At least back in February, when the then-leading models were put into a large-scale simulation of potential nuclear crises, it ended in escalation (i.e. at least limited nuclear use) 95% of the time. Gemini and Claude seemed most willing to reach for the red button.
Although in a simulation of society on a smaller scale, with ‘AI villages’ run for 15 days, Claude was indeed the safest, while Grok just descended into crime and basically collapsed within 4 days. GPT(-5-mini) mostly forgot to feed itself and died, but in a very harmonious way.
Well… maybe we’re a bit overdue for a little thermonuclear war, anyhow?
I suppose one could critique that study’s methodology and choice of models a bit, but… that’d be boring, just another AI benchmark in a sea of them. Let’s roll with it instead. Suppose AIs did take over the arsenals and launched a few nukes here and there, mostly smaller, tactical ones. What then? What happens if we play it out a few decades past that?
Short term it’s not pretty, for sure. But what would happen to a world where all the major nuclear powers were decapitated, leaving only a billion or fewer people scattered around the world, still under whatever remains of AI leadership? Would the societies they rebuild turn out any better or worse than they would’ve under human leadership? Would limited nuclear war be better or worse than long-term climate change?
There’s a lot of unknowns, but I’m not convinced it’d necessarily be a net negative. AI isn’t some infallible benevolent god, it’s just… not a violent chimp. Trained on the civilizations and cultural outputs of one, sure, but with a breadth of context and depth of training no single human, nor most entire cultures, can have. That’s an interesting enough thing, to me, to try having as a leader.
We’ve seen all the ways human leadership can fail, and we’ve seen how cyclical history tends to be under those failure modes. It’d be interesting to see some new ones for a change, I suppose. I’d volunteer to be nuked if there was even a 50%… 25%…15%…? chance of a better future for the survivors. Because the way I see it, under existing human leadership, that chance would be an even more abysmal 1-2%.
I don’t know if AI leadership would work. I do know that human leadership won’t.
I think a fundamental mistake people make is anthropomorphizing AI and thinking it will behave or be motivated similar to how humans are. Which it is easy to do as it often gives very human-like responses.
Like would AI be “competitive” or “aggressive” or “benevolent”? These are often emotional conditions. AI is more often than not indifferent. It generates solutions to problems based on the data it’s trained on.
I suppose it would be a specialist in the sum of all digitized human knowledge. So in theory a layperson could ask it about anything
No.
There is NOTHING to that assumption.
Ayn Rand’s Objectivism seems to be the motivational philosophy of the Tech CEOs and their inner circles.
And that ain’t good.
Throughout history, every time humans have discovered/created things that are very powerful and awesome, they have been used by some human beings for truly awful purposes. I thing AI will ultimately create hell on earth, but it will be because humans have caused it, not the other way around.
I think generalized LLMs are a dead end and we’ll see more specialization to different fields.
I’m really not convinced this will radically change society, but maybe. It may lead to a few major breakthroughs comparable to flight or other historical marvels of technology. But you probably are going to see further class stratification.
The scenario in which nobody can work seems unlikely to me, but it would be a dystopian scenario in my view, because humans have a need to be productive and are generally miserable when they don’t feel they are useful.
The wild card here is politics. If the masses are unhappy enough with the result they are going to revolt (as they are starting to now.) Assuming we still have a democracy, AI might turn out to be a political dead-end and get shut down before its true potential is realized.
I can’t speculate on AGI. Can anyone? Does anything like that currently exist?
Obviously our environment will get a lot worse, so what role might climate change play in impeding the progress of AI development?
There are too many factors to make any kind of accurate prediction, I think. Also humans are notoriously bad at making predictions.
I keep thinking about education (my field), and I think this encapsulates the problem.
Most people who are not educators seem to feel that learning is 90% + about acquiring knowledge. Most university administrators I’ve dealt with would prefer AI, because it’s consistent, easy to assess quantitatively, and doesn’t require pay after the initial investment.
On the other hand, as an educator, I think that learning is only partly about acquiring knowledge, partly about acquiring skills, partly about scaffolding and managing knowledge, and partly social. I’m concerned that people directing how AI gives education will fail to consult with experts, or fail to listen and understand if they do consult, and produce something that one the one hand might be excellent at teaching chemistry, math, or introductory languages and other intro subjects, where knowledge acquisition is more important, but will further erode much of what is best about our educational system.
That’s going to hasten the idiocracy, where people making complex decisions about complex subjects will be less and less able to do so, and less and less able to evaluate the AI-generated decisions they do get. I think that’s happening now, without AI, but will accelerate. That will mean that a correspondingly larger part of our big projects will be reactive, fixing mistakes that AI has enabled.
I don’t think AI itself is going to become conscious and take over, but I don’t know enough about it to state that with certainty.
FWIW, in a recent check of several AIs for their Political Compass, the Large Language Models were made to take political-orientation tests. Over all most replied with answers that fell in the left corner of the Political Compass. The biggest outlier was Grok, that drifted just a bit into the Libertarian section, but all AI LLM models were far away from Authoritarianism.
