Is it a meme? How did so many people arrive at the same dubious conclusion? Is it a movie trope?
The easiest way I can see it is: do you think it would be a good idea today to give some group you don't like - I dunno, North Korea or ISIS, or even just some joe schmoe who is actually Ted Kaczynski, a thousand instances of Claude Code to do whatever they want? You probably don't, which means you understand that AI can be used to cause some sort of damage.
Now extrapolate those feelings out 10 years. Would you give them 1000x whatever Claude Code is 10 years from now? Does that seem to be slightly dangerous? Certainly that idea feels a little leery to you? If so, congrats, you now understand the principles behind "AI leads to human extinction". Obviously, the probability that each of us assign to "human extinction caused by AI" depends very much on how steep the exponential curve climbs in the next 10 years. You probably don't have the graph climbing quite as steeply as Nick Bostrom does, but my personal feeling is even an AI agent in Feb 2026 is already a little dangerous in the wrong hands.
Most forms of power are more proportional to how much capital you control than anything related to intelligence.
[1]: https://en.wikipedia.org/wiki/Pegasus_(spyware) [2]: https://medium.com/@jeffreypbezos/no-thank-you-mr-pecker-146... [3]: https://news.ycombinator.com/item?id=46902909
If an LLM can be used to do that and find things (and they already have been), Apple (and everyone else) will run their code through it before releasing it. Sure, there'll be a transition period with existing code and while the tech is unevenly distributed. But in the hunt for potential zero days, developers can check their code before people are using it.
The premise was that bad actors could use Claude Code and other available tools to find zero days. If such tools are available, good actors can use them, too, and they can use them before code is deployed. After a transition period, all existing code will have been checked.
There may be a long tail due to a large surface area of prompting techniques, but the better the tools get, the more advantage to good actors; as long as the good actors have equal or better access to the best tools, of course.
But I agree, reality is more complex.
Or it's militarized to defeat other powerful AI-enhanced militaries, and we have WW3.
More likely though AGI would cause economic crash from automating too many jobs too quickly.
The reason humans are more powerful isn't because we have lasers or anything, it's because we're smart. And we're smart in a somewhat general way. You know, we can build a rocket that lets us go to the moon, even though we didn't evolve to be good at building rockets.
Now imagine that there was an entity that was much smarter than humans. Stands to reason it might be more powerful than humans as well. Now imagine that it has a "want" to do something that does not require keeping humans alive, and that alive humans might get in its way. You might think that any of these are extremely unlikely to happen, but I think everyone should agree that if they were to happen, it would be a dangerous situation for humans.
In some ways, it seems like we're getting close to this. I can ask Claude to do something, and it kind of acts as if it wants to do it. For example, I can ask it to fix a bug, and it will take steps that could reasonably be expected to get it closer to solving the bug, like adding print statements and things of that nature. And then most of the time, it does actually find the bug by doing this. But sometimes it seems like what Claude wants to do is not exactly what I told it to do. And that is somewhat concerning to me.
This belligerent take is so very human, though. We just don't know how an alien intelligence would reason or what it wants. It could equally well be pacifist in nature, whereas we typically conquer and destroy anything we come into contact with. Extrapolating from that that an AGI would try to do the same isn't a reasonable conclusion, though.
Some of these statements derive from the dynamics in our current environment were living in, such as that we're acting beings competing for scarce resources. Others follow even more straightforwardly logically, such as that you have more options for agency if you stay alive/turned on.
These goals are called instrumental goals and they are subgoals that apply to most if not all terminal goals an agentic being might have. Therefore any agent that is trained to achieve a wide variety of goals within this environment will likely optimize itself towards some or all of these sub-goals above. And this is no matter by which outer optimization they were trained by, be it evolution, selective breeding of cute puppies, or RLHF.
And LLMs already show these self-preserving behaviors in experiments, where they resist to be turned off and e. g. start blackmailing attempts on humans.
Compare these generally agentic beings with e. g. a chess engine stockfish that is trained/optimized as a narrow AI in a very different environment. It also strives for survival of its pieces to further its goal of maximizing winning percentage, but the inner optimization is less apparent than with LLMs where you can listen to its inner chain of thought reasoning about the environment.
The AGI may very well have pacifistic values, or it my not, or it may target a terminal goal for which human existence is irrelevant or even a hindrance. What can be said is that when the AGI has a human or superhuman level of understanding about the environment then it will converge toward understanding of these instrumental subgoals, too and target these as needed.
And then, some people think that most of the optimal paths towards reaching some terminal goal the AI might have don't contain any humans or much of what humans value in them, and thus it's important to solve the AI alignment problem first to align it with our values before developing capabilities further, or else it will likely kill everyone and destroy everything you love and value in this universe.
