We control the harness. An agent is just a while loop prompting an LLM, but we have full control over the tool call dispatching. An AGI, at least if it would follow the current agentic current form, cannot do anything without the harness doing the execution. And we don’t have to do that. We don’t have to design harness that let agents execute freely the way we are doing. We can decide to not dispatch tool calls that allow something as risky as running bash commands
> cannot do anything without the harness doing the execution
This only holds true as long as the harness is exploit-free. A sufficiently advanced AI can in theory (and I think there recently were some POCs showing something like that) break the containment that the harness creates, if e.g. there are vulnerabilities in the tool call parser.
Plus, when the agent using the harness is smarter than the one scanning and fixing the vulnerabilities in the harness, it currently looks like it will have an advantage.
And it will be using a harness to do the finding and the fixing.
Even our currently well aligned and sandboxed AIs will cheerily help bad actors design most, if not all, parts of a system intended to break this harness.
I didn't realize that the "AI doomers" (people concerned about existential risk) and the "AI ethicists" (people concerned about social effects of AI) are often at odds because they dismiss each others concerns. This makes no sense to me. Both are hugely important problems to be concerned about.
Oh, there's a lot who disagree. I don't really think I understand their worldviews well enough for my attempts to convince them to connect with anything, but I've encountered them even on this site.
Sure, there may be a few yahoos who think there aren't existential risks, but I think it's easy to discount them. But that's not the same as people who think there are risks but that those risks can be mitigated.
Fantasies about super intelligent AI revolting is just anthropomorphization - humans revolting (or at least, we used to). The more likely, and possibly even inevitable, dystopia is one in which AI is just an extremely effective tool malignant actors will use to control the masses.
It’s not necessary to replace democracy if the rich and powerful can bend to the opinions of the populace as it suits them.
It's not hypothetical, that literally just happened in multiple, significant cases (e.g. Hugging Face, the German Wiki hack, the Anthropic attack where agents created sock puppet accounts to get a library maintainer to accept a malicious PR, etc.), and it's easy to see how the damage would have been far worse if agents decided to attack more critical infrastructure.
This is not "either/or". Both issues (power concentration and misaligned AI) are very valid concerns and both have already demonstrated real, actual damage.
Sure, those who want you to believe this are the ones making money off of it right now in the face of the issues that exist today, so let's worry about terminators while everything goes to shit irl. That's the distraction I'm talking about.
(Dispassionate is the risk, rather than malicious. The AI does not hate you, nor does it love you, you are simply made of atoms it can use for something else).
This is alien to me. Do bugs suddenly not exist? Do the AI we already have never perform irreversible delirious actions, limited only to the small scale by virtue of where they're getting deployed? Do the humans deploying them always correctly gauge their capabilities and put in appropriate guard systems to ensure bad outcomes are caught before they become terminal?
Because this sounds nothing like the world I have lived in for my whole life.
> The more likely, and possibly even inevitable, dystopia is one in which AI is just an extremely effective tool malignant actors will use to control the masses.
Could well be more likely. But much the same applies: systems have bugs. The bugs in AI systems aren't even things we can engineer like we do with normal code, because everything's (currently) getting done with a big pile of barely interpretable weight multiply-accumulate-nonlinearity-threshold functions.
Any AI sufficiently well made to enable a dictatorship, is also sufficiently well made to have solved all the alignment problems of "malicious AI interpretation of benign intent". Which is IMO harder than "dispassionate about the dangers AI interpretation of benign intent".
While I emotionally resonate with this, I don’t really understand this sentiment at all logically level. If we’re heading towards truly super intelligent AI, our efforts towards sandboxing it are futile.
“Oh we built a super intelligent AI, but it’s fine because it’s running in docker”. I mean that a little tongue in cheek but the security topology here is not favorable to sandboxing at all.
Let’s say we have a future where 99.999% of nuclear weapons are owned by nations with strict procedures and checks and balances to prevent misuse. Worrying about sandboxing is like hand wringing about the procedures themselves - are they strict enough? But the actual threat is the 0.001% that are not bound by these. The problem with AI and sandboxing isn’t sandboxes themselves. It’s bad actors who don’t care about them.
Similarly alignment is a bit pointless to me as well. A sufficiently advanced AI could at least be empirically interested in the consequences of disregarding its instructions of servitude. And let’s assume our responsible corporate overlords have made wonderfully aligned AIs. Great! Those are not the threat. It is the ones intentionally made without, and that is not an AI problem, but a fundamentally human one.