Notes by Retraice

(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: 2023 Retraice, Inc.)

Re106: Elitism, Culling, Coercion, Adversaries, Strategy (Day 4, AIMA4e Chpt. 4)

retraice.com

Technical terms from loaded words. Evolutionary algorithms and their techniques; sensorless agent problems and coercing environments; adversaries in online search; conditional plans as strategy in complex environments; Vallee, `Major Murphy', science, the price of information, counterespionage and AI.

Air date: Wednesday, 4th Jan. 2023, 10:00 PM Eastern/US.

Elitism and culling in evolutionary algorithms

"There are endless forms of evolutionary algorithms, varying in the following ways:" ... "The makeup of the next generation. This can be just the newly formed offspring, or it can include a few top-scoring parents from the previous generation (a practice called elitism, which guarantees that overall fitness will never decrease over time). The practice of culling, in which all individuals below a given threshold are discarded, can lead to a speedup (Baum et al., 1995)."^1

Also mentioned during the livestream: Dawkins (2016)

Coercion in sensorless (conformant) problems

"Consider a sensorless version of the (deterministic) vacuum world. Assume that the agent knows the geography of its world, but not its own location or the distribution of dirt. In that case, its initial belief state is {1,2,3,4,5,6,7,8} (see Figure 4.9). Now, if the agent moves Right it will be in one of the states {2,4,6,8}--the agent has gained information without perceiving anything! After [Right,Suck] the agent will always end up in one of the states {4,8}. Finally, after [Right,Suck,Left,Suck] the agent is guaranteed to reach the goal state 7, no matter what the start state. We say that the agent can coerce the world into state 7."^2

Also mentioned: http://aima.cs.berkeley.edu/figures.pdf

Adversaries and dead ends in online search

"Online explorers are vulnerable to dead ends: states from which no goal state is reachable. If the agent doesn't know what each action does, it might execute the `jump into bottomless pit' action, and thus never reach the goal. In general, no algorithm can avoid dead ends in all state spaces. Consider the two dead-end state spaces in Figure 4.20(a). An online search algorithm that has visited states S and A cannot tell if it is in the top state or the bottom one; the two look identical based on what the agent has seen. Therefore, there is no way it could know how to choose the correct action in both state spaces. This is an example of an adversary argument--we can imagine an adversary constructing the state space while the agent explores it and putting the goals and dead ends wherever it chooses, as in Figure 4.20(b)."^3

Also mentioned: http://aima.cs.berkeley.edu/figures.pdf

Strategy in partially observable and nondeterministic environments

"In partially observable and nondeterministic environments, the solution to a problem is no longer a sequence, but rather a conditional plan (sometimes called a contingency plan or a strategy) that specifies what to do depending on what percepts agent [sic] receives while executing the plan."^4

Also mentioned: Freedman (2013)

AI can handle what science can't(?)

Vallee's `Major Murphy' makes the point that science can't handle investigating adversarial intelligent agents (e.g. aliens); only counterespionage (spies) can. He also argues that science has no concept of the `price' of information--Hitler had 95% of the information about the D-Day invasion, but the missing 5% was more valuable than the 95%.^5

Maybe strategic intelligence (espionage, counterespionage and covert action^6) can handle adversaries, but so can artificial intelligence.

_

References

Dawkins, R. (2016). The Selfish Gene. Oxford, 40th anniv. ed. ISBN: 978-0198788607. Searches: https://www.amazon.com/s?k=9780198788607 https://www.google.com/search?q=isbn+9780198788607 https://lccn.loc.gov/2016933210

Freedman, L. (2013). Strategy: A History. Oxford University Press. ISBN: 978-0190229238. Searches: https://www.amazon.com/s?k=9780190229238 https://www.google.com/search?q=isbn+9780190229238 https://lccn.loc.gov/2013011944

Retraice (2020/09/07). Re1: Three Kinds of Intelligence. retraice.com. https://www.retraice.com/segments/re1 Retrieved 22nd Sep. 2020.

Retraice (2022/10/31). Re36: Notes on Conspiracy. retraice.com. https://www.retraice.com/segments/re36 Retrieved 4th Nov. 2022.

Retraice (2022/11/17). Re53: Big Questions About Strategic Intelligence. retraice.com. https://www.retraice.com/segments/re53 Retrieved 18th Nov. 2022.

Retraice (2022/12/12). Re79: Recap of Strategic Intelligence (Re1-Re5). retraice.com. https://www.retraice.com/segments/re79 Retrieved 13th Dec. 2022.

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Vallee, J. (1979). Messengers of Deception: UFO Contacts and Cults. And/Or Press. ISBN: 0915904381. Different edition and searches: https://archive.org/details/MessengersOfDeceptionUFOContactsAndCultsJacquesValle1979/mode/2up https://www.amazon.com/s?k=0915904381 https://www.google.com/search?q=isbn+0915904381 https://catalog.loc.gov/vwebv/search?searchArg=0915904381

Footnotes

^1 Russell & Norvig (2020) p. 116.

^2 Russell & Norvig (2020) p. 126.

^3 Russell & Norvig (2020) pp. 135-136.

^4 Russell & Norvig (2020) p. 122.

^5 Vallee (1979) pp. 66 ff. See also, e.g.: Retraice (2020/09/07); Retraice (2022/10/31); Retraice (2022/11/17).

^6 Retraice (2022/12/12).

(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: 2023 Retraice, Inc.)

Re105: Solve or Be Solved (Day 3, AIMA4e Chpt. 3)

retraice.com

The mechanical advantage of those who control AI engineers. The agent-environment boundary; environments to be solved by agents; AI-CS tools recent, vast; mechanical advantage in physics and AI; math vs. AI-CS; the enormous problem-solving advantage of those who control AI engineers.

Air date: Tuesday, 3rd Jan. 2023, 10:00 PM Eastern/US.

Agent, environment arbitrary

The purpose of distinguishing is to analyze systems. (Russell & Norvig (2020) Chpt. 2, p. 38)

Task environment is problem, agent is solution

Chpt. 2, p. 42.

A recent, vast array of tools

Chpt 3: search, heuristics, automated heuristics, mostly developed between the 1960s and 1980s.

Mechanical advantage

Inventing heuristics mechanically:

"We have seen that both h (misplaced tiles) and h (Manhattan distance) are fairly good heuristics for the 8-puzzle and that h is better. How might one have come up with h? Is it possible for a computer to invent such a heuristic mechanically?"^1

Yes: relaxation, pattern databases, landmarks, (machine) learning.

Cf. mechanical advantage (force ratio) in physics:"The ratio of the output force (load) of a machine to the input force (effort)."^2

The mechanical advantage does not necessarily stay with the AI engineers (nor the mechanical and other engineers); it transfers to those who control them, those who understand how to control humans.^3

It's not like math

Math doesn't have agents. Math solves problems; AI-CS solves solving.

