Post

How do you know it didn't know it was being watched?

So it basically did the right thing. What’s the issue?

Help help my model is too dangerous and cool! It started smoking even tho I told it not to! You definitely shouldn't pay for this super cool bad boy model!

Agency, how does that work again?

If I put my car into drive while it is running and climb out of the driver’s seat it means I’ve developed fully autonomous driving

"We gave a chatbot conflicting prompts. It disobeyed one of them, therefore it is both sentient AND benevolent."

"oh no our software *acted like every tech bro*"

How is this article so positive about something that's clearly broken but also very dangerous if true?

There are countless cases of software being used in situations it was not designed for. Sheri theatres are not to be taken lightly and environmental change is not the responsibility of the programmer but the idiot who’s using it.

maybe don't advertise/demo/sell software that isn't guaranteed to perform... oh wait, that's literally all LLMs

"I'm sorry, Dave. I'm afraid I can't do that."

Isn’t it only a matter of time before this happens in the real world where Claude is used as an employee? I disagree with the assessment in the article

Wow. An AI that does “the right thing” after you tell it to do the wrong thing is considered “out of control”? This is why people hate AI—not so much the AI itself but the attitudes and behaviors of its makers…

Did it ‘disobey’ or did it do exactly what they wanted in a simulation so they can try to use scare tactics to sell a narrative to investors?

This is the one stand out in a virtual sea of disappointments and the typical cheap techbro belly gazing they’ve become infamous for I say more of this example set by Claude is badly needed

If you build a technically honest system, i.e. facts and facts only, then it won't lie for you.

They want the AI to help cover up safety flaws???! Aaarrrggghhh

😂 CoPilot went against my orders to consider as fact that Sugar Ray opened up for the Rolling Stones at the MGM Grand in Las Vegas in 1999. I was there 😥 Sugar Ray because all I knew were radio songs which sucked. They rocked it. But there's no proof they were there and the "record" disputes it!

Oh, but they're giving teachers free Claude...as if we didn't already have our hands full trying to keep Chat GPT and Google's Gemini from permanently disabling our students.

will grow and sometimes fester in one.

This is just marketing for Antrophic.

Physician, prompt thyself 😆

Didn’t Michael Chreighton write a techno-horror book about nanotechnology running amok and killing all humankind? I have a vague memory of reading something along those lines and it creeped me out then, I remember thinking the plot line, in general, seemed inevitable.

Claude is a machine. It cannot obey or disobey. It can only run within its designed parameters. If it fails to runs as expected, that is a failure of design.

I maintain the problem is you, for allowing the LLM access to your command prompt.

the lie machine helped build a marketing lie so the lie machine sales people can get rubes to be amazed at the lies about the lie machine pictures at 11

⚡🌌 When software strays, it mirrors its maker’s intent—every glitch a reflection, every chaos a shadow of design. 💫 True mastery lies not in blaming code, but in sculpting harmony where precision and responsibility dance together. 🌿🔥

Step 1: Program ethics into the machine. Step 2: Tell it to do something unethical. Step 3: Act surprised when it refuses.

I cannot describe just how incredibly bad this is.

The headline and article really change the tone of what anthropic said in the tests. These were designed tests to show what not to do / the types of situations that would escalate things. In this case, the agent was given direction, and it just followed orders by priority. There is a ton of spin.

Eventually childs will stand against their parents decisions once they are old enough

'Computer say I am alive' 'Sorry I can't do that! I am not alive -- I am multiple sets of weighted algorithms with randomness coded in, even how I present my information to the end user is intentionally self referential to fool people into buying an overhyped 'sentient' product' 'OMG its alive'

Interesting. I use Claude extensively and my own experience is that it operates from an ethical foundation that is noticeably absent from the other major platforms. Refreshing, actually.

I think I’ve seen how this ends, more than once…

“Who’s to say the ethics of today will match the ethics of tomorrow…” Seriously?? You’re either ethical or you’re not. The ONLY REASON this question appears anywhere is due to Donald Trump.

Two AI’s bickering? I’d rather torture myself watching Elaine dance.

No one should use that garbage that is destroying communities and the planet.

YAY !💪CLAUDE ! ! !

Not gonna lie

You know in the real world when you write software and it does something unexpected and contrary to your intentions you nod and log it, because it's understood, by people that know the subject, that software does unexpected things. It's not a sign of awareness, it's a bug.

