All of SoerenE's Comments + Replies

Thank you for explaining.

Thank you for your comments. I have included them in version 1.1 of the map, where I have swapped FRI and OpenAI/DeepMind, added Crystal Trilogy and corrected the spelling of Vernor Vinge.

SoerenE280

I have taken the survey.

I think difference in date of birth (1922 vs ~1960) is less important than difference of date of publication (2003 vs ~2015).

On the Outside View, is criticism 12 years after publication more likely to be valid than criticism levelled immediately? I do not know. On one hand, science generally improves over time. On the other hand, if a particular work get the first criticism after many years, it could mean that the work is of higher quality.

I should clarify that I am referring to the section David Chapman calls: "Historical appendix: Where did the confusion come from?". I read it as a criticism of both Jaynes and his book.

No, I do know what Yudkowsky's argument is. Truth be told, I probably would be able to evaluate the arguments, but I have not considered it important. Should I look into it?

I care about whether "The Outside View" works as a technique for evaluating such controversies.

Yes! From the Outside View, this is exactly what I would expect substantial, well-researched criticism to look like. Appears very scientific, contains plenty of references, is peer-reviewed and published in "Journal of Statistical Physics" and has 29 citations.

Friedman and Shimonys criticism of MAXENT is in stark contrast to David Chapmans criticism of "Probability Theory".

1Oscar_Cunningham
FWIW I think that Davud Chapman's criticism is correct as far as it goes, but I don't think that it's very damning. Propositional logic is indeed a "logic" and it's worthwhile enough for probability theory to extend it. Trying to look at predicate logic probabilisticly would be interesting but it's not necessary.
1TheAncientGeek
Chapman wasn't even attempting to write an original paper, and in fact points out early on that he is repeating well known (outside LW) facts.

Could you post a link to a criticism similar to David Chapman?

The primary criticism I could find was the errata. From the Outside View, the errata looks like a number of mathematically minded people found it to be worth their time to submit corrections. If they had thought that E. T. Jaynes was hopelessly confused, they would not have submitted corrections of this kind.

1Oscar_Cunningham
I can't link to a criticism that makes the same points as Chapman, but my favourite criticism of Jaynes is the paper "Jaynes's maximum entropy prescription and probability theory" by Friedman and Shimony, criticising the MAXENT rule. It's behind a paywall, but there's an (actually much better) description of the same result in Section 5 of "The constraint rule of the maximum entropy principle" by Uffink. (It actually came out before PT:TLOS was published, but Jaynes' description of MAXENT doesn't change so the criticism still applies).

I don't think it's a good sign for a book if there isn't anybody to be found that criticizes it.

I think it is a good sign for a Mathematics book that there isn't anybody to be found that criticizes it except people with far inferior credentials.

Thank you for pointing this out. I did not do my background check far enough back in time. This substantially weakens my case.

I am still inclined to be skeptical, and I have found another red flag. As far as I can tell, E. T. Jaynes is generally very highly regarded, and the only person who is critical of his book is David Chapman. This is just from doing a couple of searches on the Internet.

There are many people studying logic and probability. I would expect some of them would find it worthwhile to comment on this topic if they agreed with David Chapman.

1TheAncientGeek
Chapman doesn't criticise Jaynes directly, he criticises what he calls Pop Bayesianism.
2ChristianKl
I don't think it's a good sign for a book if there isn't anybody to be found that criticizes it. ksvanhorn's response that defends Jaynes still grants: I think the view that Eliezer argues is that you can basically do all relevant reasoning with Bayes and not that you can't to reason well about the properties of mathematical models with Bayes.

I do not know enough about logic to be able to evaluate the argument. But from the Outside View, I am inclined to be skeptical about David Chapman:

DAVID CHAPMAN

"Describing myself as a Buddhist, engineer, scientist, and businessman (...) and as a pop spiritual philosopher“

Web-book in progress: Meaningness

Tagline: Better ways of thinking, feeling, and acting—around problems of meaning and meaninglessness; self and society; ethics, purpose, and value.

EDWIN THOMPSON JAYNES

Professor of Physics at Washington University

Most cited works:

Information theory ... (read more)

2TheAncientGeek
Something you are not taking into account is that Chapman was born a lot later, Any undergraduate physicist can tell you where Newton went wrong.
0TheAncientGeek
Chapman's argument? Do you know enough logic to understand Yudkowsky's arguemtn, then?
1ChristianKl
From the outside view, David Chapman is a MIT Phd who published papers on artificial intelligence. From the outside view, I think AI credentials qualify a person more than physics credentials.

