60 Nations agree to regulate AI in the military before it destroys the world | Hacker News Reader

60 Nations agree to regulate AI in the military before it destroys the world

theregister.com

54 pointsrntn20 comments

FYI, the US, or at least the navy, has grappled with letting a computer shoot at things since the 50s, when they first computerized target tasking and prioritization and fleet networking. It was explicitly discussed that older navy folk were wary of letting the machines make decisions, so they purposely designed it in a way that a human was the one giving firing permission.

By the 80s however, we were willing to let at least one system operate entirely autonomously, the Phalanx CIWS, and that decision has already lead to a few friendly fire incidents https://en.wikipedia.org/wiki/Phalanx_CIWS#Incidents

The US military already has experience allowing computers to make death happen, and Raytheon is likely already testing an "AI pulls the trigger with no intervention" system right now. If we are unwilling to implement laws against that right now, then it will 100% happen.

Chilling.

Well I'm not holding my breath this is going to work for real. But, yeah, good initiative, at least some easy intel can be extracted from everywhere in one sweet sweap.

But again, humans do war, that's the way it works, it has been like that since when? a million of years?

Some new technology won't change all those gens and more important, the cultural memory teached by societies around the world: "war is a way to solve problems".

So, nice you did this, but better get ready for fighting involving AIs in the conflicts.

If it isn’t clear to everyone, this is a gesture meant to calm public populace fears about “AI”.

First of all, there is no definition of “AI”. It’s a running joke in the community that once an “AI” approach gets good enough to be commoditized it is no longer called “AI” and becomes simply “ML” (machine learning) that we teach to college students.

With that in mind, military drones have used “AI” since at least during the Obama administration. This is a nonstory.

I’ve been to summits like the one talked about that are a bunch of people in government, nonprofit, or for profit executive roles that get together to discuss things they don’t understand and make a resolution that looks good. They are all for show. Nothing useful, meaningful, or impactful comes from these things.

Yes, yes, we should be responsible with the use of AI. Everyone agrees without asking what that really means.

When I think war AI, I am thinking 3 levels: the global coordinating AI on a battle field, the AIs of a swarm of "killer robots"/weapons (should not work with good EM jamming), the embedded AI in each "killer robot"/weapon with a local non-jammable tactical "vision" (should not be very interesting due to the very limited "size" of an "embeddable" AI).

I am a dota2 player, and I very well remember E.Musk openAI: a team of 5 independent AIs with each a different set of skills based on the hero... and they had to slow them down and make them inaccurate...

In the end, I cannot see clearly where an AI would pertinently fit in a battle. Stategic assistance?

What matters is an AI's ability to do things that human combatants cannot, and that which human opponents cannot effectively counter.

As you refer to videogame AI, think back to the Berkeley Overmind, a Starcraft 1 Zerg AI that beat seasoned humans by the simple expedient of being able to shuffle around mutalisks [0] in such a way that no human micromanagement could efficiently kill any. The only realistic way to defeat the Berkeley Overmind, as I recall the decade-old coverage, was to attack the AI's base before it could build the tech necessary to spawn mutalisks.

[0] Mutalisks are a Zerg air unit whose attack bounces from enemy to enemy. Like all Zerg units, it regenerates health gradually, so the AI's chief tactic of moving even slightly damaged units to the back of the battlegroup allowed surprising compounding benefits.

I am cursing you right now because it is occurred to me that I have to go home and install starcraft and play. There goes my weekend.

> I cannot see clearly where an AI would pertinently fit in a battle

Here's a scenario: one F-35 together with about two dozen loyal wingman drones [1]. The drones stay within line of sight of the F-35, so they are immune to electronic warfare of the kind that lead to the capture of a US drone by Iran in 2011 [2]. The F-35 pilot assigns the drones high-level missions, like take care of the incoming bogeys in the north-east sector. The drones then use AI to coordinate their actions in an optimal way and carry out the mission. Some of the drones are lost. No biggie.

[1] https://en.wikipedia.org/wiki/Boeing_MQ-28_Ghost_Bat

[2] https://en.wikipedia.org/wiki/Iran%E2%80%93U.S._RQ-170_incid...

I cannot understand the usage of combat planes nowdays: unless in a unbalanced conflict, all combat planes are sitting ducks with modern air-ground missiles. No to mention cruise missiles should be able to do the same if it is the "delivery" some payload somewhere.

It is the same for tanks.

Maybe that's where AI could be "usefull", way more efficient autonomous guidance systems for all those missiles, actually scary efficient.

