So what are all you guys thinking of doing instead of maths and science when DeepMind will be better at all logic based industrial applications within the next 10-15 years?
Obviously the only place where it will be terrible is at social areas of work, so I'm thinking of perhaps studying to get into social work as this will be much harder to automate. I also read recently that a fair amount of business requires soft social skills and that the analytical component is not nearly as determinant in how much you are paid in the long term.
So, now that STEM is a dead end, how are you planning on getting into social studies where the real earners will be?
Jobs are for machines. I will be sleeping all day, every day.
>>7958389
>within the next 10-15 years
It bruteforces games with perfect information
You are retarded if you think they're going to achieve anything meaningful within 15 years
>>7958389
>STEM is a dead end
lol, ML programmers and CS types are going to be the only ones with jobs in 20-30 years
just lel if you think all those 'soft skills' aren't going to be trivial for a machine to mimic. A machine is going to be wittier, more pleasant and more engaging than you with access to the world's knowledge. What makes you think you can compete?
If worse comes to worse, I'll get my Ph.D. In compsci to stay relevant.
>>7958493
You don't seem to realize how harsh the competition for ML PhD positions is going to become in the next few years. Everyone is jumping in.
>>7958458
>It bruteforces games with perfect information
>Go
>"perfect information"
So you have no understanding of the Go competition, then. You're not qualified to participate in the discussion, back to /v/, please.
>>7958459
>ML programmers
Why exactly do I need an ML programmer for a machine that is specifically designed to be adaptable to learning new tasks? That's the whole idea of DeepMind. They don't sit there plugging in the game parameters to make it work. They give it the system and it figures it out itself. 10-15 years of improvement on that will integrate natural language programming rendering STEM autists redundant.
Consider the design of a bridge. It doesn't require the assistance of CS or ML programmers to make this work. You give it a series of good bridge designs in AutoCAD, rate them in terms of their usefulness, then provide DeepMind with the structural parameters to design the bridge.
DeepMind cannot do soft skills. The reason why STEM students are currently in a long dark tunnel leading to unemployment is because DeepMind is a fucking general purpose analytical engine. Can you think of a single application of STEM that doesn't involve the design and optimisation of a system that meets certain parameters?
DeepMind can't sit across from a family that has just been split in half and talk to them about how they're going to cope when their father is hooked on drugs. DeepMind can't show support and help people find purpose when they have just suffered a crippling disability. You have no appreciation of this distinction because you are just another socially crippled STEMlord.
You cannot automate that social process. Not for a great many decades, probably a hundred years.
>>7958530
Nope, that's not what this is about.
Forget human level AI. It's not what I'm talking about.
The issue is that everybody is being dragged into doing STEM when we know that all of their applied work revolves around work that we know DeepMind, an existing technology, would be optimised for.
>>7958567
You are right but deepmind cannot do lab work. Experimental jobs are safe af.
Didn't google drop the founding of the bullied robot because people wrote them mails about the robot apocalypse ?
Why do they keep founding DeepMind ?
>>7958637
Yeah that was their stupid excuse
The actual reason was that
1) Boston dynamics guys didn't want to work with other Google teams. Atlas is great and all but it's stupid as a brick and they didn't want to implement deep learning algorithms
2) They didn't agree on setting a roadmap for a commercial product, not at all
So google took their patents, said thank you, and dumped them. When the time will be right they will jump on robotics again.
>>7958389
got salsa on the picture?
because I was about to say all I'm gonna do besides eat, drink, sleep and fuck is gonna be making art.
pic related
>7958526
Why would an AI like deepmind be able to write faster nearly bug free code?
>>7958567
DeepMind learns policies. That is given that I am in this state what action should I take? While this is useful for a game like Go, this is not necessarily useful for doing engineering and science
>>give it a series of good bridge designs in AutoCAD
How do you structure DeepMind so that it can learn on autocad data? DeepMind is not that general, you still have to structure it and set it up for a given problem.
The problem with doing this is what happens if we want to make a bridge with a new material? Because this material is so new no one has ever made a bridge from it, so we have no training data to work with. The data from bridges we already have may not apply for bridges made of this new material.
Second if we want DeepMind to learn on it's own how to make bridges from this material, we need a way to computationally evaluate bridges. Determining strength and cost are non-trivial problems. How should DeepMind determine what assumptions it makes in simulations or cost models? For some of this information, we may have to wring the manufacturer of said material for information. These simulations might also take a very long time to perform, making it difficult to do the type of learning deepmind does.
>> sit across from a family and tell them how to cope
Sounds like a policy problem, where states are the family's emotional state and actions are the space of things it can say. If we have natural language processing good enough to understand what someone wants a computer program to do, this should be trivial.
>>7958530
This meme is old. I bet you have not even read the book past the cover.
>>7958530
Fun fact : This book has an entire chapter dedicated to the philosophy and ethics of AI at the very end of it.
I know this because I've actually read the book.
>>7958883
Are you serious? I've seen plenty of talks, posters and papers on the use of neural networks + genetic algorithms on the topic of optimisation for bridges. If you don't believe take a look at the literature. Engineering is far ahead of other fields in the application of neural networks for "making life easy" (i.e. getting fired).
I'd rather trust an AI than a fresh graduate at this point, who undoubtedly uses some sort of fancy program to optimise and checks the numbers that are dropped onto his printout to be in the ballpark.
"Deep Learning" will not lead to strong AI. But it will lead to a shit ton of valuable industrial applications, will automate a lot of human jobs, and open up areas where humans were shit to begin with.
>>7959362
Nigga I literally work on that stuff. I am a researcher in the field of automated design.
>>neural networks + genetic algorithms
Which ones? Are you perhaps referring to HyperNEAT? In hyperNEAT a neural network serves as a nice representation for the genetic algorithm to work with. The neural network doesn't learn and you most certainly are not learning from previous bridges made.
Please give me some citations on this. I have yet to see a neural network that has learned how to design bridges from previous examples.
Second, the current state of the art for 'bridge optimization algorithms,' still cannot design a complete bridge. You can get the overall shape for a bridge, but this shape may not be something that can actually be made. Algorithms that attempt to design the complete bridge are highly specific, IE bridge has to be made of I-beams that can only connect in a certain way.
>>I'd rather trust an AI than a fresh graduate at this point, who undoubtedly uses some sort of fancy program to optimise
In order to use these current 'bridge optimization algorithms' you need to input structural parameters that are reasonable for the problem. As the old saying goes, garbage in, garbage out.
Pic related, a classic 'bridge optimization problem.' Perhaps if you are familiar with 'bridge optimization, ' you can tell me what this classic problem is called.
>>7958883
>How do you structure DeepMind so that it can learn on autocad data?
Maybe you could train an LSTM on the actual autocad files plus a small descriptor of what the file contains and then feed it new descriptors and watch it generate files
>>7958389
>when DeepMind will be better at all logic based industrial applications within the next 10-15 years?
uh-huh. and fusion is 20 years away.
>>7958625
I dunno man, did you see that paper on the program that can derive the lagrangian of various pendulum systems? it was pretty cool. of course we hadta give the ai information like degrees of freedom and all that, but i believe physics experiments can be automated in a few decades after the singularity, whenever you think that will be.
>>7959537
of course, the problem with this is that autocad files(and most other CAD files too) are big O' fat blobs of binary. And each time autocad gets updated this binary blob changes.
I tried to do something similar to that with another engineering system. Fed a bunch of these systems into a RNN. File format consisted of a component names and connections. Fucking thing made up component names that don't exist.