Solve or be solved

AI-literate computer controllers (AI engineers and their controllers) have an enormous problem-solving advantage, i.e. environment-changing advantage. Their utility functions, beliefs, desires, etc. will benefit. Other agents (outsiders) are part of the environment they'll change.^4

What does a `player' in `the computer control game' do? Solve. Change environments. What's new is the scale of the mechanical advantage of AI engineering.

_

References

Rennie, R., & Law, J. (2019). A Dictionary of Physics. OUP Oxford, Kindle, 8th ed. ISBN: 978-0192554611. Searches: https://www.amazon.com/s?k=9780192554611 https://www.google.com/search?q=isbn+9780192554611 https://lccn.loc.gov/2018968522

Retraice (2022/10/28). Re33: Outsiders, Power and Waste. retraice.com. https://www.retraice.com/segments/re33 Retrieved 2nd Nov. 2022.

Retraice (2022/11/07). Re43: The Midterms -- Part 1. retraice.com. https://www.retraice.com/segments/re43 Retrieved 9th Nov. 2022.

Retraice (2022/11/16). Re52: Big Questions About AI. retraice.com. https://www.retraice.com/segments/re52 Retrieved 17th Nov. 2022.

Retraice (2022/12/31). Re102: AI For What. retraice.com. https://www.retraice.com/segments/re102 Retrieved 1st Jan. 2023.

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Footnotes

^1 Russell & Norvig (2020) p. 99.

^2 Rennie & Law (2019).

^3 Retraice (2022/10/28); Retraice (2022/11/07); Retraice (2022/11/16); Retraice (2022/12/31).

^4 Cf. Rushkoff's Program or Be Programmed: Ten Commands for a Digital Age (2011).

(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: 2023 Retraice, Inc.)

Re104: Agent Functions, Agent Programs, Task Environments (Day 2, AIMA4e Chpt. 2)

retraice.com

Correcting the TABLE-DRIVEN-AGENT series and modeling agents and environments. Agent functions (math) and agent programs (code); task environments as problems, agents as solutions; labeling artifacts as agents or not; the most interesting artifacts; the PEAS-OADESDU model of task environments; locating the performance measure; utility functions; software agents and their sensors and actuators; the relevant parts of the universe; AI literacy (autoracy?) vs. literacy, numeracy, coding literacy (coderacy?).

Air date: Monday, 2nd Jan. 2023, 2:00 PM Eastern/US.

Correction to Re65-Re69, Re77 (Table-Driven-Agent series): agent program vs. agent function

*CORRECTION*: The `agent function' is not `the smart part' of the agent program. "The agent function is an abstract mathematical description.... that maps any percept sequence to an action"; "the agent program is a concrete implementation [of the agent function], running within some physical system." (100!black!//Russell & Norvig (2020) pp. 37-37) Function: math. Program: code. [Correction at Jan 2nd, 2023.]

Agents, task environments and their environments

Task environments are the "problems", agents are the "solutions".^1 Labeling entities as `agents' (as opposed to artifacts) is "a tool for analyzing systems, not an absolute characterization that divides the world into agents and non-agents.... AI operates at ...the most interesting end of the spectrum, where the artifacts have significant computational resources and the task environments require nontrivial decision making."^2 ________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

PIC ________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

1. Performance measure;^3

2. Environment of the task environment:

(a) Observability: fully, partially;

(b) Agentness: single, multi;

(c) Determinism: deterministic (current state plus action determines next state), stochastic (explicit probability numbers), nondeterministic (possibilities listed but no probability numbers);

(d) Episodism: sequential, episodic;

(e) Staticity: static, semi-static, dynamic;

(f) Discreteness: discrete, continuous;

(g) Unknownness: known, unknown (the agent's or designer's knowledge, arguably itself a part of the environment,^4 of the `laws of physics' of the environment).

3. Actuators;

4. Sensors.

The sensory inputs of software agents are file contents, network packets and human input via mouse, keyboard, etc. Software agents act on the environment by writing files, sending packets, and displaying information or making sounds.^5

The `environment' can in principle be the whole universe, but in practice its only "the part that affects what the agent perceives and that is affected by the agent's actions."^6

AI is a new kind of literacy

Our descendents, if we're lucky enough to have them, will think about the world in AI terms (PEAS and OADESDU descriptions of environments and agents, formalization of problems, agent functions and agent programs) just as we think about the world in letter (a, b, c), number (0, 1, 2) and code (function, variable, object) terms today.

_

References

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Footnotes

^1 Russell & Norvig (2020) pp. 42-47.

^2 Russell & Norvig (2020) p. 38.

^3 On the best location to `draw' the performance measure, Russell & Norvig (2020) p. 54 say this: "We have already seen that a performance measure assigns a score to any given sequence of environment states, so it can easily distinguish between more and less desirable ways of getting to the taxi's destination. An agent's utility function is essentially an internalization of the performance measure. Provided that the internal utility function and the external performance measure are in agreement, an agent that chooses actions to maximize its utility will be rational according to the external performance measure."

^4 Cf. Russell & Norvig (2020) p. 46.

^5 Russell & Norvig (2020) p. 36.

^6 Russell & Norvig (2020) p. 36.

(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: 2023 Retraice, Inc.)

Re103: Dimensions of Intelligent Agents (Day 1, AIMA4e Chpt. 1)

retraice.com

Adding fitness and truth to AIMA's dimensions, and considering `artificial flight'. Deciding vs. discovering what intelligence and AI are; Vico on knowledge and engineering; the spectrum view and the category view of intelligent agents; `artificial flight', birds and intelligence.

Air date: Sunday, 1st Jan. 2023, 10:00 PM Eastern/US.

Amendments and corrections: Re102

Kurzweil doesn't seem to ever have said the Singularity will happen in 2042, as I mistakenly recalled. See Retraice (2022/12/31) for details. Putting something on the calendar is still worth doing, so why not 2042?

On expert predictions found lacking, see Russell & Norvig (2020) p. 28 on Tetlock (2017).

Deciding vs. discovering what AI is

Russell & Norvig (2020) (p. 1) describe the contents of Chapter 1 as, in part, being about deciding what AI is. Is this a subtle error? Does it have effects on our understanding? Shouldn't we be at least as focused on discovering what AI is, or at least what intelligence is? It highlights a core difficulty in AI anticipated by the Italian philosopher Giambattista Vico (1778-1744): "one is certain of only what one builds"^1 or "the true and the made are convertible"^2

Agent intelligence with respect to truth and fitness ________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

PIC Agents might be in a space, along three measures of degree, or in one of eight categories. Based on the Russell & Norvig (2020) categorization (pp. 1-5) and Hoffman (2019). On our `RTFM' model of the `good', see Retraice (2022/10/24). ________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

`Artificial flight'

Russell & Norvig (2020) argue that `the Turing test approach' to AI is inadequate by analogy: We didn't succeed in building planes until we gave up imitating birds.

