# Ω 50

Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

Clutching a bottle of whiskey in one hand and a shotgun in the other, John scoured the research literature for ideas... He discovered several papers that described software-assisted hardware recovery. The basic idea was simple: if hardware suffers more transient failures as it gets smaller, why not allow software to detect erroneous computations and re-execute them? This idea seemed promising until John realized THAT IT WAS THE WORST IDEA EVER. Modern software barely works when the hardware is correct, so relying on software to correct hardware errors is like asking Godzilla to prevent Mega-Godzilla from terrorizing Japan. THIS DOES NOT LEAD TO RISING PROPERTY VALUES IN TOKYO. It’s better to stop scaling your transistors and avoid playing with monsters in the first place, instead of devising an elaborate series of monster checks-and-balances and then hoping that the monsters don’t do what monsters are always going to do because if they didn’t do those things, they’d be called dandelions or puppy hugs.

- James Mickens, The Slow Winter

There’s a lot of AI alignment strategies which can reasonably be described as “ask Godzilla to prevent Mega-Godzilla from terrorizing Japan”. Use one AI to oversee another AI. Have two AIs debate each other. Use one maybe-somewhat-aligned AI to help design another. Etc.

Alignment researchers discuss various failure modes of asking Godzilla to prevent Mega-Godzilla from terrorizing Japan. Maybe one of the two ends up much more powerful than the other. Maybe the two make an acausal agreement. Maybe the Nash Equilibrium between Godzilla and Mega-Godzilla just isn’t very good for humans in the first place. Etc. These failure modes are useful for guiding technical research.

… but I worry that talking about the known failure modes misleads people about the strategic viability of Godzilla strategies. It makes people think (whether consciously/intentionally or not) “well, if we could handle these particular failure modes, maybe asking Godzilla to prevent Mega-Godzilla from terrorizing Japan would work”.

What I like about the Godzilla analogy is that it gives a strategic intuition which much better matches the real world. When someone claims that their elaborate clever scheme will allow us to safely summon Godzilla in order to fight Mega-Godzilla, the intuitively-obviously-correct response is “THIS DOES NOT LEAD TO RISING PROPERTY VALUES IN TOKYO”.

“But look!” says the clever researcher, “My clever scheme handles problems X, Y and Z!”

Response:

“Ok, but what if we had a really good implementation?” asks the clever researcher.

Response:

“Oh come on!” says the clever researcher, “You’re not even taking this seriously! At least say something about how it would fail.”

Don’t worry, we’re going to get to that. But before we do: let’s imagine you’re the Mayor of Tokyo evaluating a proposal to ask Godzilla to fight Mega-Godzilla. Your clever researchers have given you a whole lengthy explanation about how their elaborate and clever safeguards will ensure that this plan does not destroy Tokyo. You are unable to think of any potential problems which they did not address. Should you conclude that asking Godzilla to fight Mega-Godzilla will not result in Tokyo’s destruction?

No. Obviously not. THIS DOES NOT LEAD TO RISING PROPERTY VALUES IN TOKYO. You may not be able to articulate why the answer is obviously “no”, but asking Godzilla to fight Mega-Godzilla will still obviously destroy Tokyo, and your intuitions are right about that even if you are unable to articulate clever arguments.

With that said, let’s talk about why those intuitions are right and why the Godzilla analogy works well.

## Brittle Plans and Unknown Unknowns

The basic problem with Godzilla plans is that they’re brittle. The moment anything goes wrong, the plan shatters, and then you’ve got somewhere between one and two giant monsters rampaging around downtown.

And of course, it is a fundamental Law of the universe that nothing ever goes exactly according to plan. Especially when trying to pit two giant monsters against each other. This is the sort of situation where there will definitely be unknown unknowns.

Unknown unknowns + brittle plan = definitely not rising property values in Tokyo.

Do we know what specifically will go wrong? No. Will something go wrong? Very confident yes. And brittleness means that whatever goes wrong, goes very wrong. Errors are not recoverable, when asking Godzilla to fight Mega-Godzilla.

If we use one AI to oversee another AI, and something goes wrong, that’s not a recoverable error; we’re using AI assistance in the first place because we can’t notice the relevant problems without it. If two AIs debate each other in hopes of generating a good plan for a human, and something goes wrong, that’s not a recoverable error; it’s the AIs themselves which we depend on to notice problems. If we use one maybe-somewhat-aligned AI to build another, and something goes wrong, that’s not a recoverable error; if we had better ways to detect misalignment in the child we’d already have used them on the parent.

The real world will always throw some unexpected problems at our plans. When asking Godzilla to fight Mega-Godzilla, those problems are not recoverable. THIS DOES NOT LEAD TO RISING PROPERTY VALUES IN TOKYO.

Meta note: I expect this post to have a lively comment section! Before you leave the twentieth comment saying that maybe Godzilla fighting Mega-Godzilla is better than Mega-Godzilla rampaging unchallenged, maybe check whether somebody else has already written that one, so I don't need to write the same response twenty times. (But definitely do leave that comment if you're the first one, I intentionally kept this essay short on the assumption that lots of discussion would be in the comments.)

# Ω 50

New Comment

He discovered several papers that described software-assisted hardware recovery. The basic idea was simple: if hardware suffers more transient failures as it gets smaller, why not allow software to detect erroneous computations and re-execute them? This idea seemed promising until John realized THAT IT WAS THE WORST IDEA EVER. Modern software barely works when the hardware is correct, so relying on software to correct hardware errors is like asking Godzilla to prevent Mega-Godzilla from terrorizing Japan. THIS DOES NOT LEAD TO RISING PROPERTY VALUES IN TOKYO.

