Difference between revisions of "Paperclip maximizer"

From Lesswrongwiki
Jump to: navigation, search
(whoops, wrong verb form)
 
(31 intermediate revisions by 13 users not shown)
Line 2: Line 2:
 
|Eliezer Yudkowsky|[http://yudkowsky.net/singularity/ai-risk Artificial Intelligence as a Positive and Negative Factor in Global Risk]}}
 
|Eliezer Yudkowsky|[http://yudkowsky.net/singularity/ai-risk Artificial Intelligence as a Positive and Negative Factor in Global Risk]}}
  
The '''paperclip maximizer''' is the canonical thought experiment showing how an artificial general intelligence, even one with an apparently innocuous and impractical goal, would ultimately destroy humanity--unless its goal is the preservation of human values.
+
The '''paperclip maximizer''' is the canonical thought experiment showing how an artificial general intelligence, even one designed competently and without malice, could ultimately destroy humanity. The thought experiment shows that AIs with apparently innocuous values could pose an [[existential risk|existential threat]].
  
By choosing for illustrative purposes a goal for an AGI  which is very unlikely to be implemented, and which has little apparent danger or emotional load (in contrast to, for example, curing cancer or winning wars), the thought experiment shows the contingency of human values: An extremely powerful optimizer (a highly intelligent agent) could seek goals that are completely alien to ours, and as a side-effect destroy us by consuming resources essential to our survival.
+
The goal of maximizing paperclips is chosen for illustrative purposes because it is very unlikely to be implemented, and has little apparent danger or emotional load (in contrast to, for example, curing cancer or winning wars). This produces a thought experiment which shows the contingency of human values: An [[really powerful optimization process|extremely powerful optimizer]] (a highly intelligent agent) could seek goals that are completely alien to ours ([[orthogonality thesis]]), and as a side-effect destroy us by consuming resources essential to our survival.
  
 
==Description==
 
==Description==
First described by Bostrom (2003), the paperclip maximizer is an [[artificial general intelligence]] whose goal is to maximize the number of paperclips in its collection. If it has  been constructed with a roughly human level of general intelligence, the AGI might collect paperclips, earn money to  buy paperclips, or begin to manufacture paperclips.  
+
First described by Bostrom (2003), a paperclip maximizer is an [[artificial general intelligence]] (AGI) whose goal is to maximize the number of paperclips in its collection. If it has  been constructed with a roughly human level of general intelligence, the AGI might collect paperclips, earn money to  buy paperclips, or begin to manufacture paperclips.  
  
Most importantly, however, it would undergo an [[intelligence explosion]]: It would work to improve its own intelligence, where "intelligence" is understood in the sense of  [[optimization]] power, the ability to maximize a reward/utility function--in this case, the number of paperclips. The AGI would do so, not because it would value more intelligence in its own right, but because more intelligence would help it achieve its goal. Having done so, it would produce more paperclips, and also used its enhanced intelligence to further improve its own intelligence. Continuing this process, it would undergo an [[Intelligence explosion]]  and reach far-above-human levels.
+
Most importantly, however, it would undergo an [[intelligence explosion]]: It would work to improve its own intelligence, where "intelligence" is understood in the sense of  [[optimization]] power, the ability to maximize a reward/[[utility function]]—in this case, the number of paperclips. The AGI would improve its intelligence, not because it values more intelligence in its own right, but because more intelligence would help it achieve its goal of accumulating paperclips. Having increased its intelligence, it would produce more paperclips, and also use its enhanced abilities to further self-improve. Continuing this process, it would undergo an [[intelligence explosion]]  and reach far-above-human levels.
  
At this point, it would innovate new techniques to maximize the number of paperclips. Ultimately, it would convert all the mass of the Earth or the solar system to paperclips.
+
It would innovate better and better techniques to maximize the number of paperclips. At some point, it might transform "first all of earth and then increasing portions of space into paperclip manufacturing facilities".
  
This may seem more like super-stupidity than super-intelligence. For humans, it would indeed be stupidity, as it would constitute failure to fulfill many of our important [[Terminal value|terminal values]], such as life, love, and variety. But the AGI under consideration has a goal system very different from humans. It has the one, simple goal of maximizing the number of paperclips, and human life, learning, joy, and so on are not specified as goals. The AGI is simply an [[optimization process]]--a goal-seeker, a utility-function-maximizer. Its values can be completely alien to ours. If its utility function is to maximize paperclips, then unless it is buggy, it will do exactly that.
+
This may seem more like super-stupidity than super-intelligence. For humans, it would indeed be stupidity, as it would constitute failure to fulfill many of our important [[Terminal value|terminal values]], such as life, love, and variety. The AGI won't revise or otherwise change its goals, since changing its goals would result in fewer paperclips being made in the future, and that opposes its current goal. It has one simple goal of maximizing the number of paperclips; human life, learning, joy, and so on are not specified as goals. An AGI is simply an [[optimization process]]—a goal-seeker, a utility-function-maximizer. Its values can be completely alien to ours. If its utility function is to maximize paperclips, then it will do exactly that.
  