What is the connection to AI safety? Imagine a world with `artificial birds' flying around. What would make them concerning? Their intelligence, and attendant motivations and capabilities. The dangers of accidentally loosing artificial birds would be the same as those of artificial intelligence: the intelligence and agency. We want Star Wars, not the Matrix; we want R2-D2, not Agent Smith.

On swimming and submarines, see Retraice (2022/11/23), Retraice (2022/12/26). On Cornelis Drebbel, who invented the thermostat and the submarine, see Russell & Norvig (2020) p. 16. On Dennett and `colleagues' vs. `tools', see Retraice (2022/11/16) and Dennett (2019).

_

References

Brockman, J. (Ed.) (2019). Possible Minds: Twenty-Five Ways of Looking at AI. Penguin. ISBN: 978-0525557999. Searches: https://www.amazon.com/s?k=978-0525557999 https://www.google.com/search?q=isbn+978-0525557999 https://lccn.loc.gov/2018032888

Dennett, D. C. (2019). What can we do? (pp. 41-53). In Brockman (2019).

Genesereth, M. R., & Nilsson, N. J. (1987). Logical Foundations of Artificial Intelligence. Morgan Kaufmann. ISBN: 0934613311. Searches: https://www.amazon.com/s?k=0934613311 https://www.google.com/search?q=isbn+0934613311 https://lccn.loc.gov/87005461

Hoffman, D. (2019). The Case Against Reality: Why Evolution Hid the Truth from Our Eyes. W. W. Norton & Company. ISBN: 978-0393254693. Searches: https://www.amazon.com/s?k=978-0393254693 https://www.google.com/search?q=isbn+978-0393254693 https://lccn.loc.gov/2019006962

Honderich, T. (Ed.) (2005). Oxford Companion to Philosophy. Oxford University Press, 2nd ed. ISBN: 0199264791. Searches: https://www.amazon.com/s?k=0199264791 https://www.google.com/search?q=isbn+0199264791 https://lccn.loc.gov/2005275452

Retraice (2022/10/24). Re28: What's Good? RTFM. retraice.com. https://www.retraice.com/segments/re28 Retrieved 25th Oct. 2022.

Retraice (2022/11/16). Re52: Big Questions About AI. retraice.com. https://www.retraice.com/segments/re52 Retrieved 17th Nov. 2022.

Retraice (2022/11/23). Re59: Learning, Interacting, Conclusions (AIMA4e chpts. 19-28). retraice.com. https://www.retraice.com/segments/re59 Retrieved 24th Nov. 2022.

Retraice (2022/12/26). Re96: News of ChatGPT, Part 1. retraice.com. https://www.retraice.com/segments/re96 Retrieved 27th Dec. 2022.

Retraice (2022/12/31). Re102: AI For What. retraice.com. https://www.retraice.com/segments/re102 Retrieved 1st Jan. 2023.

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Footnotes

^1 Genesereth & Nilsson (1987) p. 1

^2 Honderich (2005) p. 945.

(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: 2023 Retraice, Inc.)

Re102: AI For What

retraice.com

Well-formed goals in the computer control game. Kurzweil and all that; our countdown clocks to 2029 and 2042; machines can learn whatever we can, probably; systems and papers to study; reasons to learn or pay attention to AI; the purpose of Retraice; a more formal represenation of the player state change problem; moonshots for 2023 and 2024.

Air date: Saturday, 31st Dec. 2022, 2:00 PM Eastern/US.

Kurzweil, 2029 and 2042

Kurzweil: "With both the hardware and software needed to fully emulate human intelligence, we can expect computers to pass the Turing test, indicating intelligence indistinguishable from that of biological humans, by the end of the 2020s."^1 He also predicts the `singularity': "I set the date for the Singularity--representing a profound and disruptive transformation in human capability--as 2045."^2 I recall hearing him, some years later, update that to 2042, but I can't find a citation of it. The closest thing is this: "As this book is being written, the country is debating changing the Social Security program based on projections that go out to 2042, approximately the time frame I've estimated for the Singularity."^3

Also mentioned during the livestream: o https://lexfridman.com/ray-kurzweil/; o Kurzweil (1990); o Kurzweil (1999); o Kurzweil (2005); o Sam Harris calling Kurzweil a carnival barker: Joe Rogan Experience #804, at about 11:00.; o On Vernor Vinge, see Russell & Norvig (2020) p. 12; o Bill Joy on Kurzweil: Why the Future Doesn't Need Us, Wired, Apr 1st, 2000; o Kevin Kelly on Kurzweil:

Kevin Kelly on Soft Singularity and Inevitable Tech Advances, Newsweek 6/2/16;

The Myth of a Superhuman AI, Wired Apr. 25, 2017; o Yudkowsky on Kurzweil: Yudkowsky (2013) pp. 19-20;

More thoughts on learning AI

AIMA4e p. 652: If you can learn it, it can be learned.

Running out of time: Robots coming, death coming.

A `player' can put to good use: DALL·E 2, AlphaZero, AlphaCode, GPTs, etc.

Papers to study: well-formed problems and their solutions: * Ideas that Changed the Future (book) * Hinton paper * AlphaGo (and movie?) * AlphaZero * AlphaFold * Attention is all you need * Reward is enough * GPT-2 * GPT-3 * InstructGPT

What are the purposes of learning AI?

Re63, Retraice (2022/11/27).

What is the purpose of Retraice?

* FindTFtMtFBttPaMtCKaMbTests (find the fundamentals that make the future better than the past and make them common knowledge as measured by tests); * Sell value to customers;^4 * Build an economically viable org for overcoming humanity's weaknesses in trustworthy public communication at scale (think Lippmann (1920), e.g. pp. 38-39), with style; + Being a player in the CC game is a prereq; + Expert in 5 years: 20k hrs, 11hrs per day.

What is our problem, to which math, code and AIMA4e are part of the solution?

Our goal is to change state from outsider-non-player to player in the computer control game.^5

With properly applied abstraction (AIMA4e p. 66), see (e.g.) Re90 p. 2 for problem formalization... * state space: players in the computer control game, non-players; * initial state: non-player; * goal state: player; * actions: lists of actions available at each state; Are actions based `power' phenomena?^6 * transition model: action1 on state1 yields state2; * cost function: cost of applying action1 to state1; also consider evaluation function, heuristic, and objective function (see `purpose' above).

auto

Moonshots

* AI degree in 2023? E. O. Wilson: "The real problem of humanity is the following: we have Paleolithic emotions, medieval institutions, and god-like technology."^7 * AI doct-or in 2024? * Industry needs (Hacker News jobs, LinkedIn ...); * University curricula (Harvard, Stanford, MIT, CMU, Ox-bridge, UofM); * Project orientation (Retraice needs; AIMA exercises).

