Every time somebody says "AI safety", I hear "perpetuum mobile". Those two things are literally on the same level of nonsensical ridiculousness, it's just that one of     See More lves information , while the other involves energy . We are now in the information (AI) stage of complete ignorance of Intelligence Laws, that is matched only by complete ignorance of the Laws of Thermodynamics (aka energy laws) of the early years of the 19th century.

The real (important) difference, however, lies in the fact that no one has ever died by trying to create an impossible "perpetuum mobile" machine, whereas literally everybody on the planet is going to die by the hands of a bunch of ignorant (and, one could argue, schizophrenic [as in completely-disconnected-from-reality]) morons trying to create a "safe AI".

Then again... everybody always gets what they deserve (especially morons), so I'm not really complaining here.

I am simply stating the facts.    See Less

- "Our mission is to ensure that artificial general intelligence benefits all of humanity"
- Literally partners with Rupert Murdoch     See Less

What I get from the Anthropic paper is the possibility to find the node patterns relevant for most, if not all, responses - the overlaying 'foundational' algorithms that says "i     See More an input of this type, do this", making it possible to scale down the models, yet getting the same intelligence out. At least I find that concept fascinating. What if we could one day scale down to the absolute minimum God algorithm - a single line of code (that probably just says "42")..? 😁    See Less

Oh no! You said asteriks!     See Less

How can we predict future scaling-performance without yet having trained on said scale...?

It's akin to saying "we haven't yet come close to the bursting point of fire e     See More bble".

You can't know until the bubble is already popped. You can't know scale:performance until it's already begun to level    See Less

great white -> orca -> humpback -> blue whale -> bloop -> your mom -> Cthulhu     See Less

It's got to be said: capitalism really does drive progress. Progress for bad and progress for good. The pressure AI technology puts on manufacturers to make faster and more efficient chi     See More t.    See Less

Finding the racism node and showing that supercharging it leads to questionable content is massive. This means that instead of training the model to be less racist (or whatever you're tr     See More ign), we can train the best model we can - identify "bad" neurons and reduce their influence. Alignment seems to really hamper model quality - perhaps this will allow full capacity and still support alignment. Also, alignment can be cheaply adjusted over time instead as a training finalization step.    See Less

Now that its becoming more evident that AI is going bring a drastic change in job markets, what would you suggest would be relatively "safe" and helpful(for humanity) career for th     See More ble future?    See Less

Arguably. You can train on the whole internet and STILL not be an Einstein or a Nietzsche or a Van Gogh of today. Why? Simply because those ideas don’t exist within the latent space, they     See More IDE of it. I’m happy to be wrong, but I’m always 10 years at least out from AI predictions, 20 if it’s from a sales pitch.    See Less