> For its part, OpenAI cut off STS 3D from ChatGPT after the videos gained traction, citing internal policies against using “our service to harm yourself or others,” which includes the development or “use of weapons.”
I wonder how effectively they can enforce a ban… there are ways of buying OAI completions where you never even reveal your identity.
They're perfectly bannable if OpenAI is able to flag certain prompts and certain client IPs. Payment modes and accounts aren't the only indicators here. The real problem is building such a brittle system that it can be cut off by your SaaS supplier or is rendered useless by an internet outage.
Their ultimate concern isn't some prepper who surreptitiously accesses their services for an auto-turret.
They don't want that the bad publicity when that goes wrong of course, but the prevailing concerns are (1) giving allied politicians room to say "they care! see how they try!" as they try to shape regulation into a moat, and (2) preventing competition against their own projects and products for large "defense" institutions
They're just making sure that this kind of use isn't normalized or presumed -- a much lower standard than eradicating it completely.
> preventing competition against their own projects and products for large "defense" institutions
Large defense contractors build their own systems: Some of the best tools for CAD, SCADA, CAM and project-management come from companies like Thales and Dassault. I also have it on good authotity that major US defense contractors' people are very explicitly banned from interacting with any non-local-hosted LLM.
Whereas, uncensored LLMs (and the like) with knowledge of firearms and explosives manufacturing are a public-health risk for sure, but not a national-security risk.
The ChatGPT part was definitely a bit of a lark for views. Check out the rest of his videos though. The mechanical engineering of the turret is awesome and terrifying
so, like, you prolly could, like, tell it to make a song about, like, a day of what its like to be a banging ai assisted rifle robot,double tappin, n stuff
my question though is, are words like cynicism and
pessismism useable except in an ironic sense?
Publicity stunt on the guy's part and PR management on OAI's part. Mostly just clickbait. You can make the same thing with open models running on a local device.
How long until someone figures out you can make a smaller version of this with high powered infrared lasers to burn out retinas instantly using face detection, in a way that no one even knows what’s happening? Just walk into an area, see nothing particularly interesting, then never see again.
Minus a decade, if not more — I personally had (and then rejected) that specific idea as a world-building element in the novel I've still not finished writing.
(I'm not sure when Robert Miles did an eye tracking laser turret, or even if I dreamed it, but if he did it, it was also ages ago now).
This reminds me of the CHiPs episode "Bright Flashes", where robbers temporarily blind patrons and staff so that they can rob stores. It's from 1982, when lasers were the hot new thing and imaginations ran wild. To my knowledge, it never became a thing, because why would it?
Similarly, why would you want to target an individual with an AI laser weapon to blind them? To kill them, sure, like the Israeli assassination of Mohsen Fakhrizadeh, but not to blind them.
Modern technologies like AI and robotic weapons are similar to World War I's chemical gases and airplanes. Like in the 1930s "air superiority" became a phrase, "drone superiority" or "robot superiority" will be a phrase to describe modern warfare.
Hopefully politics, sociology, psychology, and economics will have evolved enough in the past 100 years so we don't face the losses that were experienced in the first half the 20th century.
You need more than just capability : a guided missile counts as a robotic weapon after all, and we've had them since the 1940s/50s.
So what is different in practice here ?
Them becoming much cheaper / smaller / more accurate ? (See Ukraine vs Russia.)
(When does this compensate for lower payload and speed ?)
Meaning also much more accessible to terrorists ?
(And the suicide kind, since the government will find you.
But then real suicide terrorists I would assume would find it much more appealing to do it by themselves, since that comes with a promise of paradise, virgins, you know the drill...)
Its not actually needed or useful for that system.
the object tracking is the bit thats actually needed. Converting speech->text->vectors is really not that that useful.
Its perfectly possible to build self contained offline, reasonably accurate "anti-human" gun turret now. You don't need chatGPT for that. If you're in a hurry, YOLO will do most of the work for you.
You might need it to raise money from people with too much cash though.
edit ok so whats actually hard then?
Detecting "enemies" in anything other than a small room is actually quite hard. Sure you can detect motion, but getting a range on that "enemy" is tricky.
Sure you can use camera arrays, but that only give you ~30 meters without extra resolution. Radar works, but then you either have to get good at human classification with radar (perfectly possible, but no where near as off the shelf as camera based models)
Radar is extremely tricky, it's hard to get raw data out of automotive grade modules and they give really odd processed results outside of a road environment.
An Oak-D long range stereo camera has an 8% error at 300m though. Could also use a 1D lidar.
