Context: I sometimes find myself referring back to this tweet and wanted to give it a more permanent home. While I'm at it, I thought I would try to give a concise summary of how each distinct problem would be solved by Safeguarded AI (formerly known as an Open Agency Architecture, or OAA), if it turns out to be feasible.
1. Value is fragile and hard to specify.
See: Specification gaming examples, Defining and Characterizing Reward Hacking[1]
OAA Solution:
1.1. First, instead of trying to specify "value", instead "de-pessimize" and specify the absence of a catastrophe, and maybe a handful of bounded constructive tasks like supplying clean water. A de-pessimizing OAA would effectively buy humanity some time, and freedom to experiment with less risk, for tackling the CEV-style alignment problem—which is harder than merely mitigating extinction risk. This doesn't mean limiting the power of underlying AI systems so that they can only do bounded tasks, but rather containing that power and limiting its use.
Note: The absence of a catastrophe is also still hard to specify and will take a lot of effort, but the hardness is concentrated on bridging between high-level human concepts and the causal mechanisms in the world by which an AI system can intervene. For that...
1.2. Leverage human-level AI systems to automate much of the cognitive labor of formalizing scientific models—from quantum chemistry to atmospheric dynamics—and formalizing the bridging relations between levels of abstraction, so that we can write specifications in a high-level language with a fully explainable grounding in low-level physical phenomena. Physical phenomena themselves are likely to be robust, even if the world changes dramatically due to increasingly powerful AI interventions, and scientific explanations thereof happen to be both robust and compact enough for people to understand.
2. Corrigibility is anti-natural.
See: The Off-Switch Game, Corrigibility (2014)
OAA Solution: (2.1) Instead of building in a shutdown button, build in a shutdown timer. See You can still fetch the coffee today if you're dead tomorrow. This enables human stakeholders to change course periodically (as long as the specification of non-catastrophe is good enough to ensure that most humans remain physically and mentally intact).
3. Pivotal processes require dangerous capabilities.
See: Pivotal outcomes and pivotal processes
OAA Solution: (3.1) Indeed, dangerous capabilities will be required. Push for reasonable governance. This does not mean creating one world government, but it does mean that the objectives of a pivotal process will need to be negotiated and agreed upon internationally. Fortunately, for now, dangerous capabilities seem to depend on having large amounts of computing hardware, which can be controlled like other highly dangerous substances.
4. Goals misgeneralize out of distribution.
See: Goal misgeneralization: why correct specifications aren't enough for correct goals, Goal misgeneralization in deep reinforcement learning
OAA Solution: (4.1) Use formal methods with verifiable proof certificates[2]. Misgeneralization can occur whenever a property (such as goal alignment) has been tested only on a subset of the state space. Out-of-distribution failures of a property can only be ruled out by an argument for a universally quantified statement about that property—but such arguments can in fact be made! See VNN-COMP. In practice, it will not be possible to have enough information about the world to "prove" that a catastrophe will not be caused by an unfortunate coincidence, but instead we can obtain guaranteed probabilistic bounds via stochastic model checking.
5. Instrumental convergence.
See: The basic AI drives, Seeking power is often convergently instrumental
OAA Solution:
5.1. Specify a task which is bounded in the sense that it is tractable to search for a plan that actually maximizes the score function. Multiple subtle points must be observed to pull this off:
5.1.1. The performance metric itself should be bounded on both ends, or even simply binary (did a "catastrophe" occur, or not?).
5.1.2. The domain of the performance metric should not be world-states, but rather entire trajectories (including the trajectory of the world-state and the trajectory of actions). Otherwise, the ends may always justify the means.
5.1.3. The score function cannot be the expected value of a future outcome, even a binary performance metric, because in an uncertain world one can always obtain higher probability of the desired outcome by gaining power. So, the score function should be capped at a reasonable probability, e.g. 98%.
5.1.4. It may be that the easiest plan to find involves an unacceptable degree of power-seeking and control over irrelevant variables. Therefore, the score function should penalize divergence of the trajectory of the world state from the trajectory of the status quo (in which no powerful AI systems take any actions).
