IQ test: 'Artificial intelligence system as smart as a four year-old' Researchers at the University of Illinois at Chicago have applied an IQ test to MIT's ConceptNet 4 artificial intelligence system, and determined it's about as smart as a somewhat-challenged four-year-old child. The team used the Weschsler Preschool and Primary Scale of Intelligence Test on the system and found it performed …
What it's revealing is how little we understand of what 'intelligence' really is (let alone 'consciousness'). There's a famous story of an AI researcher from the 'golden years' (Donald Michie?) setting a doctoral candidate a summer task of developing robotic vision, thinking that it was a relatively easy problem. 50 years later, we have ANPR, but facial recognition remains pretty hit or miss. Progress is being made in restricted specialised areas, such as autonomous cars, but I think it will be decades at least before there's an AI capable of sustaining an intelligent conversation (though, again, one could say the same of many humans).
So LISP was what 1963? 50 yrs to get to 4.
And how long did it take for the human race to evolve beyond knuckle-dragging...? About a hundred thousand years or so?
I've been quite harsh on "AI" stories here in the past because they were almost all "faster recipe crunching" or "bloody great database search" stories. It's good to see some are still trying (and some are still evaluating) AI in the real conceptual and intelligence realm. It's a bloody difficult problem, and getting to 4 in 50 years is an amazing accomplishment.
The IQ test was developed by the french to identify which children needed more support at school. The author of the original test didn't think you could sum up intelligence with a single number and actually considered that the test was only useful comparing children with similar backgrounds.
Despite the tweaks to the test since, personally I think a better test of intelligence is asking peoples perception of the worth of the test across the entire population against people with radically different ages.
I personally consider people with a high opinion of the test (knowing it's history and methodology) to be remarkably unintelligent.
Nobody has perfect common sense, but people have a lot more common sense than any Artificial Intelligence program. Common sense in't a 1 or a 0 kind of thing. It's a matter of degree. Which means that Common Sense is analog and not digital.
To be able to answer the questions of 'Why?' about anything you need to have working mental models of people, things, and nature in your mind. Because to be able to answer, 'Why does ice feel cold for a person?', you need to have a working model of ice where you can mentally simulate ice touching various things and people and cooling them through physical contact.
Just remembering facts doesn't help you much to answer the question of 'Why?', because to answer such a question you need to create a mental simulation of the situation you are being asked about and observe what happens in this simulation, so that you can explain what's going on, which is the answer to 'Why?'.
Your ability to create mental simulations of various situations depends a lot on how sophisticated your metal models of various things and people are. That's why children's answers to the question of 'Why?' are a lot less explanatory than are such answers from adults. The mental models that children have are a lot simpler.
Perhaps modern computers are still too slow and have too little memory to be able to have working models of everything and be able to create computer simulations of various situations at the drop of a hat, the way people do mentally. But quantum computers might be able to do something like this in the future.
...but is this not in truth a computer-based software system developed specifically to approximate and imitate a narrow range of intelligent behaviour? If I am not mistaken, the thing is not self-aware. It is therefore an experiment in cutting-edge, clever programming but there's no real intelligence, merely pre-programmed responses, and it is certainly not at the level of a 4 year old child in reality. Of course I may be wrong. If so please set me right.
@ Anomalous Cowshed
I think the idea is simply to benchmark how far they have reached. It may be 'clever programming' etc. and 'no real intelligence' but there's a lot to be learnt by attacking the problem in that way. For example, they now know (and perhaps this was obvious to them or perhaps not) that the 'why' problems are the hard ones. So a future version might be hybrid: use the current technology for the 'what' and 'where' type questions which are factual and knowledge-based and working reasonably well; and work on something new and better or different for the 'why' questions. But at least they know where to focus their efforts.
If you've ever had the opportunity to observe it in real life, the transition between mechanical response and self-awareness is amazing. If your computer achieved this state, you would be justified in kneeling before it and welcoming it as your new overlord.
"...but is this not in truth a computer-based software system developed specifically to approximate and imitate a narrow range of intelligent behaviour? "
I think we are talking about something a bit more complex than ELIZA here. Very impressive in my opinion. The problem seems to be with giving an AI the right background parametres. To program a 4 year old would appear to take about 4 years of streaming data to it...
AI does not have to replicate human intelligence to be useful. Just being able to replicate the intelligence of a beetle would be a damn good start.
A few months back I spent many minutes trying, and eventually succeeding, to kill a fly. A fly's ability to navigate and evade danger, find food etc are quite remarkable. Being able to achieve that level of AI in a robot would lift the game considerably.
"A fly's ability to navigate and evade danger, find food etc are quite remarkable."
Ah, but is the fly remarkable, or is it able to evade you so well because its 2GHz clocking speed far outstrips your 90MHz clock? To a fly we are lumbering slow-motion jackasses. We only get them effectively when we learn how they fly and aim there rather than at the fly itself, and even then the fly can evade if it notices in time.
As for food...it's a fly. Everything is food.
This is how to kill a fly WITH YOUR BARE HANDS!!!
1. wait for the fly to land somewhere convenient
2. put your hand in "ready to flick" mode (index finger bent with nail against pad of thumb)
3. slowly slide your hand along the surface that the fly is sitting on until the fly is just within range
4. FLICK
5. go squish the fly which is now lying stunned on the floor a couple of metres away.
This gets lots of attention, theses, papers and grants. Not much talked about, though, is that no one can get money to study artificial stupidity, probably because we have too much of the natural kind already, organic -- and smelling of it.
But that's something that REALLY needs a cure.
Time to overhaul the laws, in anticipation for sentient AIs. Might as well prepare for the possibility of them becoming full citizens at some point in the future, and have the reasonings worked out in advance of how we are going to do this, then have to deal with it when our attention might be diverted elsewhere.
Instead the MSM bombards us on a daily basis with promises of 'Terminator' class technological self-awareness on the horizon. Ray Kurzweil et al, would you please STFU! Just give me a boring domestic bot! I want an easier life!!!
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
