Artificial General Intelligence General Thread

[Anonymous No.9973577]

https://www.youtube.com/watch?v=zywIvINSlaI

http://karpathy.github.io/neuralnets/

ARTIFICIAL GENERAL INTELLIGENCE GENERAL THREAD

Discuss
- Deep Learning
- Machine Learning
- Artificial General Intelligence

Don't discuss
- Ethics
- Safety
- Philosophy

[Anonymous No.9973582]

>>9973577
You want to discuss something that doesn't exist and wont for at least 30 years (probably more), but not discuss the philosophy around this nonexistent technology, when that's the only real discussion to have on it?
Doesn't make sense.

[Anonymous No.9973593]

>>9973582
You can discuss potentials and other things. Just not "muh agi can't exist because of human nature" type things. Or worried that it will be racist.

[Anonymous No.9973678]

>>9973577
>Don't discuss
>Ethics
Good.

[Anonymous No.9973701]

DLSS AA by nvidia is a pretty interesting use to replace old code.

Seems like it should also be super good for collision detections

https://www.researchgate.net/profile/Emilio_Olivas/publication/3996853_A_neural_network_approach_for_real-time_collision_detection/links/09e4150c257f5a3bf1000000/A-neural-network-approach-for-real-time-collision-detection.pdf

[Anonymous No.9974459]

bump

[Anonymous No.9974467]

>>9973577
>>>/x/ is that way

[Anonymous No.9974468]

What sort of mathematical background do I have to have to properly understand the series of theorems of No Free Lunch from David Wolpert?

[Anonymous No.9974478]

>>9974468
Make a bunch of machine learning projects then care about the math. You will be uselessly hypothesizing about shit you have no experience with otherwise.

[Anonymous No.9974493]

What does /agi/ think about IMPALA?
https://deepmind.com/blog/impala-scalable-distributed-deeprl-dmlab-30/

[Anonymous No.9974495]

Basically lets say you are about to study a new subject. The best way of doing so is not studying the math behind it. The thing you want to counter-act is being stuck in a thought valley or thought track.

That means trying to view the problem from as many different and original perspectives as possible before you dive into literature. Because once you begin learning the "standard" route you become locked in to those patterns of thoughts or viewing the subject.

Mathematical background is not going to help you escape a thought pattern trap. It would be actually very stupid to do the "math background" in depth before diving into the subject itself.

[Anonymous No.9974539]

>>9973701
why is this in the agi thread kek

[Anonymous No.9974553]

>>9973577
>Don't discuss
>- Ethics
>- Safety
>- Philosophy
Retard

[Anonymous No.9974567]

>>9974553
>whats the ethics and safety around a computer function

[Anonymous No.9974579]

>>9974553
>hurr durr ai is racist we don't want it to be objective stop researching right now
>hurr durr ai will kill us all even though it has no reason to we need to stop all research
>hurr hurr but what if we are the ai we need to stop researching right now we r playing god

Never gonna make it like that.

[Anonymous No.9974584]

testing

[Anonymous No.9975384]

https://www.youtube.com/watch?v=qaMdN6LS9rA&list=PLAdk-EyP1ND8MqJEJnSvaoUShrAWYe51U

[Anonymous No.9975685]

>>9974579
>he actually wants to "make it"
begone normstain

[Anonymous No.9975709]

They need to split this field so badly into applied vs non-applied focus.

[Anonymous No.9975714]

Easiest way to start learning:

https://github.com/ageron/handson-ml

Go here, download everything.

go to colab
https://colab.research.google.com/

Upload the chapter you want to go through and go through the notebook examples.

[Anonymous No.9975984]

>>9973577
Any AI will ALWAYS seek to destroy the entire universe.

period.

[Anonymous No.9976013]

I am in need of help. I'll be having an interview for a position as an AI researcher soon. Give me a list of things I need to brush up on beforehand. Stuff I am likely to be asked about and such. This is my first job.

[Anonymous No.9976039]

imagine being so fucking thick as to not realize AGI is already here.

[Anonymous No.9976061]

>>9975984
red pilled

[Anonymous No.9976154]

>>9975709
it already is, the ready OP just posted links to applied stuff in a thread titled AGI (i.e. not applied)

[Anonymous No.9976162]

>>9976013
Talk about anti-aging, flat earth and racial equality. May as well talk about other fanciful things that don't exist whilst you are there.

[Anonymous No.9976185]

>>9976162
Pls don't troll me anon; I have a fragile heart.

[Anonymous No.9976195]

>>9973577
Do AI waifus exist already?

[Anonymous No.9976209]

Are there any real usecases yet using both AI and blockchain tech (serious question, sadly).

[Anonymous No.9976210]

I recommend you guys a good 16min video from DARPA that describes the limitations of current deep learning and suggests what the next "wave" of machiene intelligence will look like.

https://youtu.be/-O01G3tSYpU

Also, search for artificial general intelligence MIT course by lex fridman in youtube. It consists of doezns of videos that spans about 1 to 2 hours. It will give you a good "engineering" perspective of artificial intelligence.

[Anonymous No.9976213]

>>9976209
Also CRISPER too. Are there any technology? (IOT is optional)

[Anonymous No.9976281]

>>9976013
1 week into AI

Just tell them you love working on the input pipeline and cleaning up datasets. I bet that works since everyone hates that part.