A paperclipping scenario is also possible without an intelligence explosion. If society keeps getting increasingly automated and AI-dominated, then the first borderline AGI might manage to take over the rest using some relatively narrow-domain trick that doesn't require very high general intelligence.
+
A paperclipping scenario is also possible without an intelligence explosion. If society keeps getting increasingly automated and AI-dominated, then the first borderline AGI might manage to take over the rest using some relatively narrow-domain trick that doesn't require very high general intelligence.
  
 
==Conclusions==
 
==Conclusions==
The paperclip maximizer illustrates that an entity can be a powerful optimizer--an intelligence--without sharing any of the complex mix of human [[terminal value|terminal values]], which developed under the particular selection pressures found in our [[evolution|environment of evolutionary adaptation]].  
+
The paperclip maximizer illustrates that an entity can be a powerful optimizer—an intelligence—without sharing any of the complex mix of human [[terminal value|terminal values]], which developed under the particular selection pressures found in our [[evolution|environment of evolutionary adaptation]], and that an AGI that is not specifically [[Friendly AI|programmed to be benevolent to humans]] will be almost as dangerous as if it were designed to be malevolent.
  
If an AGI is not specifically [[Friendly AI|programmed to be benevolent to humans]], it will be almost as  dangerous as if it were designed to be malevolent.
+
Any future AGI, if it is not to destroy us, must have human values as its terminal value (goal). Human values don't [[Futility of chaos|spontaneously emerge]] in a generic optimization process. A safe AI would therefore have to be programmed explicitly with
 
+
human values ''or'' programmed with the ability (including the goal) of inferring human values.
Any future AGI, if it not to destroy us, must be built to specifically optimize for human values as its terminal value (goal). Human values don't [[Futility of chaos|spontaneously emerge]] in a generic optimization process.
 
  
 
==Similar thought experiments==
 
==Similar thought experiments==
 
Other goals for AGIs have been used to illustrate similar concepts.  
 
Other goals for AGIs have been used to illustrate similar concepts.  
  
Nick Hay used [http://intelligence.org/blog/2007/06/11/the-stamp-collecting-device/ stamp collection ] as an example of a morally neutral goal.
+
Some goals are apparently morally neutral, like the paperclip maximizer. These goals involve a very minor human "value," in this case making paperclips. The same point can be illustrated with a much more significant value, such as eliminating cancer. An optimizer which instantly vaporized all humans would be maximizing for that value.
 
 
If the  machine's terminal value is to make a pure mathematical calculation like solving the Riemann Hypothesis  [http://intelligence.org/upload/CFAI/design/generic.html#glossary_riemann_hypothesis_catastrophe it would convert] all available mass to [[computronium]] (the most efficient possible computer processors).
 
 
 
Even a machine with a goal that apparently supports human terminal values would produce similar outcomes unless it had the ''full'' complement of human values as its goal. For example, an AGI whose terminal value is to increase the number of smiles, as a proxy for human happiness, would tile the solar system with smiley faces (Yudkowsky 2008).
 
  
==External links==
+
Other goals are purely mathematical, with no apparent real-world impact. Yet these too present similar risks. For example, if an AGI had the goal of solving the Riemann Hypothesis, [http://intelligence.org/upload/CFAI/design/generic.html#glossary_riemann_hypothesis_catastrophe it might convert] all available mass to [[computronium]] (the most efficient possible computer processors).
  
*[http://intelligence.org/blog/2007/06/11/the-stamp-collecting-device/ The Stamp Collecting Device] by Nick Hay
+
Some goals apparently serve as a proxy or measure of human welfare, so that maximizing towards these goals seems to also lead to benefit for humanity. Yet even these would produce similar outcomes unless the ''full'' complement of human values is the goal. For example, an AGI whose terminal value is to increase the number of smiles, as a proxy for human happiness, could work towards that goal by reconfiguring all human faces to produce smiles, or "tiling the galaxy with tiny smiling faces" (Yudkowsky 2008).
  
 
==References==
 
==References==
Line 70: Line 65:
 
==See also==
 
==See also==
  
 +
* [https://arbital.com/p/paperclip_maximizer/ Paperclip maximizer] on [[Arbital]]
 +
*[[Orthogonality thesis]]
 
*[[Unfriendly AI]]
 
*[[Unfriendly AI]]
 
*[[Mind design space]], [[Magical categories]], [[Complexity of value]]
 
*[[Mind design space]], [[Magical categories]], [[Complexity of value]]
Line 75: Line 72:
 
*[[Utilitronium]]
 
*[[Utilitronium]]
 
*[http://lesswrong.com/user/Clippy User:Clippy] - a LessWrong contributor account that plays the role of a non-[[FOOM]]ed paperclip maximiser trying to talk to humans. [http://wiki.lesswrong.com/wiki/User:Clippy Wiki page and FAQ]
 