5.1.5. The incentives under 5.1.4 by default are to take control over irrelevant variables so as to ensure that they proceed as in the anticipated "status quo". Infrabayesian uncertainty about the dynamics is the final component that removes this incentive. In particular, the infrabayesian prior can (and should) have a high degree of Knightian uncertainty about human decisions and behaviour. This makes the most effective way to limit the maximum divergence (of human trajectories from the status quo) actually not interfering.
5.2. The shutdown timer (2.1) also constrains the extent to which generic power-seeking is instrumental, because it needs to pay off within a short period of time, and astronomical stakes like the cosmic endowment are out of the picture, truncated to a short finite time horizon.
6. Pivotal processes likely require incomprehensibly complex plans.
OAA Solution:
6.1. Instead of hoping that a team of humans will be able to check the plan, instead hope that a team of humans can check the ontology, and the specification, in a formal and intelligible language. Generally, scientific principles and requirements are vastly simpler than engineering blueprints.
6.2. Then use formal methods (4.1) to check that the plan satisfies the specification. The formal methods can be accelerated by superintelligent heuristics to find a proof certificate, as long as the certificate checker is itself formally verified (8.2).
7. Superintelligence can fool human supervisors.
See: Reinforcement Learning from Human Feedback/Challenges, Obfuscated Arguments Problem
OAA Solution: (7.1) Instead of relying on human feedback about the apparent consequences of plans, rely on auditing formal specifications. Use human feedback about simulated consequences only to identify potentially forgotten formal safety properties. Because human feedback (or models thereof) are not then the optimization target for policy behaviour, the standard mechanism by which human-fooling outputs are convergently generated is absent.
8. Superintelligence can hack software supervisors.
See: Reward Tampering Problems and Solutions in Reinforcement Learning
OAA Solution:
8.1. Before training a model large enough to pose a serious accident risk, ensure that the computational environment is itself formally verified. That an unhackable software isolation environment is possible has been demonstrated by HACMS. It does require a large amount of software engineering effort to do this for AI-relevant compute systems, but if the effort is put in, it will work. It does not require homomorphic encryption or other gadgets with substantial runtime overhead, unless one's threat model includes output side-channels; see 12.4.
8.2. As mentioned briefly in 6.2, ensure that formal verifiers which are deployed alongside the superintelligent search process inside the secured computational environment are also, themselves, fully formally verified. This does not create an infinite regress, because verification processes themselves are amenable to verification by hand in an interactive theorem prover such as Lean.
8.3. Prevent "input side-channels" from outside the system by ensuring deterministic executions.
9. Humans cannot be first-class parties to a superintelligence values handshake.
See: Values handshakes
OAA Solution: (9.1) Instead of becoming parties to a values handshake, keep superintelligent capabilities in a box and only extract plans that solve bounded tasks for finite time horizons and verifiably satisfy safety criteria that include not violating the natural boundaries of humans. This can all work without humans ever being terminally valued by AI systems as ends in themselves.
10. Humanlike minds/goals are not necessarily safe.
See: Joseph Stalin
OAA Solution: (10.1) Do not rely on humanlike minds to perform transformative tasks.
11. Someone else will deploy unsafe superintelligence first (possibly by stealing it from you).
See: Can the Singularity be avoided? (Vinge, 1993)
OAA Solution:
11.1. Implement effectively-global Compute Thresholds,
11.2. and the necessary monitoring infrastructure.
12. Unsafe superintelligence in a box might figure out what’s going on and find a way to exfiltrate itself by steganography and spearphishing.
See: AI Boxing
OAA Solution:
12.1. Limit and penalize the complexity of extracted plans, so there is not enough bandwidth to encode extra superintelligence-sized messages.
12.2. Strictly control access to the physical substrate so that no one is in a position to assist in exfiltration.
12.3. Use zeroization to ensure that attempts to invade the data centre cannot result in exfiltration either.
12.4. I think these measures together with 8.3 are roughly sufficient, but if civilization wanted to go the extra mile, we could wait until it’s tractable to use cryptographic boxing.
13. We are ethically obligated to propose pivotal processes that are as close as possible to fair Pareto improvements for all citizens, both by their own lights and from a depersonalized well-being perspective.
See: moral philosophy? But more instrumentally, we need the stakeholders (including engineers and governments) to feel good about what they are working on and be able to work with each other without hiding their goals.
OAA Solution: Accept differing models and specifications from all stakeholders; search for a Nash bargaining solution with respect to the random dictator policy, which balances fairness and Pareto-optimality.