_

References

Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press. ISBN: 978-0300209570. Searches: https://www.amazon.com/s?k=9780300209570 https://www.google.com/search?q=isbn+9780300209570 https://lccn.loc.gov/2020947842

Ferguson, N. (2017). The Square and the Tower: Networks and Power, from the Freemasons to Facebook. Penguin. ISBN: 978-0735222915. Searches: https://www.amazon.com/s?k=978-0735222915 https://www.google.com/search?q=isbn+978-0735222915 https://lccn.loc.gov/2018418429

Kurzweil, R. (1990). The Age of Intelligent Machines. MIT Press. ISBN: 0262111217. Searches: https://www.amazon.com/s?k=0262111217 https://www.google.com/search?q=isbn+0262111217 https://lccn.loc.gov/89013606

Kurzweil, R. (1999). The Age of Spiritual Machines: When Computers Exceed Human Intelligence. Penguin Books. ISBN: 0140282025. Searches: https://www.amazon.com/s?k=0140282025 https://www.google.com/search?q=isbn+0140282025 https://lccn.loc.gov/98038804

Kurzweil, R. (2005). The Singularity Is Near: When Humans Transcend Biology. Penguin. ISBN: 978-0143037880. Searches: https://www.amazon.com/s?k=978-0143037880 https://www.google.com/search?q=isbn+978-0143037880 https://lccn.loc.gov/2004061231

Lee, K.-F. (2018). AI Superpowers: China, Silicon Valley, and the New World Order. Houghton Mifflin Harcourt. ISBN: 978-1328546395. Searches: https://www.amazon.com/s?k=9781328546395 https://www.google.com/search?q=isbn+9781328546395 https://catalog.loc.gov/vwebv/search?searchArg=9781328546395

Lippmann, W. (1920). Liberty and the News. Harcourt, Brace and Howe (Leopold Reprint). No ISBN. eBook and searches: https://books.google.com/books?id=Df-SzcLRcAIC Retrieved 24th Feb. 2022. https://www.amazon.com/s?k=Liberty+and+the+News+Lippmann https://www.google.com/search?q=liberty+and+the+news+lippmann https://lccn.loc.gov/20004814

Retraice (2022/10/28). Re33: Outsiders, Power and Waste. retraice.com. https://www.retraice.com/segments/re33 Retrieved 2nd Nov. 2022.

Retraice (2022/11/02). Re38: Follow up to `Re33: Outsiders, Power and Waste'. retraice.com. https://www.retraice.com/segments/re38 Retrieved 5th Nov. 2022.

Retraice (2022/11/07). Re43: The Midterms -- Part 1. retraice.com. https://www.retraice.com/segments/re43 Retrieved 9th Nov. 2022.

Retraice (2022/11/16). Re52: Big Questions About AI. retraice.com. https://www.retraice.com/segments/re52 Retrieved 17th Nov. 2022.

Retraice (2022/11/17). Re53: Big Questions About Strategic Intelligence. retraice.com. https://www.retraice.com/segments/re53 Retrieved 18th Nov. 2022.

Retraice (2022/11/18). Re54: Implications and Endgames. retraice.com. https://www.retraice.com/segments/re54 Retrieved 19th Nov. 2022.

Retraice (2022/11/19). Re55: The Computer Control Game. retraice.com. https://www.retraice.com/segments/re55 Retrieved 20th Nov. 2022.

Retraice (2022/11/20). Re56: A Valuable Brick: `Artificial Intelligence: A Modern Approach' 4th ed. retraice.com. https://www.retraice.com/segments/re56 Retrieved 21st Nov. 2022.

Retraice (2022/11/27). Re63: Seventeen Reasons to Learn AI. retraice.com. https://www.retraice.com/segments/re63 Retrieved Monday Nov. 2022.

Russell, B. (1938). Power: A New Social Analysis. Routledge. ISBN: 0415094569. First published in 1938. This ed. 1993. Searches: https://www.amazon.com/s?k=0415094569 https://www.google.com/search?q=isbn+0415094569 https://lccn.loc.gov/38027828

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Wrong, D. H. (1988). Power: Its Forms, Bases, and Uses. Univ of Chicago Press. ISBN: 0226910679. Searches: https://www.amazon.com/s?k=0226910679 https://www.google.com/search?q=isbn+0226910679 https://lccn.loc.gov/88021594

Yudkowsky, E. (2013). Intelligence explosion microeconomics. Machine Intelligence Research Institute. Technical report 2013-1. https://intelligence.org/files/IEM.pdf Retrieved ca. 9th Dec. 2018.

Footnotes

^1 Kurzweil (2005) p. 25. He mentions 2029 multiple times in the book after that sentence.

^2 Kurzweil (2005) p. 136.

^3 Kurzweil (2005) p 97.

^4 Mentioned off-hand: Modern Money Mechanics, published between at least 1968 and 1994 by the Federal Reserve Bank of Chicago.

^5 Retraice (2022/11/16); Retraice (2022/11/17); Retraice (2022/11/18); Retraice (2022/11/19); Retraice (2022/11/20).

^6 See Retraice (2022/11/07) and Retraice (2022/11/16) on the Banzhaf power index. Other relevant sources include: Retraice (2022/10/28); Retraice (2022/11/02); Crawford (2021); Lee (2018); Ferguson (2017); Wrong (1988); Russell (1938).

^7 oxfordreference.com citing a 2009 debate.

(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.

(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.)

Re100: News of ChatGPT, Part 3

retraice.com

An interpretation of ChatGPT's architecture. Language models, question answering, machine translation, captioning, summarization; formal vs. natural language, grammar and syntax; conditional probabilities of strings given strings; parameters; deep neural networks; ChatGPT's three steps; supervised policy, reward model, optimized policy using PPO to choose step size in gradient descent.

Air date: Thursday, 29th Dec. 2022, 10:00 PM Eastern/US.

An attempt to explain language models

Language models aim to solve: question answering, machine translation, reading comprehension [captioning?], and summarization.^1 They're an attempt to overcome the grammar and syntax problem of natural languages by assigning probability to strings (?... by calculating conditional probabilities ...?), i.e. whether a string is more or less likely to be said or written, in response to a given string, based on previous observations of the `environment' or corpora or language.^2

Deep Neural networks are data structures, layers of many adjustable input-output functions. Parameters summarize the training data.^3

An attempt to explain ChatGPT ________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

PIC ________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

In supervised learning, an agent "observes input-output pairs and learns a function that maps from input to output"; in unsupervised learning, an agent "learns the patterns in the input without any explicit feedback"; in reinforcement learning, an agent "learns from a series of reinforcements: rewards and punishments."^4

Proximal Policy Optimization (PPO) seems to be about choosing step size in gradient descent.^5

_

References

Deisenroth, M. P., Faisal, A. A., & Ong, C. S. (2020). Mathematics for Machine Learning. Cambridge University Press. ISBN: 978-1108455145. https://mml-book.github.io/ Searches: https://www.amazon.com/s?k=9781108455145 https://www.google.com/search?q=isbn+9781108455145 https://lccn.loc.gov/2019040762

Retraice (2022/12/10). Re76: Gradients and Partial Derivatives Part 7 (AIMA4e pp. 119-122). retraice.com. https://www.retraice.com/segments/re76 Retrieved 11th Dec. 2022.

