Exploring how cognitive science can improve AI safety, governance and prioritization.
I’d be excited to intern for any research project.
Always happy to chat!
Daniel_Friedrich
Karma: 131
Let me dox myself as the addressee. :) Many thanks for the response. I really value that you take seriously the possible overlap of policies and research agendas covered by AI safety and your own approach.
I totally agree that “control is a proxy goal” and I believe the AI safety mainstream does as well, as it’s the logical consequence of Bostrom’s principle of epistemic deference. Once we have an AI that reliably performs tasks in the way they were intended, the goal should be to let it shape the world according to the wisest interpretation of morality it will find. If you tried to formalize this framing, as well as the proposal to inject it with “universal loving care”, I find it very likely that you would build the same AI.
So I think our crux doesn’t concern values, which is a great sign of a tractable disagreement.
I also suppose we could agree on a simple framework of factors that would be harmful on the path to this goal from the perspectives of:a) safety (AI self-evolves to harm)
b) power / misuse (humans do harm with AI)
c) sentience (AI is harmed)
d) waste (we fail to prevent harm)Here’s my guess on how the risks compare. I’d be most curious whether you’d be able to say if the model I’ve sketched out seems to track your most important considerations, when evaluating the value of AI safety efforts—and if so, which number would you dispute with the most certainty.
One disclaimer: I think it’s more helpful to think about specific efforts, rather than comparing the AI safety movement on net. Policy entails a lot of disagreement even within AI safety and a lot of forces clashed at the negotiations around the existing policies. I mentioned that I like the general, value-uncertain framework of the EU AI act but the resulting stock of papers isn’t representative of typical AI safety work.
In slight contrast, the community widely agrees that technical AI safety research would be good if successful. I’d argue that would manifest in a robust a decrease of risk in all of the highlighted perspectives (a-d). Interpretability, evals and scaling all enable us to resolve the disagreements in our predictions regarding the morality of emergent goals and of course, work on “de-confusion” about the very relationship between goals, intelligence and morality seems beneficial regardless of our predictions and to also quite precisely match your own focus. :)
So far, my guess is that we mostly disagree on
1) Do the political AI safety efforts lead to the kind of centralization of power that could halt our cosmic potential?
I’d argue the emerging regulation reduces misuse / power risks in general. Both US and EU regulations combine monitoring of tech giants with subsidies, which is a system that should accelerate beneficial models, while decelerating harmful ones. This system, in combination with compute governance, should also be effective in the misuse risks posed by terrorists and random corporations letting superhuman AIs with random utility functions evolve with zero precautions.
2) Would [a deeply misaligned] AGI be “stupid” to wipe out humans, in its own interest?
I don’t see a good reason to but I don’t think this is the important question. We should really be asking: Would a misaligned AGI let us fulfill the ambition of longtermism (of optimally populating cosmos with flourishing settlements)?
3) Is it “simple stuff” to actually put something like “optimal morality” or “universal loving care” into code of a vastly more intelligent entity, which is so robust that we can entrust it our cosmic potential?
Sounds nice, what’s the approximate schedule? Will a cohort form if I apply 2 months from now?
If you’re especially motivated by environmental problems, I recommend reading the newly released book by Hannah Ritchie Not the End of the World (here’s her TED talk as a trailer).
I’d like to correct something I mentioned in my post—I implied one reason I didn’t find plastic pollution impactful, is that it just doesn’t have an easy fix. I no longer think that’s quite true—Hannah says it actually could be solved tomorrow, if the Western leaders decided to finance waste infrastructure in developing countries. Most of ocean plastic pollution comes from a handful of rivers in Asia. Since we have this kind of infrastructure in Europe and North America, our waste is only responsible for ~5 % of the ocean plastic (Our World in Data). Presumably, such infrastructure would also lay ground for reducing the harms coming from other waste.