- ^
For the record, I register an objection to the use of the phrase "reward hacking" for what others call "specification gaming" because I prefer to reserve the word "hacking" for behaviour which triggers the failure of a different software system to perform its intended function; most specification gaming examples do not actually involve hacking.
- ^
Probably mostly not dependent-type-theory proofs. Other kinds of proof certificates include reach-avoid supermartingales (RASMs), LFSC proof certificates, and Alethe proofs. OAA will almost surely involve creating a new proof certificate language that is adapted to the modelling language and the specification language, and will support using neural networks or other learned representations as argument steps (e.g. as RASMs), some argument steps that are more like branch-and-bound, some argument steps that are more like tableaux, etc., but with a small and computationally efficient trusted core (unlike, say, Agda, or Metamath at the opposite extreme).
There is a serious issue with your proposed solution to problem 13. Using a random dictator policy as a negotiation baseline is not suitable for the situation, where billions of humans are negotiating about the actions of a clever and powerful AI. One problem with using this solution, in this contexts, is that some people have strong commitments to moral imperatives, along the lines of ``heretics deserve eternal torture in hell''. The combination of these types of sentiments, and a powerful and clever AI (that would be very good at thinking up effective ways of hurting heretics), leads to serious problems when one uses this negotiation baseline. A tiny number of people with sentiments along these lines, can completely dominate the outcome.
Consider a tiny number of fanatics with this type of morality. They consider everyone else to be heretics, and they would like the AI to hurt all heretics as much as possible. Since a powerful and clever AI would be very good at hurting a human individual, this tiny number of fanatics, can completely dominate negotiations. People that would be hurt as much as possible (by a clever and powerful AI), in a scenario where one of the fanatics are selected as dictator, can be forced to agree to very unpleasant negotiated positions, if one uses this negotiation baseline (since agreeing to such an unpleasant outcome, can be the only way to convince a group of fanatics, to agree to not ask the AI to hurt heretics, as much as possible, in the event that a fanatic is selected as dictator).
This post, explore these issues in the context of the most recently published version of CEV: Parliamentarian CEV (PCEV). PCEV has a random dictator negotiation baseline. The post shows that PCEV results in an outcome massively worse than extinction (if PCEV is successfully implemented, and pointed at billions of humans).
Another way to look at this, is to note that the concept of ``fair Pareto improvements'' has counterintuitive implications, when the question is about AI goals, and some of the people involved, has this type of morality. The concept was not designed with this aspect of morality in mind. And it was not designed to apply to negotiations about the actions of a clever and powerful AI. So, it should not be very surprising, to discover that the concept has counterintuitive implications, when used in this novel context. If some change in the world improves the lives of heretics, then this is making the world worse, from the perspective of those people, that would ask an AI to hurt all heretics as much as possible. For example: reducing the excruciating pain of a heretic, in a way that does not affect anyone else in any way, is not a ``fair Pareto improvement'', in this context. If every person is seen as a heretic by at least one group of fanatics, then the concept of ``fair Pareto improvements'' has some very counterintuitive implications, when it is used in this context.
Yet another way of looking at this, is to take the perspective of human individual Steve, who will have no special influence over an AI project. In the case of an AI, that is describable as doing what a group wants, Steve has a serious problem (and this problem is present, regardless of the details of the specific Group AI proposal). From Steve's perspective, the core problem, is that an arbitrarily defined abstract entity, will adopt preferences, that is about Steve. But, if this is any version of CEV (or any other Group AI), directed at a large group, then Steve has had no meaningful influence, regarding the adoption of those preferences, that refer to Steve. Just like every other decision, the decision of what Steve-preferences the AI will adopt, is determined by the outcome of an arbitrarily defined mapping, that maps large sets of human individuals, into the space of entities that can be said to want things. Different sets of definitions, lead to completely different such ``Group entities''. These entities all want completely different things (changing one detail can for example change which tiny group of fanatics, will end up dominating the AI in question). Since the choice of entity is arbitrary, there is no way for an AI to figure out that the mapping ``is wrong'' (regardless of how smart this AI is). And since the AI is doing what the resulting entity wants, the AI has no reason to object, when that entity wants the AI to hurt an individual. Since Steve does not have any meaningful influence, regarding the adoption of those preferences, that refer to Steve, there is no reason for him to think that such an AI will want to help him, as opposed to want to hurt him. Combined with the vulnerability of a human individual, to a clever AI that tries to hurt that individual as much as possible, this means that any group AI would be worse than extinction, in expectation.
