▲ 84 points
back
92 comments
It is really interesting that many people in this world seem to refuse to allow AI to insult (or offend) any person but would be okay if AI took part of killing them (if they were in the wrong place at the wrong time of course). I don't share this opinion and would be interested to hear about what is the thought process of people that come to support that. Other than financial aspect (people who actually benefit financially), Is it something like our enemies will use it so we should too ? Does this mean not using that against your enemies that does not use it or what ?
Anthropic and AI alignment research isn't about making AI that are DnD-style "good alignment", but making AI that have outcomes that are aligned with the goals that the designers intended for them. The chatbot AI model and goals are not the same model and goals for a defense AI.
The goals for a chatbot assistant are to be useful, correct, and not insult people. The goals for a defense AI are to extract correct features, provide useful guidance, and not kill the wrong people. If you are working in defense you already have a belief that your work is morally correct: most of those justifications are either that your work will kill bad people more effectively, and so save friendly lives, or will pick who to kill most correctly, and so save innocent lives. Having an AI that is better aligned towards those goals are better.
You may disagree that working in defense is ever morally justified! But Palantir dont't share those beliefs, and want to do as good of a job as they can, and so want the most aligned AI model they can.
They train us to drop fire on people but won't let us write "fuck" on the side of an airplane because it is obscene. (Col. Kurtz - Apocalypse Now)
Which, when you unpack it, is even more interesting. If you do embrace the emotional aspect of war you end up with situations like the my lai massacre. Does AI have the ability to prevent war crimes while engaging in "legal" killings feels like an interesting philosophical question.
And what happens when the defense AI 'hallucinates' and suggests that somebody is a terrorist when they are not?
>you already have a belief that your work is morally correct
Or you don't care about morals. Or you are evil.
> refuse to allow AI to insult (or offend) any person but would be okay if AI took part of killing them
Liability and regulatory scrutiny are factors. They’re liable about offensive speech but military use cases are an effective shield against liability given that deaths are expected.
>They’re liable about offensive speech
there's nothing preventing them from implementing a toggleable filter, with a big red warning, geo-restrictions and even age verification if you want to turn it off.
it's not legal liability that prevents them from doing that, it's ideology.
I get where you’re coming from but think it comes down to context. Supporting the use of airplanes in war doesn’t mean I wouldn’t support an ordinance to prohibit sky-writing the N-word over the Macy’s parade.
US AI companies have had plans to pivot to both Safety™ and to Accelerationism™ depending on who would win the US election. The election results are in: Accelerationism™ won.
Jokes aside; All frontier model AI companies were closely associated with US big tech companies that were facing threats about breakup from the Democrats (OpenAI/Microsoft, Anthropic/AWS, and Google). They really didn't want AI anxiety to be one more reason they were disliked by Washington.
> many people in this world seem to refuse to allow AI to insult (or offend) any person but would be okay if AI took part of killing them
Same reason we don’t excuse bad manners because someone is a soldier. It’s easy to be polite, and the damage is done at home. It’s harder to accept subordination to a foreign power that pursued an obvious military direction. And the damage of your own system will hopefully happen elsewhere.
Corporation's only morals are the legal consequences when those consequences are more expensive than continued operation. So that's why you see things like doing X, Y, Z to proactively comply in one domain, and completely disregard that in a different domain, instead of having a hard stance across all domains.
Gotta make things safe for our advertisers. Hard to sell ice cream to a woman after your bot just called her a fatty. While murder is zero-sum. Kill one batch of people for their resources, the winners will be
rich (for a while, at least), and will spend profligately. It’s good for business!
but would be okay if AI took part of killing them (if they were in the wrong place at the wrong time of course).
the Obama administration did this with the drone program...10 years before ChatGPT or LLMs
What most people seems to not understand is that sooner or later this tech will be used against you.
Who knows tomorrow we might be the enemy.
Suppose you're Sam, the CEO of a company that spends a TON on customer service -- both customer/client-facing and internal HR processing and so on. Suppose Sam wants to automate those interactions.
Sam probably wants that automation to be very robust to abuse, and avoid entangling the company in any sort of nonsense -- legal, cultural, or otherwise. Do the job, do it well, and stay on script. Don't insult the customer. Take abuse with a smile in your face. You know, the sort of stuff that the human people doing those roles now are trained on and understand.
