Slashdot Mirror
Software Is Eating the World, But AI Is Going To Eat Software, Nvidia CEO Says (technologyreview.com)
An anonymous reader writes: Nvidia's revenues have started to climb in the recent quarters as it looks at making hardware customized for machine-learning algorithms and use cases such as autonomous cars. At the company's annual developer conference in San Jose, California last week, the company's CEO Jensen Huang spoke about how the machine-learning revolution is just starting. "Very few lines of code in the enterprises and industries all over the world use AI today. It's quite pervasive in Internet service companies, particularly two or three of them," Huang said. "But there's a whole bunch of others in tech and other industries that are trying to catch up. Software is eating the world, but AI is going to eat software."
You really shouldn't write headlines when you're stoned.
Slashdot is the only place I know of that has such a ridiculously restrictive, illogical definition of "AI." To Slashdot, something is only "really" AI if it works exactly like the "AI" in TV and movies. What if it translates speech into words, like a human can? Not AI. What if it can identify objects in an image, and write unique natural language sentences about them, like a human can? Not AI.
The reason I say it's illogical is because it's not even a difference in kind, it's a difference of degree. If we make an android that works just like TV and movies, it is very likely going to be using techniques like these for its vision and language components. It's like saying a model rocket isn't a rocket at all, because it can't go to space like other heavier-duty rockets. The fact that it looks like a rocket or functions similarly to a rocket does not qualify it to be called a model "rocket".
Indeed, most software is governed by business rules and being able to explain exactly what the system will do and why is essential. If say budgets over $100k need board approval, the someone has to program exactly that and nothing else. AI is great when the outcome is more important than the reasons behind it, like does this patient have cancer? If it can consult a huge database of cases and make millions of statistical weights we don't really care how it arrives at 83% as long as roughly 83 out of 100 patients end up actually having cancer. Then it's usually back to business rules for further examination/treatment though. More AI = more software work, not less.
Live today, because you never know what tomorrow brings
Nothing we're seeing these days is actually AI.
You should try to learn what "AI" actually means. Lookup "Strong AI" and "Weak AI", also referred to as "Hard AI" and "Soft AI".
We're seeing complex decision trees based on statistics, not AI.
No, what NVidia is talking about is not "decision trees".
Handwriting, speech recognition, and image processing along with their machine learning foundations do not impress the older /. crowd because they are not new technologies.
Dragon has been doing speech recognition better than Siri for almost 20 years. Simple command-based systems that only recognize a few words have been around longer than that.
Handwriting recognition for constrained tasks is also not new. The US Postal Services has had zip-code OCR systems since the 1980s.
Feature detection in images is not new, either. The only thing that's really changed there is we have the processing power to do it at scale.
Going beyond the applications, all the modern "AI" systems are simply classifiers on steroids. Processing power and great storage capacity allows us work on larger data sets, but in the end, we're just creating complex hyper-planes to bin data in one bucket or another.
Machine learning algorithms are great tools and it's great that we have the compute resources to really leverage them, but there's nothing really new that wasn't obvious 30 years ago. The only question was when we'd have the compute power to start doing the cool things we knew they could be used for.
(ok, I'll give a little credit to the deep learning researchers for bringing neural nets back into vogue, since those were written off 30 years ago during the AI winter, but they're still just classifiers from the mathematical perspective).
-Chris
And I should reply directly the the parent: statistically stringing together text is also not new. We just have better collections to start training algorithms with.
Why do you false dichotomy guys always show up in these threads? Of course pattern recognition is AI, and of course it is just an advanced program. Are a lot of programmers dualists instead of physicalists? Your brain went through the same brute force feedback loop when it was wiring itself before you were conscious enough to recognize it. It is also far better than anything we'll be able to make artificially in a variety of ways, but it isn't qualitatively different.
Siri uses the Dragon engine for speech to text. It's literally impossible for Dragon to be doing it better.