OpenAI No Longer Takes Safety Seriously

[0] https://deepmind.google/research/publications/35829/

This debate is probably mostly just an argument over semantics. What exactly do we mean by "Large Language Model"? What exactly do we mean by "Artifical General Intelligence"?

On the strict end of things, if "LLM" the current state-of-the-art generative transformer architecture, and if we think that "general intelligence" requires something like sentience, or at least a certain capacity for flexible, iterative symbolic reasoning, then I think that most people would agree that LLMs clearly aren't any more capable of achieving artificial intelligence than a stock 1989 Buick LeSabre is of achieving sustained flight.

But the more flexible you're willing to be about how you define "large language model" and "artificial general intelligence", the more room there is for it to be possible. But also, the less practical value there is in making such a statement. Beyond some hand-waviness event horizon, you can probably just declare that any model capable of handling natural language inputs and outputs is a language model, take a nice slow toke, and begin pontificating about the Church-Turing thesis.

the section: "No, AI Safety isn't a fringe concern by sci-fi weebs"

> AI Safety / AI Risk used to be less mainstream, but now in 2024, the US & UK governments now have AI Safety-specific departments![10]10 This resulted from many of the top AI researchers raising alarm bells about it. These folks include: > Geoffrey Hinton and Yoshua Bengio > Stuart Russell and Peter Norvig > Paul Christiano

> I'm aware "look at these experts" is an appeal to authority, but this is only to counter the idea of, "eh, only sci-fi weebs fear AI Risk". But in the end, appeal to authority/weebs isn't enough; you have to actually understand the dang thing. (Which you are doing, by reading this! So thank you.)

We’ve seen a huge shift of wealth in the last twenty years in no small part passing through and at the insistence of Venture Capital and Private Equity. AI will be used to again accelerate this with more layoffs more outsourcing and lower quality services captured.

Imagine an AI agent which manipulates exposed SCADA systems. That could be incredibly destructive!

“Can we just turn it off?"

"It has thought of that. It will not give you a sign that makes you want to turn it off before it is too late to do that."

https://youtu.be/_8q9bjNHeSo?si=a7PAHtiuDIAL2uQD&t=4817

If AI ever replaces white collar jobs then that will be at an unprecedented scale and with essentially nowhere to shift to.

But so far it doesn‘t seem to happen.

Our parents and grandparents getting scammed by deepfakes seems like a very real threat. Not only do anecdotes abound (plus FBI's database of reports) but the elderly were already vulnerable to these scams since before the telephone was invented. LLMs and all the voice cloning tech really turns up scammers' psychological weapons to 11 and scale up processes that were previously very manual.

> AGI taking over the world? Does no one realize that without people going out every day and doing work the world ends in a week... Does no one remember when we had the spate of transformer shootings? A rifle or a bucket of water ends AGI threats

This isn't a real threat and I think most people are missing the more practical threat: someone - either a nation state or a small group of well resourced individuals - hitting an industrial singularity, allowing them to rapidly scale up self-improving robotics and automation to the point where their only limit is their ability to acquire raw materials. I think by definition, long before we hit an AI singularity with runaway super intelligence that can take over the world, we'll have a much dumber AI capable of maximizing killer paperclips for their human overlords to take over the world (that's my plan anyway :-)).

I don't think we're anywhere near the latter yet, but the language in the OpenAI blogpost announcing Sora points in that direction. The first step to an AI that can orchestrate a factory is learning the physical world through video. Once an AI model can follow instructions and directly control a robotic arm via camera, we'll be entering the next stage of the industrial revolution.

So, some organization deliberately plugging the technology into something dangerous like a defense computer network that includes computers that can be used to control weapons systems. That would potentially be more worrisome because one would assume that the organization setting it up, being a military agency, would be building it to be more resistant to attack than your average municipal power grid.

Also, we don't necessarily need to posit that it deliberately tries to take over the world, Skynet-style. Just that it does something surprising and unintended. Which is not necessarily all that far-fetched? Despite the leaps and bounds we've made in raw performance at certain tasks, this whole branch of technology continues to be notoriously flaky.

Militaries using AI checkpoints, guards, patrols, and loitering munitions (from a Humanitarian / rules of war perspective, not necessarily existential risk). Possibly to continue regimes where human soldiers would defect or differentiate civilians from armed opponents.

AI-reinforced media bubbles promoting alternative realities (most people blame this on social media, but from QAnonCasualties it seems like YouTube and Facebook recommendations hooked a lot of people before mainstream companies assessed whether they could be harmful). I think this is more core than "deepfakes" because people believe even fake Tweet screenshots if it backs the narrative.

On this site, of all sites, it's remarkable how many people overlook that scale and automation lead to important qualitative shifts.

This is like asking a dodo in 1598 to be able to explain what the threat of humans are. Their inability to do so isn't related to the probability that the threat is real.

Likewise, our inability to articulate the threat posed by AI isn't evidence that a threat does not exist. We simply don't know that threat is and even if we knew there was a threat, there's no guarantee that we would be able to comprehend it enough to describe it. Rather, our inability to comprehend super-intellegent threats make us more susceptible to identifying and stopping them.

Why don't we apply the same logic to the status quo? Because we are, for the most part, exposed to shocks and change within the regime of normal variance and have continued to exist which is evidence for continued existence. Planetary co-habitation with super-intelligence represents a many-times-beyond the standard variance of shocks to our natural order.

If any of the lifeforms we've extincted already were as intelligent as humans, would that have increased their chances of survival? Personally, I see that as an easy yes which implies that differences in intelligence are at least somewhat correlated with extinction outcomes. But perhaps others see it otherwise.

"Your problem has been found statistically unlikely and the projected calculations of the loss of your future business does not justify the risk to fix it. This is our final determination. There is no manager to speak to, and no higher tier to escalate to. Have a nice day."

- Gandalf the Grey