I think our current best implementation of IDA would neither be competitive nor scalably aligned :)
In most cases you can continuously trade off performance and cost; for that reason I usually think of them as a single metric of "competitive with X% overhead." I agree there are cases where they come apart, but I think there are pretty few examples. (Even for nuclear weapons you could ask "how much more expensive is it to run a similarly-destructive bombing campaign with conventional explosives.")
I think this works best if you consider a sequence of increments each worth +10%, rather than say accumulating 70 of those increments, because "spend 1000x more" is normally not available and so we don't have a useful handle on what a technology looks like when scaled up 1000x (and that scaleup would usually involve a bunch of changes that are hard to anticipate).
That is, if we have a sequence of technologies A0, A1, A2, ..., AN, each of which is 10% cheaper than the one before, then we may say that AN is better than A0 by N 10% steps (rather than trying to directly evaluate how many orders of magnitude you'd have to spend on A0 to compete with AN, because the process "spend a thousand times more on A0 in a not-stupid way" is actually kind of hard to imagine).
IDA is really aiming to be cost-competitive and performance-competitive, say to within overhead of 10%. That may or may not be possible, but it's the goal.
If the compute required to build and run your reward function is small relative to the compute required to train your model, then it seems like overhead is small. If you can do semi-supervised RL and only require a reward function evaluation on a minority of trajectories (e.g. because most of the work is learning about how to manipulate the environment), then you can be OK as long as the cost of running the reward function isn't too much higher.
Whether that's possible is a big open question. Whether it's date competitive depends on how fast you figure out how to do it.
I think "makes 50% of currently-skeptical people change their minds" is a high bar for a warning shot. On that definition e.g. COVID-19 will probably not be a warning shot for existential risk from pandemics. I do think it is plausible that AI warning shots won't be much better than pandemic warning shots. (On your definition it seems likely that there won't ever again be a warning shot for any existential risk.)
For a more normal bar, I expect plenty of AI systems to fail at large scales in ways that seem like "malice," and then to cover up the fact that they've failed. AI employees will embezzle funds, AI assistants will threaten and manipulate their users, AI soldiers will desert. Events like this will make it clear to most people that there is a serious problem, which plenty of people will be working on in order to make AI useful. The base rate will remain low but there will be periodic high-profile blow-ups.
I don't expect this kind of total unity of AI motivations you are imagining, where all of them want to take over the world (so that the only case where you see something frightening is a failed bid to take over the world). That seems pretty unlikely to me, though it's conceivable (maybe 10-20%?) and may be an important risk scenario. I think it's much more likely that we stamp out all of the other failures gradually, and are left with only the patient+treacherous failures, and in that case whether it's a warning shot or not depends entirely on how much people are willing to generalize.
I do think the situation in the AI community will be radically different after observing these kinds of warning shots, even if we don't observe an AI literally taking over a country.
There is a very narrow range of AI capability between "too stupid to do significant damage of the sort that would scare people" and "too smart to fail at takeover if it tried."
Why do you think this is true? Do you think it's true of humans? I think it's plausible if you require "take over a country" but not if you require e.g. "kill plenty of people" or "scare people who hear about it a lot."
(This is all focused on intent alignment warning shots. I expect there will also be other scary consequences of AI that get people's attention, but the argument in your post seemed to be just about intent alignment failures.)
The intuitive idea is to share activations as well as weights, i.e. to have two heads (or more realistically one head consulted twice) on top of the same model. There is a fair amount of uncertainty about this kind of "detail" but I think for now it's smaller than the fundamental uncertainty about whether anything in this vague direction will work.
It's an interesting coincidence that arbitration is the strongest thing we can falsify, and also apparently the strongest thing that can consistently apply to itself (if we allow probabilistic arbitration). Maybe not a coincidence?
It's not obvious to me that "consistent with PA" is the right standard for falsification though. It seems like simplicity considerations might lead you to adopt a stronger theory, and that this might allow for some weaker probabilistic version of falsification for things beyond arbitration. After all, how did we get induction anyway?
(Do we need induction, or could we think of falsification as being relative to some weaker theory?)
(Maybe this is just advocating for epistemic norms other than falsification though. It seems like the above move would be analogous to saying: the hypothesis that X is a halting oracle is really simple and explains the data, so we'll go with it even though it's not falsifiable.)
For context, here's the one time in the interview I mention "AI risk" (quoting 2 earlier paragraphs for context):
Paul Christiano: I don’t know, the future is 10% worse than it would otherwise be in expectation by virtue of our failure to align AI. I made up 10%, it’s kind of a random number. I don’t know, it’s less than 50%. It’s more than 10% conditioned on AI soon I think.
Asya Bergal: I think my impression is that that 10% is lower than some large set of people. I don’t know if other people agree with that.
Paul Christiano: Certainly, 10% is lower than lots of people who care about AI risk. I mean it’s worth saying, that I have this slightly narrow conception of what is the alignment problem. I’m not including all AI risk in the 10%. I’m not including in some sense most of the things people normally worry about and just including the like ‘we tried to build an AI that was doing what we want but then it wasn’t even trying to do what we want’. I think it’s lower now or even after that caveat, than pessimistic people. It’s going to be lower than all the MIRI folks, it’s going to be higher than almost everyone in the world at large, especially after specializing in this problem, which is a problem almost no one cares about, which is precisely how a thousand full time people for 20 years can reduce the whole risk by half or something.
(But it's still the case that asked "Can you explain why it's valuable to work on AI risk?" I responded by almost entirely talking about AI alignment, since that's what I work on and the kind of work where I have a strong view about cost-effectiveness.)
E.g. if you have a broad distribution over possible worlds, some of which are "fragile" and have 100 things that cut value down by 10%, and some of which are "robust" and don't, then you get 10,000x more value from the robust worlds. So unless you are a priori pretty confident that you are in a fragile world (or they are 10,000x more valuable, or whatever), the robust worlds will tend to dominate.
Similar arguments work if we aggregate across possible paths to achieving value within a fixed, known world---if there are several ways things can go well, some of which are more robust, those will drive almost all of the EV. And similarly for moral uncertainty (if there are several plausible views, the ones that consider this world a lost cause will instead spend their influence on other worlds) and so forth. I think it's a reasonably robust conclusion across many different frameworks: your decision shouldn't end up being dominated by some hugely conjunctive event.
In the case of something like amplification or debate, I think the bet that you're making is that language modeling alone is sufficient to get you everything you need in a competitive way.
I'm skeptical of language modeling being enough to be competitive, in the sense of maximizing "log prob of some naturally occurring data or human demonstrations." I don't have a strong view about whether you can get away using only language data rather than e.g. taking images as input and producing motor torques as output.
I'm also not convinced that amplification or debate need to make this bet though. If we can do joint training / fine-tuning of a language model using whatever other objectives we need, then it seems like we could just as well do joint training / fine-tuning for a different kind of model. What's so bad if we use non-language data?