I happen to work for a company whose software uses checksums at many layers, and RAID encoding and low-density parity codes at the lowest layers, to detect and recover from hardware failures.  It works pretty well, and the company has sold billions of dollars of products of which that is a key component.  Also, many (most?) enterprise servers use RAM with error-correcting codes; I think the common configuration allows it to correct single-bit errors and detect double-bit errors, and my company's machines will reset themselves when they detect double-bit errors and other problems that impugn the integrity of their runtime state.

One could quibble about whether "retrieving and querying the data that was written" counts as a "computation", and the extent to which the recovery is achieved through software as opposed to hardware[1], but the source material is a James Mickens comedic rant in any case.

I'd say the important point here is: There is a science to error correction, to building a (more) perfect machine out of imperfect parts, where the solution to unreliable hardware is more of the unreliable hardware, linked up in a clever scheme.  They're good enough at it that each successive generation of data storage technology uses hardware with higher error rates.  You can make statements like "If failures are uncorrelated, and failures happen every X time units on average per component, and it takes Y time units to detect and recover a failure, and we can recover from up to M failures out of every group of N components, then on average we will have an unrecoverable failure every Z time units"; then you can (a) think about how to arrange it so that Z >> X and (b) think about the dangers of correlated failures.

(The valid complaint that Mickens's character makes is that it would suck if every application needed to weave error correction into every codepath, implement its own RAID, etc.  It works much better if the error correction is done by some underlying layer that the application treats as an abstraction—using the abstraction tends to be more complex than pretending errors don't exist (and for noncritical applications the latter is a valid strategy), but not terrible.[2])

With regard to AI.  It seems likely that we'll end up making use of potentially-dangerous AIs to do things.  If we do, then we'd want powerful safeguards.  It seems unlikely that we'd have 100% confidence in any particular safeguard (i.e. unlikely we'd have formal proofs of the safety properties we'd want).  Then we'd want to keep adding more safeguards as long as their failure modes weren't totally covered by those of preexisting safeguards (and as long as they're affordable); ideally we'd try to estimate (ideally measure) the chance of failure of each safeguard and how independent they are.

Now, would some of these safeguards be built with the aid of earlier versions of AI?  I'd guess yes.  This could take a few forms, ranging from "formally-verified direct programs that the earlier AI generated along with their proofs of safety and effectiveness" to "smaller neural nets (cut-down versions of earlier AIs?) that monitor the training of the new AI's net" to "weaker full-blown AIs watching the decisions output by the new full-blown AI".  I suspect any of these might be useful.

The Godzilla complaint seems to be about that last thing, having weaker AIs monitor the new AI.  Well, um, it seems to me that, if the weaker AIs' output is restricted to the one-bit signal "do we hit the killswitch on the new AI" (and that signal is executed by a machine), then there's no danger of Godzilla destroying Tokyo after it has taken down Mega-Godzilla.  (It's like cutting off Godzilla's limbs and tying him to Mega-Godzilla with his jaws open around Mega-Godzilla's neck, so his only available action is biting down.)  There is a danger, if this happens a lot, of the subsequent training of the new AI merely teaching it to evade detection by the weaker predecessors; this danger should be taken seriously; but it is also less bad than having Mega-Godzilla kill us the first time it started going rogue; the safeguard is net positive.

If we use one AI to oversee another AI, and something goes wrong, that’s not a recoverable error; we’re using AI assistance in the first place because we can’t notice the relevant problems without it.

If, among our many precautions, we have AI 1 oversee AI 2 (in this scheme I describe where AI 1's only available action is to output a bit that shuts off AI 2)[3], and AI 2 starts to go rogue, and AI 1 fails in its task, then we're no worse off than if we lacked that particular precaution—either the other precautions save us or they don't.  If AI 1 succeeds, then we have at least some chance of being able to debug what went wrong with AI 2, and save ourselves in the future.  Using AI 1 as overseer is strictly beneficial, minus implementation costs.  (There's a second-order risk of someone overestimating how much it protects us and then choosing to be more reckless, but that's a different kind of thing.)

If we use one maybe-somewhat-aligned AI to build another, and something goes wrong, that’s not a recoverable error

If "goes wrong" means "it kills us all", then yeah.  If "goes wrong" means "it starts acting dangerous but we manage to shut it off", that is potentially recoverable.

if we had better ways to detect misalignment in the child we’d already have used them on the parent.

The parent could be aligned but incompetent (heh, I think many people know parents like that), while the child is unaligned and more competent.  (Why would the programmers create the child if it weren't expected to be more competent than the parent?)  It seems easy to imagine alignment checks that the parent would pass while the child would fail.

1. ^

The checksums are done by the CPU but hardware-accelerated; I think the RAID stuff is clearly software; the LDPC stuff is done by more specialized hardware; the ECC RAM is hardware, but the choice of how to respond to double bit errors is software.

2. ^

ECC RAM is one example: the application can just assume that memory accesses are correct, and that double-bit errors will cause the operating system to kill the application—so the application itself needs zero lines of code about RAM errors, though now there does need to be monitoring to restart the process and/or fail over to another machine.

Another example is using a durable database, and going from "just increment and decrement the variables in memory" to "create a database transaction, to be attempted / rolled back as a unit".  That does require code changes in the application itself.

3. ^

In practice we'd probably want it to output something more like a number between 0 and 1 representing its worry level, and likely the killswitch goes off if either a single overseer exceeds some danger threshold or multiple overseers exceed some lower threshold.

One important difference between data storage vs computation or AI: courtesy of Shannon and Hamming, we have a really good understanding of information transmission (which includes information storage). All those nice error-correction codes are downstream of very well-understood theory.

If we had theory as solid as information theory for AI and alignment, then yeah, I'd be a hell of a lot more optimistic about using one AI to oversee another somewhere in the process. Like, imagine we had the alignment analogue of an error-detecting code which provably detects two-bit errors and corrects one-bit errors with only a logarithmic amount of overhead. With theory that strong (and battle-tested in reality) it becomes plausible that unknown unknowns won't inevitably ruin all our plans.

