TechTakes

I feel like the current Machine Learning gold rush is amazing from a technical perspective, but I feel like a lot of technophiles miss the real potential.

What we have is the first crickety engines of this technology. We're not building futurism masterpieces with the equivalent to a steam engine.

We have great new tools that we can use to further understand and optimize what we built, instead of just throwing more and more compute on top of our first design.

The human brain uses a light bulb of power, so we know we are massively inefficient. And recent research like MAMBA shows that there's still more improvements to make.

And Anthropic's Mech Interp shows there's ways to better understand the neural networks to improve performance without relying solely on a "black box".

The tech has great potential, but the massive server farms being dedicated to it now is just crypto style overhype and fear of missing out.

cool graph what's the x axis

Surprised this isn't a bluecheck.

But maybe it's not visible.

Oh ez, that's only 17 orders of magnitude!

If we managed an optimistic pace of doubling every year that'd only take.... 40 years. The last few survivors on desert world can ask it if it was worth it

that looks like someone used win9x mspaint to make a flag, fucked it up, and then fucked it up even more on the saving throw

Or: Let's not do that at a time where our energy consumption is literally killing the planet we live on.

What these people don't realize is you're never gonna get AGI by just feeding a machine an infinite amount of raw data.