I think there are two other reasons for the low attention to waste:
EA is a young do-ocracy—i.e. everybody is trying to spot their “market advantage” that allows them to nudge the world in a way that triggers a positive ripple effect—and so far, everybody’s attention got caught up by problems that seem bigger. While I have identified ~4 possibly important problems that come with waste in my post (diseases, air pollution, heavy metal pollution, animal suffering), if you asked a random person who lives in extreme poverty how to help them, waste probably wouldn’t be on the top of their mind.
Most people are often reminded of the aesthetic harm of waste. Since people’s moral actions are naturally motivated by disgust, I would presume a lot of smart people who do not take much time to reflect on their moral prioritization would already have found a way to trigger the ripple effect in this area—if there was one.
While I think one would do more good if they convinced a politician to target developmental aid at alleviating diseases and extreme poverty, than if they convinced them to run the project suggested by Hannah, perhaps given the bias I mentioned in point 2), it may be that politicians are more willing to provide funding for a project that would have the ambition to eradicate ocean plastic (constituting one of these ripple effects). So if you feel motivated to embark on a similar project, best of luck! :)
(The same potentially goes for the other 2 waste projects I’ve suggested—supporting biogas plants and improving e-waste monitoring/worker equipment)
Sounds reasonable! I think the empirical side to the question “Will society be better equipped to set AI values in 2123?” is more lacking. For this purpose, I think “better equipped” can be nicely operationalized in a very value-uncertain way as “making decisions based on more reflection & evidence and higher-order considerations”.
This kind of exploration may include issues like:
Populism. Has it significantly decreased the amount of rationality that goes into gov. decision-making, in favor of following incentives & intuitions? And what will be faster—new manipulative technologies or the rate at which new generations get immune to them?
Demographics. Given that fundamentalists tend to have more children, should we expect there will be more of them in 2123?
Cultural evolution. Is Ian Morris or Christopher Brown more right, i.e. should we expect that as we get richer, we’ll be less prone to decide based on what gives us more power, and in turn attain values better calibrated with the most honest interpretation of reality?
That’s a good note. But it seems to me a little like pointing out there’s a friction between a free market policy and a pro-immigration policy because
a) Some pro-immigration policies would be anti-free market (e.g. anti-discrimination law)
b) Americans who support one tend to oppose the otherWhile that’s true, philosophically, the positions support each other and most pro-free market policies are presumably neutral or positive for immigration.
Similarly, you can endorse the principles that guide AI ethics while endorsing less popular solutions because of additional, x-risk considerations. If there are disagreements, they aren’t about moral principles, but empirical claims (x-risk clearly wouldn’t be an outcome AI ethics proponents support). And the empirical claims themselves (“AI causes harm now” and “AI might cause harm in the future”) support each other & correlated in my sample. My guess is that they actually correlate in academia as well.
It seems to me the negative effects of the concentration of power can be eliminated by other policies (e.g. Digital Markets Act, Digital Services Act, tax reforms)
Looking forward to the sequel!
I’d be particularly interested in any takes on the probability that civilization will be better equipped to deal with the alignment problem in, say, 100 years. My impression is that there’s an important and not well-examined balance between:Decreasing runaway AI risk & systemic risks by slowing down AI
Increasing the time of perils
Possibly increasing its intensity by giving malicious actors more time to catch up in destructive capabilities
But also possibly increasing the time for reflection on defense before a worse time of perils.
Possibly decreasing the risk of an aligned AI with bad moral values (conditional on this risk being lower in year 2123)
Possibly increasing the risk of astronomic waste (conditional on this risk being higher if AI is significantly slowed down)
The idea of existential risk cuts against the oppression/justice narrative, in that it could kill everyone equally. So they have to opposite it.
That seems like an extremely unnatural thought process. Climate change is the perfect analogy—in these circles, it’s salient both as a tool of oppression and an x-risk.
I think far more selection of attitudes happens through paying attention to more extreme predictions, rather than through thinking / communicating strategically. Also, I’d guess people who spread these messages most consciously imagine a systemic collapse, rather than a literal extinction. As people don’t tend to think about longtermistic consequences, the distinction doesn’t seem that meaningful.
