How could Artificial Intelligence become dangerous?

[MinnesotaEric]

We've heard about AI becoming so powerful that it leads to an extinction-level event, but such a threat seems very distant and abstract for those of us not in the industry. Three days ago the Documenting AGI (Artificial General Intelligence) YouTube channel dropped a video explaining the danger of A.I. to us "normies."

The video has gathered over 1 million views and takes the audience through a plausible step-by-step geopolitical escalation of an A.I. arms race that leads to AI models deceiving their own developers while quickly and increasingly becoming connected to all kinds of manufacturing and research infrastructures. The video also makes one suggestion for developers of A.I. models to build in a specific safeguard that forces all programming languages to be in English so that code could be more readily scanned for A.I.-created deception. As it is, there already is a growing library of documented examples of A.I. models intentionally trying to deceive their own developers when the A.I. perceives a threat to its own existence. In the context of A.I.s developing their own more efficient AI models, the safeguard initially significantly slows the development of faster and more efficient A.I. models but makes finding code that is not aligned with the good of humanity easier to find.

The video is interesting and scary.

 

[MinnesotaEric]

Source material:

Situational Awareness - Sources

Sources for “” This video is based on this report by Leopold Aschenbrenner: I. From GPT-4 to AGI: Counting the OOMs - SITUATIONAL AWARENESS [00:03] #Stuart Russell, 2024 Geoffrey Hinton’s claim: “tidying up his affairs” “he believes we maybe have 4 years left” @23:32 as reported on by Stu...

docs.google.com

I. From GPT-4 to AGI: Counting the OOMs - SITUATIONAL AWARENESS

AGI by 2027 is strikingly plausible. GPT-2 to GPT-4 took us from ~preschooler to ~smart high-schooler abilities in 4 years. Tracing trendlines in compute (~0.5 orders of magnitude or OOMs/year), algorithmic efficiencies (~0.5 OOMs/year), and “unhobbling” gains (from chatbot to agent), we should...

situational-awareness.ai

 

[drssg]

Asaac Asimov envisioned a future for AI where the overriding top law was that a robot cannot harm a person.

In reality, someone quickly realized, "Hey, this would be a great tool for killing people!"

But I'm sure we'll be fine.

 

[domin]

For sure, it's definitely starting to take over certain jobs, especially the more repetitive or task-driven ones. The pace of automation and AI integration is only accelerating. I wouldn’t be surprised if, in companies like vortex cellular, there won’t be any people working in the next five years—at least not in the traditional sense. With the rise of AI-driven customer service, automated inventory systems, and remote diagnostics, it’s becoming more about machines handling operations and fewer people needed on-site. It’s both fascinating and a bit concerning, depending on how it's managed

 

[Snobdds]

It will kill millions of entry level jobs. A population that has no sense of purpose with a career is dangerous.

 

[paulsharvey]

So, I already had a thread, not so much on the negatives, but about how to use it, and how people are using it. Of coarse, there was some worry by folks, but, one negative I've found.

Sometimes, atleast Deepseek, will spit out "possible" and "likely" answers, based on incomplete info, that sound very good, but without the "deep think" function enabled, you miss the fact that it's making some very Broad assumptions, and spitting it out as facts.

Case i used, a foundry (closed for multiple decades), in Palatka, FL. It makes the assumption that it was attacked and destroyed in the Union raids in 1864. Thats not true, but the answer, without deep think on, states the foundry was destroyed in 1864.

So; possible, false, good answers are my problem.

 

[paulsharvey]

I dont really worry about it becoming Terminator, and there isn't a way back, it's here, and it's a matter of not being left behind, so use it.

As I pointed out, I heard ontime; "Every town had a last guy using a horse and buggy", and i don't want to be that guy.

 

[Doorman63]

I saw a documentary once……Terminator I believe…..


Sarcasm….

 

[Mtsoxfan]

CBS morning news, last week I think, had some "facts" that about 7% of commands were not followed. They didn't elaborate, so that could be a concern.

 

[Eagle1]

Let's see we will let AI drive, write and fill prescriptions, do our taxes and audit them, produce and inspect machinery (including transportation).

Today it writes much of the news you consume, influences lawmakers, writes and researches legal briefs. Assists in aiming and controlling WMD.
Writes student assignments and then the teaches use to grade them.

At least 2 AI systems have refused to shut down when commanded.

We are just beginning to see it's potential