Open AI – Hiding in plain sight

Open AI is still using brute force.

• Open AI’s latest paper makes its algorithms look very clever, but they remain subject to the limitations of deep learning making Open AIs goal of artificial general intelligence almost impossible to achieve.
• Open AI’s latest paper (see here) creates a very simple world with a few moveable objects and obstacles where two AIs play a game of hide and seek.
• AI agents are placed in this world with knowledge of the surroundings and that they can move and lock objects.
• The aim of the game is for the red agents to find the blue agents and points (which create the incentive) are awarded based on success.
• The red agents count for a few seconds while the blue agents have time to move or interact with the environment before the game begins.
• Over the course of millions and millions of games, stages emerged where one side would get the upper hand which would then be adapted to by the other side which would then force further adaption (see here).
• This ended with the hiders locking down every object in the area and then building a fort meaning where there was no way that the seekers could find them.
• Open AI compares this reinforcement-based learning approach to evolution and in many ways, it is right to do so.
• This is because evolution is a process where very occasionally a random gene mutation produces a characteristic that confers such a great advantage that individuals that possess that gene are more successful.
• This results in the gene being passed on more frequently, leading to it spreading to all individuals within the gene pool over course of many generations.
• In Open AI’s world, the agents try various actions randomly and once a pathway of random actions produces a successful result, the incentive to continue that behaviour is delivered.
• It was these random actions that revealed some of the physical inaccuracies of the world Open AI created as agents were able to climb onto boxes using ramps but then to move the boxes while still standing on top of them.
•  
• Open AI goes on in the discussion section to state that because the world it has created is grounded in the laws of physics there could eventually be applications in the real world.
• This is where I see the shortcomings of this research.
• This experiment worked beautifully with a deep learning approach because the world Open AI created was both finite and stable in terms of the data set that defined it.
• Unfortunately, as autonomous vehicle creators are finding out, reality is both infinite and unstable meaning that deep learning systems become very unreliable.
• Open AI’s approach to solving this problem is massive amounts of compute power which makes Open AI Microsoft’s best friend but I think the approach is flawed and won’t work (see here).
• I can see very specific applications for this development such as debugging computer software, but it will not advance Open AI’s stated goal of AGI in my opinion.
• Hence, I still see AI winter coming sometime in the next two to two and a half years.
• The cold forecast remains in place.