Don't fall for the hype around OpenAI's Rubik's Cube playing robot, Berkeley bans facial recognition, and more Just in case you're addicted to the world of AI, here's more news beyond what we have already covered this week. Yay or nay - OpenAI’s robot Rubik’s Cube hand: Some in the AI community have been gushing over OpenAI’s most recent video showing a mechanical hand deftly solving a Rubik’s Cube, even when it was being disturbed by …
Solving the cube is no where near as important as the algorithms being used to manipulate the cube.
Personally, I find the hand to be incredible in it's capacity to "realise" the 3d positioning of the cube in 3d space and therefore it's capacity to subsequently manipulate it...
I loved the video even though it's yet another step towards scary terminator bots....
"Solving the cube not so important"
Spot on! Rubik's cube is 'solved' in the sense that you could feed the robot a specific algorithm to get it done - but that's not AI, it's a bog-standard 'factory' robot that has been around for decades.
The great thing about this research is that there is actual learning taking place. The robot is told what it needs to achieve and what it needs to avoid (through reward / penalty system which is actually how humans learn), but not how. That it can even solve the cube at all is important in the sense that you can see that the approach works.
There are, at maximum, 28 steps needed to go from any of those 43 quintillion permutations to being solved. If it takes significantly more than that, then thats evidence that the robot isn't actually solving it but trying random moves until it gets lucky.
> There are, at maximum, 28 steps needed to go from any of those 43 quintillion permutations to being solved. If it takes significantly more than that, then thats evidence that the robot isn't actually solving it but trying random moves until it gets lucky.
Strictly speaking, any sequence longer than than God's Number (ie. 20, which tells us that after 21 moves any pattern that results has been observed before in the sequence at some point) means the recommendation isn't optimal*. This has the potential to be useful: if a robot betters the recommendation then it has taught itself that; if it's consistently worse (than "an average of 28" if it's the algorithm discussed at https://www.theregister.co.uk/2019/07/16/ai_rubiks_cube/) then it's not taking the hint any algorithm should be followed.
(* possibly, as in the Thistlethwaite algorithm, a human observer might more readily judge whether progress is being made throughout ... but I speculate)
If it takes the equivilent of 10,000 years to 'train' this robot to a 20% success then it's slow machine learning not AI. Having said that, if it can be improved to the point where retasking to any other 'one-hand problem' reaches a similar level success within a few weeks then we're well on the way to a general purpose robot that will render the need for human manual interaction with any environment optional.
As the only useful purpose of a robot hand is to manipulate the items we've designed to be used by hands (quite literally, everything) I see these robots main use in and around the home, "Robot, Make me a cup of tea, butter some scones then mow the lawn and prune the roses" or more likely "Help me out of bed, help me in/out of my stairlift, give me my medication..."
Everything outside the home environment will be better served by custom built designs as is happening now.
I can see why the Rubik's Cube solver here is a great achievement, but I still prefer the custom machine that solves the cube in less than 0.4 seconds.
Since the particular biological neural network writing this post can solve the cube zero percent of the time.
I never did manage to get past the 2nd layer when I was a kid, then I just lost interest anyway. Its one of those puzzles that you either really get into and spend years playing with it, or mess about with it for a few weeks then toss it into a drawer only to be rediscovered when you finally move out of your parents place, have a 2 minute nostalgia moment, then toss it back in again and leave :)
I think I got to the point where I could reliably solve all but the final four corners, then got distracted by other things. (Of course by that time I'd also learned to take it apart and reassemble it in the solved configuration.)
On the other hand, I had a friend who was ... not generally noted for his intellectual achievements, shall we say ... who within a year or so of the Cube's introduction to our set could solve it reliably. He wasn't terribly fast, as these things go; I think it generally took him 5-10 minutes. But he wasn't focused on it either. He'd do it while carrying on a conversation or eating lunch or walking somewhere, occasionally glancing down to see how it was going. It was just A Thing He Could Do.
being used by government agencies. However, they did not vote to ban the buying of face recognition data from private companies who can't be bothered with such mundane topics as ethics. Am I the only one seeing the difference here ?
Solving a Rubick's isn't all that impressive, especially with the seer volume of computational power that was put into it. I've seen Rubick's Cube solving devices built out of a couple servos and a color sensor wired up to a Arduino. Sure, the system didn't "Learn" to perform the task, but it was programmed by a hobbyist working in their spare time. And I am willing to bet that that programmer didn't expend nearly the level of resources 10,000 years of CPU time would burn.
A human produces approximately 15 tons of CO2 per year to support. Assuming that the AI burned 50 Watt-hours per hour of compute time, you are still talking about 4383 tons of carbon dioxide, or 300x the amount a human would expend. These are extremely rough calculations based on averages and assuming that that human expended 8766 labor hours to producing the code. This is also based on per capita emissions, so includes industrial emissions in that calculation (Although some of those industrial emissions would be from producing necessities for that human. Really, there are just far too many variables to deal with here). And this is even being generous in assuming that the AI can be considered as complete rather than only 20% accurate, so we are ignoring the future computational hours needed to solve the problem (Which I think fairly balances out the somewhat pessimistic wattage calculations of the machines in the AI cluster)
I feel that we see the inefficiency of trying to teach one neural net to do two very different things simultaneously, i.e. solve the rubiks cube, and manipulate the hand/cube system. My guess is that if you used two different networks and limited them to one of the problems each, trained them and then fed the ouput from the solver i.e. do this move to the manipulator it would have been far superior at both tasks with far less time expended.
I'm assuming the real goal of this is not to make a rubiks cube solving robot but to investigate how to solve a much more complex multidimensional problem in this case solve the cube AND manipulate the arm and how to make it more efficient at those kinds of tasks.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