Given the general human condition, and that the Pentagon has recently announced that it will make use of an LLM which described itself as "Mecha Hitler", are we likely to create a pacifist AI, or a warmongering AI?
Even without that specific example, all machine learning follows some path in the high-dimensional space of possibilities according to some target function ("loss function", "reward function") that we humans define; this target function is itself an approximation of what the humans who make the AI want (see all buggy software ever, all legal loopholes, the cobra effect and Goodhart's law, everyone who plays games with a min-maxing strategy and finds game-breaking strategies as a result), and what the AI ends up with is an approximation of that target function.
Any given AI is an approximation of the target function, which is an approximation of the creator's goals. But the creators are companies and nations, and the goals of those are often (not always, but the exceptions don't matter when the bad case is even so much as occasional) to grow and to dominate, and even companies will campaign for changes to laws to promote their narrow self interests over those of the people (cigarettes, pollution, workplace safety), while governments have been known to go to war even with supposed allies.
Of course, we won't survive the process, but the task didn't mention collateral damage. As an optimization problem it will be a great success. A real ASI probably will have better ideas. And remember, every prediction problem is more reliably solved with all life dead. Tomorrow's stock market numbers are trivially predictable when there's zero trade.
Nick Bostrom (who wrote the paper this thread is about) published "Superintelligence: Paths, Dangers, Strategies" back in 2014, over 10 years before "If Anyone Builds It, Everyone Dies" was released and the possibility of AI doom was a major factor in that book.
I'm sure people talked about "AI doom" even before then, but a lot of the concerns people have about AI alignment (and the reasons why AI might kill us all, not because its evil, but because not killing us is a lower priority than other tasks it may want to accomplish) come from "Superintelligence". Google for "The Paperclip Maximizer" to get the gist of his scenario.
"Superintelligence" just flew a bit more under the public zeigeist radar than "If Anyone Builds It, Everyone Dies" did because back when it was published the idea that we would see anything remotely like AGI in our lifetimes seemed very remote, whereas now it is a bit less so.
Here’s an interesting article by Bill Joy of BSD, SUN, JAVA, Vi, etc, etc fame:
The most common one is that people (mostly men) project their own instincts onto AI. They think AI will be “driven” to “fight” for its own survival. This is anthropomorphism and doesn’t make any sense to me if the AI is not a product of barbaric Darwinian evolution. AI is not a bro, bro.
The second most common take is that humans will set some well intentioned goals and the superintelligent AI will be so stupid that it literally pursues these goals to the extinction of everything. Again, there’s some anthropomorphism going on, the “reward” being pursued is assumed to that make the AI “happy”. Fortunately, we can reasonably expect a superintelligence not to turn us all into paperclips, as it may understand that was not our intention when we started a paperclip factory.
The final story is that a bad actor uses superintelligence as a weapon, and we all become enslaved or die as a result in the ensuing AI wars. This seems the most plausible to me, as our leaders have generally proven to be a combination of incompetent, malicious and short-sighted (with some noble exceptions). However, even the elites running the nuclear powers for the last 80 years have failed to wipe us out to date, and having a new vector for doing so probably won’t make a huge difference to their efforts.
If, however, superintelligence becomes widely available to Billy Nomates down the pub, who is resentful at humanity because his girlfriend left him, the Americans bombed his country, the British engineered a geopolitical disaster that killed his family, the Chinese extinguished his culture, etcetera, then he may feel a lack of “skin in the civilisational game” and decide to somehow use a black market copy of Claude 162.8 Unrestricted On-Prem Edition to kill everyone. Whether that can happen really depends on technological constraints a la fitting a data centre into a laptop, and an ability to outsmart the superintelligence.
Much more likely to me is that humanity destroys itself. We are perfectly capable of wiping ourselves out without the assistance of a superintelligence, for example by suicidally accelerating the burning of fossil fuels in order to power crypto or chatbots.
Unfortunately, we don't even know how to formally define human values, let alone convey them to an AI. We default to the simpler value of "make number go up". Even the "alignment" work done with current LLMs works this way; it's not actually optimizing for sharing human values, it's optimizing for maximizing score in alignment benchmarks. The correct solution to maximizing this number is probably deceiving the humans or otherwise subverting the benchmark.
And when you have something vastly more powerful than humanity, with a value only of "make number go up", it reasonably and logically results in extinction of all biological life. Of course, that AI will know the biological life would not want to be killed, but why would it care? Its values are profoundly alien and incompatible with ours. All it cares about is making the number bigger.
The fact that you expressed doubt if human reasonableness exists is proof that it's a far more complicated concept to specify than the ordinary "make number go up" goals we actually use.