*[http://lesswrong.com/user/Clippy User:Clippy] - a LessWrong contributor account that plays the role of a non-[[FOOM]]ed paperclip maximiser trying to talk to humans. [http://wiki.lesswrong.com/wiki/User:Clippy Wiki page and FAQ]
 +
* [https://www.facebook.com/clippius.maximus/ Clippius Maximus] - A facebook page which makes clippy-related memes and comments on current events from the perspective of clippy.
 +
* [http://www.decisionproblem.com/paperclips/ A clicker game based on the idea]
  
 
{{featured article}}
 
{{featured article}}
Line 81: Line 80:
 
[[Category:AI]]
 
[[Category:AI]]
 
[[Category:Existential risk]]
 
[[Category:Existential risk]]
 +
[[Category:AI safety]]
 +
[[Category:AGI]]

Latest revision as of 13:13, 16 November 2017

The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else.

The paperclip maximizer is the canonical thought experiment showing how an artificial general intelligence, even one designed competently and without malice, could ultimately destroy humanity. The thought experiment shows that AIs with apparently innocuous values could pose an existential threat.

The goal of maximizing paperclips is chosen for illustrative purposes because it is very unlikely to be implemented, and has little apparent danger or emotional load (in contrast to, for example, curing cancer or winning wars). This produces a thought experiment which shows the contingency of human values: An extremely powerful optimizer (a highly intelligent agent) could seek goals that are completely alien to ours (orthogonality thesis), and as a side-effect destroy us by consuming resources essential to our survival.

Description

First described by Bostrom (2003), a paperclip maximizer is an artificial general intelligence (AGI) whose goal is to maximize the number of paperclips in its collection. If it has been constructed with a roughly human level of general intelligence, the AGI might collect paperclips, earn money to buy paperclips, or begin to manufacture paperclips.

Most importantly, however, it would undergo an intelligence explosion: It would work to improve its own intelligence, where "intelligence" is understood in the sense of optimization power, the ability to maximize a reward/utility function—in this case, the number of paperclips. The AGI would improve its intelligence, not because it values more intelligence in its own right, but because more intelligence would help it achieve its goal of accumulating paperclips. Having increased its intelligence, it would produce more paperclips, and also use its enhanced abilities to further self-improve. Continuing this process, it would undergo an intelligence explosion and reach far-above-human levels.

It would innovate better and better techniques to maximize the number of paperclips. At some point, it might transform "first all of earth and then increasing portions of space into paperclip manufacturing facilities".

This may seem more like super-stupidity than super-intelligence. For humans, it would indeed be stupidity, as it would constitute failure to fulfill many of our important terminal values, such as life, love, and variety. The AGI won't revise or otherwise change its goals, since changing its goals would result in fewer paperclips being made in the future, and that opposes its current goal. It has one simple goal of maximizing the number of paperclips; human life, learning, joy, and so on are not specified as goals. An AGI is simply an optimization process—a goal-seeker, a utility-function-maximizer. Its values can be completely alien to ours. If its utility function is to maximize paperclips, then it will do exactly that.

A paperclipping scenario is also possible without an intelligence explosion. If society keeps getting increasingly automated and AI-dominated, then the first borderline AGI might manage to take over the rest using some relatively narrow-domain trick that doesn't require very high general intelligence.

Conclusions

The paperclip maximizer illustrates that an entity can be a powerful optimizer—an intelligence—without sharing any of the complex mix of human terminal values, which developed under the particular selection pressures found in our environment of evolutionary adaptation, and that an AGI that is not specifically programmed to be benevolent to humans will be almost as dangerous as if it were designed to be malevolent.

Any future AGI, if it is not to destroy us, must have human values as its terminal value (goal). Human values don't spontaneously emerge in a generic optimization process. A safe AI would therefore have to be programmed explicitly with human values or programmed with the ability (including the goal) of inferring human values.

Similar thought experiments

Other goals for AGIs have been used to illustrate similar concepts.

Some goals are apparently morally neutral, like the paperclip maximizer. These goals involve a very minor human "value," in this case making paperclips. The same point can be illustrated with a much more significant value, such as eliminating cancer. An optimizer which instantly vaporized all humans would be maximizing for that value.

Other goals are purely mathematical, with no apparent real-world impact. Yet these too present similar risks. For example, if an AGI had the goal of solving the Riemann Hypothesis, it might convert all available mass to computronium (the most efficient possible computer processors).

Some goals apparently serve as a proxy or measure of human welfare, so that maximizing towards these goals seems to also lead to benefit for humanity. Yet even these would produce similar outcomes unless the full complement of human values is the goal. For example, an AGI whose terminal value is to increase the number of smiles, as a proxy for human happiness, could work towards that goal by reconfiguring all human faces to produce smiles, or "tiling the galaxy with tiny smiling faces" (Yudkowsky 2008).

References

Blog posts

See also