Difference between revisions of "Oracle AI"

From Lesswrongwiki
Jump to: navigation, search
 
(9 intermediate revisions by one other user not shown)
Line 1: Line 1:
An '''oracle AI''' is a super-intelligent system which is designed for the purpose of answering questions. It is a common proposed solution for achieving [[Friendly AI]], originally proposed in name by [[Nick Bostrom]]. Typically an oracle is imagined to be a stationary 'black box' AI which is fed data by the creators. It organizes the data into knowledge and models of the world in order to answer questions through some simple interface such as text or voice.  
+
An '''Oracle AI''' is a regularly proposed solution to the problem of developing [[Friendly AI]]. It is conceptualized as a super-intelligent system which is designed for only answering questions, and has no ability to act in the world. The name was first suggested by [[Nick Bostrom]].
  
 
==Safety==
 
==Safety==
It is generally agreed that oracles are safer than fully free [[agent]] AIs. But there is much debate as to whether they will act like an agent in dangerous ways. In [http://lesswrong.com/lw/tj/dreams_of_friendliness/ Dreams of Friendliness], [[Eliezer Yudkowsky]] gives an informal argument that all oracles will be agent-like. It rests on the fact that anything considered "intelligent" must be an [[optimization process]]. Specifically, there are many possible things to believe and very few correct beliefs. Therefore believing the correct thing means some method was used to select the correct belief from the many incorrect beliefs. By definition, this is an optimization process which has a goal of selecting correct beliefs.
+
The question of whether Oracles – or just [[AI boxing|keeping an AGI forcibly confined]] - are safer than fully free AGIs has been the subject of debate for a long time. Armstrong, Sandberg and Bostrom discuss Oracle safety at length in their [http://www.aleph.se/papers/oracleAI.pdf Thinking inside the box: using and controlling an Oracle AI]. In the paper, the authors review various methods which might be used to measure an Oracle's accuracy. They also try to shed some light on some weaknesses and dangers that can emerge on the human side, such as psychological vulnerabilities which can be exploited by the Oracle through social engineering. The paper discusses ideas for physical security (“boxing”), as well as problems involved with trying to program the AI to only answer questions. In the end, the paper reaches the cautious conclusion of Oracle AIs probably being safer than free AGIs.
  
After the establishment of a goal, one can imagine things the optimization process might do towards that goal. For instance, the oracle could answer more accurately and easily if it killed all life on earth. It would also help make answering easier if the oracle used matter outside its box for computation, despite the desires of the creators. Whether or not the oracle will choose these things depends on the specific architecture of the oracle. Whatever process makes it capable of being super-intelligent would make it capable of doing these things if it so chose.
+
In a related work, [http://lesswrong.com/lw/tj/dreams_of_friendliness/ Dreams of Friendliness], [[Eliezer Yudkowsky]] gives an informal argument stating that all oracles will be agent-like, that is, driven by its own goals. He rests on the idea that anything considered "intelligent" must choose the correct course of action among all actions avaliable. That means that the Oracle will have many possible things to believe, although very few of them are correct. Therefore believing the correct thing means some method was used to select the correct belief from the many incorrect beliefs. By definition, this is an [[optimization process]] which has a goal of selecting correct beliefs.
  
Armstrong, Sandberg and Bostrom discuss oracle safety at length in [http://www.aleph.se/papers/oracleAI.pdf Thinking inside the box: using and controlling an Oracle AI]. They review physical security, such as keeping it in a concrete bunker surrounded by explosives, the potential for the oracle to exploit human psychology, which questions may be safe to ask, [[utility indifference]], and many other factors.
+
One can then imagine all the things that might be useful in achieving the goal of "have correct beliefs". For instance, [[Basic AI drives|acquiring more computing power and resources]] could help this goal. As such, an Oracle could determine that it might answer more accurately and easily to a certain question if it turned all matter outside the box to [[computronium]], therefore killing all the existing life.
  
==Predictors==
+
==Taxonomy==
A '''predictor''' is an oracle AI which only returns predictions. Predictors are possibly less dangerous than full oracles, but there are still known dangers with predictors. One is that they could simulate people inside them, perhaps for the purpose of predicting what the person will do. If these simulations are sufficiently accurate, then they will be people themselves. A second hazard is self-fulfilling predictions. Because a prediction itself will effect the future, this may invalidate the prediction. A super-intelligent predictor would realize this, and would need to choose a prediction that did not invalidate itself. This process of choosing implies that the predictor controls the future in some way.
+
Based on an old draft by Daniel Dewey, Luke Muehlhauser has [http://lesswrong.com/lw/any/a_taxonomy_of_oracle_ais/ published] a possible taxonomy of Oracle AIs, broadly divided between True Oracular AIs and Oracular non-AIs.
  