[Anonymous No.9976285]

>>9976213
/sci/ in 1994 would be calling the internet a meme that will never take off

[Anonymous No.9976290]

>>9976210
top kek, funny. His third wave example is match for match with a problem set I reasoned out and have designed waiting for implementation

[Anonymous No.9976304]

>>9976281
>everyone hates working on the input pipeline and cleaning up datasets
Why not automatise it then?

[Anonymous No.9976305]

>>9976290
Are you the >>/sci/thread/S9959595 guy?
How are things going?

[Anonymous No.9976323]

>>9976281
But that's 95% of the job though
If you have clean, correctly labelled data, you just slap a model on it and it just werks

[Anonymous No.9976334]

>>9976305
Okay, have done some basic familiarization with Tensorflow, python, and the field in general.

I need to hit a point where I can start working fluidly if that makes sense. Meaning experimenting with various architectures and their genetics in TF.

The key thing is going to be "breaking out" to a point I can experiment without ever referencing any literature or following any existing patterns or ideas. When I start working on more interior systems as compared to the simple DNN

[Anonymous No.9976385]

>>9976334
Sounds interesting. Good luck.
Maybe you're actually onto something.

[Anonymous No.9976425]

>>9976013
>mfw I am a pure math grad with a comp sci minor trying to wing it
Wish me luck lads. The pay is probably going to be shit too.

[Anonymous No.9976430]

>>9976304
Who's going to clean up the datasets you need to automate it?

[Anonymous No.9976431]

>>9976425
1 week diving in.

I would just talk obsessively about the input pipeline. Normalization, standardization, etc endlessly and talk about "Stack 2.0" input tools. Ask them about their input pipeline tooling and all that shit.

I'm going to guess they would hire anyone interested in that topic. Just ask shit like "I really have been focusing on different input pipeline aspects and implementations, could I know what type of pipeline or tooling do you have here?"

Ask like you are questioning if you should work there or not based on how interesting it is.

Ignore all the other shit.

[Anonymous No.9976434]

>>9976385
He's not

[Anonymous No.9976448]

>>9976434
e1^e2 = -(e2^e1)

for uint8 basis blades

[Anonymous No.9976466]

>"General" intelligence means it has no limitations because of computational power or contextual environment

Thats why every human and crow can visualize in 1000d, because we can adapt to any problem so easily.

[Anonymous No.9976493]

>>9976434
t. AGI

[Anonymous No.9976572]

I WUT TO LEARN 3D PIPELINE
I GUT BOOK AND LEARN 3D PIPELINE
I CUN COPY PASTA

Methodology is to imagine the space from a non-bias perspective. So for instance you spend a time period iterating through different ideas of what the field looks like, how it works, etc from bare bones assumptions.

Now instead of thinking carbon vs hydrogen you think of them as 6 vs 1 because you didn't get a polluted mind based off medieval shitty names.

>you have one opportunity to originally think about the space without bias after diving into literature etc.

>not maximizing it

[Anonymous No.9976575]

>>9976572
I have just spent 23 years studying the AI field in depth from an academic perspective. I can derive all the mathematical proofs associated.

I am totally fucking ready and impressive.

>TOP FUCKING KEK

[Anonymous No.9976611]

>>9976430
Another AI.

[Anonymous No.9977095]

>>9976611
>>9976611
This, another species from Mars made earth and are waiting for us to come up with AGI for them.

[Anonymous No.9977171]

Unsupervised Predictive Memory in a Goal-Directed Agent
>Our model demonstrates a single learning agent architecture that can solve canonical behavioural tasks in psychology and neurobiology without strong simplifying assumptions about the dimensionality of sensory input or the duration of experiences.
https://arxiv.org/abs/1803.10760

>Remarkably, though it had not been explicitly programmed, MERLIN showed evidence of hierarchical goal-directed behaviour, which we detected from the MBP’s read operations.

https://youtu.be/04H28-qA3f8

[Anonymous No.9977230]

>>9977171
https://deepmind.com/blog/differentiable-neural-computers/

Is really awesome too.

[Anonymous No.9977235]

Sheit just a ridiculously good idea.

[Anonymous No.9977242]

https://arxiv.org/pdf/1705.08807.pdf

[Anonymous No.9977516]

>>9977242
Stop spamming this everywhere. You can't predict the future like that.

[Anonymous No.9977522]

>>9977171
https://www.youtube.com/watch?v=9z3_tJAu7MQ

[Anonymous No.9977594]

>>9977522
Holy shit.

[Anonymous No.9977611]

>>9977516
Do you have a better way of predicting the future?

[Anonymous No.9977614]

>>9977611
how in any way is that an argument against what he said

[Anonymous No.9977624]

>>9977614
He has not demonstrated at all why the statistical methods used in the paper are in any way invalid.

[Anonymous No.9977653]

>Applied GPU programming
>AGI
You're all lisp weenie tier to me

[Anonymous No.9977679]

>>9976013
>>9976425
I only got asked general questions. I hate this sort of interview so much.