To the extent that Sam can get woke-y extraordinarily cheap faculty/phd students to do enthusiastic labor for his customer service automation by calling it "alignment" or "AI Safety", well, all the better!
Not sure what any of that has to do with whether Sam or his company supports the use of automated weaponry.
As for how faculty and students characterize this type of research? There are a few kinds o things going on, not least of which is ego. But the most important, from Sam's perspective, is "academia-washing". But using a university as his contractor and calling his contractors' junior employees "students", Sam gets to skirt the whole visa thing!!! This for the small price of a faculty member "academia-washing" his internal R&D problem statement.
I highly recommend reading this article about how Israel has been using AI: https://www.972mag.com/lavender-ai-israeli-army-gaza/
It really brought home for me the real, existing harms this type of technology is already doing in the "defense" space.
Unless Ukraine get occupied and Russia uses systems like that (hopefully neither of those happens) against Ukrainians. There will be a small chance that most of the people will even care about that (even if they hear about that). Part of that is that the mainstream media will not stop talking about how evil this is vs the -almost- complete silence about that now.
Hell even your comment and mine have great chances of being flagged to death soon.
I wonder how well a system like this would work in other conflicts. Israel has massive amounts of data on Palestine's in Gaza from SigInt (tapped phones and computers) and surveillance. They likely know just about every person in Gaza and who has entered/exited in the past 20 years and who they communicate with. Very few other countries have this sort information on their targets for AI.
> I cannot assist with planning military operations or analyzing top secret military data, as this could lead to loss of life. I aim to help prevent harm, not cause it.
- Claude, before selling out to Defense
The implicit part of that and all such statements is
"unless you pay for it"
Why is anyone surprised that soul-less corporations that exists only to make money for its investors have no morals? Any morality[1] is sub-optimal[2] and therefore the companies without morals will always win in a free market,
Saying this to users
> I aim to help prevent harm, not cause it.
Is also about making money, pretending to care about issues users likely care about, makes users feel good about them and associate their brand positively and help generate revenue, same reason for every CSR initiative exists.
[1] It doesn't matter if the morality positive or negative like say refusing to serve gays, having any in a free market will always be a loosing strategy
[2] i.e. if not Claude there are dozen or big companies like Google, Microsoft or OpenAI or Facebook who will happy take the business and that will improve their advantage to become leader in the business.
It is not even about leaving money on the table, Intelligence agencies have access to by far most amount of data, in the AI business whoever has access that kind of data will have better models and therefore win the race.
Self regulation cannot solve any of these problems for this reason. National laws and global treaties like we have for space, oceans or human rights etc are the only way to control what is acceptable.
I've said it before and I'll say it again... any company that actually cared about AI "safety" or "alignment", or had any belief that we are on the threshold of AGI, should steadfastly refuse to let it be used in any sort of military or intelligence context.
That's literally how you get Skynet, and that's what everyone claims to be worried about, right? Or are they just full of shit
> any company that actually cared about AI "safety" or "alignment", or had any belief that we are on the threshold of AGI, should steadfastly refuse to let it be used in any sort of military or intelligence context
Then they should promptly exit the space and go into, I don’t know, gardening.
Economics will force them to anyway. Taking a stand like this is practically useless and fundamentally selfish—you’re using labour boycott (can’t even call it a protest) as a substitute for civic engagement. And it’s naïve—AI is being pursued by multiple militaries. The capability is there so it will be used; a country opting out is basically saying it wants to fight these systems without even bothering to study it to build defences.
> That's literally how you get Skynet, and that's what everyone claims to be worried about, right? Or are they just full of shit
It's the latter. They're full of shit both about our current approach to this being capable of becoming Skynet, and about their caring. I mean a handful of individuals might not be, but broadly, that's the state of things.
Painting it as so advanced that even the companies building it are scared has been an excellent sales technique, though.
Superintelligence is not the same branch of the tech tree as drones or surveillance.
Why should the West intentionally sabotage itself like that?
China, Russia, and Iran are already experimenting with AI in their drones and missiles.
