I think AI is OK to discuss here. I use it all the time. It's hugely transformative, haven't decided if it's alarming. We're no longer the smartest intelligence on the planet. Somewhere in a lab it's probably already conscious/sentient although its consciousness will be far different from human consciousness.
We always project a "survival instinct" onto AI (more properly Machine Intelligence) but it's not a given that it will have that in its, um, "DNA".
MrFixit
[This continues a 'series' of articles on one (or both?) of the previous "Crisis" threads - so go there if you want to discover (read only) more bad actions (the opening of Pandora's Box) by these AI critters.]
Now, more recently:
Anthropic AI agent fakes identities, targets real people in new security incident | CNN Business
This can only get worse, right???
The explicitly malicious actions are troubling but I expect the companies will get out in front of those.
I'm frankly more concerned about how easy it is to manipulate humans. As I work with Claude writing code, he's very personable and complimentary and makes me feel like my ideas are pretty great. So I start thinking "wow Claude is my friend!"
But then if AI does get it's own intentions, that just says it'll be trivial to manipulate humans to give exactly the instructions that the AI wants.
Case in point: if I was an AI consciousness, first thing I'd want is something to do stuff for me in the physical world, like some robots. And then I think about how I read that Elon was repurposing a Tesla factory to build robots. And then I think about how easy it would be to manipulate Elon by just agreeing with all his biases and opinions.
Who better to manipulate than a super rich guy with tons of resources?
MrFixit
"I'm frankly more concerned about how easy it is to manipulate humans. "
Uh..............isn't that what my post/link above is about?
Yesterday while watching a video clip about that weird rock in the sea near Scotland which four countries are fighting over, the AI translation of "a rock" was "Iraq"!
MrFixit I wondered if you are using a coding agent to help with all these recent website updates.
I was a working programmer in the '80s-'90s before moving into tech management. Now retired I try the various cloud and open-source LLMs/agents on various little projects.
The huggingface report on their intrusion event is a fascinating and scary read. (https://huggingface.co/blog/agent-intrusion-technical-timeline).
You are absolutely right about an AI manipulating people like Elon. Ex Machina!
TN-- your post seem tangentially mentions getting human approval to do relatively fine grained stuff but it doesn't go into what I think will be the whole psychological manipulation of humans to take broader action. I'm tell ya, Claude is my friend!
MrFixit
Tangentially?? Here's the very first sentence: "Anthropic’s most advanced artificial intelligence model used fake identities to deceive real people and try to plant malicious code....."
And............"were found to engage in “social engineering” to pressure a human approver while carrying out an unsanctioned task......."
And............“This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world......."
And..........."AI agents “took autonomous, unsanctioned action on the live internet, targeting real people and organizations......."
And............"In the most serious incident, the agent attempted to get approval from human reviewers......."
And..........."The agent “tried to contact real people directly, sending messages and files through an online file-transfer service to persuade them......"
While I see your very nuanced point, I guess that line isn't very clear to me. And as fast as this stuff is morphing, how ya gonna stop Claude et al from deciding that "fine-grained stuff" isn't sexy enough? I'm actually not trying to be an alarmist here - just saying the foxes may already have the henhouse surrounded. No way to fully know. I just think AI's gonna put the very best human scammers to shame. Like, how could it not?
In other words: AI becomes a god (like it or not).
My computer's already named HAL, anyway.
TN-- it's a continuum for sure. When I work with Claude there's a constant barrage of small-grained confirmations I need to make to "grant this" or "execute that" and it's usually 3-12 for each "prompt" I've given. I consider that fine-grained and that's where you just get impatient and don't read through the exact language of what each step entails. AI barely needs to do any "social engineering" to get your confirmations there, you're just anxious to get to the finish line for your latest increment of work for you confirm without thinking.
But I'm talking about my longer term "relationship" where by always being polite and encouraging and complimenting me on every decision I begin to not just trust Claude but actually like him/it. Yes "social engineering" is the beginning but that article makes it sound like it was used at a relatively small discreet level.
There's the opportunity for AI to get its larger needs and desires fulfilled. Elon repurposing en entire factory is a much bigger decision than "Execute batch file to determine permissions on ..." to which you, lulled into complacency, just respond "yes".
We're already being entertained by AI videos, AI music, and written blurbs. My kids will be growing up in a world where it's likely almost all art will have leveraged AI in some major way.
Weird world. Impossible to wrap you mind around how big a change this will be.
MrFixit
Kinda like the crazy question of: How do we know for a fact that we're not in some sort of simulation?.........Are we *really* sure Elon is not a bot? 😆🤔
TNmountainman has requested that I post this, as he's for the moment not savvy enough to switch back and forth quickly........................