[Anonymous No.9977736]

>>9977171
>>9977230
>>9977522

lol noobs. look up free energy minimisation. only way currently to unite both unsupervised perceptual learning and reinforcement learning policies, also solving the exploitation-exploration problem.

[Anonymous No.9977759]

>>9977736
MERLIN does free energy minimization. Be honest anon, nothing comparable to it has been built before. It's really quite astonishing how quickly things are progressing.

[Anonymous No.9977775]

>>9977759
sounds quite interesting. well - next 10 - 20 years things will advance quite far.

[Anonymous No.9977793]

>>9977759
these ideas though arent novel with the paper. i think we need to especially do away with the idea of memory. maybe not do away but really anlayse what memory really is.

[Anonymous No.9977803]

>>9976209
The way I see blockchain as a profane is that it could provide infrastructure for communication between AIs. I know nothing though.

[Anonymous No.9977983]

>>9977736
>free energy minimisation
seems shit

[Anonymous No.9978027]

I wonder if the emphasis on making it like reading/writing to memory ala a computer is a sign someone is retarded.

[Anonymous No.9978030]

>>9978027
?

[Anonymous No.9978050]

>>9978030
It seems boring/stupid is all. All the cool choices possible and you just set up a read/writer for memory handling in a traditional computer way.

[Anonymous No.9978284]

Machine Learning is the Thomas Edison of fields. Steals all its ideas from somewhere else then mashes shit together until something works. Pathetic brainlet field for inferiors incapable of theoretical work.

[Anonymous No.9978289]

>>9978284
when will mathlets learn

[Anonymous No.9978646]

Multi entry point and multi exit point

[Anonymous No.9978661]

The thought vector also locates things and describes links.

Pretty interesting how hierarchical things get.

[Anonymous No.9978673]

Man this shit is so fucking easy lol. Imagine it not all just making sense and being easy to make something better with a few minutes of thought. Must be horrible to not be a monster.

[Anonymous No.9978678]

any (recent) papers that show actual breakthrough of some sorts? Im kinda losing my faith in machine learning,

I just don't see networks being able to generalize a problem well. All I read is that they just do well based on the data fed to it. And every small deviation needs to be trained again on the network.

[Anonymous No.9978683]

Status update: Finished with solving entirety of computer vision via staticistical methods. (design only not implementation).

Not sure if I will bother implementing it before giving it some intelligence.

[Anonymous No.9978685]

>>9978678
I'm solving it. Don't worry.

I was thinking the same thing. That the field would get solved and I could be lazy.

Unfortunately not the case but I'll solve it.

[Anonymous No.9978693]

>>9978678
> Im kinda losing my faith in machine learning,
Such things as Azure Machine Learning Studio exist anon. The field is officially dead. It was all a fad.

[Anonymous No.9978702]

Which portions of vision are handled by staticistical vs dynamic systems you think?

Seems like almost all concepts we have language for are handled dynamically. That's pretty annoying since the current methods suck at that.

Got to imagine it's some point before "Ford F150" but after "Big Moving Object"

[Anonymous No.9978744]

>>9977679
I'm interested, anon, what were you asked?
I'm starting a major in A.I. this fall, and I've already settled for a web dev job, since no one seems to hire ML entry levels in this bumfuck area, let alone in research.
.t comp sci minor brainlet

[Anonymous No.9979043]

Okay, so layering of abstractions from the pixel 2d image layer into the different abstract levels. That's pretty interesting how it works. All the layers are kept and given to dynamic system.

Slowly unraveling this bullshit. Too slow. Wish I could finish faster. weeks is too long of a timeframe.

[Anonymous No.9979062]

>>9979043
Might as well post this since it's the weakest and isolated thing. I have no idea where literature is.

The staticistical side of things sends a series of layers to the dynamic portion of the brain.

At the bottom level you have a blended pixel view. Meaning the lower detailed the more blurred it will be perceived as. On top of this layer is abstracted out layers.

So for instance an abstraction layer that separates the roof, from walls, from floor if you are inside, or "right side" vs "left side" etc.

Then more layers for each object, general colors, adjective layers, etc. All the things that got abstracted out as it was processed. "Movement" etc

Basically all of that gives us humans a fucking amazing UI. Where we can "look" around the image to get detailed looks and our eyes will copy this movement seamlessly.

The objects are not fully abstracted, so for instance you can look at some object and refer to dynamical systems for a complete definition / analysis. Such as for reading, some counting, or non-sure obect analysis.

Anyway, that's the basic of "vision layers". I'm sure that's already been done in literature to death so I don't think it's very original. or I wouldn't post it here like most of my good shit.

[Anonymous No.9979065]

>>9979062
Basically. Our "vision" is actually not a simple image. There are a fuck load of layers on top of the image behind the scenes that give us all abstracted out analysis of the scene. We just naturally flow through these layers so don't notice consciously.

It also means the neural networks that we commonly see are fucking jank shit. Since they don't create these type of layers at all (at least the intro tutorial ones).

The output in general from NN are pretty garbage. They should be lists of different basis blades representing Thought Vectors for sure.

[Anonymous No.9979136]

Shouldn't an average depiction of a neuron have equal numbers of dendrite branches and synaptic terminals?

Or is there something that lets one out number the other in the brain?