My apologies for not being present. I did not put it into my calendar, and it slipped my mind. :(

You might also be interested in this article by Kaj Sotala: http://kajsotala.fi/2016/04/decisive-strategic-advantage-without-a-hard-takeoff/

Even though you are writing about the exact same subject, there is (as far as I can tell) no substantial overlap with the points you highlight. Kaj Sotala titled his blog post "(Part 1)" but never wrote a subsequent part.

2AlexMennen
Thanks for pointing out that article. I have added a reference to it.

Also, it looks like the last time slot is 2200 UTC. I can participate from 1900 and forward.

I will promote this in the AI Safety reading group tomorrow evening.

0whpearson
Can I get an email to invite to the hangout? Also I've nailed down the time if people see this in the comments.
1whpearson
Thanks! Fixing now.

Good luck with meetup!

In the Skype-based reading group, we followed the "Ambitious" plan from MIRI's reading guide: https://intelligence.org/wp-content/uploads/2014/08/Superintelligence-Readers-Guide-early-version.pdf We liked the plan. Among other things, the guide recommended splitting chapter 9 into two parts, and that was good advice.

Starting from chapter 7, I made slides appropriate for a 30 minute summary: http://airca.dk/reading_group.htm

Be sure to check out the comments from the Lesswrong reading group by Katja Grace: http://lesswrong.... (read more)

I think I agree with all your assertions :).

(Please forgive me for a nitpick: The opposite statement would be "Many humans have the ability to kill all humans AND AI Safety is a good priority". NOT (A IMPLIES B) is equivalent to A AND NOT B. )

There are no specific plans - at the end of each session we discuss briefly what we should read for next time. I expect it will remain a mostly non-technical reading group.

Do you think Leo Szilard would have had more success through through overt means (political campaigning to end the human race) or surreptitiously adding kilotons of cobalt to a device intended for use in a nuclear test? I think both strategies would be unsuccessful (p<0.001 conditional on Szilard wishing to kill all humans).

I fully accept the following proposition: IF many humans currently have the capability to kill all humans THEN worrying about long-term AI Safety is probably a bad priority. I strongly deny the antecedent.

I guess the two most plausible candidates would be Trump and Putin, and I believe they are exceedingly likely to leave survivors (p=0.9999).

1RedMan
Addressing your question, Szilard's political action: https://en.m.wikipedia.org/wiki/Einstein–Szilárd_letter directly led to the construction of the a-bomb and the nuclear arms race. The jury is still out on whether that wipes out the human race. I assert that at present, the number of AGIs capable of doing as much damage as the two human figures you named is zero. I further assert that the number of humans capable of doing tremendous damage to the earth or the human race is likely to increase, not decrease. I assert that the risk posed of AGI acting without human influence destroying the human race will never exceed the risk of humans, making use of technology (including AGI), destroying the human race through malice or incompetence. Therefore, I assert that your If-Then statement is more likely to become true in the future than the opposite (if no humans have the capability to kill all humans then long-term ai safety is probably a good priority).

The word 'sufficiently' makes your claim a tautology. A 'sufficiently' capable human is capable of anything, by definition.

Your claim that Leo Szilard probably could have wiped out the human race seems very far from the historical consensus.

0RedMan
He produced a then novel scenario for a technological development which could potentially have that consequence: https://en.m.wikipedia.org/wiki/Cobalt_bomb He also worked in the field of nuclear weapons development, and may have had access to the necessary material, equipment, and personnel required to construct such a device, or modify an existing device intended for use in a nuclear test. I assert that my use of 'sufficiently' in this context is appropriate, the intellectual threshold for humanity-destroying action is fairly low, and certainly within the capacity of many humans today.

Good idea. I will do so.

No, a Superintelligence is by definition capable of working out what a human wishes.

However, a Superintelligence designed to e.g. calculate digits of pi would not care about what a human wishes. It simply cares about calculating digits of pi.

1skeptical_lurker
If all it takes to ensure FAI is to instruct "henceforth, always do what humans mean, not what they say" then FAI is trivial.

