To Build An AI Like GlaDOS, You’d Need A Bitcoin Mining Operation The Likes Of Which The World Has Never Seen

Portal‘s GlaDOS is clearly a sophisticated artificial intelligence. It can speak. It can think. It can, uh, bake. But, using the technology of today, would it be possible to build GlaDOS, AI and all? In the video above, Game Noggin tries to answer this very question. Spoilers ahead if you haven’t played either game!

To build GlaDOS — or any AI for that matter — we have to figure out the whole “intelligence” part first. In terms of current science, the best way to do that is with a neural network and deep learning.

As the video explains, up until 2011, we were bashing away at the problem with CPUs. Then we tried it with GPUs.

It was better. Like, much better.

Graphics chips are insanely good at running the same numbers again and again, pumping out shader-manipulated pixels. This makes them hardware perfect for AI:

This parallel computing nature also makes the GPU ideally suited for training AI. When creating artificial intelligence, instead of rigid human-written software, programmers create a neural network … making a CPU run a neural network that needs to perform millions of simultaneous calculations would take forever … just 12 GPUs [from 2011 delivered] the same performance as 2000 CPUs.

The problem with neural networks today is that they’re designed for a single task. So, to get to GlaDOS levels, you’d need a lot of video cards. And if you had that many GPUs, chances are you’d be crypto-mining your way to financial independence (and a monster power bill) instead.

But, theoretically? It’s possible.

Can We Build Portal’s GLaDOS? [YouTube]


  • “But, theoretically? It’s possible.”

    Not if you believe this: “In terms of current science, the best way to do that is with a neural network and deep learning.”

    Deep learning is great for things like speech recognition, image recognition and game playing (DeepMind), but that does nothing for making (what we assume) is a general AI like GlaDOS.

    At a fundamental level, we still don’t even understand how we’d even design a general AI. So no, its not theoretically possible (yet).

Show more comments

Log in to comment on this story!