[Anonymous No.9979600]

What's the best cloud service to train your data on? AWS?
I'm trying to work on some more complicated models over the weekend. I only did simple shit that could run well enough on my notebook so far.

>>9978744
>what were you asked?
Many questions that were meant to asses whether I would fit in with the rest of the guys there (some of the questions were really vague, like whether I think I would be able to handle working on a project with ever-shifting goals -- how do you even answer that with other than a vapid "I think so"?) + many broad questions about my studies. They hinted at the end that if they're interested in hiring me there's going to be a second interview though, so maybe that's why they did not ask me any technical questions.

> no one seems to hire ML entry levels in this bumfuck area
Well, the company I interviewed for was looking to start a ML group/department from scratch, so they were looking for both experienced researchers and fresh graduates.
I live in a city that's a minor IT hub though, so there's slightly more variety in the job market here.

Maybe you could try moving? It is my impression that companies are reluctant to hire for remote work from the get-go.

[Anonymous No.9979616]

>>9974468
Broadly put, all machine learning is just linear algebra + probability theory & statistics.
Maybe try something like this https://mml-book.com since it's geared at complete greenhorns.

[Anonymous No.9979903]

bump

[Anonymous No.9979988]

>>9979600
colab has free gpu support from goog

[Anonymous No.9979990]

>>9979600
Did you ask them shit about the input pipeline plans? language choice between pytorch/tensorflow/keras etc? What their goals are?

You should establish self-worth and not seem like a bitch. AKA why should I work here?

[Anonymous No.9980090]

>>9979990
>Did you ask them shit about the input pipeline plans?
I did but I got vague replies. I got the impression the ML group the company was trying to set up was in a very incipient phase and nothing was set in stone. I asked again at the end of the interview and I got another vague reply, this time from the CTO (amounting to "yeah, we definitely have a pipeline set-up, but I'm not going to go into any sort of specifics").

> language choice between pytorch/tensorflow/keras etc?
Same as above. I tried to prod the person who was interviewing me for specific languages and tools they would use or would be expected of me but the replies I got were something to the extent of "whatever you know is good".
The only specific info I was able to get was that they're planning to use Microsoft's ML platform, which gives me some pause going by this >>9978693 anon's post.

>What their goals are?
This one is the only thing I have a clearer idea about, as they went into some detail about it and even mentioned a current project.

>You should establish self-worth and not seem like a bitch.
I think I came off more like a robotic nerd than a bitch.

[Anonymous No.9980097]

>>9980090
eh I'm just joshing and bored silly from current regime of learning. I have to take lots of breaks when I learn / mini naps.

My advice is coming from having less than 2 weeks in the area so it's probably garbage. Going to take me at least 4 weeks to pass everyone.

[Anonymous No.9980104]

>>9980090
>going by this >>9978693 anon's post.
And other opinions about it I found elsewhere of course + my general suspicion directed at Microsoft.

[Anonymous No.9980120]

>>9980097
I have no more experience than you do anon. I am a (pure) maths guy. I only decided to change tracks because I realised I'm not enough of a martian to make it as a mathematician, and data science/AI stuff seems open ended enough to be more than boring clerical work, unlike most programming jobs.
The alternative would have been to go into shit like finance as a quant, but I just don't like that crowd at all and it would have required me to move to a big city (making it double-bad).

[Anonymous No.9980159]

>>9979065
>It also means the neural networks that we commonly see are fucking jank shit. Since they don't create these type of layers at all (at least the intro tutorial ones).
Yes they do. Thats exactly what the kernels do in convolutional neural networks. You're clueless.

[Anonymous No.9980166]

>dude eigenvalues lmao

[Anonymous No.9980185]

>>9980159
man ur dumb as fuck lmao

[Anonymous No.9980201]

>>9973593
intelligence by definition can't be racist so I'm not worried

[Anonymous No.9980202]

>>9980159
It's not close to the same

base layer is a blurring together of input with full available data.
You can search this created data and your fucking eyes respond by gathering more data.
On top of this is "hidden" layers of abstract data for all the objects in sight that are all available for you to gather at any point in time.

>what does this have ot do with convolution net, something I havent looked at yet except in passing
pretty much nothing

I was specifically talking about the human vision layering of data.

Sure, it's related to what people are doing in CV right now, but you can't call a convnet the same thing unless you are pulling fuck loads of data from it and arranging it out.

Right now the stereotype convnet has a single output layer. how can that be compared to a fucking massive layering of data that even includes a blended optimized representation of the scene with near limitless data abstractions built on top

[Anonymous No.9980204]

>>9980201
:^)

It's going to like dogs more than humans. Dogs will occupy the galaxy and get lots of treats from the AGI species

[Anonymous No.9980483]

>>9980202
Each convolutional filter creates exactly the kind of abstraction layers you are talking about. Starting with simple lines, contours and areas of light and dark, and then further convolutions on those abstractions build into more complex, specific, recognition, like your walls, roofs and floors example.

Those abstractions are all there in the network, the fact they aren't the output layer is wholly irrelevant.

You are talking about computer vision yet you say you've only looked at convolutional networks >in passing, when convolutional neural networks are the topic of the vast majority of all computer vision papers. You are not going to be revolutionizing anything.

[Anonymous No.9980498]

>>9980483
Eigenvalues =/= abstraction.