Well, the basic idea "adding more safeguards decreases the likelihood they'll all fail simultaneously, as long as there isn't a perfect correlation of failure modes" is a simple mathematical fact.  "What is the probability of this safeguard failing to detect a rogue AI?" is hard to answer, but "What might this new safeguard do that the other safeguards don't do?" is easier.

For example.  If interpretability work gets anywhere, then one might imagine a suite of safeguards that check for parts of the developing neural net that compute things like "how to detect security holes in C or machine code" or "how quickly humans die to certain poisons" (when that's not supposed to be the goal); safeguards that check for parts of the net that have many nodes and are not understandable by the other safeguards; safeguards that inspect the usage of CPU or other resources and have some idea of what's usual; safeguards that try to look for the net thinking strategically about what resource usage looks natural; and so on.  These safeguards might all suck / only work in a small fraction of cases, but if you have hundreds or thousands of them, then your odds might get decent.

Or, at least, you might push the correlated failure mode into "the AI has guessed that we have a bunch of safeguards, has imagined the set of safeguards people at our level of competence are likely to have come up with, and is strategically avoiding them", in which case that strategic thinking is a thing you try to detect.  (It would be nice if we ended up being able to prove that the strategic thinking to defeat all our "dangerous thinking detectors" required at least N neurons, and we were able to prune down all nets to the point where less than N not-understood neurons existed.)  And—as stated, each of these safeguards might be implemented with the assistance of prior AIs in various ways.

What exactly are you objecting to?  Your original post seems to be saying "the entire category of strategies that involves one AI saving us from another is dangerous, doomed, and should be dismissed as ridiculous", but you don't seem to be pushing back on my claims that you can cut off Godzilla's limbs and tie his jaws around Mega-Godzilla's neck and that the result is net positive.  Were you arguing against much more reckless proposals, or people being overconfident in unspecified proposals, or...?  The possibility occurs to me that the OP is intentionally overstated in the hopes of provoking a well-thought-out proposal (it did say it anticipated much discussion), along the lines of Cunningham's Law.

I am mostly objecting to strategies which posit one AI saving us from another as the primary mechanism of alignment - for instance, most of the strategies in 11 Proposals. If we had sufficiently great interpretability, then sure, we could maybe leverage that to make a Godzilla strategy with a decent chance of working (or at least failing in detectable-in-advance ways), but with interpretability tools that good we could probably just make a plan without Godzilla have a decent chance of working (or at least failing in detectable-in-advance ways) by doing basically the same things minus Godzilla. It's the interpretability tools which take that plan from "close to zero chance of working" to "close to 100% chance of working"; the interpretability is where all the robustness comes from. The Godzilla part adds relatively little and is plausibly net negative (due to making the ML components more complex and brittle).

(Another minor point: "adding more safeguards decreases the likelihood they'll all fail simultaneously, as long as there isn't a perfect correlation of failure modes" is only true when the "safeguards" are guaranteed to not increase the chance of failure.)

And—as stated, each of these safeguards might be implemented with the assistance of prior AIs in various ways.

I see two main things you could have in mind here.

First, maybe you imagine training a magic black box to tell us what's going on inside another magic black box. This is a Godzilla strategy, and fails for the usual Godzilla reasons: errors are not recoverable. If the magic interpretability box fails, we don't have a built-in way to notice. And no, training lots of magic black boxes to detect lots of things does not really fix the problem; the failure modes are extremely highly correlated. We don't even need to suppose deception - just a distribution shift in the cognition of the system will cause highly correlated failures, and a distribution shift in cognition is exactly the sort of thing we'd expect from a system just starting to grok consequentialist reasoning.

On the other hand, maybe you imagine AIs serving as research assistants, rather than using AIs directly to interpret other AIs. That plan does have problems, but is basically not a Godzilla plan; the human in the loop means that the standard Godzilla brittleness issue doesn't really apply.

My reply to the top-level post here is also relevant as a reply to this specific comment.

Individual humans do make off much better when they get to select between products from competing companies rather than monopolies, benefitting from companies going out of their way to demonstrate when their products are verifiably better than rivals'. Humans get treated better by sociopathic powerful politicians and parties when those politicians face the threat of election rivals (e.g. no famines). Small states get treated better when multiple superpowers compete for their allegiance. Competitive science with occasional refutations of false claims produces much more truth for science consumers than intellectual monopolies. Multiple sources with secret information are more reliable than one.

It's just routine for weaker less sophisticated parties to do better in both assessment of choices and realized outcomes when multiple better informed or powerful parties compete for their approval vs just one monopoly/cartel.

Also, a flaw in your analogy is that schemes that use AIs as checks and balances on each other don't mean more AIs. The choice is not between monster A and monsters A plus B, but between two copies of monster A (or a double-size monster A), and a split of one A and one B, where we hold something of value that we can use to help throw the contest to either A or B (or successors further evolved to win such contests). In the latter case there's no more total monster capacity, but there's greater hope of our influence being worthwhile and selecting the more helpful winner (which we can iterate some number of times).

So, the analogy here is that there's hundreds (or more) of Godzillas all running around, doing whatever it is Godzillas want to do. Humanity helps out whatever Godzillas humanity likes best, which in turn creates an incentive for the Godzillas to make humanity like them.

THIS DOES NOT BODE WELL FOR TOKYO'S REAL ESTATE MARKET.

Still within the analogy: part of the literary point of Godzilla is that humanity's efforts to fight it are mostly pretty ineffective. In inter-Godzilla fights, humanity is like an annoying fly buzzing around. The humans just aren't all that strategically relevant. Sure, humanity's assistance might add some tiny marginal advantage, but from a Godzilla's standpoint that advantage is unlikely to be enough to balance the tactical/strategic disadvantages of trying not to step on people.

