In science fiction, imagination has always had no limits. Writers could invent anything, and for decades the machines that spoke did so in a deliberately mechanical way, like with clipped, monotone, full of "affirmative" and "does not compute."
That robotic voice was not a technical limitation so much as a storytelling choice. Creators wanted the audience to know, instantly and without doubt, that this was not a person.
The artificial had to sound artificial so the human could remain clearly, comfortingly human.
That distinction has collapsed.
The newest language models (LLMs) no longer announce themselves with mechanical cadence. They hedge, digress, joke, admit uncertainty, and chase metaphors the way people do. The old sonic marker that once separated machine from mind has quietly disappeared.
And not only that, LLMs are also smarter than any single human being.
Marc Andreessen believes that higher threshold has already been reached:
AGI (Artificial General Intelligence) is already here.
He further explained this in an appearance on the Joe Rogan Experience:
[AI] was a baby. It was a little, little, yes, a little tiny baby learning how the world works. The technology advances in the last three years have been mind-boggling. Like, crazy. Amazing. Impressive.
I actually, people talk about this concept called AGI, which means artificial general intelligence, which basically means an AI that's as smart as a person. And I actually think we crossed that about three months ago.
I think it was with the very latest versions of the leading models.
One of the reasons people are having a hard time understanding what's happening in AI is because it's moving so fast that if you don't use the latest thing, you don't understand what's happening because you're not seeing it.
[...] in my line of work, 99% of the time the answer that I'm getting from the AI from those from the most advanced models is better than I would get from talking to basically almost any expert I have access to.
The statement is based on simple facts.
Alan Turing proposed a test in 1950 that asked whether a machine could hold a conversation indistinguishable from a human's.
By the original terms of that test, today's leading models should pass without serious difficulty. They already do in most casual settings. The problem is that the test itself has lost its status as a gold standard.
Passing it no longer feels like a decisive achievement; it feels like table stakes. Something more demanding is required before anyone is willing to call the threshold crossed.
For nearly any topic, top models now outperform world-class specialists. While Andreessen has not published systematic data behind this observation, and that LLMs do make mistakes, it's safe to say that the AI can answer all questions asked with expertise that sometimes can go beyond human comprehension.
LLMs are designed to be this way.
One of the really fun things I do with AI is, you know, I'll ask it a question. I'll get back this complicated answer, and I'll just be like, "This is too complicated for me. I don't know something in quantum physics or something." And I'll say, "Explain it to me like I'm 10."
And it gives you the answer. It's like all of a sudden it's talking to you in terms you understand. Then you're like, "All right, this is still confusing. Explain it to me like I'm five."
At night, what I'll do is keep going all the way back. I'll say, "Explain it to me like I'm two."
His broader point is that, the moment people think "I don’t know how to figure this out," is exactly when most people give up. And that is exactly when they should open the AI.
The only real skill left, in his view, is knowing what to ask. The bottleneck is in the users' head, not the model.
The productivity shock is most visible in coding.
Virtually every serious AI-using programmer Andreessen knows is working more hours than ever, not fewer. The expected outcomes, either mass exit from the profession or dramatically shorter workweeks, have not materialized.
Instead a new term of art has appeared in Silicon Valley: the "AI vampire," a term to describe AI turning humans into a vampire.
So everybody I know who uses AF for coding, you would think basically one of two things would have happened. One is they just would be out of the profession entirely because there's no point anymore. Or you would think, well, maybe they just have a better life now because they're working less, right?
And so if AI coding makes them four times more productive, if they can write four times the amount of code in the same amount of time because they've got AI helping them, then maybe they're working only a fourth the time and now they've got a great life.
What's actually happened is, virtually to a person, they're all working more hours than ever, to the point where there is a new term of art that's used in the Valley called the "AI vampire," which is when AI turns you into a vampire. You're up all night doing AI coding because you are so productive.
No matter how much programmers complain about their work, there's a moment when everything finally clicks.
The code runs, the bug disappears, and the result is almost euphoric. Once they experience the leverage AI provides, the ability to build faster, solve harder problems, and accomplish more than they could before, it's difficult to go back.
It's difficult to stop.
And this is where Andreessen’s take on AI risk is characteristically dry.
Unlike AI doomers, who believe AI will replace humans and eventually take over the world, creating a dystopian future that has long been a staple of science fiction, Andreessen sees things very differently.
He positions himself more like a techno-optimist.
According to him, it's all humanity's doing. AI doesn't invent goals of its own. Instead, it simply follows the instructions it's given, and following instructions is something it does exceptionally well.
[AI doomers are] people who were worried about AI doing bad things were writing about AI doing bad things, which has given the AI the training data to be able to write the Netflix scripts in which AIs do bad things, right?
And so as we say, the call is coming from inside the house, right? Like, if you're worried about bad AI, rule number one is: stop writing internet posts about bad AI, right?
But of course, number one, of course people are going to do that because people are going to write everything. And then, as I like to say, number two is every bad thing you can imagine is in a novel somewhere or in a movie, right?
Or has been discussed in an internet forum. And so it's all in there. These are powerful things, and it's all in there. And a fully unconstrained one will plan a bank robbery. It will do it.
If someone asks a fully unconstrained system to plan a bank robbery, it will do it. The scripts, in other words, are still being written by humans.
The models amplify, remix, and execute what we have already put into the collective record. Whether that produces abundance, addiction, or catastrophe depends less on the sudden arrival of AGI than on the questions people choose to ask and the constraints they decide to keep (or remove).
AGI, if Andreessen is right, is already here. The vampires are already coding through the night. And the stories the machines tell are still, for now, the ones we fed them.