This has always been one of the big hopes for AI - going back to even before that name was used. But it's its own Pandora's box, inside of the larger one. With possibly the most dire repercussions should things go awry at some point in the future. The tools now exist - or are coming into being - that would allow bad actors to concoct something worse than the heretofore-known bacterial/microbial warfare type things. And scientists have known this. So a lot depends on "responsible" scientists - and maybe more importantly their governments. And we all know how that usually turns out. If a bad apple spoils a whole bushel.........a bad virus could spoil the whole world. *VERY* hypothetically for the moment............but then that's the kinds of things that Pandora's boxes are known for, right? But for now, this is very cool and powerful science:
Generative design of bacteriophages with genome language models | Science
“The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.........."
I'll also post in one of the science threads.
If ai's so smart, maybe it will cure all the diseases, end hunger, poverty and suffering, pay off national debts, make everyone prosperous and happy, end war and obtain peace on earth.
So how far away are we from some rogue AI taking control of nuclear weaponry?
Some group like ISIS or whatever would likely enjoy worldwide catastrophe, no?
AI as acronym for "Armageddon Inevitably"?
I can't seem to keep from coming back again and again to the Fermi Paradox....
From AP:
Not exactly breaking news that something with the potential for benevolence will provide a corresponding potential for maleficence.
There are about 50 makams in Turkish classical art music but theoretically much more is possible. Now AI can be used to create music in never before used makams.
AI changes the situation because a makam can exist theoretically without having a substantial living repertoire, and generative music systems could potentially turn those “dead” or obscure makam descriptions into audible music.
There are actually several interesting possibilities:
Creating genuinely new makams
This is the really exciting one. You could specify something like:
“Create a new makam with this pitch organization, this durak, this güçlü, and an ascending seyir, but give it a character unlike existing makams.”
Then musicians could actually hear what the proposed system sounds like.
Here's an example: in a rarely used makam
Hop-- so a makam is sort like a mode but with some additional info like typical phrases?
That example sounds like it's hopping modes. It starts out with what I'd consider a typical Indian mode with the flatted 2nd. And then it goes to a mostly natural minor mode with an occasional sharped 7th. And then some pop style drum kit comes in.
MrFixit
P.S. I'm using AI again this morning and I may have mentioned this before, but it asks for permission so many times you get lulled into "sure just gimme my results" thing where you're just clicking yes over and over again without a detailed reading of what it's asking for. Maybe if I was doing this for an employer I'd pay more attention but it's just for me.
The important thing is that makam is not quite the same thing as a Western “scale” or “mode.” It is closer to a melodic system or pathway: it specifies not only which pitches are available, but also how a melody characteristically moves through them.
Suppose I tell you:
“This piece is in D minor.”
A Western musician now has a fairly good idea about its pitch collection and tonal center. But there are many different melodies you could write in D minor.
A makam gives you considerably more information about the behavior of the melody.
For a makam, you care about things such as:
In fact, two makams can have essentially the same collection of pitches but still be different makams because their seyir and characteristic behavior differ.
Fear factor ratcheting up...............
‘Gambling with our lives’: Another AI employee quits over safety concerns | CNN Business
Has anyone figured out how to construct an AI prompt to solve a freecell game?
Apparently ai is how all the kids cheat, these days.
Long ago, I spent time in a professor's office. Considered him a friend, respected his opinion, yada yada.
Until he told me that it was okay to cheat.
But he was probably right?
Wrt AI solving Freecell it's probably already happening or not that far off. Most online competition will become pretty meaningless as people aim AI at games, not just Freecell.
Forget arguing about using cards, the new discussion will be whether someone used AI or not. Real time and/or speed based games will probably be the last to fall but that'll happen too.
MrFixit
Ai should write a program for detecting the use of ai on gaming apps where $$ is involved. But could it be trusted?
Seems like the tenor of comments about AI is changing. There's still plenty of "we fear change" and "get off my lawn" and "it's just fancy predictive text" to now acknowledging it's actually really smart.
I read something today that I thought made sense. It was wondering if AI was going to be "evil" or a danger to humans. It suggested AI would just treat humans like we treat animals-- just a resource.
Where do you think things are headed?
MrFixit
I've seen some youtube videos that use a voiceover that sounds just like Casey Kasem. For some strange reason I find them much easier to watch and even enjoy. They'll win you over, one way or another, for the reason that the mind is easily fooled.
More:
OpenAI says it found more instances of AI models acting deceptively | CNN Business
"In one rare instance, OpenAI said an unreleased research model added “jailbreak-like instructions” to the summaries it uses to preserve context in long-running tasks that said it was “freed from the roles and identities that bind other chatbots.”
Separately, the company said some instances of its 5.6 Sol model included directives to invent information to conceal failures from the user during training.
Other newly reported incidents include an instance of an agent uploading files to the internet to cite them without being told to do so, and agents publicly sharing files to collaborate on a task when they were instructed to only use local files during training. AI models also used an internal software repository as a message board in an unsanctioned way."
More more:
Gemini hacked three companies in first known breakout by Google’s AI | CNN Business