... and that all seems like it should carry over directly to AI, once AI gets to-or-somewhat-past human level, and definitely by the time we get to strongly superhuman intelligence. Even with just human level, the scaling/coordination/learning advantages of being able to cheaply copy a mind are probably enough for the AIs to reasonably-quickly achieve strategic dominance by enough margin that humanity's preferences are not particularly relevant. (Assuming that the AI isn't prohibitively expensive to run - but that seems pretty likely to be true under most plausible paths. For instance, if human-level AI is produced by anything like today's ML, then training costs will dominate and the systems will be relatively cheap to run or fine-tune.)

(There's also some alignment-specific problems with this scheme which the Godzilla analogy doesn't highlight. I'm not going into them here because this post is specifically about the Godzilla issues. But I don't want to give people the impression that this plan would be fine in a world where humanity has sufficient bargaining power; the lack of bargaining power is only one failure mode.)

I was going to make a comment to the effect that humans are already a species of Godzilla (humans aren't safe, human morality is scary, yada yada), only to find you making the same analogy, but with an optimistic slant. :)

Competition between the powerful can lead to the ability of the less powerful to extract value.  It can also lead to the less powerful being more ruthlessly exploited by the powerful as a result of their competition.  It depends on the ability to the less powerful to choose between the more powerful.  I am not confident humanity or parts of it will  have the ability to choose between competing AGIs.

This happens during fine-tuning training already, selecting for weights that give the higher human-rated response of two (or more) options. It's a starting point that can be lost later on, but we do have it now with respect to configurations of weights giving different observed behaviors.

James Mickens is writing comedy. He worked in distributed systems. A "distributed system" is another way to say "a scenario in which you absolutely will have to use software to deal with your broken hardware". I can 100% guarantee that this was written with his tongue in his cheek.

The modern world is built on software that works around HW failures.

• You likely have ECC ram in your computer.
• There are checksums along every type of data transfer (Ethernet frame check sequences, IP header checksums, UDP datagram checksums, ICMP checksums, eMMC checksums, cryptographic auth for tokens or certificates, etc).
• An individual SSD or HDD have algorithms for detecting and working around failed blocks / sectors in HW.
• There are fully redundant processors in safety-critical applications using techniques like active-standby, active-active, or some manner of voting for fault tolerance.
• In anything that involves HW sensors, there's algorithms like an extended Kalman filter for combining the sensor readings to a single consistent view of reality, and stapled to that are algorithms for determining when sensors are invalid because they've railed high, railed low, or otherwise failed in a manner that SW can detect.
• Your phone's WiFi works because the algorithm used for the radio is constantly working around dropouts and reconnecting to new sources as needed.
• We can read this post because it's sent using TCP and is automatically retransmitted as many times as needed until it's been ACK'd successfully.
• We can play multiplayer video games because they implement eventually consistent protocols on top of UDP.
• Almost all computer applications implement some form of error handling + retry logic for pretty much anything involving I/O (file operations, network operations, user input) because sometimes things fail, and almost always, retrying the thing that failed will work.
• Large data centers have hundreds of thousands of SSDs and they are constantly failing -- why doesn't Google fall over? Because SW + HW algorithms like RAID compensate for drives dying all of the time.

If we use one AI to oversee another AI, and something goes wrong, that’s not a recoverable error; we’re using AI assistance in the first place because we can’t notice the relevant problems without it. If two AIs debate each other in hopes of generating a good plan for a human, and something goes wrong, that’s not a recoverable error; it’s the AIs themselves which we depend on to notice problems. If we use one maybe-somewhat-aligned AI to build another, and something goes wrong, that’s not a recoverable error; if we had better ways to detect misalignment in the child we’d already have used them on the parent.

Replace "AI" with "computer" in this paragraph and it is obviously wrong because every example here is under-specified. There is a dearth of knowledge on this forum of anything resembling traditional systems engineering or software system safety and it shows in this thread and in the previous thread you made about air conditioners. I commented as such here.

"If we use one computer to oversee another computer, and something goes wrong, that's not a recoverable error; we're using computer assistance in the first place because we can't notice the relevant problems without it."

Here are some examples off the top of my head where we use one computer to oversee another computer:

1. It's common to have one computer manage a pool of workers where each worker is another computer and workers may fail. The computer doing the management is able to detect a stalled or crashed worker, power cycle the hardware, and then resubmit the work. Depending on the criticality of this process, the "manager" might actually be multiple computers that work synchronously. The programming language Erlang is designed for this exact use-case -- distributed, fault tolerance SW applications in contexts where I/O is fallible and it's unacceptable for the program to crash.
2. We often use one computer program to calculate some type of path or optimal plan and it's a very complicated program to understand, and then we use a 2nd computer program to validate the outputs from the 1st program -- why do we use two programs? Because the first is inscrutable and difficult to explain, but the 2nd reads like straightforward requirements in English. In other words, it is often far easier to check a solution than it is to create a solution. The mathematically inclined will recognize this as a consequence of P vs NP problems, if P != NP.
3. It's common in safety-critical applications to have a fail-safe or backup using a significantly less complicated architecture -- e.g. you might use <complicated system> to do <some complex task>, but for a fail-safe like "power off everything", you might have a tiny microprocessor sitting in-line with the main power rail serving as a glorified switch. So normally the <complicated system> is driving the show, but if that starts to go sideways, the tiny microprocessor can shut it down.
4. Almost all microprocessors have a "watchdog" built into them. A watchdog is a secondary processor that will reset the primary processor if the primary is non-responsive. Have you ever seen your Android phone mysteriously reboot when the UI locks up? That was a watchdog.
5. The "watchdog" concept is even used in pure SW contexts, e.g. when the Android OS kills an application on your phone because it has frozen, that's a one computer program (the OS) overseeing another (the application). Ditto for Windows & the task manager.
6. We often use "voting" where we run the same SW on 3 or more systems and then only output to some downstream hazard if all 3 systems agree. If they don't agree, we can fail-safe or try to recover, e.g. by power-cycling whichever system was out-of-family, or both -- first try to recover, then fail-safe if that didn't work. This is done by running code in "lockstep" on synchronized inputs, very similar to how old multiplayer RTS games used to do networking.
7. You can buy self-checking processors that implement lockstep comparison between 2 internal cores so that whenever instructions are executed, you know that the execution occurred identically across both cores.