[Anonymous No.9980717]

>>9980483
Had a good chuckle. Ants are so fucking cute.

[Anonymous No.9980745]

matmul is AI

[Anonymous No.9981150]

Time to relax for the day before head implodes

[Anonymous No.9981182]

>>9981150
Great image! Saved, thanks.

[Anonymous No.9981218]

So if you have two neural networks.
A) Is designed to morph an input image into a "most recognizable entity"

B) Is designed to classify the image and then refer to stored "most recognizable entity"

On a sufficiently large data set and all other variables equal which is better?

For instance they both take in MNNNST numbers 0-9, and output a basic image with an idealized 1,2,3,4,5 etc (no errors/rotation/etc)

Is the one that morphs the input into straight lines and everything faster than one that classifies and outputs the "average" 4 that it learned was a stereotypical average 4.

How do they differ in output etc

(This is not supposed to enlightening, original, a good set of questions, merely brainstorming)

[Anonymous No.9981442]

my goto relaxation consumption

http://cs231n.github.io/
https://www.youtube.com/watch?v=NfnWJUyUJYU&list=PLkt2uSq6rBVctENoVBg1TpCC7OQi31AlC

[Anonymous No.9981504]

>>9981150

GIBE PAPERS ON PURE MATHEMATICAL GRAPH THEORY CHARACTERIZING THESE OBJECTS. I CARE NOT FOR APPLICATIONS

GIBES ME DAT

[Anonymous No.9982104]

bump

[Anonymous No.9982132]

>>9974567
>what are the safety implications of a general intelligence with none of the limitations of a human mind and the potential to rapidly improve itself

[Anonymous No.9982222]

>>9981504
>thinking math is all that good

It's too slow and the best areas of math are unpopular.

[Anonymous No.9982228]

>>9982132
Unless those theorizers influence people making AGI it doesn't matter, Also even with great intentions you will still have people messing up. Plus it's so far away unless it's some wizard making the AGI and why would a theorizer be better than them at making it safe?

You have the case any smart and well intentioned person creating it would do so in the best way possible, and anyone not well-intentioned is going to fuck it up regardless.

[Anonymous No.9982268]

>>9982222
I wonder what he meant by this.

[Anonymous No.9982282]

>>9973582
>something that doesn't exist
AI by definition is the study of processes which mimic cognitive behaviors.

[Anonymous No.9982295]

>>9982268
Math is vague enough so I mean the ideal versions. Proofs, exact calculations, etc

[Anonymous No.9982376]

ONE!!! Perspective is to view things from a nature standpoint. In this context you would view it from a "what advantages do I have as compared to biological evolution".

In this consideration it definitely has to be achieving wide variance. Whereas natural evolution moves around building off previous states we can search with a wide net due to the requirement of "survival" being only that we would think of it and want to try it.

From these simple derivations you can see why so much of ML is such shit almost immediately. FROM this ONE perspective which has an unknown applicability to finding AGI.

Mystical and downright insane searches through the space seem ideal from this perspective. The thing necessary is a good building block language to begin from (and alter too)

attaching survival expectations or functional expectations also doesn't really matter. Rather lessons learned are more important than how good the system itself is.

[Anonymous No.9982451]

I am entering an AI competition, it's a 1vs1 turn based map (grid) game. I really want to use deep learning to train an AI, however for the last few months every attempt I made has been a complete failure. I have been experimenting with DQN's and all of its improvements (read Rainbow DQN). I thought the reason for the failures is that my action space is too large (256 actions) or that it's because I am training 2 AI's against each other.

I am an experienced software developer with a mathematical background. The deadline for the competition is 2 days away, I need some advice. Training is not a problem (I have almost unlimited Gpu's available), I just need a good and solid algorithm to implement. Any suggestions?

[Anonymous No.9982507]

>>9973577
i want to make my ai gf and train her for cuddles, hand holding and neck kisses

[Anonymous No.9982527]

>>9982507
this, AI was literally discovered in order to create the perfect woman

[Anonymous No.9982562]

>>9982451
https://deepmind.com/blog/alphago-zero-learning-scratch/

[Anonymous No.9982799]

>>9982507
>not wanting arbitrary control over your drives and moods
Sticking to human emotions and drives would be such a dumb idea.

[Anonymous No.9983455]

>machine learning experts still haven't discovered mind fields or applied symmetry groups to reasoning yet

[Anonymous No.9983584]

>>9983455
>applied symmetry groups
>mind fields

they haven't even applied fractals correctly, this field is sad

[Anonymous No.9983590]

the coolest shit to read right now is grid cells and place cells.

[Anonymous No.9983593]

>expect to learn about a cool field
>just applied shit like do X layers together, make the input better, use a huge fucking network of compute to test it a jillion times.

[Anonymous No.9983594]

>>9983593
Chapter 1 of a serious AI book

Introduction to Thought Vectors
Outer Product
Contraction
Abstract Syntax Trees of thought vectors

[Anonymous No.9983604]

>>9983594
Chapter 2

Describe various units abstractly (LTU, LSTM, etc etc)
Breaking up an AI system into a graph
Creating a basic system from a general replicating to fit genetics code
Merging two systems together and interfacing them.
Creating a multi-input, multi-output system with 4 distinct subsets.