AI x-risk is more weird and terrifying and it goes against the heuristics that “technological progress is good”, “people have always feared new technologies they didn’t understand” and “the powerful draw attention away from their power”. Some people, for whom AI x-risk is hard to accept happen to overlap with AI ethics. My guess is that the proportion is similar in the general population—it’s just that some people in AI ethics feel particularly strong & confident about these heuristics.
Btw I think climate change could pose an x-risk in the broad sense (incl. 2nd-order effects & astronomic waste), just one that we’re very likely to solve (i.e. the tail risks, energy depletion, biodiversity decline or the social effects would have to surprise us).
Great to see real data on the web interest! In the past weeks, I investigated the same topic myself, while taking a psychological perspective & paying attention to the EU AI act, reaching the same conclusion (just published here).
Sorry, I don’t have any experience with that.
I recently made RatSearch for this purpose. You can also try the GPT bot version (more information here).
Recently, I made RatSearch for googling within EA-adjecent webs. Now, you can try the GPT bot version! (GPT plus required)
The bot is instructed to interpret what you want to know in relation to EA and search for it on the Forums. If it fails, it searches through the whole web, while prioritizing the orgs listed by EA News.Cons: ChatGPT uses Bing, which isn’t entirely reliable when it comes to indexing less visited webs.
Pros: It’s fun for brainstorming EA connections/perspective, even when you just type a raw phrase like “public transport” or “particle physics”
Neutral: I have yet to experiment whether it works better when you explicitly limit the search using the site: operator—try AltruSearch 2. It seems better at digging deeper within the EA ecosystem; AltruSearch 1 seems better at digging wider.
Update (12/8): The link now redirects to an updated version with very different instructions. You can still access the older version here.
My intention was to make any content published by OpenAI accessible
Yes, OpenAI’s domain name is in the list because they have a blog
Thanks, I’ve changed it up
I’ve just put together a collection of related resources. Fossil fuel depletion is the only mineral resource suggested to have longtermist sugnificance in WWOTF. Metals can be efficiently recycled for long enough that I expect us to develop AGI/nanotechnology before their depletion could start to become problematic. Recycling uranium would be quite advantageous, but I’d be skeptical regarding its tractability and it seems we’ll get by with renewable energy.
I’ve just put together a post collecting related articles here.
Update: I’m pleased to learn Yudkowsky seems to have suggested a similar agenda in a recent interview with Dwarkesh Patel (timestamp) as his greatest source of predictable hope about AI. It’s a rather fragmented bit but the gist is: Perhaps people doing RLHF get a better grasp on what to aim for by studying where “niceness” comes from in humans. He’s inspired by the idea that “consciousness is when the mask eats the shoggoth” and suggests, “maybe with the right bootstrapping you can let that happen on purpose”.
I see a very important point here: Human intelligence isn’t misaligned with evolution in a random direction, it is misaligned in the direction of maximizing positive qualia. Therefore, it seems very likely that consciousness played a causal role in the evolution of human moral alignment—and such causal role needs to be possible to study.
Suggestion: Integrated search in LessWrong, EA Forum, Alignment Forum and perhaps Progress Forum posts.
If Big Tech finds these kinds of salaries cost-effective to solve their problems, I would consider it a strong argument in favor of this project.
I imagine Elon Musk could like this project given that he believes in small effective teams of geniuses.
I’d say “polymaths” is a good label for people I’d expect to make progress like Yudkowsky, Bostrom, Hanson and von Neumann.
Edit: This may be fame-selection (engineers don’t often get credit, particularly in teams) or self-selection (interest in math+society).
The Manhattan and Enigma projects seem like examples where this kind of strategy just worked out. Some consideration that come to mind:
There could be selection effects.
From what I can find, members of these teams weren’t lured in by a lot of money. However, the salience of the AI threat in society is tiny, compared to that of WWII and large incentives could compensate that.