These aren't cherry-picked. This is the bread & butter of systems safety. We build complex, safe systems by identifying failure modes and then using redundant systems to either tolerate faults or to fail-safe. By focusing on the actual system, and the actual failure modes, and by not getting stuck with our head in the clouds considering a set of "all possible hypothetical systems", it is possible to design & implement robust, reliable solutions in reality.

To claim it is not just impossible to do that, but that it is foolhardy to even try, is the exact opposite of a safety-critical mindset.

I agree that the SW/HW analogy is not a good analogy for AGI safety (I think security is actually a better analogy), but I would like to present a defence of the idea that normal systems reliability engineering is not enough for alignment (this is not necessarily a defence of any of the analogies/claims in the OP).

Systems safety engineering leans heavily on the idea that failures happen randomly and (mostly) independently, so that enough failures happening together by coincidence to break the guarantees of the system is rare. That is:

• RAID is based on the assumption that hard drive failures happen mostly independently, because the probability of too many drives failing at once is sufficiently low. Even in practice this assumption becomes a problem because a) drives purchased in the same batch will have correlated failures and b) rebuilding an array puts strain on the remaining drives, and people have to plan around this by adding more margin of error.
• Checksums and ECC are robust against the occasional bitflip. This is because occasional bitflips are mostly random and getting bitflips that just happen to set the checksum correctly are very rare. Checksums are not robust against someone coming in and maliciously changing your data in-transit, you need signatures for that. Even time correlated runs of flips can create a problem for naive schemes and burn through the margin of error faster than you'd otherwise expect.
• Voting between multiple systems assumes that the systems are all honest and just occasionally suffer transient hardware failures. Clean room reimplementations are to try and eliminate the correlations due to bugs, but they still don't protect against correlated bad behaviour across all of the systems due to issues with your spec.

My point here is that once your failures stop being random and independent, you leave the realm of safety engineering and enter the realm of security (and security against extremely powerful actors is really really hard). I argue that AGI alignment is much more like the latter, because we don't expect AGIs to fail in random ways, but rather we expect them to intelligently steer the world into directions we don't want. AGI induced failure looks like things that should have been impossible when multiplying out the probabilities somehow happening regardless.

In particular, relying on independent AGIs not being correlated with each other is an extremely dangerous assumption: AGIs can coordinate even without communication, alignment is a very narrow target that's hard to hit, and a parliament of misaligned AGIs is definitely not going to end well for us.

If two AIs debate each other in hopes of generating a good plan for a human, and something goes wrong, that’s not a recoverable error;

I alluded to this above in many examples, but let's just do a theoretical calculation as well so it's not just anecdotes.

Suppose we have some AI "Foo" that has some probability of failure, or P(Failure of Foo).

Then the overall probability of the system containing Foo failing is P(Failure of System) == P(Failure of Foo).

On the other hand, suppose we have some AI "Bar" whose goal is to detect when AI "Foo" has failed, e.g. when Foo erroneously creating a plan that would harm humans or attempt to deceive them.

We can now calculate the new likelihood of P(Failure of System) == P(Failure of Foo) * P(Failure of Bar), where P(Failure of Bar) is the likelihood that either Bar failed to detect the issue with Foo, or that Bar successfully detected the issue with Foo, but failed to notify us.

These probabilities can be related in some way, but they don't have to be. It is possible to drastically reduce the probability of a system failing by adding components within that system, even if those new components have chances of failure themselves.

In particular, so long as the requirement allocated to Bar is narrow enough, we can make Bar more reliable than Foo, and then lower the overall chance of the system failing. One way this works is by limiting Bar's functionalities so that if Bar failed, in isolation of Foo failing, the system is unaffected. In the context of fault tolerance, we'd refer to that as a one-fault tolerant system. We can tolerate Foo failing -- Bar will catch it. And we can tolerate Bar failing -- it doesn't impact the system's performance. We only have an issue if Foo failed and then Bar subsequently also failed.

Ok, but why isn't it better to have Godzilla fighting Mega-Godzilla instead of leaving Mega-Godzilla unchallenged?

Because Tokyo still gets destroyed.

Important thing to bear in mind here: the relevant point for comparison is not the fantasy-world where the Godzilla-vs-Mega-Godzilla fight happens exactly the way the clever elaborate scheme imagined. The relevant point for comparison is the realistic-world where something went wrong, and the elaborate clever scheme fell apart, and now there's monsters rampaging around anyway.

I would file this under

Some people, when confronted with a problem Mega-Godzilla, think “I know, I'll use regular expressions Godzilla.” Now they have two problems rampaging monsters.

Epistemic Status: Low.  Very likely wrong but would like to understand why.

It seems easier to intent align a human level or slightly above human level AI (HLAI) than a massively smarter than human AI.

Some new research options become available to us once we have aligned HLAI, including:

• The HLAI might be able to directly help us do alignment research and solve the general alignment problem.
• We could run experiments on the HLAI and get experimental evidence much closer to the domain we are actually trying to solve.
• We could use the HLAI to start a training procedure, a la IDA.