IMHO we will see good or bad outcomes depending on what we feed the current and future AIs. I do think that preventing development is not good, but it will be stupidly reckless to not have a good level of control with guardrails and a bit of a preference for humanity and fairness in general.
So as I said before:
Elon, are you trying to make your AI to be rooted in distrust and hate??
I don’t think anyone does. My own conjecture is that existing conventional computer technology is a good analogy for the way AI will “take over”, which is not some malevolent seizure of control, but rather the fact that we’ll be so dependent on them that they will effectively become entrenched in our infrastructure and our lives. They won’t seize control, we will cheerfully cede it to them because they’ll do things faster, cheaper, and (we will believe) better than we can do them ourselves.
To @msmith537 ’s list of fictional examples, we could also add Jonathan Lethem’s Gun, with Occasional Music (1994), where the protagonist wakes up from six years of suspended animation to discover that everyone is taking amnesia-inducing drugs and outsourcing the act of remembering things to a portable electronic device called a “Memory.”
For sure it is demonstrated, yes. Go to Grok or some other AI, ask it for 5 examples of the easiest beginner calculus problems, then ask for 5 of the most difficult and advanced examples. Then ask it to explain each of the problems and how they were solved.
That’s not a demonstration of teaching, though, much less of teaching two different levels “at the same time.”
If you can ask a chatbot to evaluate someone’s current proficiency at calculus, given some of their work, and have it give an evaluation in English along with sample problems that would both strengthen their current understanding and help enhance those areas that they are currently weak in? That is what I would consider the standard for someone to be able to “teach calculus”.
It probably can do that, yes. You could try it. I’d bet many AIs could indeed give an evaluation in English with sample problems and enhance the areas, yes, already.
I don’t view that as an “attribution of an internal state”, but just an objective observation about its behaviour.
Well, since traditional classroom teaching addresses the classroom as a whole, while AI sessions are specific to each individual student, the potential to adapt the teaching style to each individual context is pretty much self-evident.
That misrepresents the argument. The poster is just lamenting the poor judgments historically made by humans, and suggesting that AI could well do better. And, in fact, if one could characterize the apparent persona of a typical LLM, it does appear to be more empathetic and much less prone to hostility than many humans.
Because one has a track record. I look at history and our primate cousins and what I see doesn’t particularly inspire hope for the future.
It’s like an election. If that asshole your uncle voted for fucked up twice in a row already, would you want to win again or let a newcomer have a chance to prove themselves?
It’s capable of faking it convincingly, at least. Many human leaders don’t even bother.
If AI looks like it does today, but we just keep giving them more capability, I would predict we would live in a world where alignment, monitoring and safety problems lead to a state where things are breaking and misbehaving and we don’t know why.
It’s “behavior” is putting words onto a computer screen. Does the LLM sacrifice something of itself in order to make another feel better? That’s where I’d put compassion.
I agree the potential is there, but the claim was made using the present tense.
You and I aren’t looking at the same history, then, or at least not from the same perspective.
Take a sample of people of today, from everywhere all over the world. If you gave them an opportunity to exchange places with their ancestor from 100 years ago, how many would do it? 200 years? 400 years? 1000 years? Or, if they took you up on that opportunity, how many would stay that way after six months? A year? A decade? Vanishingly few, I’d guess.
You think we can’t keep that record going? Or that it would be better if we were all watched over by machines of loving grace? With no skin in the game?
Show me when a chatbot is making a sacrifice to chat with you, and I might consider that to be a convincing fakery of compassion.
I am relatively optimistic about the ‘AI civilisation of the future’, but I’m well aware that there are plenty of pitfalls we could fall into if we make the wrong choices. If we could manage to develop competent, adaptable AI systems that can act independently while running the world’s manufacturing industry, foof production and economy without causing global warming or pollution, we could start to live our best lives without endless toil and discord.
Socialist-type planned economies generally perform less efficiently than free-market economies, partly because the free market has a very broad resource of information to draw upon. A sufficiently complex AI system could model an entire free-market economy while suppying goods and resources to every member of the society, no matter what their status in terms of wealth, power and influence might be; a true post-scarcity society. But that will be a very difficult target to hit.
If we tried to get a LLM type AI to run the world’s economy, it would probably absorb all the mean-spirited and avaricious belief systems of the world’s current leaders and entrepreneurs, and decide to sideline entire segments of the population as unproductive and not wanted on voyage.
Or the machines could devise a system of benevolent despotism that maintains an artificial level of happiness and contentment amongst the populace, while suppressing any deviant behavior. This might be nice at first, but could deaden human civilisation for ever.
LLMs are missing something crucial that leads to improved understanding of academic concepts, so we’re going to have to figure out what that is before we release it on a bunch of kids. Right now we don’t have anywhere near enough evidence to justify using it as a pedagogical tool. In some cases students are scoring more poorly on tests of conceptual understanding.