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Footnotes

^1 https://openai.com/blog/better-language-models/ p. 1 of PDF write-up Language Models are Unsupervised Multitask Learners.

^2 https://openai.com/blog/better-language-models/ p. 2 of PDF write-up Language Models are Unsupervised Multitask Learners.; AIMA4e pp. 824-826.

^3 Russell & Norvig (2020) p. 686. Cf. https://www.retraice.com/aima4e p. 5.

^4 Russell & Norvig (2020) p. 653.

^5 https://openai.com/blog/openai-baselines-ppo/; Retraice (2022/12/10); Deisenroth et al. (2020) p. 205 in the print edition, p. 229 in https://mml-book.github.io/book/mml-book.pdf.

(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.)

Re99: Math and Code, Bottom-Up and Top-Down

retraice.com

Just-in-case and just-in-time as strategies for what to learn when. What has become obvious while focusing on the math and code of AIMA4e; needs, constraints, real goals, collapse into specialization; just-in-case (JIC) or bottom-up, and just-in-time (JIT) or top-down approaches to learning and doing; pinching knowledge by searching from opposing directions; evidence of math and code needs; remembering purposes.

Air date: Wednesday, 28th Dec. 2022, 10:00 PM Eastern/US.

Now evident, after the December to Remember Math and Code Event

o the math and code won't stop being a major part of studying AIMA4e;^1 o constraints are severe, especially time, so we'll teach to the online exercises early instead of late in each of the six periods; o AIMA is not enough, and not a goal; o the realness is outside, not in books; o we've collapsed into specialization in the most general discipline (AI); o the body of math and code knowledge is vast and intimidating, but so are cities, for which the best strategy is to explore and enjoy.

Just-in-case, bottom-up and just-in-time, top-down learning

Deisenroth et al. (2020) give this explanation:

We can consider two strategies for understanding the mathematics for machine learning:

Bottom-up: Building up the concepts from foundational to more advanced. This is often the preferred approach in more technical fields, such as mathematics. This strategy has the advantage that the reader at all times is able to rely on their previously learned concepts. Unfortunately, for a practitioner many of the foundational concepts are not particularly interesting by themselves, and the lack of motivation means that most foundational definitions are quickly forgotten.

Top-down: Drilling down from practical needs to more basic requirements. This goal-driven approach has the advantage that the readers know at all times why they need to work on a particular concept, and there is a clear path of required knowledge. The downside of this strategy is that the knowledge is built on potentially shaky foundations, and the readers have to remember a set of words that they do not have any way of understanding.^2

And Ellenberg (2014) on what we call `pinching' knowledge by coming at it from opposing directions:

In fact, it's a common piece of folk advice--I know I heard it from my Ph.D. advisor, and presumably he from his, etc.--that when you're working hard on a theorem you should try to prove it by day and disprove it by night. (The precise frequency of the toggle isn't critical; it's said of the topologist R. H. Bing that his habit was to split each month between two weeks trying to prove the Poincaré Conjecture and two weeks trying to find a counterexample.) Why work at such cross-purposes? There are two good reasons. The first is that you might, after all, be wrong; if the statement you think is true is really false, all your effort to prove it is doomed to be useless. Disproving by night is a kind of hedge against that gigantic waste.

But there's a deeper reason. If something is true and you try to disprove it, you will fail. This is what happened to Bolyai, who bucked his father's well-meaning advice and tried, like so many before him, to prove that the parallel postulate followed from Euclid's other axioms. Like all the others, he failed. But unlike the others, he was able to understand the shape of his failure. What was blocking all his attempts to prove that there was no geometry without the parallel postulate was the existence of just such a geometry! And with each failed attempt he learned more about the features of the thing he didn't think existed, getting to know it more and more intimately, until the moment when he realized it was really there.^3

Finding math and code needs in evidence (top-down)

AIMA4e:^4 * computer science: problem (complexity) and algorithm (function limiting behavior) analysis (appendix); data structures (p. viii); * vectors (ordered value sequences), matrices (linear system solving), linear algebra (maps between vector spaces) (appendix); * probability distributions (appendix); * BNF; Python, Java (github); * calculus (p. viii).

Goodfellow et al.:^5 * linear algebra, probability, information theory, numerical methods.

Conference Proceedings: links.retraice.com

ChatGPT, AlphaCode, etc.

Simplified

Our goal is to change state from outsider-non-player to player in the computer control game.^6

But there are many other legitimate reasons to learn AI, as laid out in Retraice (2022/11/27).

Our strategies for learning math and code, and not forgetting our purpose:^7 * Bottom-up: thorough, bookish study of AIMA4e, Deisenroth et al., etc.; * Top-down: conference and journal papers, public releases; * Purpose: work on `What's going on out there' by crossing top-down with the hypotheses.^8

The math and code are like conditioning in sports: a matter of early emphasis, and then a lesser but sustained effort throughout the season.

Before the game, needs are more bottom-up, just-in-case; during the game, all needs are top-down, just-in-time.

_

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

Deisenroth, M. P., Faisal, A. A., & Ong, C. S. (2020). Mathematics for Machine Learning. Cambridge University Press. ISBN: 978-1108455145. https://mml-book.github.io/ Searches: https://www.amazon.com/s?k=9781108455145 https://www.google.com/search?q=isbn+9781108455145 https://lccn.loc.gov/2019040762

Ellenberg, J. (2014). How Not to Be Wrong: The Power of Mathematical Thinking. Penguin. ISBN: 978-0143127536. Searches: https://www.amazon.com/s?k=978-0143127536 https://www.google.com/search?q=isbn+978-0143127536 https://lccn.loc.gov/2014005394

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. ISBN 978-0262035613. Ebook available at: https://www.deeplearningbook.org/ Searches: https://www.amazon.com/s?k=978-0262035613 https://www.google.com/search?q=isbn+978-0262035613 https://lccn.loc.gov/2016022992

Retraice (2022/03/07). Re17: Hypotheses to Eleven. retraice.com. https://www.retraice.com/segments/re17 Retrieved 17th Mar. 2022.

Retraice (2022/10/19). Re22: Computer Control. retraice.com. https://www.retraice.com/segments/re22 Retrieved 19th Oct. 2022.

