CarlShulman

Posts

Sorted by New

Comments

2019 AI Alignment Literature Review and Charity Comparison
MIRI researchers contributed to the following research led by other organisations
MacAskill & Demski's A Critique of Functional Decision Theory

This seems like a pretty weird description of Demski replying to MacAskill's draft.

What failure looks like

OK, thanks for the clarification!

My own sense is that the intermediate scenarios are unstable: if we have fairly aligned AI we immediately use it to make more aligned AI and collectively largely reverse things like Facebook click-maximization manipulation. If we have lost the power to reverse things then they go all the way to near-total loss of control over the future. So i would tend to think we wind up in the extremes.

I could imagine a scenario where there is a close balance among multiple centers of AI+human power, and some but not all of those centers have local AI takeovers before the remainder solve AI alignment, and then you get a world that is a patchwork of human-controlled and autonomous states, both types automated. E.g. the United States and China are taken over by their AI systems (inlcuding robot armies), but the Japanese AI assistants and robot army remain under human control and the future geopolitical system keeps both types of states intact thereafter.

What failure looks like
Failure would presumably occur before we get to the stage of "robot army can defeat unified humanity"---failure should happen soon after it becomes possible, and there are easier ways to fail than to win a clean war. Emphasizing this may give people the wrong idea, since it makes unity and stability seem like a solution rather than a stopgap. But emphasizing the robot army seems to have a similar problem---it doesn't really matter whether there is a literal robot army, you are in trouble anyway.

I agree other powerful tools can achieve the same outcome, and since in practice humanity isn't unified rogue AI could act earlier, but either way you get to AI controlling the means of coercive force, which helps people to understand the end-state reached.

It's good to both understand the events by which one is shifted into the bad trajectory, and to be clear on what the trajectory is. It sounds like your focus on the former may have interfered with the latter.

What failure looks like
I think we can probably build systems that really do avoid killing people, e.g. by using straightforward versions of "do things that are predicted to lead to videos that people rate as acceptable," and that at the point when things have gone off the rails those videos still look fine (and to understand that there is a deep problem at that point you need to engage with complicated facts about the situation that are beyond human comprehension, not things like "are the robots killing people?"). I'm not visualizing the case where no one does anything to try to make their AI safe, I'm imagining the most probable cases where people fail.

Haven't you yourself written about the failure modes of 'do things predicted to lead to videos that people rate as acceptable' where the attack involves surreptitiously reprogramming the camera to get optimal videos (including weird engineered videos designed to optimize on infelicities in the learned objective? E.g. in your discussion of an unaligned benchmark here:

We train an autoregressive generative model, to maximize the log probability assigned to the actual sequence of actions and observations produced by the AI (with each observation conditioned on the past actions). This isn’t actually a good way to train the generative model, but it’s not really central to the discussion.
We train the reward function by showing humans sequences of actions and predicted observations, asking them to assign scores, then predicting those scores with supervised learning. We show humans the sequences of actions that look most promising to the system...
I am quite confident that “world which looks appealing on screen” is not the same as “world which is good;” the gap is already large today, and it will grow much larger over time.
Maintaining a fiction. Ultimately the goal of this AI is to make the sensors look like good things are happening in the world. This can be done either by making the world good or by carefully controlling what gets fed to the sensors. In general, there are many fewer constraints on “what gets fed to the sensors” than “what the world is actually like,” and so it’s going to be possible to produce much more desirable-looking outcomes, much more easily, by controlling the sensors.
If the humans understand what the AI is doing in detail then they can observe the sequence of actions to notice any actions that might corrupt the sensors. But as soon as the AI is engaging in complicated plans this appears to be totally intractable. Even if the AI was human-level, there are many extremely hard-to-detect plans: for example, the AI could collaborate with an external attacker, who will surreptitiously corrupt the sensor. Or the AI could create a copy of itself on an external server who then helped orchestrate a hard-to-detect corruption.

Are you now very confident that no AI company would implement something with this vulnerability? Or does the paragraph below just mean that the part where the AI gets control of the camera has to happen before the robot armies (or even robot armies identified as traceable to the AI in question), which then happen?

Part I has this focus because (i) that's where I think the action is---by the time you have robot armies killing everyone the ship is so sailed, I think a reasonable common-sense viewpoint would acknowledge this by reacting with incredulity to the "robots kill everyone" scenario, and would correctly place the "blame" on the point where everything got completely out of control even though there weren't actually robot armies yet (ii) the alternative visualization leads people to seriously underestimate the difficulty of the alignment problem, (iii) I was trying to describe the part of the picture which is reasonably accurate regardless of my views on the singularity.

Because it definitely seems that Vox got the impression from it that there is never a robot army takeover in the scenario, not that it's slightly preceded by camera hacking.

Is the idea that the AI systems develops goals over the external world (rather than the sense inputs/video pixels) so that they are really pursuing the appearance of prosperity, or corporate profits, and so don't just wirehead their sense inputs as in your benchmark post?

What failure looks like

I think the kind of phrasing you use in this post and others like it systematically misleads readers into thinking that in your scenarios there are no robot armies seizing control of the world (or rather, that all armies worth anything at that point are robotic, and so AIs in conflict with humanity means military force that humanity cannot overcome). I.e. AI systems pursuing badly aligned proxy goals or influence-seeking tendencies wind up controlling or creating that military power and expropriating humanity (which eventually couldn't fight back thereafter even if unified).

E.g. Dylan Matthews' Vox writeup of the OP seems to think that your scenarios don't involve robot armies taking control of the means of production and using the universe for their ends against human objections or killing off existing humans (perhaps destructively scanning their brains for information but not giving good living conditions to the scanned data):

Even so, Christiano’s first scenario doesn’t precisely envision human extinction. It envisions human irrelevance, as we become agents of machines we created.
Human reliance on these systems, combined with the systems failing, leads to a massive societal breakdown. And in the wake of the breakdown, there are still machines that are great at persuading and influencing people to do what they want, machines that got everyone into this catastrophe and yet are still giving advice that some of us will listen to.

The Vox article also mistakes the source of influence-seeking patterns to be about social influence rather than systems that try to increase in power and numbers tend to do so, so are selected for if we accidentally or intentionally produce them and don't effectively weed them out; this is why living things are adapted to survive and expand; such desires motivate conflict with humans when power and reproduction can be obtained by conflict with humans, which can look like robot armies taking control.takes the point about influence-seeking patterns to be about. That seems to me just a mistake about the meaning of influence you had in mind here:

Often, he notes, the best way to achieve a given goal is to obtain influence over other people who can help you achieve that goal. If you are trying to launch a startup, you need to influence investors to give you money and engineers to come work for you. If you’re trying to pass a law, you need to influence advocacy groups and members of Congress.
That means that machine-learning algorithms will probably, over time, produce programs that are extremely good at influencing people. And it’s dangerous to have machines that are extremely good at influencing people.

Two Neglected Problems in Human-AI Safety

I think this is under-discussed, but also that I have seen many discussions in this area. E.g. I have seen it come up and brought it up in the context of Paul's research agenda, where success relies on humans being able to play their part safely in the amplification system. Many people say they are more worried about misuse than accident on the basis of the corruption issues (and much discussion about CEV and idealization, superstimuli, etc addresses the kind of path-dependence and adversarial search you mention).

However, those varied problems mostly aren't formulated as 'ML safety problems in humans' (I have seen robustness and distributional shift discussion for Paul's amplification, and daemons/wireheading/safe-self-modification for humans and human organizations), and that seems like a productive framing for systematic exploration, going through the known inventories and trying to see how they cross-apply.