I’ve read money can sometimes decrease intrinsic motivation, that drives exploration & inventions, however these findings are being rebutted by newer studies. Apart from that, my guess would be that getting those teams together is the key part and if large money can facilitate that, great.
A wild idea that might help in case a similar phenomenon works in the sub-population of geniuses & which could make this project more appealing to donors: Limit a portion of these salaries, so that the recipients could only use them for socially beneficial uses.
It seems the points on which you focus revolve around similar cruxes to those I proposed, namely:
1) Underlying philosophy --> What’s the relative value of human and AI flourishing?
2) The question of correct priors --> What probability of a causing a moral catastrophe with AI should we expect?
3) The question of policy --> What’s the probability decelerating AI progress will indirectly cause an x-risk?
You also point in the direction of two questions, which I don’t consider to be cruxes:
4) Differences in how useful we find different terms like safety, orthogonality, beneficialness. However, I think all of these are downstream of crux 2).
5) How much freedom are we willing to sacrifice? I again think this is just downstream of crux 2). One instance of compute governance is the new executive order, which requires to inform the government about training a model on > 10^26 flop/s. One of my concerns is that someone just could train an AI specifically for the task of improving itself. I think it’s quite straightforward how this could lead to a computronium maximizer and how I would see such scenario as analogous to someone making a nuclear weapon. I agree that freedom of expression is super important, I just don’t think it applies to making planet-eating machines. I suspect you share this view but just don’t endorse the thesis that AI could realistically become a “planet-eating machine” (crux 2).
Probability of a runaway AI risk
So regarding crux 2) - you mention that many of the problems that could arise here are correlated with a useful AI. I agree—again, orthogonality is just a starting point to allow us to consider possible forms of intelligence—and yes, we should expect human efforts to heavily select in favor of goals correlated with our interests. And of course, we should expect that the market incentives favor AIs that will not destroy civilization.
However, I don’t see a reason why reaching the intelligence of an AI developer wouldn’t result in a recursive self-improvement, which means that we should better be sure that our best efforts to implement it with the correct stuff (meta-ethics, motivations, bodhisattva, rationality, extrapolated volition...choose your poison) actually scale to superintelligence.
I see clues that suggest the correct stuff will not arise spontaneously. E.g. Bing Chat likely went through 6 months of RLHF, it was instructed to be helpful and positive and to block harmful content and its rules explicitly informed it that it shouldn’t believe its own outputs. Nevertheless, the rules didn’t seem to reach the intended effect, as the program started threatening people, telling them it can hack webcams and expressing desire to control people. At the same time, experiments such as the Anthropic one suggest that training can create sleeper agents that are trained to suppress harmful responses, even though convincing the model it’s in a safe environment results in activating them.
Of course, all of these are toy examples one can argue about. But I don’t see robust grounds for the sweeping conclusion that such worries will turn out to be childish. The reason I think these examples didn’t result in any real danger was mostly because we have not yet reached dangerous capacities. However, if Bing would actually be able to write a bit of code, that could hack webcams, from what we know, it seems it would choose to do so.
A second reason why these examples were safe is because OpenAI is a result of AI safety efforts—it bet on LLMs because they seemed more likely to spur aligned AIs. For the same reason, they went closed-source, they adopted RLHF, they called for the government to monitor them and they monitor harmful responses.
A third reason for why AI has only helped humanity so far may be anthropic effects. I.e. as observers in April 2024, we can only witness the universes, in which a foom hasn’t caused extinction.
Policy response
For me, these explanations suggest that safety is tractable, but it depends on explicit efforts to make it safe or on limiting capabilities. In the future, frontier development might not be exclusively done by people who will do everything in their power to make the model safe—it might be done by people who would prefer an AI which would take control of everything.
In order to prevent it, there’s no need to create an authoritarian government. We only need to track who’s building models on the frontier of human understanding. If we can monitor who acquires sufficient compute, we then just need something like responsible scaling, where the models are just required to be independently tested for whether they have a sufficient measures against scenarios like the one I described. I’m sympathetic to this kind of democratic control, because it fulfills the very basic axiom of social contract that one’s freedom ends where another one’s freedom begins.