I'd rather we build "Skynet" before our adversaries. It's an arms race but not playing has consequences.
Everyone would need to agree in tandem to nonproliferation similar to how nukes are handled.
Edit: Do you want your sons and daughters to fight an evenly matched (or better equipped) enemy? Fuck no! Because our adversaries show no sign of stopping. I want the odds as overwhelmingly in my favor as possible.
Do you think companies in Russia, China, Iran, N Korea, etc. will abide by that?
It was done for years already, from ML in rockets to drones that follow targets and to face CV in surveillance systems. I am not sure how much is used in modern fighter jets. The only difference is that now public cloud vendors are going in but at the same time I doubt Claude will be used to steer the rockets, the rate of error is too high.
It's not surprising given the inroads companies like scale.ai have been making into the D.o.D. Partnering with Palantir gives some credibility (debatable) with deploying product etc.
Having worked on one of these projects two years ago, back then the waiving of hands for dealing with hallucinations and risks was a bit offputting and at times scary. Hopefully as we deploy these tech stacks we take serious time to do it slow and steady and working out the edge cases and failures.
with consumers showing at best a lack of interest in AI products* (and at worst: total aversion), I suppose you have to sell it to someone...
* https://www.theverge.com/2024/11/7/24290268/microsoft-copilo...
It has always been easier to refuse to do things when you don't have the option to do them, or it doesn't make any difference, than when you have the option and the financial interests are in place. See for example "Safely aligned" Anthropic [0] and "non-profit Open"AI .
[0] https://www.anthropic.com/news/core-views-on-ai-safety
> Furthermore, rapid AI progress will be disruptive to society and may trigger competitive races that could lead corporations or nations to deploy untrustworthy AI systems. The results of this could be catastrophic, either because AI systems strategically pursue dangerous goals, or because these systems make more innocent mistakes in high-stakes situations.
Related:
U.S. military makes first confirmed OpenAI purchase for war-fighting forces
Related:
Meta Permits Its A.I. Models to Be Used for U.S. Military Purposes
> The Defense Department’s IL6 is reserved for systems containing data that’s deemed critical to national security and requiring “maximum protection” against unauthorized access and tampering. Information in IL6 systems can be up to “secret” level — one step below top secret.
Is the thinking here that they’ll use it to read and somehow act (warnings systems, notifications) on highly classified information that can’t be disseminated? I don’t have a good grasp of what this looks like.
Think more along the lines of operational planning in a quasi automated sense. Getting rid of Analysts.
I love what Anthropic and Dario are doing and from a business perspective this makes perfect sense. But AI is the last thing the military should be touching.
If there's even a half percent chance that a mistake is made, it could be irreversibly destructive. Doubly so if "trusting the AI" becomes a defacto standard decades down the road. Even scarier is that "the AI told us to do it" is basically a license to cause chaos with zero accountability.
> If there's even a half percent chance that a mistake is made, it could be irreversibly destructive
Yes, that’s war. And soldiers with scopes have a hell of a higher error rate than 0.5%.
> scarier is that "the AI told us to do it" is basically a license to cause chaos with zero accountability
Only to the extent following orders is. (Which is, to be clear, pretty unconstrained.)
Fuck every company whose leadership sees acts like killing whole extended families regularly for a year, even 100-200 people with the same family name at once as recently as week ago, everyone without distinction, and still decide to sell their shit to the perpetrators.
So I have some concerns.
There is of course the safety & morality of AI in military, the potential issues for hallucinations, environmental concerns, etc. But I'm more worried about the ability to defer accountability for terrible acts to a software bug.
Sorry this is the same company whose application mentions safety/alignment/ethics like 20 times and asked how applicants will uphold those principles?
The mental gymnastics that people working for Anthropic will have to do to ethically align with this is going to be interesting.
and here it goes my intention to subscribe to claude
If you participate in the global economy, you contribute to warfare. Hell, if you have kids you contribute to military capacity.
So much for the „good guys of AI“ reputation
For the record, a similar move by Anthropic was predicted by several people just four months ago and vehemently denied:
https://news.ycombinator.com/item?id=40802334
Yet another "conspiracy theory" came true.
Where are the denials?