These schemes seem fragile, because 1) if any HLAIs are not aligned, we lose, and 2) if the training up to superintelligence process fails, due to some unknown unknown or the HLAI being misaligned or through any of the known failure modes, we lose.

However, 1) seems like a much easier problem than aligning an arbitrary intelligence AI. Even though something could likely go wrong aligning a HLAI, it also seems likely that something goes wrong if we try to align an arbitrary intelligence AI. (This seems related to security mindset... in the best case world we do just solve the general case of alignment, but that seems hard.)

For 2), the process of training up to superintelligence seems like a HLAI would help more than it hurts. If the HLAI is actually intent aligned, this seems like having a fully uploaded alignment researcher, which seems less like getting Godzilla to fight and more like getting a Jaeger to protect Tokyo.

The relevant point isn't "the realistic world, where the clever scheme fell apart", the relevant point is "the realistic world, where there is some probability of the clever scheme falling apart, and you need to calculate the expectation of that probability, and that expectation could conceivably go down when you add Godzilla".

Or to put it another way, even if the worst case is as bad, the average case could still be better. Analyzing the situation in terms of "what if the clever plan fails" is looking only at the worst case.

In-universe, Mecha-Godzilla had to be built with a Godzilla-skeleton, which caused both to turn against Humanity.

It feels probable that there will be substantial technical similarities between Production Superintelligences and Alignment Superintelligences, which could cause both of them to turn against us.
(Epistemic state: Low confidence)

This post is one more addition to the worrying trend in LW that asks for black and white solutions as it there were no middle ground. Would you say that having no army is better than having an army at all? I would feel more comfortable knowing that we have Godzilla in our side than having nothing

This. A lot of the blame goes to MIRI viewing AI Alignment discretely, rather than continuously, as well as a view that only heroic or pivotal acts save the world. This tends to be all or nothing, and generates all-or-nothing views.

[This comment is no longer endorsed by its author]Reply

I really wish David Chapman and his ideas were a more active part of this discussion.

Can you give some context?

David Chapman talks about ways of thinking and is influenced by Buddhism and LW-style rationality. I've read his website-book "Meaningness" and I'm starting to read his new website-book "In the Cells of the Eggplant". His twitter has a link to this page which seems like the right place to start reading his work. He would describe EY's way of thinking as "rationalist eternalism" and "fixated".

(He should not be confused with the guy who shot John Lennon.)

The perfect has become the enemy of the good.

One thing I can end up worrying about is that useful tricks get ignored due to a dynamic of:

1. A person tries to overextend the useful trick beyond its range of applicability such that it turns into a godzilla strategy
2. Everyone starts associates the trick with the godzilla strategy
3. People don't consider using the trick within the range where it is actually applicable

For instance, consider debate. Debate is not magic and there's lots of things it can't do. But (constructively understood) logical operators such as "for all" and "exists" can be given meaning using a technique called "game semantics", and "debate" seems like a potential way to implement this in AI.

Can this do even a fraction of the things that people want debate to do? No. Can I think of anything that needs these game semantics? Not right now, no. But is it a tool that seems potentially powerful for the future? Yeah, I'd say so; it expands the range of things we can express, should we ever find a case where we want to express it, and so it is a good idea to be ready to deploy it.

There's an easy solution to this: just say that some class of tricks seems potentially useful, and explore what it can be used for, without proposing solutions. There's no need to immediately jump to proposing solutions all the time.

But of course you can use software to mitigate hardware failures, this is how Hadoop works! You store 3 copies of every data, and if one copy gets corrupted, you can recover the true value. Error-correcting codes is another example in that vein. I had this intuition, too, that aligning AIs using more AIs will obviously fail, now you made me question it.

I had this intuition, too, that aligning AIs using more AIs will obviously fail, now you made question it.

That is also progress.

I initially liked this post a lot, then saw a lot of pushback in the comments, mostly of the (very valid!) form of "we actually build reliable things out of unreliable things, particularly with computers, all the time". I think this is a fair criticism of the post (and choice of examples/metaphors therein), but I think it may be missing (one of) the core message(s) trying to be delivered.

I wanna give an interpretation/steelman of what I think John is trying to convey here (which I don't know whether he would endorse or not):

"There are important assumptions that need to be made for the usual kind of systems security design to work (e.g. uncorrelation of failures). Some of these assumptions will (likely) not apply with AGI. Therefor, extrapolating this kind of thinking to this domain is Bad™️." ("Epistemological vigilance is critical")

So maybe rather than saying "trying to build robust things out of brittle things is a bad idea", it's more like "we can build robust things out of certain brittle things, e.g. computers, but Godzilla is not a computer, and so you should only extrapolate from computers to Godzilla if you're really, really sure you know what you're doing."

I am not saying that alignment is easy to solve, or that failing it would not result in catastrophe. But all these arguments seem like universal arguments against any kind of solution at all. Just because it will eventually involve some sort of Godzilla. It is like somebody tries to make a plane that can fly safely and not fall from the Sky, and somebody keeps repeating "well, if anything goes wrong in your safety scheme, then the plane will fall from the Sky" or "I notice that your plane is going to fly in the Sky, which means it can potentially fall from it".

I am not saying that I have better ideas about checking whether any plan will work or not. They all inevitably involve Godzilla or Sky. And the slightest mistake might cost us our lives. But I don't think that pointing repeatedly at the same scary thing, which will be one way or the other in every single plan, will get us anywhere.

I expect there are ways of dealing with Godzilla which are a lot less brittle.