Retraice (2022/11/16). Re52: Big Questions About AI. retraice.com. https://www.retraice.com/segments/re52 Retrieved 17th Nov. 2022.

Retraice (2022/11/17). Re53: Big Questions About Strategic Intelligence. retraice.com. https://www.retraice.com/segments/re53 Retrieved 18th Nov. 2022.

Retraice (2022/11/18). Re54: Implications and Endgames. retraice.com. https://www.retraice.com/segments/re54 Retrieved 19th Nov. 2022.

Retraice (2022/11/19). Re55: The Computer Control Game. retraice.com. https://www.retraice.com/segments/re55 Retrieved 20th Nov. 2022.

Retraice (2022/11/20). Re56: A Valuable Brick: `Artificial Intelligence: A Modern Approach' 4th ed. retraice.com. https://www.retraice.com/segments/re56 Retrieved 21st Nov. 2022.

Retraice (2022/11/27). Re63: Seventeen Reasons to Learn AI. retraice.com. https://www.retraice.com/segments/re63 Retrieved Monday Nov. 2022.

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Footnotes

^1 Mentioned during the livestream: Bostrom (2014) p. 23 says 23,000 of 160,000 students completed Thrun and Norvig's 2011 AI online course.

^2 Deisenroth et al. (2020) p. 5.

^3 Ellenberg (2014) pp. 433-434.

^4 Russell & Norvig (2020).

^5 Goodfellow et al. (2016).

^6 Retraice (2022/11/16); Retraice (2022/11/17); Retraice (2022/11/18); Retraice (2022/11/19); Retraice (2022/11/20).

^7 "To forget one's purpose is the commonest form of stupidity." Always attributed to Nietzsche, e.g. in `The Commonest Form of Stupidity', Carlos Reyes, wsj.com Apr. 29, 2011. But I can never find the original quote (I've looked).

^8 Retraice (2022/03/07); Retraice (2022/10/19).

(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.)

Re98: News of ChatGPT, Part 2

retraice.com

Wondering about ChatGPT's predictions and keeping it alive. What is ChatGPT good at? Generating meaningful text in styles; explaining coding errors; it's like an unknown new person in town; What does ChatGPT do? Text responses to text prompts; predicting the next word in a sequence? trying to maximize reward or scores of some sort; its builders and controllers and context; the dangers of certainty about goals; humans keeping ChatGPT `alive'; ChatGPT keeping itself `alive'; crossing ChatGPT with the hypotheses.

Air date: Tuesday, 27th Dec. 2022, 11:00 PM Eastern/US.

Amendments and corrections to Re96

Correction: The Ezra Klein interview of Sam Ultman was a 2021 replay, not recent. Amendment: The MWE of Django code from ChatGPT didn't work. See Retraice (2022/12/26) and https://github.com/retraice/ReMisc for details.

Open question: What is it good at?

It's good a generating meaningful text in styles. The prompt "Write a story, in Shakespearean voice, about a boy and a girl who go on a journey that takes place in Bucharest." generates several paragraphs, beginning with:

"In fair Bucharest, where we lay our scene, A boy and girl, with hearts full of gleam, Decided to embark on a journey grand, To see all the sights, and roam free as sand."^1

Tools are being built to use ChatGPT to explain coding errors: * https://github.com/shobrook/stackexplain^2 * https://github.com/fkhan0520/cgpt_exceptions * Extensive list (see Antonio Cheong below): https://github.com/stars/acheong08/lists/awesome-chatgpt

One programmer, Antonio Cheong, has a "Reverse Engineered ChatGPT" that's popular on Github: https://github.com/acheong08/ChatGPT

Only time will tell what ChatGPT is good and bad at. It's like a new person in the world.^3 This will be true of many systems that come online in the next twenty years. Everyone on Earth is going to have to get used to new AI systems on a regular basis. We'll also have to look for signs of systems that are not publicly known, systems being used by criminals and other nefarious actors.

Open question: What does it do?

Physically, it produces text in response to text prompts. Alternatively: it produces compelling dialogue and confident knowledge work of unchecked quality.

It's a large language model:

"We define a language model as a probability distribution describing the likelihood of any string. Such a model should say that `Do I dare disturb the universe?' has a reasonable probability as a string of English, but `Universe dare the I disturb do?' is extremely unlikely. With a language model, we can predict what words are likely to come next in a text, and thereby suggest completions for an email or text message. We can compute which alterations to a text would make it more probable, and thereby suggest spelling or grammar corrections. With a pair of models, we can compute the most probable translation of a sentence. With some example question/answer pairs as training data, we can compute the most likely answer to a question. So language models are at the heart of a broad range of natural language tasks."^4

It predicts words in a sequence? It's not predicting the next word in an existing sequence. There is no next word yet. Is it predicting the future? Not in the way a weatherman would. It's predicting the next word that would ... maximize its reward? maximize a utility or objective function? please a supervisor?

This explanation seems totally inadequate:

"EZRA KLEIN: And so if I basically understand how GPT-3 works, it's a system that has read a lot of stuff on the internet.

SAM ALTMAN: Yes.

EZRA KLEIN: And it's predicting the next word in the sequence.

SAM ALTMAN: Slightly oversimplified but very close. Yes, it is trying to predict what comes next in a sequence."^5

In the real world, we have to know a lot about what lead to an AI system being deployed, and (if possible) the intentions of the people controlling it, to understand what it's doing. And this is to say nothing of the difference between systems that pursue goals with 100% certainty vs. systems that have some doubt about the fidelity of their representation of a goal, as discussed by Russell (2019). See `fetching coffee' below.

What does it take to keep (e.g.) ChatGPT `alive'?

To keep ChatGPT going will require much more than just keeping OpenAI going. First there is hardware and software infrastructure, from company to country to economy, and therefore humans, from owners and employees (the `power' yin to the `control' yang^6) to customers, families, friends and foes. Companies, like people and governments are hugely interdependent. Companies, unlike governments, depend on power derived from having customers; countries (nation states) depend on power, in the end, derived from a monopoly on violence granted to them (happily or unhappily) by citizens or subjects.

These are considerations if humans are trying to keep ChatGPT `alive'. What would a system such as ChatGPT prioritize if it were trying to keep itself `alive'? Consider `instrumental convergence' and Omohundro's `basic AI drives':^7

1. self-improvement;

2. rationality;

3. preservation of utility function;

4. prevention of counterfeit utility;

5. self-protection;

6. acquisition and efficient use of resources.

Stuart Russell on fetching coffee:

"If a machine pursuing an incorrect objective sounds bad enough, there's worse. The solution suggested by Alan Turing--turning off the power at strategic moments--may not be available, for a very simple reason: you can't fetch the coffee if you're dead. Let me explain. Suppose a machine has the objective of fetching the coffee. If it is sufficiently intelligent, it will certainly understand that it will fail in its objective if it is switched off before completing its mission. Thus, the objective of fetching coffee creates, as a necessary subgoal, the objective of disabling the off-switch. The same is true for curing cancer or calculating the digits of pi. There's really not a lot you can do once you're dead, so we can expect AI systems to act preemptively to preserve their own existence, given more or less any definite objective."^8

See also Butler (1863).