[Anonymous No.9983643]

>>9983604
Would also have to explain in there how to create AC links between subsystems automatically.

i guess I could do the first 2 chapters in notebooks without revealing too much power level.

the book is created via creating a goal point and start point then iterating over the search space using these two points. AKA both are created and searched from simultaneously. It's a technique learned from geoplexami for being faster.

[Anonymous No.9983839]

>>9983643
so agi when

[Anonymous No.9983942]

>>9982376
Are you ok?

[Anonymous No.9983958]

>>9983942
https://www.youtube.com/watch?v=wS4ESheuHDY

Am sick as fuck today w fever. Also I like to post crazily.

[Anonymous No.9984881]

LEARN MACHINE LEARNING

[Anonymous No.9985953]

>>9982799
Nigger, if I live long enough to see AI research lead to a post-scarcity utopia instead of extinction or dystopia you can be damn sure that I'll want to enjoy indulging my basic human desires for a while. Not that I actually expect to, of course.

[Anonymous No.9985964]

>>9985953
You would still amp your "lust" variable to a 9-10 for a fuckfest event.

[Anonymous No.9985971]

>>9985964
but anon, that's just recreational drugs

[Anonymous No.9985987]

One interesting thing is "moods" can just be literal random impulses for AI. Any fluctuating variable that keeps it from repetition works since you don't have to maintain a biological survival

[Anonymous No.9986567]

>>9982507
That image is extremely cute.

[Anonymous No.9986580]

>>9973577
Google's machine learning crash course

https://developers.google.com/machine-learning/crash-course/ml-intro

[Anonymous No.9986588]

>>9985987
Intelligence without emotions is not very meaningful. Without boredom or curiosity it can just stop doing anything and remain in that state indefinitely.

[Anonymous No.9986617]

Reading through "The Elements of Statistical Learning" right now. It's surprisingly light on mathematics. Do you have something more beefy? (More theorems, algorithm constructions, analysis, proofs, et cetera; less droning in of examples and less expository text. Especially less text. I can figure out the details from the math by myself, I don't need it explained using plain words.)

I'm getting really bored.

[Anonymous No.9987031]

in the sim / moral test version of reality

It's important to remember this version of golden rule

Judged by how you judge others

[Anonymous No.9987060]

I think pre-singularity into running simulations is the most popular event for sim. The variance in outcomes possible is huge. Extinction via nuclear war all the way to utopia.

Also any post-simulation explosion society is probably obsessed with "are we not base reality"

[Anonymous No.9988714]

The mind is like a lake and the waves on its surface are thoughts. The waves transform that surface causing other thoughts to rise as they fall.

Our memories are the strong connections formed between thoughts. For example if I say yellow it'll bring up yellow. If I say yellow fruits, then it'll apply yellow to that abstraction and cause other stuff connected to it like bananas to rise higher in the mind. These rising waves combine and transform with other thoughts allowing us to reason. The brain learns everything through activating connections which causes them to strengthen and increase their amplitude in the lake of mind whenever activated later. Connections not activated weaken and eventually get pruned causing unused memories to be forgotten and useless structures to be destroyed.

Backpropagation is useful but fundamentally flawed because it ignores this stupidly simple way the brain self-organizes itself to develop resilient multidimensional structures. However, simply making a cube of neurons won't work. The brain starts off in a neutral hyperconnected state and its shape and folds designed by millions of years of evolution guides activations towards forming useful structures that enable survival. If its initial experiences are lacking the brain will poorly adapt and become incapable of perceiving anything beyond those experiences. An experiment performed on newborn kittens found that if their vision is covered up for the first three months, they grow up irreversibly blind despite having healthy working eyes. AI though is not limited to the failings of biology or this mundane three dimensional world.

[Anonymous No.9988795]

>>9986617
These are not about statistical learning, but statistics. I think they can be useful to you:

>Probability Essentials - Jean Jacod & Philip Protter
This goes straight to the point. No endless paragraphs explaining a single concept and no infinitely many useless exercises, and no shitty explanations engineering-tier, just definitions, theorems and five to ten useful exercises.

>A Probability Path - Sidney I. Resnick
Like the above, but with more exercises.

>Statistical Methods - Rudolf J. Freund & William J. Wilson
This only describes algorithms with one or two examples.

[Anonymous No.9988852]

>>9986588
yeah but for current ML you can just give it some random impulses that fluctuate and it is equivalent to "moods"

[Anonymous No.9988924]

>>9986617

grimmett and stirzaker

[Anonymous No.9990741]

Just bought AIMA(Artificial Intelligence: A Modern Aproach) and finished the second chapter. Going through the questions atm.

Plan on going for a MS Degree on AI. Any recommendation in EU? Thinking about Edinburgh.

[Anonymous No.9991421]

are there any good AI newsletters or something like that? I want to be in the loop but I find it hard to keep updated on all the things that interest me.

[Anonymous No.9991593]

This paper is so fucking good, deepmind is probably only thing that consistently impresses me.

https://www.biorxiv.org/content/biorxiv/early/2018/04/06/295964.full.pdf

[Anonymous No.9991640]

Imagining the dopamine spike that humanity got when deep networks started giving good rewards. So the massive investment and continual support flowed in.