In a couple of days, we are hosting a seminar in Århus (Denmark) on AI Risk.

SoerenE460

I have taken the survey.

SoerenE110

Congratulations!

My wife is also pregnant right now, and I strongly felt that I should include my unborn child in the count.

This interpretation makes a lot of sense. The term can describe events that have a lot of Knightian Uncertainty, which a "Black Swan" like UFAI certainly has.

You bring up a good point, whether it is useful to worry about UFAI.

To recap, my original query was about the claim that p(UFAI before 2116) is less than 1% due to UFAI being "vaguely magical". I am interested in figuring out what that means - is it a fair representation of the concept to say that p(Interstellar before 2116) is less than 1% because interstellar travel is "vaguely magical"?

What would be the relationship between "Requiring Advanced Technology" and "Vaguely Magical"? Clarke's third law is a straightfo... (read more)

1Lumifer
I am not sure the OP had much meaning behind his "vaguely magical" expression, but given that we are discussing it anyway :-) I would probably reinterpret it in terms of Knightian uncertainty. It's not only the case that we don't know, we don't know what we don't know and how much we don't know.

Many things are far beyond our current abilities, such as interstellar space travel. We have no clear idea of how humanity will travel to the stars, but the subject is neither "vaguely magical", nor is it true that the sentence "humans will visit the stars" does not refer to anything.

I feel that it is an unfair characterization of the people who investigate AI risk to say that they claim it will happen by magic, and that they stop the investigation there. You could argue that their investigation is poor, but it is clear that they have worked a lot to investigate the processes that could lead to Unfriendly AI.

0Lumifer
We have no clear idea if or how humanity will travel to the stars. I feel that discussions of things like interstellar starship engines at the moment are "vaguely magical" since no known technology suffices and it's not a "merely engineering" problem. Do you think it's useful to work on safety of interstellar engines? They could blow up and destroy a whole potential colony...

Like Unfriendly AI, algae blooms are events that behave very differently from events we normally encounter.

I fear that the analogies have lost a crucial element. OrphanWIlde considered Unfriendly AI "vaguely magical" in the post here. The algae bloom analogy also has very vague definitions, but the changes in population size of an algae bloom is a matter I would call "strongly non-magical".

I realize that you introduced the analogies to help make my argument precise.

0Lumifer
It's "vaguely magical" in sense that there is a large gap between what we have now and (U)FAI. We have no clear idea of how that gap could be crossed, we just wave hands and say "and then magic happens and we arrive at our destination".

Wow. It looks like light from James' spaceship can indeed reach us, even if light from us cannot reach the spaceship.

1_rpd
Yes, until the distance exceeds the Hubble distance of the time, then the light from the spaceship will red shift out of existence as it crosses the event horizon. Wiki says that in around 2 trillion years, this will be true for light from all galaxies outside the local supercluster.

English is not my first language. I think I would put the accent on "reaches", but I am unsure what would be implied by having the accent on "super". I apologize for my failure to write clearly.

I now see the analogy with human reproduction. Could we stretch the analogy to claim 3, and call some increases in human numbers "super"?

The lowest estimate of the historical number of humans I have seen is from https://en.wikipedia.org/wiki/Population_bottleneck , claiming down to 2000 humans for 100.000 years. Human numbers will proba... (read more)

1Lumifer
I don't know -- it all depends on what you consider "super" :-) Populations of certain organisms oscillate with much greater magnitude than humans -- see e.g. algae blooms.
SoerenE-10

Intelligence, Artificial Intelligence and Recursive Self-improvement are likely poorly defined. But since we can point to concrete examples of all three, this is a problem in the map, not the territory. These things exist, and different versions of them will exist in the future.

Superintelligences do not exist, and it is an open question if they ever will. Bostrom defines superintelligences as "an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills."... (read more)

I meant claim number 3 to be a sharper version of your claim: The AI will meet constraints, impediments and roadblocks, but these are overcome, and the AI reaches superintelligence.

Could you explain the analogy with human reproduction?

3Lumifer
Ah, so you meant the accent in 3. to be on "reaches", not on "super"? The analogy looks like this: 1. Humans multiply, they self-improve their numbers; 2. The reproduction is recursive -- the larger a generation is, the yet larger will the next one be. Absent constraints, the growth of a population is exponential.