War is a game of cat an mouse. First, F-35 and the MQ-28 drone I mentioned are stealth aircraft. It does not mean they can't be found by ground-based radar, but it's not going to be very easy, plus, that radar exposes itself. Just like you can shoot a missile from the ground to try and kill the F-35, the F-35 can shoot a missile down at you. And while the F-35 can move around and try to evade the incoming missile, the land-based platform will be much more sluggish.

Why not cruise missiles? Cruise missiles will continue to have a role in combat, but they are too expensive to use them for everything. A stealth bomber like F-35 or B2 or the future B-21, can ferry a lot of relatively inexpensive munitions and drop them on target. A Tomahawk costs north of $1 million, while a JDAM can be procured for about $25k and a Hellfire for about $150k.

This is not what I "heard" from a much more reliable source than HN: once a missile of "proper tech" is shot at an aircraft, there is no evading it whatever the tech of the aircraft, unless luck modulo altitude. Maybe this is not true anymore, but that would probably be classified information. Classification of a boggey on a radar is usually a matter of seconds, if not less than 1 second, and once the missiles are in flight, it is very probably over for the combat aircraft. And I have not heard of any efficient "stealth" of combat aircraft on the whole range of existing radars, and I would not bet on this.

And it seems there is often more than one missile shot at a combat aircraft.

That would mean, before any combat aircraft can fly, all ground base radars would have to be disabled... and how do you get rid of a dynamic mesh of mobile ground/sea based radars (or in a place you cannot really "reach")?

In the end, the tech of the opponent in a "standard" conflict should be assessed before anything can fly, and get ready for some very bad surprises (high altitude cruise missiles...).

drones with payloads may be able to do "more" for "cheaper" than a cruise missile: I guess a drone is "cheaper" than a combat aircraft but more "expansive" than a cruise missile.

"AI" may find its way in the autonomous guidance systems of a cruise missile, like drones which would only need to receive some sort of high level order to perform some "tasks" and that nearly completely autonomously, but I don't think you can "embed" such "big" AI in a drone though, you would need a near "realtime" link to some AI center, then it would be "space" war, war satellites destruction and jamming.

I think you will enjoy reading about Zoltan Dani [1]. He managed to take down a stealth F-117 during the war in Yugoslavia. But it was not easy. I would go as far as to say that Zoltan Dani was a military genius for figuring out how to achieve what he achieved.

A repeat is quite unlikely though, as the US military is very good of learning lessons in war.

[1] https://en.wikipedia.org/wiki/Zolt%C3%A1n_Dani

You are talking about low tech developing countries a long time ago (an eternity for tech), namely extremely uneven conflict.

That's why it is very important to assess the balance of a conflict on recent data in order to deploy the right equipment. Of course, there will be bluf, bad surprises, etc.

> an eternity for tech

Not really. As everyone could see, Russia is still using tanks made 50 years ago.

I may be talking about old tech, but you are talking hypothetical tech. "I don't understand why anyone would invest in combat aircraft when this hypothetical tech could be developed by a peer nation and render the aircraft obsolete". Does that sum up the argument?

Well, I know with better source than HN that combat aircrafts are mostly useless due to "proper" air defence missiles.

But combat aircrafts are still usefull in uneven conflicts, namely when one side has really weak air defence, or even none. It is a similar issues with tanks.

> It is the same for tanks.

And helicopters, and trucks, and humvees, and missiles themselves, and drones, and humans...

We have very good and usually very cheap (compared to the target's price) ways to destroy things.

Looks like planes and tanks are still useful, Ukraine wouldn't ask for them otherwise

That would means they are super omega sure that putin's regime invaders are very limited in their capacity to destroy combat aircrafts and tanks, basically have very low tech missiles and/or very little high tech missiles. It seems putin's regime is "making" room in it prisons sending political dissidents and "unwanted" people to be slaughtered by the ukrainian army.

Wild guess, the USA intelligence services know and have green-lighted the usage of combat aircrafts/tanks to friendly countries. Hope it is not the same type of intelligence used for the weapons of mass destruction in irak... ahem...

Each of the AIs you mentioned may have a dedicated counter AI fielded by the opponent.

Notably https://arxiv.org/abs/2211.00241

> We attack the state-of-the-art Go-playing AI system, KataGo, by training adversarial policies that play against frozen KataGo victims. Our attack achieves a >99% win rate when KataGo uses no tree-search, and a >77% win rate when KataGo uses enough search to be superhuman. Notably, our adversaries do not win by learning to play Go better than KataGo -- in fact, our adversaries are easily beaten by human amateurs. Instead, our adversaries win by tricking KataGo into making serious blunders. Our results demonstrate that even superhuman AI systems may harbor surprising failure modes

The only reason we call those failure modes surprising is that we anthropomorphise the AI. Entirely understanble, we all do it and it may become a serious issue.

Reply on news.ycombinator.com