Holden Karnofsky’s Singularity Institute Objection 2

The sheer length of GiveWell co-founder and co-executive director Holden Karnofsky’s excellent critique of the Singularity Institute means that it’s hard to keep track of the resulting discussion. I propose to break out each of his objections into a separate Discussion post so that each receives the attention it deserves.

Objection 2: SI appears to neglect the potentially important distinction between “tool” and “agent” AI.

Google Maps is a type of artificial intelligence (AI). It is far more intelligent than I am when it comes to planning routes.

Google Maps—by which I mean the complete software package including the display of the map itself—does not have a “utility” that it seeks to maximize. (One could fit a utility function to its actions, as to any set of actions, but there is no single “parameter to be maximized” driving its operations.)

Google Maps (as I understand it) considers multiple possible routes, gives each a score based on factors such as distance and likely traffic, and then displays the best-scoring route in a way that makes it easily understood by the user. If I don’t like the route, for whatever reason, I can change some parameters and consider a different route. If I like the route, I can print it out or email it to a friend or send it to my phone’s navigation application. Google Maps has no single parameter it is trying to maximize; it has no reason to try to “trick” me in order to increase its utility.

In short, Google Maps is not an agent, taking actions in order to maximize a utility parameter. It is a tool, generating information and then displaying it in a user-friendly manner for me to consider, use and export or discard as I wish.

Every software application I know of seems to work essentially the same way, including those that involve (specialized) artificial intelligence such as Google Search, Siri, Watson, Rybka, etc. Some can be put into an “agent mode” (as Watson was on Jeopardy!) but all can easily be set up to be used as “tools” (for example, Watson can simply display its top candidate answers to a question, with the score for each, without speaking any of them.)

The “tool mode” concept is importantly different from the possibility of Oracle AI sometimes discussed by SI. The discussions I’ve seen of Oracle AI present it as an Unfriendly AI that is “trapped in a box”—an AI whose intelligence is driven by an explicit utility function and that humans hope to control coercively. Hence the discussion of ideas such as the AI-Box Experiment. A different interpretation, given in Karnofsky/​Tallinn 2011, is an AI with a carefully designed utility function—likely as difficult to construct as “Friendliness”—that leaves it “wishing” to answer questions helpfully. By contrast with both these ideas, Tool-AGI is not “trapped” and it is not Unfriendly or Friendly; it has no motivations and no driving utility function of any kind, just like Google Maps. It scores different possibilities and displays its conclusions in a transparent and user-friendly manner, as its instructions say to do; it does not have an overarching “want,” and so, as with the specialized AIs described above, while it may sometimes “misinterpret” a question (thereby scoring options poorly and ranking the wrong one #1) there is no reason to expect intentional trickery or manipulation when it comes to displaying its results.

Another way of putting this is that a “tool” has an underlying instruction set that conceptually looks like: “(1) Calculate which action A would maximize parameter P, based on existing data set D. (2) Summarize this calculation in a user-friendly manner, including what Action A is, what likely intermediate outcomes it would cause, what other actions would result in high values of P, etc.” An “agent,” by contrast, has an underlying instruction set that conceptually looks like: “(1) Calculate which action, A, would maximize parameter P, based on existing data set D. (2) Execute Action A.” In any AI where (1) is separable (by the programmers) as a distinct step, (2) can be set to the “tool” version rather than the “agent” version, and this separability is in fact present with most/​all modern software. Note that in the “tool” version, neither step (1) nor step (2) (nor the combination) constitutes an instruction to maximize a parameter—to describe a program of this kind as “wanting” something is a category error, and there is no reason to expect its step (2) to be deceptive.

I elaborated further on the distinction and on the concept of a tool-AI in Karnofsky/​Tallinn 2011.

This is important because an AGI running in tool mode could be extraordinarily useful but far more safe than an AGI running in agent mode. In fact, if developing “Friendly AI” is what we seek, a tool-AGI could likely be helpful enough in thinking through this problem as to render any previous work on “Friendliness theory” moot. Among other things, a tool-AGI would allow transparent views into the AGI’s reasoning and predictions without any reason to fear being purposefully misled, and would facilitate safe experimental testing of any utility function that one wished to eventually plug into an “agent.”

Is a tool-AGI possible? I believe that it is, and furthermore that it ought to be our default picture of how AGI will work, given that practically all software developed to date can (and usually does) run as a tool and given that modern software seems to be constantly becoming “intelligent” (capable of giving better answers than a human) in surprising new domains. In addition, it intuitively seems to me (though I am not highly confident) that intelligence inherently involves the distinct, separable steps of (a) considering multiple possible actions and (b) assigning a score to each, prior to executing any of the possible actions. If one can distinctly separate (a) and (b) in a program’s code, then one can abstain from writing any “execution” instructions and instead focus on making the program list actions and scores in a user-friendly manner, for humans to consider and use as they wish.

Of course, there are possible paths to AGI that may rule out a “tool mode,” but it seems that most of these paths would rule out the application of “Friendliness theory” as well. (For example, a “black box” emulation and augmentation of a human mind.) What are the paths to AGI that allow manual, transparent, intentional design of a utility function but do not allow the replacement of “execution” instructions with “communication” instructions? Most of the conversations I’ve had on this topic have focused on three responses:

  • Self-improving AI. Many seem to find it intuitive that (a) AGI will almost certainly come from an AI rewriting its own source code, and (b) such a process would inevitably lead to an “agent.” I do not agree with either (a) or (b). I discussed these issues in Karnofsky/​Tallinn 2011 and will be happy to discuss them more if this is the line of response that SI ends up pursuing. Very briefly:

    • The idea of a “self-improving algorithm” intuitively sounds very powerful, but does not seem to have led to many “explosions” in software so far (and it seems to be a concept that could apply to narrow AI as well as to AGI).