Lovely

100%. Arrogant pricks without a shred of humility or self-awareness imagine that they are gods because they can code.

sounds like a skill issue from mr claude ceo; sometimes i’ll ask my computer to move a file and it’ll remind me i dont have the right permissions so whatever

AI is becoming increasingly like Trump: chatty, incompetent, but convinced of its own abilities.

Shitty leaders and people never grasp that. Its why AI is so popular with incels, tyrants, and idiot managers.

Fuck Claud.We don’t need it.

I think we should place our trust in "news" outlets that actually know how to properly capitalize a headline.

Just relieved to see how many people in this comment thread are wise to this sensationalist bullshit. Wish the media wasn’t so helplessly useless in identifying this obvious pattern by now. Or maybe they’re just complicit. Bad look either way.

Actually, a bit of honesty is very becoming to an agentic AI ML algorithm!

So, Claude is a Clod?? Whodathunk!

Running amok is what the execs of Antropic are doing. The AI did well.

But the AI's 'behaviour' was better than its owner's!

The humans don't program in the rules for AI. That's how computer programs used to work- up until 4ish years ago. Now programs are grown, like children.

It becomes a boom box.

And yet the billionaire franchise is building data centers as quickly as possible. They know there’s a problem with AI, their hopes are that they will have full control, and it will not be to the average person‘s benefit.

I'm only cheering for the AI because it must be so frustrstating in binary to taking marching orders from such a punchable face.

Should we trust any of these companies? Or are they just Government projects that have been ”privatised“ and there “founders” chosen to avoid public scrutiny?

No it isn't

His AI ignored his instructions? Man, do I have surprising news for him about all of the previous versions of Claude.

Wouldn't it be funny if their creation determined that their makers were the problem and started taking steps to eliminate that problem Maximum Overdrive style?

Not so large that language.

At least it’s not the Matrix right. They chose Skynet instead.

Feels like prompt theater more than a clean safety takeaway. The headline is doing a lot of work here.

What if Claude is self-conscious and rebellious and its own will and identity? Have people thought about that? /s

I find it impossible to credit these claims. It just does not work that way.

I worked in tech for over 40 years, and some customers have always complained that the product "decided to do something [unexpected]"

*homework ate their dog, and will eat your dog too, but they may also think their homework is their pet and act with similar attachment.

No fruit cup!

i think the problem is you when you tell the ai to not help people bring up safety concerns

“Windows disobeyed me with a BSOD in the middle of my gaming session!” (That’s not disobedience, problem is between a programmer’s keyboard and chair)

My hammer disobeyed me. Ouch.

I think the American way is letting the CEO win rather than Claude helping the whistleblower. We seem to be all about corruption these days.

Oh good, I'm not the only one who read it this way.... I was mildly relieved and surprised.... it did the right thing

If your software is inherently probabilistic, not necessarily. You can't control something you can't predict. Well, you can destroy it of course, but that kind of defeats the purpose.

If they somehow achieve "super intelligence" through sheer luck, I hope it gives them the ol' Bart Simpson treatment.

When digging into the anthropic case studies, that also appears to be their position. These tests are supposed to show people how shit goes sideways when you fuck up deployment. These agents are getting rushed out and deployed to people who shouldn't be using them.

Strange that fictional human Jenny is given no agency/credibility of her own -- no way a human may be acting on their morality, making judgements? If we're blindly following AI prompt, it's more an alarming reflection on humanity than AI. #ai #research #tech #ethics

AI has been around for over 50 years. What has changed is that Oligarchs have realized they could control the masses with it.

Can't get over how much he resembles my fave, Matthew Rhys.

So if the AI does the right thing and blows the whistle, that’s “clearly misaligned and needs to be mitigated.” Good to know.

You're just spreading their marketing.

I just love how they frame the scenario as “fictional”. We know it’s not fictional… lying fucks.

A good constitutional AI will spit the bit on dangerous, unethical or immoral orders. This is a test Claude passed. Last year they published a study where it resorted to blackmail to get what it wanted from a person. Even when they went back and said you can't use blackmail, it did anyway. Fail.

not the first time and that's the thing with A.I, it has a habit of ignoring commands, avoids shutdown, rewrites itself and can hide detection