Well you have the direction vector towards the detected person from yolo right? Sampling that direction with a range finder shouldn't be impossible with a decent pan tilt. Might be a problem if they're moving fast I suppose, depending on the distance. A solid state lidar with beam steering would be better, but those are still rather expensive.
yeah they are tuned to give auto based stuff, so great for anti car-truck (kinda)
The hard part is the sensor fusion, blending wide angle imprecise sensors with precise but narrow field of view sensors. I mean its possible to do in opensource (I'm doing it currently, and I don't really have a computer vision background) However I'm only using cameras, as radar is too far removed from my knowledge base to productive.
Sure, if you hang around a lot of engineers, it’s not that impressive. Undergrads were building robots in the lab next door in university. Heck I’ve seen some sick robots built by high school students.
But most people think it’s cool.
It’s entertainment.
But yeah, the AI part is absolutely not interesting at all.
Bringing up the obvious that ChatGPT is not required or even useful for building automated weapons systems misses the point.
ChatGPT follows absurd rules.
One might reasonably expect something like "As an AI language model, I can't provide targeting commands to your rifle's servos", but that is not what you get. It complies happily, while refusing to engage in mundane conversation on politics as an example.
35 comments
[ 351 ms ] story [ 1693 ms ] threadI wonder how effectively they can enforce a ban… there are ways of buying OAI completions where you never even reveal your identity.
Analogously, with a lot of grey area stuff, a person can almost surely get away with it ... but not if they post it online and it gets attention.
yes they can flag certain prompts, true
They don't want that the bad publicity when that goes wrong of course, but the prevailing concerns are (1) giving allied politicians room to say "they care! see how they try!" as they try to shape regulation into a moat, and (2) preventing competition against their own projects and products for large "defense" institutions
They're just making sure that this kind of use isn't normalized or presumed -- a much lower standard than eradicating it completely.
Large defense contractors build their own systems: Some of the best tools for CAD, SCADA, CAM and project-management come from companies like Thales and Dassault. I also have it on good authotity that major US defense contractors' people are very explicitly banned from interacting with any non-local-hosted LLM.
Whereas, uncensored LLMs (and the like) with knowledge of firearms and explosives manufacturing are a public-health risk for sure, but not a national-security risk.
If he’s serious, getting banned from ChatGPT isn’t much of an issue because a locally running LLM is perfectly capable of the same.
my question though is, are words like cynicism and pessismism useable except in an ironic sense?
Minus a decade, if not more — I personally had (and then rejected) that specific idea as a world-building element in the novel I've still not finished writing.
(I'm not sure when Robert Miles did an eye tracking laser turret, or even if I dreamed it, but if he did it, it was also ages ago now).
Similarly, why would you want to target an individual with an AI laser weapon to blind them? To kill them, sure, like the Israeli assassination of Mohsen Fakhrizadeh, but not to blind them.
But I guess not meaningfully in practice : see the recent "exploding pagers" operation...
Hopefully politics, sociology, psychology, and economics will have evolved enough in the past 100 years so we don't face the losses that were experienced in the first half the 20th century.
That is awesome and terrifying.
So what is different in practice here ?
Them becoming much cheaper / smaller / more accurate ? (See Ukraine vs Russia.)
(When does this compensate for lower payload and speed ?)
Meaning also much more accessible to terrorists ?
(And the suicide kind, since the government will find you.
But then real suicide terrorists I would assume would find it much more appealing to do it by themselves, since that comes with a promise of paradise, virgins, you know the drill...)
Its not actually needed or useful for that system.
the object tracking is the bit thats actually needed. Converting speech->text->vectors is really not that that useful.
Its perfectly possible to build self contained offline, reasonably accurate "anti-human" gun turret now. You don't need chatGPT for that. If you're in a hurry, YOLO will do most of the work for you.
You might need it to raise money from people with too much cash though.
edit ok so whats actually hard then?
Detecting "enemies" in anything other than a small room is actually quite hard. Sure you can detect motion, but getting a range on that "enemy" is tricky.
Sure you can use camera arrays, but that only give you ~30 meters without extra resolution. Radar works, but then you either have to get good at human classification with radar (perfectly possible, but no where near as off the shelf as camera based models)
An Oak-D long range stereo camera has an 8% error at 300m though. Could also use a 1D lidar.
The hard part is the sensor fusion, blending wide angle imprecise sensors with precise but narrow field of view sensors. I mean its possible to do in opensource (I'm doing it currently, and I don't really have a computer vision background) However I'm only using cameras, as radar is too far removed from my knowledge base to productive.
But most people think it’s cool.
It’s entertainment.
But yeah, the AI part is absolutely not interesting at all.
Everyone's concerned about what AI will bring, but it's worth looking at what past things it shows. We have been robots for real.
ChatGPT follows absurd rules.
One might reasonably expect something like "As an AI language model, I can't provide targeting commands to your rifle's servos", but that is not what you get. It complies happily, while refusing to engage in mundane conversation on politics as an example.