Updates: harnessing the power of the GPU for deep learning

https://openai.com/blog/openai-pytorch/

Robot learned via OpenAI to solve rubic cube single handed

will AI share our values?

“some robots will eat some jobs and some robots will kill humans” (for example self driving cars)

Mr Altman hopes: AI will “redo the socio economic contract”

… at the same time he says only massive amount of computing resources will be needed to get it working costing a lot of money.

Endowment
  • $1 billion pledged to the non-profit (2015)
  • $1 billion invested into the corporation by Microsoft (2019) [2] (The New York Times)
    • <- oh shit the killer robot that will wipe out mankind will have Bill Gate’s face:
    • “Most of that $1 billion, Mr. Altman said, will be spent on the computing power OpenAI needs to achieve its ambitions.”
Website openai.com

why do people working on robots look like robots themselves? X-D

why do people working on robots look like robots themselves? X-D

is this a move closer to … when the “product” will outperform the creator… what if OpenAI actually gains consciousness and escapes the cage/sandbox?

Will it be the best future one can imagine – or the end of mankind? (product kills it’s creator?)

risks of Super AI: very depressing TED Talk: when humans want to stop self improving Super AI robots to self improve… they might just wipe us out like ants

precaution rule of thumb:

  • before starting the SuperAI

    • make sure there is no PHYSICAL connection to the internet possible (completely isolated machine, no wifi or bluetooth interfaces)

    • exchanging files with the machine shall be done with extreme caution, because SuperAI would find a way to modify/infect USB Stick with (self starting?) virus to free itself from the cage and upload itself to the internet.

      • 90% of all money is digital and a SuperAI could (probably) hack any system in short period of time and in the worst case simply steal all digital money/delete all bank accounts plus: “buy” technicians to work for it

    • extreme caution in terms of file-transfer would mean: one time usage USB Stick only: any USB stick that was connected to the SuperAI machine must be regarded dangerous

      • virus can also hide in files such as jpg doc docx and so on

      to any other computer and must be physically destroyed immediately after usage (file upload to SuperAI machine).

    • also: any computer that the SuperAI was running on must be regarded as “infected” (could modify hide code anywhere)

    • Virtual Machine isolation “unplugging the virtual cable/interface” not enough – because virtual machine escapes are possible from any hypervisor.

once far fetched sci fi - inspiring future - where cars are farmed and buildings are grown?

once far fetched sci fi – inspiring future – where cars are farmed and buildings are grown? or terminator 2 like AI nightmare?

are we all already smart phone enabled augmented cyborgs?

  • AI designed Aircraft walls
  • AI designed pavillion (#auhive)
  • whats next?

Elon Musk: 2018: “Ai is far more dangerous than nukes” and should be regulated. “We have to ensure that the advent of digital super intelligence is symbiotic with mankind”

get the source: https://github.com/openai (mostly Python?)

“Deep Reinforcement Learning via Bayesian Neural Networks”

https://arxiv.org/abs/1605.09674

Gym Retro, a new platform integrating classic games into Gym, starting with 30 SEGA Genesis games. (src)

MarI/O: neural network learning bot playing: Mario comes with source code: https://pastebin.com/ZZmSNaHX

This manual explains, roughly, how to install, train and run MariFlow, a recurrent neural network for playing Super Mario Kart. If you want to try anything but the sample, I recommend learning about neural networks, there’s a lot to know.

My video about MariFlow: https://www.youtube.com/watch?v=Ipi40cb_RsI

Introduction

Installation

Requirements

Placing Files

Capturing Training Data

Modifying the Neural Network Inputs

Additional Capture Features

L and R

+Screen

Capturing Mixed Data

Training a Neural Network

Config Files

Validation

Running a Neural Network

The Neural Network Server

The Emulator Client

Capturing Mixed Data

https://docs.google.com/document/d/1p4ZOtziLmhf0jPbZTTaFxSKdYqE91dYcTNqTVdd6es4/edit

https://openai.com/the-international/

OpenAI at The International

We’ve created an AI which beats the world’s top professionals at 1v1 matches of Dota 2.

Watch us play on main stage during The International.

PS: to be fair – bots are better at chess – why? because they know EVERY single possible move.

bots can be set to “impossible” to beat at counter strike ( ego shooters ) – because they exactly always KNOW where you are are and just pretend to not know – otherwise it would be just unfair because the bot “knows” more than the human.

I remember that back in half life 1 times (1998) – CounterStrike was an modification of half life 1 – and you could add bots.

The bots had different strategies/implementations – of course – it is only fun to play against a bot that behaves as human as possible and actually.

Programming AI is not an easy task.

I remember that the neural network bot was actually the best – but could not navigate the map without waypoints – so you first had to put some waypoints on the map.

http://www.angelfire.com/ex/cs_light/WaypointsHowToDo.html

OpenAI team

August 2018: https://liquipedia.net/dota2/The_International/2018

PS: guess all one can do is have a cup of tea (or coffee… after all humans are said to have “free will” and (still) live in a democracy and can make (hopefully good) decisions for themselves) and watch a amok running digital Frankenstein kill mankind.

admin