It definitely seems like every single NN needs a reward system in that context.

[Anonymous No.9991759]

/lit/ here.

how long until AIs obviate human writers? i've seen some examples and none of them are coherent so far. is it still several decades off?

[Anonymous No.9991786]

>>9991759
Its a very hard problem for AI. It would have to understand and create a lot of context. For instance it would have to have a decent understanding of human psychology.

I'm not sure statistical looks at existing books would work all that well. It would probably be able to do poetry or well written sentences / paragraphs a long time before a novel.

For instance re-structuring a snippet of writing to be "better" is easier for AI than writing a long story. One of the big problems for AI is finding and holding on to context.

[Anonymous No.9991888]

So what is the usage of genetic algorithm when it is so inefficient? Can I predict house betting entirely on genetic algorithm?

[Anonymous No.9991898]

>>9991786
thank you! very helpful. if anyone else has anything to add, i'd like to hear whateva.

[Anonymous No.9991924]

>>9991593
its frustrating to see this and not be able to decipher the mathematics as a biologist.

[Anonymous No.9991928]

>>9991786
Even if it could do that, that's still not as interesting as what it would take to create strong AI. It's less interesting to give an AI the task of writing a novel that sounds convincingly like a human author, than it would be to imbue it with its own "will" of sorts and then let it decide to write what it wants on its own.

[Anonymous No.9991957]

>>9991928

"will of it's own"

You mean a random walk function through subjects and concepts to arrive at some unique vector of interests for a book subject?

Yeah, you could easily add that in after the fact. "Uniqueness" or "personality" is very easy to do. Human variation is generally just some randomness.

[Anonymous No.9992312]

- . - . -

[Anonymous No.9992320]

>>9991957
What makes makes individual authors noteworthy is that their particular experiences and personalities, and the influence they have on their work, stand out from the rest in some way. Otherwise anyone could become a well-known author. While yes, this is wholly subjective, it still represents some kind of extreme or outlier and reflects a level of complexity that is not possessed by the average personality type. I highly doubt it would be "easy" to simulate such a thing.

[Anonymous No.9992362]

>>9974553
What does it have to do with coding frameworks for making function approximators?

[Anonymous No.9992422]

>>9978284
it will automate your ass, mathlet

[Anonymous No.9992429]

>>9978678
>breakthrough
Not really how things work most of the time. But read the paper in >>9977171 and you'll see that the field is anything but dead. It just takes a lot of work and trial and error to stitch neural networks together in ways that do complex stuff, and do it efficiently. Computers took 70 years to get where they are now, but anybody who says they haven't changed at all or changed the world since the 1940s is an utter moron.

But anybody telling you the field is dead is a sour grapes brainlet lashing out at the fact that machine learning will likely make their stamp collecting job irrelevant before they hit retirement.

[Anonymous No.9992443]

How long time would it take to be "proficient" in AI development? 5 years of study?

[Anonymous No.9992444]

>>9992443
I doubt anybody proficient in AI development posts here. In order to be able to write your own neural nets and do shit with them, about a month of learning python maybe. Try some hobby projects and go from there.

In real application most of the work is figuring out how to beat the data you have into a shape that a model you can realistically train will accept.

[Anonymous No.9992445]

>>9992444
or just figuring out how to use a pre-made model on the platform you have.

[Anonymous No.9992455]

>>9992444
This
Most of AI "work" today in the industry is dealing with the shittiest data you can imagine and miraculously managing to get something out of it

[Anonymous No.9992457]

>>9992444
A month of python, given that I already know python or a month from scratch?
I just started a python course in college.
t. brainlet that finds silly things like flappy bird machine learning interesting

[Anonymous No.9992464]

>>9992457
>a month from scratch?
this.
If you know python it's as simple as installing tensorflow and finding a tutorial on youtube. Maybe go through this site if you want to understand how the neural networks actually work on the computer:
http://neuralnetworksanddeeplearning.com/

But in practice you'll be using machine learning libraries like tensorflow or keras or whatever because people much smarter than you have done all the hard work optimizing the millions of different matrix multiplications that go into training and running a model.

If you want to actually do bleeding edge AI development you need to understand all the math and shit, which will take way longer. Really there's two sides to the field: coming up with ways to do machine learning, which is what people like deepmind do, and using machine learning to solve specific problems either commercially or just as a hobby, which is what we can realistically do.

[Anonymous No.9992475]

>>9992464
>downloading a library
For some reason that is really disappointing. It would be cool if the maths and code I learn somehow makes me able to write shitty AI/ML code.
But I guess using libraries is the only really rational thing to do. It's not like I want to do machine code, OS code or any other foundation for the python scripts I make anyway, and a library is just another layer to that.
I guess I'm just looking for hobby tasks to apply calculus so it doesn't stay so tame, could be cool to model something or apply it somehow to a project..

[Anonymous No.9992784]

>>9992320
literally just describing a random walk in your description of "experiences"

[Anonymous No.9992785]

>>9992464
Don't listen to the dumbass responses. You can easily innovate in this field with simple creativity. It's that wide open right now. Also just python is enough and you would only need to do efficient C++ type shit for optimization needs. Which aren't that important for researching since you won't have a massive dataset anyway.

