Video yükleniyor...
Video Yüklenemedi
RSI section from the AI documentary Machine God The next threshold is Recursive Self-Improvement: the moment when AI can improve itself without human assistance. For decades this sounded like science fiction. Intelligence explosion scenarios imagined a system rewriting its own code, becoming smarter, then using that new intelligence to... show more
62,550 görüntüleme • 2 ay önce •via X (Twitter)
17 Yorum

Machine God trailer

Great video. I wrote about this topic earlier today.

One of the most overlooked aspects of neural networks is the trade-off between efficiency, security, and cost under resource constraints.

In the last 2 years, we only see bigger models, trained on more data, and overfit on more tasks. The recent Astra is an example. It had PhD level of mathematics, physics and industrial design knowledge. After you teach it how to use CAD. It still can’t do real engineer work

👀⚡️

"Capital buys compute; compute trains models." That's in the text, and then it's dropped. But it's the whole steering mechanism. 🙂 The loop doesn't optimise in a vacuum. It optimises toward whatever the benchmark rewards — and someone writes the benchmark, funded by institutions with a quarter to report. The scary word in the piece isn't autocatalytic. It's "improvement." Nobody in the loop has to define it. Which means it's already defined. Just not by anyone who'd say so out loud.

What is guaranteed to happen is the world will be unrecognizable after a few cycles of RSI. But then, if we compare to history, relative stability is an exception to the rule.

why are you promoting doomer shit?

Fascinating insights, Steve! It’s incredible to think about where AI might take us in the future.

when will the documentary be released?

The whole argument is an inductive fallacy. You can either have perpetual betterment (optimization) or break away from old categories, not both.

There’s a big assumption. The current underlying deep learning technology has no ceilings. Meaning this is the correct path to AGI, RSI can only speed up this process. (Do more trials and errors). But, this hasn’t been proved yet.

When will the full documentary be available?

the part you skip is verification once the loop actually speeds up human review scales linearly, a broken eval compounds exponentially across ten generations of models built on top of it that's the boring unglamorous chokepoint, not compute, not raw model quality worth the full watch

And what is really, really scary about all of this is that it is in the hands of extremely greedy people who will stop at nothing so that they get their hands on those billions of dollars.

There is an entire section of the film about "techno-capital" 👍

"better models attract more capital" There's an obvious human bottle neck. Unless AGI/ASI starts front running global energy and currency markets and crypto to accrue its FU money?

