At this point in time, we don't know enough about brains and high level cognition to understand what is missing and still needs to be done in order to make a recognisable and generally accepted "true" AI.
It could be that we need several key breakthroughs in understanding. Those could happen at any time, making it possible that we are now within a decade of creating an Ex-Machina type brain. But equally it could be a century or more, we simply don't know how hard these problems are.
Alternatively, it might "just" be a matter of increased computing power and engineering using already-understood components - given that there are successful simulations of nervous systems in very basic creatures such as nematode worms and the visual processing of the bee. Scaling these up to human-level brain power (ignoring whether or not this in itself will lead to an intelligent system) is an engineering feat that will likely take many decades. Bullish predictions by Ray Kurzweil based on Moores law suggest 2040 as date where we might have the raw computing power available.
Probably we need both things to happen. Simply scaling up our existing work will allow better resolution and faster training for vision processes, but for example a classifier that can recognise objects is not an intelligent being. Also, just combining a bunch of AI code that we already have and letting it run faster with more memory and better sensors doesn't seem that promising - it is very likely that we will need time to learn how to combine all the parts to be successful, even assuming we have most of the individual parts understood by the time it is possible.
Here are links to some recent AI projects that might be thought-provoking, and give some sense of how far we've come since the early days of computing:
Robobees - drones can be flown using a vision system based on analysis of real bees' neural networks.
Describing images in English - a neural network can describe the contents of an image using natural language (note this is not the same as understanding such an image and the network has no "agency").
Deep dreaming, a neural network vision system run in reverse with feedback, a bit of fun, but also gives a sense of how robotic vision systems work. Although there are analogs to human perception (perhaps in this case involving LSD), this and other analyses of state-of-the-art vision systems are showing that we've got something a bit wrong. Computer vision networks seem to require different architecture to biological ones, and can fail in different ways, implying we have missed something about how real brains work.
"Big Dog" robot is an advanced design for military kit-carrying device that can work in rough terrain alongside infantry. Gives a sense of how robotic movement and locomotion are doing.
COG is a research robot looking at many aspects of robotics and artificial intelligence. Take a look at the capabilities page to get a sense of the level at which the research looks into components of AI.
The Mitsuki chatbot is 2013 winner of the Loebner Prize. A quick conversation with it shows that although it can figure out a realistic response to single questions, it has real trouble with memory, common sense logic and following a conversation beyond sentence-by-sentence responses (e.g. I told it my favourite colour was a mix of red and blue, and asked it what colour that was, and it said 'Orange?')
I am not an AI researcher, so I have picked the above examples purely because I have heard about them recently. If any other project demonstrates a completely different AI or robotic capability, add a comment and I'd be pleased to add it to the list.