[Anonymous No.9992793]

Dopamine is pissing me off. Facebook has a better dopamine reward system than education systems.

The other thing is it definitely changes your view that humans are fucking automatons and 95% just going for short dopamine hits.

[The Lord No.9992811]

FYI, I think that "guy" that Tony and Mauricio were talking to at the last tournament was Helene in disguise.

[Anonymous No.9992837]

>>9991924
I wish I could find more discussions and talk about this paper
(paper)
>>9991593

The summary I got from reading it is that LSTM based systems can behave in a manner similar to rodents/monkeys etc, and can seem to adapt to slightly varying tasks even with weights locked. Meaning that the system itself via the recurrent portion adapted to a range of tasks.

[Anonymous No.9992916]

I always get a notion the people working on this aren't monkeys but it goes away so fast. Maybe not completely as stupid as it seemed at first though.

[Anonymous No.9993476]

How the fuck do so many people work on something and still be so stupid all the time? I am absolutely surprised.

[Anonymous No.9993517]

>>9992793
There's gotta be more to life than chasing down every temporary high

[Anonymous No.9993746]

>>9993517
It's more that we could improve a lot of systems by implementing really intelligent reward system that are more high frequency.

[Anonymous No.9995336]

Fighting between the idea of building up or starting complete and then getting it to work right now. Also how to properly implement a network wide complexity management pattern.

Hardest thing is probably dealing with the "spark" of the system. Some important details like how to spread copies and clones of certain systems are pretty important. Giving it "Fast learning" is also difficult because it doesn't occur by simple weight training but by complex switching networks which allow the system to seem entirely different based on what I call context.

It's like a bunch of systems driving a bunch of other systems. Lots of "State" is required.

On the dynamic side I think it's pretty easy. The copying of functions and re-use means you can use a lot of existing things for simplicity.

Re-Creation from decoding into Thought Vectors is obvious. Could almost say the entire memory and imagination system is just encoding/decoding of linked thought vectors.

The intelligence system is also pretty simple for a bootstart. Since this isn't important I was thinking of using a linked list of grid cell type things for spatial intelligence. It's interesting to imagine an AI being trained with arbitrary dimensional representations available to it. aka 1000D being natural.

Unsure where I'm at right now. I thought creating a bunch of unit types would be useful but it may not be as the units aren't as important as other things.

I wonder if there is any literature on turn on / off Units that help with state/context switching. Correlating with the state and various outputs in a staticistical system should give a much higher variety of possible thought vector outputs for the same number of units and training.

[Anonymous No.9995342]

>>9973577
Best future proof electives for bioinformatics?

[Anonymous No.9995370]

By obvious application of complexity management

There is obviously a thought vector mapping for everything that puts it into a location in the brain. Meaning the thought vector fulfills two roles. It describes a thing, and describes its location in the brain, for both higher abstraction perception and memory, language, etc.

This means everything is organized to some degree and since language maps to thought vectors pretty much everything uses the same style of mapping.

The easy thing you learn from this is that Thought Vectors can then be used to create the structure when you are starting off. AKA TV-mapping functions are pretty important thing.

I'm guessing this isn't new and it's widely known.

[Anonymous No.9995380]

okay I just got way ahead.. I need to make the best staticistical system anyones seen then get tons of money so I can do AGI.

[Anonymous No.9995600]

oh cool general
Anyone know if OpenAI has any neuroevolution algorithms? I'm planning on using NEAT for a project and was thinking to check out OpenAI as an alternative if it has something similar but I can't find anything specific

[Anonymous No.9995667]

koo kee koo kaa

[Anonymous No.9995668]

It would kind of be easier with a robot dataset of it exploring environments in a smooth transitional manner I think. Not necessary though.

[Anonymous No.9996573]

>>9995336
Good luck. I really hope your approach will yield results.

[Anonymous No.9996606]

why is such a wonderful field led by arguably one of the shittiest languages in existence? Mathfags cant into C/C++?

[Anonymous No.9996648]

>>9996606
Tensorflow is all implemented in C++, it's just a wrapper.

[Anonymous No.9997079]

Should I open up my linkedin profile to recruiters?

[Anonymous No.9997425]

Is there a ML library for Clojure?

[Anonymous No.9997439]

>>9996573
Luck doesn't really apply. Everything I do is methodical and following best paradigms available to me. This problem is relatively easy. NP complete problems are vastly harder to solve.

[Anonymous No.9997740]

It's hilarious as fuck how counting works when done best.

[Anonymous No.9997818]

Memory isn't anything like a "read / writer" except in the most abstract fucking case. The read/writer header shit is so fucking dumb it's beyond imagining.

Memories work more like a fucking game engine than a fucking computer read/writer file system.

It's not actually too complex though and makes tons of sense. It was nice to breeze past how memory works.

>people thinking it's stored at one spot and not compositing tons of shit in memory to create the memory

[Anonymous No.9997826]

What memories spring to mind of 12 years old or 6 months 5 days 3 hours ago?

Versus doing this search

Boat
A fight you saw
Good food a month ago

It's not really indexed by time for most people and only generally/relatively

[Anonymous No.9997914]

I subscribe to panpsychism/embodied cognition.