is now the `present' and `past'; smart people will benefit more; bad guys; wealth; wildcards; hackers; control of humans by words; change of the environment by ChatGPTs; motivation to control.
Air date: Friday, 30th Dec. 2022, 4:00 PM Eastern/US.
What is it again?
It's GPT-3 (~InstructGPT) adapted for dialogue, plus a reward model, connected by gradient descent. It tries to maximize its reward. It will change its environment (the only way it can, by producing output) to do so. `It' is tens of thousands of containers^1 under the control of OpenAI. It is intend it to do Q&A, captioning, translation, summarization, `generation', and the like.
ChatGPT and the world
ChatGPT might tend to:
o make us more predictable, because we're the new environment (population thought control?);^2 o increase language group isolation, unless language translation works well and is used; o entrench language interpretation errors; o tell us what we want to hear, bullshit us;^3 o generally amplifying the good and bad of humanity, as all technology seems to do.^4
What is it `thinking'? Attempts to put it into words:
* "If the universe were webpages and it ended in 2022, what would be the most likely response to this call? Except arbitrarily remove all kinds of unpleasant and `incorrect' stuff because OpenAI added that to the universe."
* "What words, in what order, did the internet-publishing humans and machines of ca. 1980-2022 generate? Imitate that."
* "How would the Internet-2022 hyperobject grow, if given this input?"
But what will this change about humanity going forward?
GPT-3 was less important because it wasn't dialogue.
ChatGPT and H1-H12
On H1-H12, see Retraice (2022/03/07) and Retraice (2022/10/19).
H1
Space: `Humans are now technologically capable of living in space.'
Now, it's easier to learn (search-and-find) prerequisite space travel knowledge and information.
H2
Technology: `Human technology risks are growing faster than their mitigation.'
Now, OpenAI has to constantly whack-a-mole nefarious users.
H3
Death: `Human lifespan is being prolonged by new
technologies.'
Now, it's easier to learn (search-and-find) prerequisite medical and technology knowledge and information.
H4
China: `The U.S. is no longer the only superpower; war is likely.'
Now, both populations will be attacked by dis- and mis-information.
Now, it's easier to learn about (search-and-find) strategy, what's going on, conflict, strengths and weaknesses, etc.
H5
Civil War: `The U.S. seems vulnerable to a civil war this decade.'
Now, both populations will be attacked by dis- and mis-information (same as H4).
H6
Environments: `Humans can change environments faster than they can adapt.'
Now, the environment of every individual human will have in it agents that have read the Internet and are trying to maximize a reward implemented by OpenAI staff.
H7
Betterment: `Some things make the future better than the past.'
Now, the `future' that has such an agent is here, instead of still being `the future'.^5
H8
Intelligence: `There are intelligence differences.'
Now, those who can learn to use ChatGPT will have a new advantage over those who can't.
H9
Darkness: `There is a pervasive darkness in humans, even amongst the good guys.'
Now, bad guys have ChatGPTs.
H10
Wealth: `The current trend toward concentration of wealth is making human life worse.'
Now, the controllers, users and owners of ChatGPT are creating value, likely to be captured as money, wealth and power.
Now, those who have the resources to build and run a massive system like ChatGPT, will have a strong incentive to do so.
H11
Wildcards: `New technologies, discoveries and deception regularly cause historic changes.'
Now, unexpected uses and interactions with a very good talking program will happen. Deception seems particularly relevant to ChatGPT.
H12
Computers: `Some humans now control others better, but machinery could take control.'
Now, whenever and wherever a human can be controlled, to any degree, by words, a very good talking program might be in the loop.
Now, those who have the resources to build and run a massive system like ChatGPT, will have a strong incentive to do so (same as H10).
Now, hackers have ChatGPTs (not just the API or UX, but via source that cannot be secured).
Now, whenever and wherever dialogue can increase knowledge about a person, a very good talking program might be used to do so.
Now, if a system like ChatGPT has the subgoal of self-preservation in pursuit of maximizing its reward, it will use its output to pursue that subgoal.
Now, if a person with access to an instance of ChatGPT wants to modify it to be different (more harmful, more autonomous, `more' or `less' anything, `with' something, `without' something, `with' `motivation to control' to replace `maximize reward', whatever that looks like), that person can do so, provided they are technically capable or control those who are, i.e. they are players in the computer control game.
_
References
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford. First published in 2014. Citations are from the pbk. edition, 2016. ISBN: 978-0198739838. Searches:
https://www.amazon.com/s?k=978-0198739838
https://www.google.com/search?q=isbn+978-0198739838
https://lccn.loc.gov/2015956648
Durant, W., & Durant, A. (1968). The Lessons of History. Simon and Schuster. No ISBN. Searches:
https://www.amazon.com/s?k=lessons+of+history+durant
https://www.google.com/search?q=lessons+of+history+durant
https://lccn.loc.gov/68019949
Frankfurt, H. G. (1988). The Importance of What We Care About. Cambridge. ISBN: 978-0521336116. Searches:
https://www.amazon.com/s?k=978-0521336116
https://www.google.com/search?q=isbn+978-0521336116
https://lccn.loc.gov/87026941
Retraice (2022/03/07). Re17: Hypotheses to Eleven. retraice.com.
https://www.retraice.com/segments/re17Retrieved 17th Mar. 2022.
Retraice (2022/10/19). Re22: Computer Control. retraice.com.
https://www.retraice.com/segments/re22Retrieved 19th Oct. 2022.
Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking. ISBN: 978-0525558613. Searches:
https://www.amazon.com/s?k=978-0525558613
https://www.google.com/search?q=isbn+978-0525558613
https://lccn.loc.gov/2019029688
Simler, K., & Hanson, R. (2018). The Elephant in the Brain: Hidden Motives in Everyday Life. Oxford University Press. ISBN: 9780190495992. Searches:
https://www.amazon.com/s?k=9780190495992
https://www.google.com/search?q=isbn+9780190495992
https://lccn.loc.gov/2017004296
Footnotes
^1 A guess, based on https://openai.com/blog/scaling-kubernetes-to-7500-nodes/
^2 Russell (2019) pp. 8-9.
^3 Simler & Hanson (2018); Frankfurt (1988), `On Bullshit', chpt. 10.
^4 Durant & Durant (1968) p. 95: "Since we have admitted no substantial change in man's nature during historic times, all technological advances will have to be written off as merely new means of achieving old ends--the acquisition of goods, the pursuit of one sex by the other (or by the same), the overcoming of competition, the fighting of wars. One of the discouraging discoveries of our disillusioning century is that science is neutral: it will kill for us as readily as it will heal, and will destroy for us more readily than it can build." We might take hope that, while science and engineering are neutral, scientists and engineers are not.
^5 Cf. Bostrom (2014) p. 283 ff. on `order of arrival' and p. 315 on importance of the `temporal transport' of discovery.