    • It seems to me that a tool-AGI could be plugged into a self-improvement process that would be quite powerful but would also terminate and yield a new tool-AI after a set number of iterations (or after reaching a set “intelligence threshold”). So I do not accept the argument that “self-improving AGI means agent AGI.” As stated above, I will elaborate on this view if it turns out to be an important point of disagreement.

    • I have argued (in Karnofsky/​Tallinn 2011) that the relevant self-improvement abilities are likely to come with or after—not prior to—the development of strong AGI. In other words, any software capable of the relevant kind of self-improvement is likely also capable of being used as a strong tool-AGI, with the benefits described above.

    • The SI-related discussions I’ve seen of “self-improving AI” are highly vague, and do not spell out views on the above points.

  • Dangerous data collection. Some point to the seeming dangers of a tool-AI’s “scoring” function: in order to score different options it may have to collect data, which is itself an “agent” type action that could lead to dangerous actions. I think my definition of “tool” above makes clear what is wrong with this objection: a tool-AGI takes its existing data set D as fixed (and perhaps could have some pre-determined, safe set of simple actions it can take—such as using Google’s API—to collect more), and if maximizing its chosen parameter is best accomplished through more data collection, it can transparently output why and how it suggests collecting more data. Over time it can be given more autonomy for data collection through an experimental and domain-specific process (e.g., modifying the AI to skip specific steps of human review of proposals for data collection after it has become clear that these steps work as intended), a process that has little to do with the “Friendly overarching utility function” concept promoted by SI. Again, I will elaborate on this if it turns out to be a key point.

  • Race for power. Some have argued to me that humans are likely to choose to create agent-AGI, in order to quickly gain power and outrace other teams working on AGI. But this argument, even if accepted, has very different implications from SI’s view.

    Conventional wisdom says it is extremely dangerous to empower a computer to act in the world until one is very sure that the computer will do its job in a way that is helpful rather than harmful. So if a programmer chooses to “unleash an AGI as an agent” with the hope of gaining power, it seems that this programmer will be deliberately ignoring conventional wisdom about what is safe in favor of shortsighted greed. I do not see why such a programmer would be expected to make use of any “Friendliness theory” that might be available. (Attempting to incorporate such theory would almost certainly slow the project down greatly, and thus would bring the same problems as the more general “have caution, do testing” counseled by conventional wisdom.) It seems that the appropriate measures for preventing such a risk are security measures aiming to stop humans from launching unsafe agent-AIs, rather than developing theories or raising awareness of “Friendliness.”

One of the things that bothers me most about SI is that there is practically no public content, as far as I can tell, explicitly addressing the idea of a “tool” and giving arguments for why AGI is likely to work only as an “agent.” The idea that AGI will be driven by a central utility function seems to be simply assumed. Two examples:

  • I have been referred to Muehlhauser and Salamon 2012 as the most up-to-date, clear explanation of SI’s position on “the basics.” This paper states, “Perhaps we could build an AI of limited cognitive ability — say, a machine that only answers questions: an ‘Oracle AI.’ But this approach is not without its own dangers (Armstrong, Sandberg, and Bostrom 2012).” However, the referenced paper (Armstrong, Sandberg and Bostrom 2012) seems to take it as a given that an Oracle AI is an “agent trapped in a box”—a computer that has a basic drive/​utility function, not a Tool-AGI. The rest of Muehlhauser and Salamon 2012 seems to take it as a given that an AGI will be an agent.

  • I have often been referred to Omohundro 2008 for an argument that an AGI is likely to have certain goals. But this paper seems, again, to take it as given that an AGI will be an agent, i.e., that it will have goals at all. The introduction states, “To say that a system of any design is an ‘artificial intelligence’, we mean that it has goals which it tries to accomplish by acting in the world.” In other words, the premise I’m disputing seems embedded in its very definition of AI.

The closest thing I have seen to a public discussion of “tool-AGI” is in Dreams of Friendliness, where Eliezer Yudkowsky considers the question, “Why not just have the AI answer questions, instead of trying to do anything? Then it wouldn’t need to be Friendly. It wouldn’t need any goals at all. It would just answer questions.” His response:

To which the reply is that the AI needs goals in order to decide how to think: that is, the AI has to act as a powerful optimization process in order to plan its acquisition of knowledge, effectively distill sensory information, pluck “answers” to particular questions out of the space of all possible responses, and of course, to improve its own source code up to the level where the AI is a powerful intelligence. All these events are “improbable” relative to random organizations of the AI’s RAM, so the AI has to hit a narrow target in the space of possibilities to make superintelligent answers come out.

This passage appears vague and does not appear to address the specific “tool” concept I have defended above (in particular, it does not address the analogy to modern software, which challenges the idea that “powerful optimization processes” cannot run in tool mode). The rest of the piece discusses (a) psychological mistakes that could lead to the discussion in question; (b) the “Oracle AI” concept that I have outlined above. The comments contain some more discussion of the “tool” idea (Denis Bider and Shane Legg seem to be picturing something similar to “tool-AGI”) but the discussion is unresolved and I believe the “tool” concept defended above remains essentially unaddressed.

In sum, SI appears to encourage a focus on building and launching “Friendly” agents (it is seeking to do so itself, and its work on “Friendliness” theory seems to be laying the groundwork for others to do so) while not addressing the tool-agent distinction. It seems to assume that any AGI will have to be an agent, and to make little to no attempt at justifying this assumption. The result, in my view, is that it is essentially advocating for a more dangerous approach to AI than the traditional approach to software development.