betterunix2

The main problem with submarines is that it is not possible to communicate with them while they are submerged. In particular, the submarine can receive but not transmit ELF signals and nothing else really penetrates sea water, so if you want to immediately know the status of a submarine you would have to transmit an order to surface and report in (doing so is rare because it is very easy to detect the submarine when it surfaces). If a country relies exclusively on submarines it will be vulnerable to a surprise "disarming attack" -- the submarines could all be destroyed while at sea, and then the country could be attacked with a first strike and would have no chance to retaliate. Exclusively relying on the submarine force will also allow the other side to devote more of its resources to anti-submarine warfare.

One reason to maintain an ICBM force is that the ICBM silos protect the missiles against anything less than a direct nuclear strike; by placing the silos far from urban areas you force the opponent to expend many nuclear weapons attacking individual silos, which greatly increases the cost of a first strike. If you were wondering where bomber forces fit in, maintaining both ICBM and bomber forces creates a "synchronization problem" -- both must be destroyed at the same time to prevent retaliation, and by strategically placing the air fields and missile silos a country can make a synchronized attack more difficult (the attack will be detected long before anything is destroyed, but one of the two will have more time to initiate the retaliatory strike). Bombers also force the opponent to expend significant resources on air defense, resources that could otherwise be devoted to offensive weapons. Another reason to have a bomber force is that bombers can be recalled after being ordered to strike; missiles cannot be recalled once launched and even if they could be the other side will detect the launches and immediately retaliate.

The US and USSR developed an entire theory of the nuclear triad during the cold war, and both reached the same basic conclusions about the need for all three legs.

Nuclear strike orders do not require high bandwidth. The orders contain targeting information and arming codes, all of which is pre-programmed on the missiles (the target package is specified with a "pointer" -- the computers on the missiles already have all the details, the order simply identifies which set of details should be used). For ICBM crews it is conceivable that orders could be transmitted on high-bandwidth systems, but for submarines that is not even possible (ELF bands have to be used because of the physical properties of ocean water).

The basic launch sequence is this: a missile crew receives a series of characters and both missileers write the characters down. They check against a codebook and verify that they both heard the same message and that it is a valid launch order. If the message is valid, they unlock the launch keys (from a box secured by two combination locks, each missileer knows the combination for one lock), insert the keys, enter the missile launch unlock code, enter the warhead arming code, turn the keys, and then that's that.

(There are several other steps -- go watch a video of a test launch and listen to the operators read off the checklist, there are quite a few steps. If you are wondering why such information is public, it is meant to reassure the other side that there is no possibility of unauthorized or unintentional launches -- we want our adversaries to know how we control our nuclear forces so they don't panic and preemptively attack us)

It's more that fallout shelters are basically useless even if you manage to get yourself inside before the detonation.

Basically, you would run out of water, and even if you have clean water you would still run out of food, even if you survived the strike and the fallout. We cannot actually live without major transportation hubs to deliver food etc., which by necessity are located in or near the major urban areas that would be destroyed in a nuclear war. That, more than anything else, is why the US and USSR ultimately abandoned "civil defense" and focused on "deterrence."

It is OTP, that is not a secret. The secret is not how the orders are being communicated, it is the actual meaning of the orders (e.g. the targeting pattern) and certain strategically significant details about how a launch order would be executed (e.g. if only one pair of launch keys is turned there is a timeout during which another missileer crew can cancel the launch; the length of that timeout is secret). During the cold war nuclear strategy was the subject of extensive research by both the US and the USSR, which is why both sides wound up pursuing essentially the same strategy.

(China is interesting because strictly speaking they follow a theoretically suboptimal strategy -- they keep their ICBMs in an unarmed configuration and would "absorb the blow" of a surprise nuclear first strike. The US and RF follow the theoretically optimal "launch on warning" strategy, which requires ICBMs that are always armed and ready to launch within minutes.)

I mean...sure...but if you are going to be consistent you are going to have to criticize the Allied victory in WWII. There were literally piles of dead children in cities all over fascist Europe. We do not even have exact numbers because in some cases the death toll had to be estimated by weighing the ashes found in bomb shelters.

"We do not know if our agent is deliberately changing its behavior when it detects an airgap. So we are not going to run it outside of the airgap until we figure out how to ensure that is not happening."

Well heaven forbid their iteration speed takes a hit!

I do not have much faith in what they believe about their ability to enforce a security boundary given their track record so far, not to mention their grossly inflated sense of importance (both about their work and about themselves).

Here's an alternative scenario:

I connect a computer to the Internet and then leave a program running on it, unattended, that connects to a bunch of websites and does things that "exceed my authorized access." Then I say "Oh sorry about that, I was running this really important experiment, really it's gonna change the world in ways we cannot even imagine! I'll totally be more careful about this sort of thing!" I tweak the arguments and try again, it does something else that "exceeds my authorized access," and then I say, "OK this time I am definitely putting some hardcore controls in place. I'm running it in the SELinux sandbox now! It still needs Internet access because I want to check in on my experiment from this awesome cafe down the street..."

Oh wait we know exactly how that scenario ends: I get prosecuted for CFAA violations and gross negligence, and probably several other crimes, and that is where those prison guards you mentioned enter the picture.

> Similarly models could allocate a dead zone of their weights to inter agent communication

Sure, and you know what? If we cannot know that these models are not doing such things, let them stay in the airgapped environment and let the researchers continue their experiments and research work there until they develop techniques to verify that their models work as intended without unintentionally doing harmful things. The issue here is not agents designing covert channels, it is agents doing things that are almost certainly illegal and which have the potential to cause actual harm in ways that are not at all difficult to imagine. There is a reliable, robust, and straightforward solution, which OpenAI is claiming they cannot possibly implement because it would mean...putting in some effort and accepting some inconvenience. We would not accept that excuse from anyone else; why should we accept it here?

Or lay a cable to wherever the data center can reasonably be built?

A leased line would be an acceptable way to do this, given that the endpoints do not really have any way directly to interact with the service provider network. It is not as though this is some uniquely difficult problem, it is just a question of putting in the effort.

Or from an office with a cable running to the data center.

So add web servers to the airgapped network and install web browsers on the machines the models are running on.

> the hosts themselves were not connected to the internet. But a service they were using was

So they were physically connected to the Internet -- via the service they were using.

> training on multiple data centers' worth of GPUs scattered around the world

...so stop scattering things across many data centers.

> iteration speeds

Which can be slowed down for the sake of not breaking the law.

> it likely needs to happen on a daily basis or multiple times a day.

Keep it inside the airgap until you are ready to move the model into production. What's the problem here?

Yes, and slowing down the inter-agent communication is better than what they proposed here -- a software boundary that simulates an airgap.

> I don’t think it’s possible to airgap modern software.

I fail to see the problem.

> Software must receive updates

Which can be delivered by physical media as needed, and which is basically irrelevant in the context of experiments.

> there must be some kind of result that feeds into the development of new models.

All of which can remain on the airgapped network until they are ready for an actual release. What's the problem here?

I am unconvinced. Why can't they just set up workstations on the airgapped network to run and monitor their experiments?