We're frustratingly close on voice recognition, and I know people using computers for day to day dictation. The interesting thing is that the problem is training and practice, not algorithmic, and Google's wrapping that up now. (As far as I can tell, they're on the forefront of computational linguistics.)

We have the ability to parse out questions and return answers at the level of Jeopardy champions with Watson. Today, Watson would have been #113 on the supercomputer 500. In 2005, it would have been #3. #113 in 2005 was a cluster of 1024 Xeon 2.4Ghz.

I hate doing apples to oranges, but the only comparison I could find was from cpubenchmark.net, which puts the Xeon 2.5ghz as 381 on their propriatary PassMark. A $335 E3-1275 runs a 9,109 with 4 cores. That's back of the napkin 20-25 times faster. Let's go with 20, extrapolate and say within six years a n equivalent 64-processor machine would be sufficient for Watson-level performance, well within the range of an individual lab, with "good-enough" voice recognition!

The last piece here is conversational flow. Chatbots have demonstrated that we're nowhere near yet. That's algorithmic as much as anything, but it's a pretty big topic in computational linguistics, so I'm hoping we'll have that cracked soon.

Expert systems look to be reasonable in 5-10 years with fluency, with conversational AI not far behind. We know what we need. It's not an unknown problem at this point, for the most part.

A large part of Watson's performance on Jeopardy came from its buzzer timing.

Most human contestants (at the champion level) know the answers to most the questions. The gamesmanship, according to the all-time Jeopardy champion Ken Jennings in a recent Fresh Air interview (Jennings returned to the show to take on Watson) is in the button press timing.

Watson has superb accuracy.

The computer actually knew the answers to fewer questions than the humans, but when Watson did know the answers, it always got the buzzer.

Helps to pull back the curtain a bit at times.

True AI and speech are hard. They're getting good. Scary good at times. There's an Android voice-to-voice translation app I've used to communicate with a Mandarin speaker (I manage just a few phrases). After doing that, I spent about fifteen minutes just staring at my phone and realizing that another huge piece of my future was now my present.

And I know the pieces behind it: voice recognition (cloud-assisted), translation, speech synthesis. But it's still mind-bogglingly cool. To me at least.

For my kids, probably not so much.

Yes, Watson had a major advantage at buzzing in but he was also capable of understanding or computing the questions. This is significant as the questions on Jeopardy aren't simply x + y = z but contain the complexities of the English language.

"True" AI is a misnomer, though, and is still far away, as we still don't know what we don't know. But conversational AI? The processing power will be there and we know where the algorithms need to improve, and we have ideas on how to do it.

My point was more that we know what's left. That hasn't been the case before the last five years or so.

But wasn't HAL ready on 2001?

Sorry for the off-topic, but you’ve seen that HAL INITIALIZATION FAILED is one of the error codes 2011 builds of Windows 8 spit out? http://cl.ly/3z3t360U203B2g303i3Y

But HAL in that context is "Hardware Abstraction Layer.