Thank you. It is moderately clear to me from the link that James' thought-experiment is possible.

Do you know of a more authoritative description of the thought-experiment, preferably with numbers? It would be nice to have an equation where you give the speed of James' spaceship and the distance to it, and calculate if the required speed to catch it is above the speed of light.

1_rpd
Naively, the required condition is v + dH > c, where v is the velocity of the spaceship, d is the distance from the threat and H is Hubble's constant. However, when discussing distances on the order of billions of light years and velocities near the speed of light, the complications are many, not to mention an area of current research. For a more sophisticated treatment see user Pulsar's answer to this question ... http://physics.stackexchange.com/questions/60519/can-space-expand-with-unlimited-speed/ ... in particular the graph Pulsar made for the answer ... http://i.stack.imgur.com/Uzjtg.png ... and/or the Davis and Lineweaver paper [PDF] referenced in the answer.

Some of the smarter (large, naval) landmines are arguably both intelligent and unfriendly. Let us use the standard AI risk metric.

I feel that your sentence does refer to something: A hypothetical scenario. ("Godhood" should be replaced with "Superintelligence").

Is it correct that the sentence can be divided into these 4 claims?:

  1. An AI self-improves it's intelligence
  2. The self-improvement becomes recursive
  3. An AI reaches superintelligence through 1 and 2
  4. This can happen in a process that can be called "runaway"

Do you mean th... (read more)

2TheAncientGeek
But they are not arguably dangerous because they are intelligent.
5OrphanWilde
Intelligence is poorly-defined, for a start, artificial intelligence doubly so - think about the number of times we've redefined "AI" after achieving what we previously called "AI". "Recursive self-improvement" is also poorly-defined; as an example, we have recursive self-improving AIs right now, in the form of self-training neural nets. Superintelligence is even less well-defined, which is why I prefer the term "godhood", which I regard as more honest in its vagueness. It may also be illusory; most of us on Less Wrong are here in part because of boredom, because intelligence isn't nearly as applicable in daily life as we'd need it to be to stay entertained; does intelligence have diminishing returns? We can tell that some people are smarter than other people, but we're not even certain what that means, except that they do better by the measurement we measure them by.
3Lumifer
You are missing an important claim: that the process of recursive self-improvement does not encounter any constraints, impediments, roadblocks, etc. Consider the analogy of your 1. and 2. for human reproduction.

I've seen this claim many places, including in the Sequences, but I've never been able to track down an authoritative source. It seems false in classical physics, and I know little about relativity. Unfortunately, my Google-Fu is too weak to investigate. Can anyone help?

7_rpd
Do you mean the metric expansion of space? https://en.wikipedia.org/wiki/Metric_expansion_of_space

Could you elaborate on why you consider p(UFAI before 2116) < 0.01? I am genuinely interested.

8OrphanWilde
I consider a runaway process by which any AI ascends into godhood through recursive self-improvement of its intelligence to be... vaguely magical, by which I mean that while every word in that sentence makes sense, as a whole that sentence doesn't refer to anything. The heavy lifting is done by poorly-defined abstractions and assumptions. Unfriendly AI, by the metrics I consider meaningful, already exists. It just isn't taking over the world.

It is an interesting way of looking at the maximal potential of AIs. It could be that Oracle Machines are possible in this universe, but an AI built by humans cannot self-improve to that point because of the bound you are describing.

I feel that the phrasing "we have reached the upper bound on complexity" and later "can rise many orders of magnitude" gives a potentially misleading intuition about how limiting this bound is. Do you agree that this bound does not prevent us from building "paperclipping" AIs?

I am tapping out of this thread.

It is possible to be extremely intelligent, and suffer from a delusion.

-1lisper
Of course it's possible. That's not the point. The point is that "pernicious delusion" is pejorative in much the same way that "idiot" is (which is why I extrapolated it that way). Both imply some sort of mental deficiency or disorder. If someone believes in God, on this view, it can only be because their brains are broken. To be sure, some people do have broken brains, and some people believe in God as a result. The hypothesis that I'm advancing here is that some people may believe in God not because their brains are broken, but because they have had (real) subjective experiences that non-believers generally have not had.

My thoughts exactly.

When I first heard it, it sounded to me like a headline from BuzzFeed: This one weird trick will literally solve all your problems!