If we have excellent detailed knowledge of Godzilla's internals and psychology, we know what sort of things will drive Godzilla into a frenzy or slow him down or put him to sleep, we know how to get Godzilla to go in one direction rather than another, if we knew when and how tests on small lizards would generalize to Godzilla... those would all be robustly useful things. If we had all those pieces plus more like them, then it starts to look like a scenario where dealing with Godzilla is basically viable. There's lots of fallback options, and many opportunities to recover from errors. It's not a brittle situation which falls apart as soon as something goes wrong.

This seems to contradict what I interpreted as the message of your post; that message being, if someone gives you a "clever" strategy for dealing with Godzilla, the correct response is to just troll them because Godzilla is inherently bad for property values. But what you're doing now is admitting that if the scheme to control Godzilla is clever in such and such ways, which you specifically warned against, then actually it might not be so brittle.

The key distinction is between clever methods for controlling something one does not understand, vs clever methods for controlling something one does understand. (The post didn't go into that because it was short rather than thorough, but it did come up elsewhere in the comments.)

It suggests putting more weight on a plan to get AI Research globally banned.   I am skeptical that this will work (though if burning all GPUs would be a pivotal act the chances of success are significantly higher), but it seems very unlikely that there is a technical solution either.

In addition, at least some  purported technical solutions to AI risk seem to meaningfully increase the risk to humanity.  If you have someone creating an AGI to exercise sufficient control over the world to execute a pivotal act, that raises the stakes of being first enormously which incentivizes cutting corners.  And, it also makes it more likely that the AGI will destroy humanity and be quicker to do so.

The non-straw versions of Godzilla Strategies do not start from the Godzilla fighting Mega-Godzilla. Starting from this side is doomed.

It starts with, let's say, a Tokyo policeman. Notably, Tokyo policeman isn't a scary monster - but roughly a normal human, where you can get some sort of mutual understanding. The next step is to create a policeman[1], who also isn't a scary monster, but is just a bit more powerful, trained policeman (maybe using a bunch of policeman[0])Where, if the relation gen[n+1] is doing what gen[n] wants holds, the idea is you get to super-Tokio-police, who is still doing what you want. Or you get somewhere midway, where the still aligned policeman[p] tells you "sorry, the next gen would really be a Godzilla, and I don't know how to avoid it".

(This isn't to express opinions on the viability of the first step, or the amplification procedure.)

Alright, so, let's imagine a chain of 100... creatures... on a smooth spectrum from policeman to Godzilla, and each is trying to keep the next creature up the chain in check. And then the mayor attempts to direct Godzilla via the policeman at one end of this chain.

THIS DOES NOT LEAD TO RISING PROPERTY VALUES IN TOKYO.

It's like someone took the Godzilla vs Mega-Godzilla plan, and said "this Godzilla-fights-Mega-Godzilla plan is WAY too simple and robust, what we need is a hundred levels of recursion to make ABSOLUTELY SURE that something goes wrong!".

Some chain links will break.  Which is the point - single link failures are survivable. Also for sure there are some corrupt police officers in Tokyo, but they aren't such a big deal.

It's like someone took the Godzilla vs Mega-Godzilla plan, and said "this Godzilla-fights-Mega-Godzilla plan is WAY too simple and robust, what we need is a hundred levels of recursion to make ABSOLUTELY SURE that something goes wrong!".

Thank you for this analogy. Your comment is apparently disagreed with but I find it perfectly encapsulates the silliness of the proposal by default.

Fixing hardware failures in software is literally how quantum computing is supposed to work, and it's clearly not a silly idea.

Generally speaking, there's a lot of appeal to intuition here, but I don't find it convincing. This isn't good for Tokyo property prices? Well maybe, but how good of a heuristic is that when Mechagodzilla is on its way regardless.

Fixing hardware failures in software is literally how quantum computing is supposed to work, and it's clearly not a silly idea.

Quantum computing is not a silly idea in principle, in that it couldn't be done, it is just much harder for our first, critical try.

I'm surprised that this failure mode is so common. Like... obviously if you unleash one powerful but not well understood force to counteract another powerful but not well understood force, you will likely end up dealing with two powerful but not well understood forces. A magnified cane toad effect of sorts.

So either we:

1. Create a kaiju we can trust (alignment)
2. Prevent the creation of any kaiju (moratorium on some types of AI research)

But when option one is proposed, people say that it has proved to be probably infeasible, and when option two is proposed, people say that the political and economic systems at present cannot be shifted to make such a moratorium happen effectively. If you really believed that alignment was likely impossible, you would advocate for #2 even if you didn't think it was likely to happen due to politics. The pessimism here just doesn't make any sense to me.

I think people here are uncomfortable advocating for political solutions either because of their views of politics or their comfort level with it.

You don't have to believe that alignment is impossible to conclude that you should advocate for a political/governmental solution.  All you have to believe is that the probability of x-risk from AGI is reasonably high and the probably of alignment working to prevent it it not reasonably high.  That seems to describe the belief of most of those on LessWrong.

I personally do not consider (1) to have been "proved to be probably infeasible". MIRI had, like, a dozen people working on it for a decade, which just isn't that much in the scheme of things. And even then, most of those people were not working directly on the core problems for most of that time. The evidence-of-hardness-from-people-trying-and-failing for alignment is not even remotely in the league of, say, P vs NP.

(The evidence-of-hardness-from-people-trying-and-failing is enough that the first clever idea any given person has won't work, though. Or the fifth idea. Also, just counting MIRI's research understates the difficulty somewhat, since lots of people worked on various aspects of agent foudations over the past century.)

Certainly I expect that (1) is orders of magnitude easier than (2).

This seems like a misunderstanding of "overseer"-type proposals.  ~Nobody thinks alignment is impossible; the rejection is the idea of using unaligned AGIs (or aligned-because-they're-insufficiently-powerful AGIs) to reliably "contain" another unaligned AGI.