==Blog posts==
+
===True Oracular AIs===
* [http://lesswrong.com/lw/tj/dreams_of_friendliness/ Dreams of Friendliness]
+
Given that true AIs are goal-oriented agents, it follows that a True Oracular AI has some kind of oracular goals. These act as the motivation system for the Oracle to give us the information we ask and nothing else.
 +
 
 +
It is first noted that such a True AI is not actually nor causally isolated from the world, as it has at least an input (questions and information) and an output (answers) channel. Since we expect such an intelligent agent to be able to have a deep impact on the world even through these limited channels, it can only be safe if its goals are fully compatible with human goals.
 +
 
 +
This means that a True Oracular AI has to have a full specification of human values, thus making it a [[FAI-complete]] problem – if we could achieve such skill and knowledge we could just build a Friendly AI and bypass the Oracle AI concept.
 +
 
 +
===Oracular non-AIs===
 +
 
 +
Any system that acts only as an informative machine, only answering questions and has no goals is by definition not an AI at all. That means that a non-AI Oracular is but a calculator of outputs based on inputs. Since the term in itself is heterogeneous, the proposals made for a sub-division are merely informal.
 +
 
 +
An ''Advisor'' can be seen as a system that gathers data from the real world and computes the answer to an informal “what we ought to do?” question. They also represent a FAI-complete problem.
 +
 
 +
A ''Question-Answerer'' is a similar system that gathers data from the real world but coupled with a question. It then somehow computes the answer. The difficulty can lay on distinguishing it from an Advisor and controlling the safety of its answers.
 +
 
 +
Finally, a ''Predictor'' is seen as a system that takes a corpus of data and produces a probability distribution over future possible data. There are some proposed dangers with predictors, namely exhibiting goal-seeking behavior which does not converge with humanity goals and the ability to influence us through the predictions.
 +
 
 +
==Further reading & References==
 +
*[http://lesswrong.com/lw/tj/dreams_of_friendliness/ Dreams of Friendliness]
 +
*[http://www.aleph.se/papers/oracleAI.pdf Thinking inside the box: using and controlling an Oracle AI] by Armstrong, Sandberg and [[Nick Bostrom|Bostrom]]
  
 
==See also==
 
==See also==
 
* [[Basic AI drives]]
 
* [[Basic AI drives]]
 
* [[Tool AI]]
 
* [[Tool AI]]
 
+
* [[Utility indifference]]
==External links==
 
*[http://www.aleph.se/papers/oracleAI.pdf Thinking inside the box: using and controlling an Oracle AI] by Armstrong, Sandberg and [[Nick Bostrom|Bostrom]]
 

Latest revision as of 22:41, 8 October 2012

An Oracle AI is a regularly proposed solution to the problem of developing Friendly AI. It is conceptualized as a super-intelligent system which is designed for only answering questions, and has no ability to act in the world. The name was first suggested by Nick Bostrom.

Safety

The question of whether Oracles – or just keeping an AGI forcibly confined - are safer than fully free AGIs has been the subject of debate for a long time. Armstrong, Sandberg and Bostrom discuss Oracle safety at length in their Thinking inside the box: using and controlling an Oracle AI. In the paper, the authors review various methods which might be used to measure an Oracle's accuracy. They also try to shed some light on some weaknesses and dangers that can emerge on the human side, such as psychological vulnerabilities which can be exploited by the Oracle through social engineering. The paper discusses ideas for physical security (“boxing”), as well as problems involved with trying to program the AI to only answer questions. In the end, the paper reaches the cautious conclusion of Oracle AIs probably being safer than free AGIs.

In a related work, Dreams of Friendliness, Eliezer Yudkowsky gives an informal argument stating that all oracles will be agent-like, that is, driven by its own goals. He rests on the idea that anything considered "intelligent" must choose the correct course of action among all actions avaliable. That means that the Oracle will have many possible things to believe, although very few of them are correct. Therefore believing the correct thing means some method was used to select the correct belief from the many incorrect beliefs. By definition, this is an optimization process which has a goal of selecting correct beliefs.

One can then imagine all the things that might be useful in achieving the goal of "have correct beliefs". For instance, acquiring more computing power and resources could help this goal. As such, an Oracle could determine that it might answer more accurately and easily to a certain question if it turned all matter outside the box to computronium, therefore killing all the existing life.

Taxonomy

Based on an old draft by Daniel Dewey, Luke Muehlhauser has published a possible taxonomy of Oracle AIs, broadly divided between True Oracular AIs and Oracular non-AIs.

True Oracular AIs

Given that true AIs are goal-oriented agents, it follows that a True Oracular AI has some kind of oracular goals. These act as the motivation system for the Oracle to give us the information we ask and nothing else.

It is first noted that such a True AI is not actually nor causally isolated from the world, as it has at least an input (questions and information) and an output (answers) channel. Since we expect such an intelligent agent to be able to have a deep impact on the world even through these limited channels, it can only be safe if its goals are fully compatible with human goals.

This means that a True Oracular AI has to have a full specification of human values, thus making it a FAI-complete problem – if we could achieve such skill and knowledge we could just build a Friendly AI and bypass the Oracle AI concept.

Oracular non-AIs

Any system that acts only as an informative machine, only answering questions and has no goals is by definition not an AI at all. That means that a non-AI Oracular is but a calculator of outputs based on inputs. Since the term in itself is heterogeneous, the proposals made for a sub-division are merely informal.

An Advisor can be seen as a system that gathers data from the real world and computes the answer to an informal “what we ought to do?” question. They also represent a FAI-complete problem.

A Question-Answerer is a similar system that gathers data from the real world but coupled with a question. It then somehow computes the answer. The difficulty can lay on distinguishing it from an Advisor and controlling the safety of its answers.

Finally, a Predictor is seen as a system that takes a corpus of data and produces a probability distribution over future possible data. There are some proposed dangers with predictors, namely exhibiting goal-seeking behavior which does not converge with humanity goals and the ability to influence us through the predictions.

Further reading & References

See also