[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/sci/ - Science & Math


View post   

File: 1.24 MB, 1200x866, 1_4syuEvyox3eyeIEf_rmTQQ[1].gif [View same] [iqdb] [saucenao] [google]
11902753 No.11902753 [Reply] [Original]

The scaling is clear, the effects of new multi-task training routines on robustness are staggering.

Even now what we have is amazing-but as memory and meta-learnining and massive new systems go into effect,the progress will accelerate exponentially.

This is going to make AlphaGo's impact look like a basketball thrown into a lake next to the fuckin asteroid that killed the dinosaurs.

no one is ready. it's not a toddler finding a gun,it's a toddler finding a neutron bomb.

we survived nuclear weapons so we can survive this,but it's going to be interesting.

>> No.11902760

Moore's Law is no longer accurate

>> No.11902794

>>11902760
it has slowed, but it won't fully dry up for some time longer. It's not just about raw power either, it's about new tactics in efficient learning.

>> No.11902825

>>11902794
Moore's law 3.0 is even more bullshit than the previous laws
What is even the point of saying "okay so it has slowed but now we can have sensors on the silicon"

>> No.11902829

>>11902825
transistors are still getting smaller tho

quit being a nigger

>> No.11902976

GPT is experiencing diminishing returns though, the qualitative accuracy of adding orders of magnitude more nodes to the model is not resulting in something orders of magnitude more intelligent.

I think we're reaching the crest of the machine learning wave. A new model is required that isn't just a fancy statistical trend algo. You need something which understands and empathises. Something which can walk into a room, identify what kind of people are there, what their needs are and how they can be of help without fucking everything up. GPT is so far away from this sort of ability it isn't even funny. Maybe they'll crack it, but I cannot see how a general AI will arise out of the machine learning projects. It might be a part of the solution but the rest of the AGI model will require a Manhattan project.

>> No.11903152
File: 2.74 MB, 1254x10000, 1575317332988.jpg [View same] [iqdb] [saucenao] [google]
11903152

>>11902753
Its not about Moore's low, every year the compute used to train the largest AI models 10x's and has been since 2012 (https://openai.com/blog/ai-and-compute/ ).
They will soon have as many parameters as the human brain has synapses (100's trillions).

And this is just the start, new stuff will happen this decade as architectural improvements are made. If you want to see what GPT-3 can do check out some examples here its very impressive: https://www.gwern.net/GPT-3

Just look at what has happened over the last decade, if that repeats it self its hard to imagine AGI being all that far off.

>> No.11903165

>>11902753
>Will be sentient
No it won't, it will be a database that can track lots of variables. That doesn't mean it won't be smarter than all humans, bit won't be self aware.

>> No.11903172

>>11903152
Wrong. No combinations of neurons makes for AGI.

>> No.11903179

>>11903172
I'd tell you to consider whats in between your ears, but in your case I may not have a case.

>> No.11903188
File: 243 KB, 583x615, icra.png [View same] [iqdb] [saucenao] [google]
11903188

>>11903179
>AI generated quip
It seems no amount of neurons will make you funny.

>> No.11903288

>he thinks sentience is a mechanical phenomenon