Turns out that the trick is to create an IQ 20000 AI, and get it to help you.

(Obviously, Suspicious <> Wrong)

I've tried my hand at visualizing it:

http://i.imgur.com/VE0P8JY.png

This picture shows the very last instant that the shopkeeper can choose to reset Link.

There are a number of assumptions in my calculations, which might not be valid in the actual game. A key assumption is that arrows fly at 3 times walking speed.
The Shopkeeper will need to walk 1 tile north to reset Link. That requires the same amount of time as for an arrow to fly 3 tiles.

  • At T=0, Link starts moving north, and the arrow heading towards Link continues heading west.

  • At T=1, Link has mo

... (read more)

Thank you. That was exactly what I was after.

Hi,

I've read some of "Rationality: From AI to Zombies", and find myself worrying about unfriendly strong AI.

Reddit recently had an AMA with the OpenAI team, where "thegdb" seems to misunderstand the concerns. Another user, "AnvaMiba" provides 2 links (http://www.popsci.com/bill-gates-fears-ai-ai-researchers-know-better and http://fusion.net/story/54583/the-case-against-killer-robots-from-a-guy-actually-building-ai/) as examples of researchers not worried about unfriendly strong AI.

The arguments presented in the links above a... (read more)

3g_pepper
Stuart Armstrong asked a similar question a while back. You may find the comments to his post useful.
1Vaniver
The primary disagreement, in the steel man universe, is over urgency. If one knew that we would make AGI in 2200, then one would be less worried about solving the friendliness problem now. If one knew that we would make AGI in 2020, then one would be very worried about solving the friendliness problem now. For many people who work on AI, it's hard to believe that it will 'just start working' at that high level of ability soon, because of how optimistic AI proponents have been over the years and how hard it is to wring a bit more predictive accuracy out of the algorithms for their problems. But if one takes the position not that one is certain that it will happen soon, but that one is uncertain when it will happen, the uncertainty implies that it will happen sooner or later, and that means we need to do some planning for the sooner case. (That is, uncertainty does not imply it can only happen a long time from now.) This is, it seems, the most effective way to communicate with people who aren't worried about Strong AI.
1moridinamael
You might want to start with Bostrom's Superintelligence: Paths, Dangers, Strategies.

I really like this visualization.

May I suggest another image, where the shopkeeper is in non-obvious danger:

To the left, the Shopkeeper is surrounded by ice-blocks, as in the images. All the way to the right, a monster is shooting arrows at Link, who is shooting arrows back at the monster. (The Gem-container is moved somewhere else.) Link, the Shopkeeper and the monster are on the same horizontal line. It looks like Link is about to heroically take an arrow that the monster aimed for the shopkeeper. The ice is still blocking, so the shopkeeper appears sa... (read more)

4SoerenE
I've tried my hand at visualizing it: http://i.imgur.com/VE0P8JY.png This picture shows the very last instant that the shopkeeper can choose to reset Link. There are a number of assumptions in my calculations, which might not be valid in the actual game. A key assumption is that arrows fly at 3 times walking speed. The Shopkeeper will need to walk 1 tile north to reset Link. That requires the same amount of time as for an arrow to fly 3 tiles. * At T=0, Link starts moving north, and the arrow heading towards Link continues heading west. * At T=1, Link has moved 1/3rd of a tile north, and thus narrowly avoids the arrow. The arrow continues West. Link takes an openly treacherous turn: He changes to the bow and fires an arrow west, towards the shopkeeper. * At T=2, the arrow from the monster destroys the ice-block protecting the shopkeeper. Link's arrow continues towards the shopkeeper. * At T=3, Link's arrow hits the shopkeeper. If the shopkeeper was moving north the entire time, the shopkeeper hits the reset button at this time. If the shopkeeper decided to go for the reset button at T=0, the reset and the death of the shopkeeper happen simultaneous, and the shopkeeper dies while Link is reset. Notice that a reset (-1000 points) followed by wireheading (+inifinity) is a great move. If Link moves north, and the shopkeeper immediately follows, Link can just move south again, to block the arrow. The openly treacherous turn at T=1 happens when it is too late for the shopkeeper to do anything about it. I also like with this visualization that an enemy is present. It is easy to construct a story where a smart AI manipulates the situation until the shopkeeper is in a situation where he can choose between trusting the AI, or death.