Smarter-than-Human Intelligence & The Singularity Summit

runamock writes "Brilliant technologists like Ray Kurzweil and Rodney Brooks are gathering in San Francisco for The Singularity Summit. The Singularity refers to the creation of smarter-than-human intelligence beyond which the future becomes unpredictable. The concept of the Singularity sounds more daunting in the form described by statistician I.J Good in 1965: 'Let an ultra-intelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultra-intelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultra-intelligent machine is the last invention that man need ever make.'"

Comments

We still have no clue how to do strong AI

[Animats]

OK. here's where we are:

• Logic-based AI AI looked so close in the 1960s, once it was realized that you could get a computer to do mathematical logic. All that was necessary was to express the real world in predicate calculus and prove theorems. After all, that's how logicians and philosophers all the way back to Aristotle said thinking worked. Well, no. We understand now that setting up the problem in a formal way is the hard part. That's the part that takes intelligence. Crunching out a solution by theorem proving is easily mechanized, but not too helpful. That formalism is too brittle, because it deals in absolutes.
• Expert systems Today, it's clear that they're no smarter than the rules somebody puts in. But back in the 1980s, when I went through Stanford, people like Prof. Ed Feigenbaum were promising Strong AI Real Soon Now from rule based systems. The claims were embarrassing; at least some of that crowd knew better. All their AI startups went bust, the "AI Winter" of low funding followed, and the whole field was stuck until that crowd was pushed aside.
• Neural nets / genetic algorithms / learning systems These all belong to the family of hill-climbing optimizers. These approaches work on problems where continuous improvement via tweaking is helpful, but usually max out after a while. We still don't really understand how evolution makes favorable jumps. I once said to Koza's crowd that there's a Nobel Prize waiting for whomever figures that out. Nobody has won it yet.
• Bayesian statistics Now used to do many of the things that used to be done with neural nets, but with a better understanding of what's going on inside. Lots of practical problems in AI, from spam filtering to robot navigation, are yielding to modern statistical approaches. Compute power helps here; these approaches take much floating point math. These methods also play well with data mining. Progress continues.

AI is one of those fields, like fusion power, where the delivery date keeps getting further away. For this conference, the claim is "some time in the next century". Back in the 1980s, people in the field were saying 10-15 years.

We're probably there on raw compute power, even though we don't know how to use it. Any medium-sized server farm has more storage capacity that the human brain. If we had a clue how to build a brain, the hardware wouldn't be the problem.