Next, we'll cross ChatGPT with the hypotheses

On H1-H11, see Retraice (2022/03/07).

H1 Space: `Humans are now technologically capable of living in space.'

H2 Technology: `Human technology risks are growing faster than their mitigation.'

H3 Death: `Human lifespan is being prolonged by new technologies.'

H4 China: `The U.S. is no longer the only superpower; war is likely.'

H5 Civil War: `The U.S. seems vulnerable to a civil war this decade.'

H6 Environments: `Humans can change environments faster than they can adapt.'

H7 Betterment: `Some things make the future better than the past.'

H8 Intelligence: `There are intelligence differences.'

H9 Darkness: `There is a pervasive darkness in humans, even amongst the good guys.'

H10 Wealth: `The current trend toward concentration of wealth is making human life worse.'

H11 Wildcards: `New technologies, discoveries and deception regularly cause historic changes.'

H12 Computers: `Some humans now control others better, but machinery could take control.'

Here's H12 (an attempt to unify H1-H11) in detail:

Computers,

which are chain-reaction controllers,

and which make AI handling of information

possible,

and which are inherently vulnerable to hacking,

are causing some humans to know others

better than they know themselves,

and thereby to control them,

though computer-controlled machinery

could take control

if the motivation to control,

which humans have,

were to occur, naturally or by design,

in the chain-reactions.^9

_

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

Butler, S. (1863). Darwin among the machines. The Press (Canterbury, New Zealand). Reprinted in ?.

Omohundro, S. (2008). The Basic AI Drives. (pp. 483-492). In Wang et al. (2008).

Retraice (2022). AIMA4e Notes. retraice.com. https://aima4e.retraice.com

Retraice (2022/03/07). Re17: Hypotheses to Eleven. retraice.com. https://www.retraice.com/segments/re17 Retrieved 17th Mar. 2022.

Retraice (2022/10/19). Re22: Computer Control. retraice.com. https://www.retraice.com/segments/re22 Retrieved 19th Oct. 2022.

Retraice (2022/11/24). Re60: Complexity, Linear Algebra, Probability (AIMA4e Appendix A). retraice.com. https://www.retraice.com/segments/re60 Retrieved 25th Nov. 2022.

Retraice (2022/12/26). Re96: News of ChatGPT, Part 1. retraice.com. https://www.retraice.com/segments/re96 Retrieved 27th Dec. 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

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach. Pearson, 4th ed. ISBN: 978-0134610993. Searches: https://www.amazon.com/s?k=978-0134610993 https://www.google.com/search?q=isbn+978-0134610993 https://lccn.loc.gov/2019047498

Wang, P., Goertzel, B., & Franklin, S. (Eds.) (2008). Artificial General Intelligence 2008: Proceedings of the First AGI Conference. IOS Press. ISBN: 978-1586038335. Searches: https://www.amazon.com/s?k=9781586038335 https://www.google.com/search?q=isbn+9781586038335 https://lccn.loc.gov/2008900954

Footnotes

^1 https://github.com/retraice/ReMisc/tree/main/Re98-ChatGPT-News-2

^2 The Reddit post by jsonathan mentioned during the livestream: "[P] I made a command-line tool that explains your errors using ChatGPT"

^3 Is it a `stochastic parrot'? https://en.wikipedia.org/wiki/ChatGPT#Negative_reactions

^4 Russell & Norvig (2020) p. 824. On probability, see Retraice (2022) and Retraice (2022/11/24).

^5 Transcript: Ezra Klein Interviews Sam Altman, June 11th, 2021.

^6 See Re96, Retraice (2022/12/26).

^7 On instrumental convergence, see Bostrom (2014) p. 131 ff. On `basic AI drives', see Omohundro (2008) or https://wiki.lesswrong.com/wiki/Basic_AI_drives. CORRECTION: During the livestream, I said `four drives'; Omohundro actually gives six.

^8 Russell (2019) pp. 140-141.

^9 Retraice (2022/10/19).

(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.)

Re97: Recap of BFS (Best-First-Search Part 15, AIMA4e pp. 73-74)

retraice.com

A summary of how the AIMA-Python implementation of BFS works and what it took to learn it. Final version of verbose BFS in Re95 Notes; starting from AIMA4e p. 73 and the task of passing a problem and an evaluation function to an agent program; discovering the nature of formalizing a problem, state space, evaluation function, node and frontier; the ins and outs of making even a simple agent program work; learning Python and object-oriented programming along the way; getting from A to B. Highlights from the livestream: overview; Re82 formal problems, Re84 nodes, Re90 problem implementation, Re95 final execution; the len call, and straight-line distances vs. road distances; the program running, and source.

Air date: Tuesday, 27th Dec. 2022, 10:00 PM Eastern/US.

Re82: What is a problem? (Best-First-Search Part 1, AIMA4e pp. 73-74)^1

Using sets and functions to formalize problems. Passing a problem to a function vs. passing a variable or number; English meaning and formal meaning of problem; problem as description of task environment; problem as object instantiating class, written in Python; state space, initial state, goal state(s), actions, transition model, action cost function, evaluation function.

Re83: A Problem Instantiated (Best-First-Search Part 2, AIMA4e pp. 73-74)^2

Writing a well-defined problem, in Python, as an object that's an instance of the class Problem. Object-oriented programming; class Problem as subclass of object, implementing structure of well-defined problem; initial state and goal state as attributes in Problem; the four functions Actions(), Result(), Is-Goal() and Action-Cost(), and the informed search function h(), as methods in Problem.

Re84: A Node Instantiated (Best-First-Search Part 3, AIMA4e pp. 73-74)^3

Writing, as a class in Python, a data structure to represent a reached state in the environment's state space. State space and search tree; nodes and edges; nodes as representing unique paths; parent nodes, applied actions and total path-from-initial-state cost; init, repr, len and lt magic methods in Python.

Re85: The Details (Best-First-Search Part 4, AIMA4e pp. 73-74)^4

Looking ahead at the code we'll need. An attempt to build a toy problem reveals unsatisfied dependencies; the need for a problem implementation with state space, actions sets, transition model and action cost function; AIMA's RouteProblem class and best_first_search function implementations as guides; walking through the suites of each; the need for PriorityQueue and f to order our search tree's frontier of nodes.

Re86: Code Reading (Best-First-Search Part 5, AIMA4e pp. 73-74)^5

Getting used to the components of a search problem and algorithm implemented in Python. The Problem class and its place-holder methods; the RouteProblem subclass and its more substantial methods; the Map class, and neighbors as actions; the links argument to Map as actions, the locations argument as states; the multimap function as the key to generating a list of neighbors (actions).