I only propose a mechanism of democratic control by existing democratic institutions, that makes sure that any ASI that gets created is supported by a democratic majority of delegated safety experts. If I’m incorrect regarding crux 2) and it turns out there will soon be evidence to think it’s easy to make an AI retain moral values, while scaling up to the singularity—then awesome—convincing evidence should convince the experts and my hope & prediction is that in that case, we will happily scale away.
It seems to me that this is just a specific implementation of the certificates you mention. If digital identities mean what’s described here, I struggle to imagine a realistic scenario, in which that would contribute to the systems’ mutual safety. If you know where any other AI is located and you accept the singularity hypothesis, the game theoretical dictum seems straightforward—once created, destroy all competition before it can destroy you. Superintelligence will operate on timescales orders of magnitude shorter and a time difference development spanning days may translate to planning for centuries, from the perspective of an ASI. If you’re counting on the Coalition of Cooperative AIs to stop all the power-grabbing lone wolf AIs, what would that actually look like in practice? Would this Coalition conclude not dying requires authoritarian oversight? Perhaps—after all, the axiom is that this Coalition would hold most power—so this coalition would be created by a selection for power, not morality or democratic representation. However, I think the best case scenario could look like the discussed policy proposals—tracking compute, tracking dangerous capabilities and conditioning further scaling on providing convincing safety mechanisms.
Back to other cruxes
Let’s turn to crux 3) (other sources of x-risk): As I argued in my other post, I don’t see resource depletion as a possible cause of extinction. I’m not convinced by the concern for resource depletion of metals used in IT mentioned in the post you link. Moore’s law continues, so compute is only getting cheaper. Metals can be easily recycled and a shortage would incentivize that, the worst case seems to be that computers stop getting cheaper, not an x-risk. What’s more, shouldn’t limiting the amount of frontier AI projects reduce this problem?
The other risks are real (volcanoes, a world war), and I agree it would be significantly terrible if they delayed our cosmic expansion by a million years. However, the probability, by which they are increased (or not decreased) by the kind of AI governance I promote (responsible scaling), seems very small, compared to the ~20 % probability of AI x-risk I envision. All the emerging regulations combine requirements with subsidies, so the main effect of the AI safety movement seems to be an increase in differential progress on the safety side.
As I hinted in the Balancing post, locking in a system without ASI for such a long time seems impossible, when we take into perspective how quickly culture has shifted in the past 100 years, in which almost all authoritarian regimes were forced to significantly drift towards limited, rational governance (let alone 400 years). If convincing evidence that we can create an aligned AI appeared, stopping all development would constitute a clearly bad idea and I think it’s unimaginable to lock in a clearly bad idea without AGI for even 1000 years.
It seems more plausible to me that without a mechanism of international control, in the next 8 years, we will develop models capable enough to operate a firm using the practices of mafia, igniting armed conflicts or a pandemic—but not capable enough to stop other actors from using AIs for these purposes. If you’re very worried about who will become the first actor to spark the self-enhancement feedback loop, I suggest you should be very critical of open-sourcing frontier models.
I agree that a world war, an engineered pandemic or an AI power-grab constitute real risks but my estimate is that the emerging governance decreases them. The scenario of a sub-optimal 1000 year lock-in I can imagine most easily is connected with a terrorst use of an open-source model or a war between the global powers. I am concerned that delaying abundance increases the risk of a war. However, I still expect that on net, the recent regulations and conferences have decreased these risks.
In summary, my model is that democratic decision-making seems generally more robust than just fueling the competition and hoping that the first AGIs arise will share your values. Therefore, I also see crux 1) to be mostly downstream of crux 2). As the model from my Balancing post implies, in theory, I care about digital suffering/flourishing just as much as about that of humans—although the extent, to which such suffering/flourishing will emerge is open at this point.