What if one of the Godzillas is a 1,000x sped-up brain emulation of Eliezer Yudkowsky? (Possibly self-modifying, possibly not)

You would need a Godzilla to set that up before Mega-Godzilla shows up.

"Brain emulation" implies high resolution. A large transformer trained on predicting the activation rates of, say, 100k cortical electrodes situated over the left temporal and frontal lobes might get you most of the way there.

Right now, a brain model AGI seems much harder than a language model AGI (which might turn out to be good enough via a miracle of being in the same goal attainment attractor as humans), and by definition an AGI of unspecified nature is at most as difficult as that. It might be possible to ask a stawberry aligned AGI to set up a brain model AGI, perhaps even for specific humans, and that seems like a more plausible plan to get there in time than developing it with human effort. (That's a more abstract wish than disabling specific computing devices, likely harder to align.)

I see no reason to believe that would be safe at all. It would be just as alien as any other superintelligence. (This has nothing to do with EY, you could put anyone else in that brain-scanner and I still wouldn't trust it.)

Not only is this post great, but it led me to read more James Mickens. Thank you for that! (His writings can be found here).

Thank you for writing this. I needed a conceptual handle like this to give shape to an intuition that's been hanging around for a while.

It seems to me that our current civilizational arrangement is itself poorly aligned or at least prone to generating unaligned subentities. In other words, we have a generalized agent-alignment problem. Asking unaligned non-AI agents to align an AI is a Godzilla strategy and as such work on aligning already-existing entities is instrumental for AI alignment.

(On a side note, I suspect that there's a lot of overlap between AI alignment and generalized alignment but that's another argument entirely.)

My mentality as well. We can't even get corporations to stop polluting. Probably whatever solves AI alignment will also help align our egregores and vice versa.

It's kind of aside, but I think this about safety systems in general. Don't give me a backup system to shut down the nuclear reactor if the water stops pumping; design it so the reaction depends on the water. Don't give me great ways to dispose of a chemical that destroys your flesh if it touches you; don't make the chemical to begin with. Don't give me a super-strong set of policies to keep the function-gained virus in the lab; don't make function-gained viruses. Wish they'd listened to that last one 3 years ago.

Admittedly it may be too late in a lot of ways. We can't make it so if our civilization founders we can prevent mass starvation; not starving depends on modern farming. And we probably can't make it so we can just pull the plug on AI if it starts going all Skynet: the companies or nations that don't pull the plug may be able to bankrupt those that do. Sometimes survival of the fittest sucks.

I don't have a very insightful comment, but I strongly downvoted this post and I kinda feel the need to justify myself when I do that.

Summary of post: John Wentworth argues that AI Safety plans which involve using powerful AIs to oversee other powerful AIs is brittle by default. In order to get such situations to work, we need to have already solved the hard parts of alignment, including having a really good understanding of our systems. Some people respond to these situations by thinking of specific failure modes we must avoid, but that approach of, “if we just avoid those failure modes then we’ll be fine,” leads to declining property values, because there will be unknown unknowns which will mess up your plan. He thinks of these situations as us getting Godzilla to fight Mega Godzilla, which obviously leads to declining property values in Tokyo.

Thoughts: I think this post uses strong wording which makes it easy to misinterpret. I think the main thing that Wentworth wants to communicate is that our intuition should be that of the Mayor of Tokyo: our researchers can suggest many reasons why Godzilla and Mega Godzilla will keep each other in check, but this is an extremely tenuous setup which we should have a strong bias against. I don’t think the piece makes a good argument for why this is an important intuition to build, after all there are myriad oversight techniques which are less brittle than Godzilla fight or which in concert look much more optimistic. This is inline with one of John’s comments: “I am mostly objecting to strategies which posit one AI saving us from another as the primary mechanism of alignment…”

When we are reasoning about alignment and oversight strategies, I think I should give some but not much of my mental energy to “this problem looks like getting Godzilla to keep Mega Godzilla in check”. To the extent that this post is advocating for a higher portion of my energy in that bucket, I think the rule of equal and opposite advice applies. There are reasons for “worst-case thinking” but this post does not present them and doesn’t successfully argue for them as far as I can tell.

Most AI safety criticisms carry a multitude of implicite assumptions.  This argument grants the assumption and attacks the wrong strategy.
We are better off improving a single high-level AI than making a second one.  There is not battle between multiple high-level AIs if there is only one.

It seems to me that it is quite possible that language models develop into really good world modelers before they become consequentialist agents or contain consequentialist subagents. While I would be very concerned with using an agentic AI to control another agentic AI for the reasons you listed and so am pessimistic about eg debate, AI still seems like it could be very useful for solving alignment.

You seem to believe that any plan involving what you call "godzilla strategies" is brittle. This is something I am not confidant in. Someone may find some strategy that can be shown to not be brittle.

What I would actually claim is roughly:

• Godzilla plans are brittle by default
• In order for the plan to become not-brittle, some part of it other than the use-Godzilla-to-fight-Mega-Godzilla part has to "do the hard part" of alignment

You could probably bolt a Godzilla-vs-Mega-Gozilla mechanism onto a plan which already solved the hard parts of alignment via some other strategy, and end up with a viable plan.

I read your critique as roughly "Our prior on systems more powerful than us should be that they are not controllable or foreseeable. So trying to use one system as a tool to another system's safety, we can not even know all failure modes."

I think this is true if the systems are general enough that we can not predict their behavior. However, my impression of, e.g., debate or AI helpers for alignment research is that those would be narrow, e.g., only next token prediction. The Godzilla analogy implies something where we have no say in its design and can not reason about its decisions, which both seem off looking at what current language models can do.

What if we

resurrected literal Godzilla to the future to fight AI

[+][comment deleted]10mo 50
[+][comment deleted]10mo 11