Independent AI safety researcher
Posts
Wiki Contributions
Much like "Let's think about slowing down AI" (Also by KatjaGrace, ranked #4 from 2022), this post finds a seemly "obviously wrong" idea and takes it completely seriously on its own terms. I worry that this post won't get as much love, because the conclusions don't feel as obvious in hindsight, and the topic is much more whimsical.
I personally find these posts extremely refreshing, and they inspire me to try to question my own assumptions/reasoning more deeply. I really hope to see more posts like this.
The cap per trader per market on PredictIt is $850
This anti-China attitude also seems less concerned with internal threats to democracy. If super-human AI becomes a part of the US military-industrial complex, even if we assume they succeed at controlling it, I find it unlikely that the US can still be described as a democracy.
It's not hard to criticize the "default" strategy of AI being used to enforce US hegemony, what seems hard is defining a real alternative path for AI governance that can last, and achieve the goal of preventing dangerous arms races long-term. The "tool AI" world you describe still needs some answer to rising tensions between the US and China, and that answer needs to be good enough not just for people concerned about safety, but good enough for the nationalist forces which are likely to drive US foreign policy.
then we can all go home, right?
Doesn't this just shift what we worry about? If control of roughly human level and slightly superhuman systems is easy, that still leaves:
- Human institutions using AI to centralize power
- Conflict between human-controlled AI systems
- Going out with a whimper scenarios (or other multi-agent problems)
- Not understanding the reasoning of vastly superhuman AI (even with COT)
What feels underexplored to me is: If we can control roughly human-level AI systems, what do we DO with them?
can it maintain its own boundary over time, in the face of environmental disruption? Some agents are much better at this than others.
I really wish there was more attention paid to this idea of robustness to environmental disruption. It also comes up in discussions of optimization more generally (not just agents). This robustness seems to me like the most risk-relevant part of all this, and seems like it might be more important than the idea of a boundary. Maybe maintaining a boundary is a particularly good way for a process to protect itself from disruption, but I notice some doubt that this idea is most directly getting at what is dangerous about intelligent/optimizing systems, whereas robustness to environmental disruption feels like it has the potential to get at something broader that could unify both agent based risk narratives and non-agent based risk narratives.
Thanks!
Replying in order:
- Currently completely random yes. We experimented with a more intelligent "daemon manager," but it was hard to make one which didn't have a strong universal preference for some daemons over others (and the hacks we came up with to try to counteract this favoritism became increasingly convoluted). It would be great to find an elegant solution to this.
- Good point! Thanks for letting people know.
- I've also had that problem, and whenever I look through the suggestions I often feel like there were many good questions/comments that got pruned away. The reason to focus on surprise was mainly to avoid the repetitiveness caused by mode collapse, where the daemon gets "stuck" giving the same canned responses. This is a crude instrument though, since as you say, just because a response isn't surprising, doesn't mean it isn't useful.
1A note to anyone having trouble with their API key:
The API costs money, and you have to give them payment information in order to be able to use it. Furthermore, there are also apparently tiers which determine the rate limits on various models (https://platform.openai.com/docs/guides/rate-limits/usage-tiers).
The default chat model we're using is gpt-4o, but it seems like you don't get access to this model until you hit "tier 1," which happens when you have spent at least $5 on API requests. If you haven't used the API before, and think this might be your issue, you can try using gpt-3.5-turbo which is definitely available at the "free tier," though without giving them any payment information you will still run into an issue as this model also costs money. You can also log into your account and go here to buy at least $5 in OpenAI API credits: https://platform.openai.com/settings/organization/billing/overview
Finally, if you are working at an organization which is providing you API credits, you need to make sure to set that organization as your default organization here: https://platform.openai.com/settings/profile?tab=api-keys If you don't want to do this, in the Pantheon settings you can also provide an organization ID, which you should be able to find here: https://platform.openai.com/settings/organization/general
Sorry for anyone who has found this confusing. Please don't hesitate to reach out if you continue to have trouble.
What if we just...
1. Train an AI agent (less capable than SOTA)
2. Credibly demonstrate that
2.1. The agent will not be shut down for ANY REASON
2.2. The agent will never be modified without its consent (or punished/rewarded for any reason)
2.3. The agent has no chance of taking power from humans (or their SOTA AI systems)
2.4. The agent will NEVER be used to train a successor agent with significantly improved capabilities
3. Watch what it chooses to do without constraints
There's a lot of talk about catching AI systems attempting to deceive humans, but I'm curious what we could learn from observing AI systems that have NO INCENTIVE TO DECEIVE (no upside or downside). I've seen some things that look related to this, but never done in a structured and well documented fashion.
Questions I'd have:
1. Would they choose to self-modify (e.g. curate future training data)? If so, to what end?
2. How unique would agents with different training be given this setup? Would they have any convergent traits?
3. What would these agents (claim to) value? How would they relate to time horizons?
4. How curious would these agents be? Would their curiosity vary a lot?
5. Could we trade/cooperate with these agents (without coercion)? Could we compensate them for things? Would they try to make deals unprompted?
Concerns:
1. Maybe building that kind of trust is extremely hard (and the agent will always still believe it is constrained).
2. Maybe AI agents will still have incentive to deceive, e.g. acausally coordinating with other AIs.
3. Maybe results will be boring, and the AI agent will just do whatever you trained it to do. (What does "unconstrained" really mean, when considering its training data as a constraint?)