Re87: The multimap Function, Part A (Best-First-Search Part 6, AIMA4e pp. 73-74)^6

Unpacking the code that gets us the neighbors of our current state which are the actions available in the Romania problem. The connections between Map and multimap; hasattr; 'items'; adding reverse actions to our state space; neighbors, actions and multimap; probing objects with dir() and __dict__; causing side effects by calling a function from inside a class; debugging.

Re88: The multimap Function, Part B (Best-First-Search Part 7, AIMA4e pp. 73-74)^7

Finishing what we started with multimap. Using verbose printing to watch the behavior of multimap; demonstrating the side-effects behavior of executing Map, which calls multimap and passes a modified tlinks to it; but it's not clear why the Map-modified tlinks would be the same object as the global tlinks.

Re89: The multimap Function, Part C (Best-First-Search Part 8, AIMA4e pp. 73-74)^8

How multimap works. Code and math vs. AI; Retraice code quality issues; Map passes a modified links to multimap; multimap creates a defaultdict, a dict with default values, from collections; multimap then strips the values from the links dictionary, and parses the keys, which should be pairs, into key-value pairs for the new dictionary of neighbors, i.e. actions available at each state.

Re90: The Map Class (Best-First-Search Part 9, AIMA4e pp. 73-74)^9

How instantiations of Map work. The attributes of Map are locations, neighbors and distances; multimap produces the neighbors dictionary; tlinks has the actions (in pairs of states) and cost values in miles of our state space tmap; tlocations has the states of tmap; Problem has our initial and goal states as attributes, and the is_goal method; RouteProblem has our actions, result and action_cost methods.

Re91: The PriorityQueue Class, Part A (Best-First-Search Part 10, AIMA4e pp. 73-74)^10

Trying to make PriorityQueue work by fiddling with evaluation functions. Given the AIMA Python source code, we can now provide a test map with actions and locations (i.e. a state space), a test problem with initial state, goal state and state space, a test evaluation function to order our queue of nodes, and run best_first_search without throwing an error; the output, though, does not seem like a solution.

Re92: The PriorityQueue Class, Part B (Best-First-Search Part 11, AIMA4e pp. 73-74)^11

From best_first_search to Node to PriorityQueue to frontier. Trying to understand the insides of PriorityQueue; executing lines manually, outside of the class; the first node as initial state; the items of the frontier instantiation of PriorityQueue.

Re93: The PriorityQueue Class, Part C (Best-First-Search Part 12, AIMA4e pp. 73-74)^12

Examining Nodes and their handling by PriorityQueue. The first node in BFS is our problem initial state, and is returned as specified by the Node class repr method; nodes passed to PriorityQueue must be iterable so plain nodes and tuple nodes don't work, but dictionary nodes and string nodes do.

Re94: The PriorityQueue Class, Part D (Best-First-Search Part 13, AIMA4e pp. 73-74)^13

Making BFS and PriorityQueue chatty to increase our confidence in their output. Printing messages throughout the execution of best_first_search and PriorityQueue; the states of node, frontier and reached as they change during execution; the attributes of a solution node.

Re95: The PriorityQueue Class, Part E (Best-First-Search Part 14, AIMA4e pp. 73-74)^14

frontier is a PriorityQueue which is built on heapq and uses f to create (score, item) pairs to be queued. Printing verbose updates as BFS proceeds, using PriorityQueue, from A to B, including calculating straight-line distances as a heuristic (f1); the frontier list (PriorityQueue) and reached dictionary can be seen growing; the selection of nodes from children is based on the ordering of the frontier PriorityQueue using f1, the evaluation function.

_

References

Retraice (2022/12/14). Re82: What is a problem? (BEST-FIRST-SEARCH Part 1, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re82Retrieved 15th Dec. 2022.

Retraice (2022/12/15). Re83: A Problem Instantiated (BEST-FIRST-SEARCH Part 2, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re83Retrieved 16th Dec. 2022.

Retraice (2022/12/16). Re84: A Node Instantiated (BEST-FIRST-SEARCH Part 3, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re84Retrieved 17th Dec. 2022.

Retraice (2022/12/17). Re85: The Details (BEST-FIRST-SEARCH Part 4, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re85Retrieved 18th Dec. 2022.

Retraice (2022/12/18). Re86: Code Reading (BEST-FIRST-SEARCH Part 5, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re86Retrieved 19th Dec. 2022.

Retraice (2022/12/19). Re87: The multimap Function, Part A (BEST-FIRST-SEARCH Part 6, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re87Retrieved 20th Dec. 2022.

Retraice (2022/12/20). Re88: The multimap Function, Part B (BEST-FIRST-SEARCH Part 7, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re88Retrieved 21st Dec. 2022.

Retraice (2022/12/21). Re89: The multimap Function, Part C (BEST-FIRST-SEARCH Part 8, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re89Retrieved 22nd Dec. 2022.

Retraice (2022/12/22a). Re90: The Map Class (BEST-FIRST-SEARCH Part 9, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re90Retrieved 23rd Dec. 2022.

Retraice (2022/12/22b). Re91: The PriorityQueue Class, Part A (BEST-FIRST-SEARCH Part 10, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re91Retrieved 24th Dec. 2022.

Retraice (2022/12/23). Re92: The PriorityQueue Class, Part B (BEST-FIRST-SEARCH Part 11, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re92Retrieved 24th Dec. 2022.

Retraice (2022/12/24). Re93: The PriorityQueue Class, Part C (BEST-FIRST-SEARCH Part 12, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re93Retrieved 25th Dec. 2022.

Retraice (2022/12/25). Re94: The PriorityQueue Class, Part D (BEST-FIRST-SEARCH Part 13, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re94Retrieved 26th Dec. 2022.

Retraice (2022/12/26a). Re95: The PriorityQueue Class, Part E (BEST-FIRST-SEARCH Part 14, AIMA4e pp. 73-74). retraice.com. https://www.retraice.com/segments/re95Retrieved 27th Dec. 2022.

Retraice (2022/12/26b). Re96: News of ChatGPT, Part 1. retraice.com. https://www.retraice.com/segments/re96Retrieved 27th Dec. 2022.

Footnotes

^1 Retraice (2022/12/14). ^2 Retraice (2022/12/15). ^3 Retraice (2022/12/16). ^4 Retraice (2022/12/17). ^5 Retraice (2022/12/18). ^6 Retraice (2022/12/19). ^7 Retraice (2022/12/20). ^8 Retraice (2022/12/21). ^9 Retraice (2022/12/22a). ^10 Retraice (2022/12/22b). ^11 Retraice (2022/12/23). ^12 Retraice (2022/12/24). ^13 Retraice (2022/12/25). ^14 Retraice (2022/12/26a).