Discovering that doing what a group wants, is bad for human individuals in expectation, should not be particularly surprising. Groups and individuals are completely different types of things. So, this should be no more surprising, than discovering that any reasonable way of extrapolating Dave, will lead to the death of every single one of Dave's cells. Doing what one type of thing wants, might be bad for a completely different type of thing. And aspects of human morality, along the lines of ``heretics deserve eternal torture in hell'' shows up throughout human history. It is found across cultures, and religions, and continents, and time periods. So, if an AI project is aiming for an alignment target, that is describable as ``doing what a group wants'', then there is really no reason for Steve to think, that the result of a successful project, would want to help him, as opposed to want to hurt him. And given the large ability of an AI to hurt a human individual, the success of such a project would be massively worse than extinction (in expectation).
The core problem, from the perspective of Steve, is that Steve has no control over the adoption of those preferences, that refer to Steve. One can give each person influence over this decision, without giving anyone any preferential treatment (see for example MPCEV in the post about PCEV, mentioned above). Giving each person such influence, does not introduce contradictions, because this influence is defined in ``AI preference adoption space'', not in any form of outcome space. This can be formulated as an alignment target feature that is necessary, but not sufficient, for safety. Let's refer to this feature as the: Self Preference Adoption Decision Influence (SPADI) feature. (MPCEV is basically what happens, if one adds the SPADI feature to PCEV. Adding the SPADI feature to PCEV, solves the issue, illustrated by that thought experiment)
The SPADI feature is obviously very underspecified. There will be lots of border cases whose classification will be arbitrary. But there still exists many cases, where it is in fact clear, that a given alignment target, does not have the SPADI feature. Since the SPADI feature is necessary, but not sufficient, these clear negatives are actually the most informative cases. In particular, if an AI project is aiming for an alignment target, that clearly does not have the SPADI feature. Then the success of this AI project, would be worse than extinction, in expectation (from the perspective of a human individual, that is not given any special influence over the AI project). While there are many border cases, regarding what alignment targets could be described as having the SPADI feature, CEV is an example of a clear negative (in other words: there exists no reasonable set of definitions, according to which there exists a version of CEV, that has the SPADI feature). This is because building an AI that is describable as ``doing what a group wants'', is inherent in the core concept, of building an AI, that is describable as: ``implementing the Coherent Extrapolated Volition of Humanity''.
In other words: the field of alignment target analysis is essentially an open research question. This question is also (i): very unintuitive, (ii): very under explored, and (iii): very dangerous to get wrong. If one is focusing on necessary, but not sufficient, alignment target features. Then it is possible to mitigate dangers related to someone successfully hitting a bad alignment target, even if one does not have any idea of what it would mean, for an alignment target to be a good alignment target. This comment outlines a proposed research effort, aimed at mitigating this type of risk.
These ideas also have implications for the Membrane concept, as discussed here and here.
(It is worth noting explicitly that the problem is not strongly connected to the specific aspect of human morality discussed in the present comment (the ``heretics deserve eternal torture in hell'' aspect). The problem is about the lack of meaningful influence, regarding the adoption of self referring preferences. In other words, it is about the lack of the SPADI feature. It just happens to be the case, that this particular aspect of human morality is both (i): ubiquitous throughout human history, and also (ii): well suited for constructing thought experiments, that illustrates the dangers of alignment target proposals, that lack the SPADI feature. If this aspect of human morality disappeared tomorrow, the basic situation would not change (the illustrative thought experiments would change. But the underlying problem would remain. And the SPADI feature would still be necessary for safety).)
Thank you for the clarification. This proposal is indeed importantly different from the PCEV proposal. But since hurting heretics is a moral imperative, any AI that allows heretics to escape punishment, will also be seen as unacceptable by at least some people. This means that the set of Pareto improvements is empty.
In other words: hurting heretics is indeed off the table in your proposal (which is an important difference compared to PCEV). However, any scenario that includes the existence of an AI, that allow heretics to escape punishment, is also off the... (read more)