‘Don’t worry about existential threat AI might pose’: futurist
- 22 August, 2018 12:55
UNSW professor Toby Walsh
CIOs need to put aside the Hollywood-induced fear that machines will take over the world and that humanity’s days are numbered, according to AI expert Toby Walsh.
“We’re a long way from building machines that match human brains,” said Walsh, UNSW professor and research group leader at Data61 (CSIRO), during his keynote address at the CIO Summit in Sydney.
“We can build machines that do narrow focused tasks - and they can do those tasks often at super-human level - but it’s yet to be (maybe 50 or 100 years, or ever) before we can build machines that match the full capabilities of humans. And we certainly don’t build machines that have any consciousness, sentience, or desires of their own.
“They do exactly what we tell them to do. That’s the problem in fact,” he said, explaining computers are frustratingly literal devices.
“There are far more pressing problems facing the planet, like climate change, that we really need to deal with before we have to worry about the existential threat that artificial intelligence might pose.”
That said, Walsh said he’s thrilled the topic is now on the radar of the CIO community and starting to show up in strategic business plans.
“AI is everyone’s favourite subject these days. It was pretty much my favourite subject since I was a young boy. I was reading too much science fiction and dreaming about a future full of intelligent computers and robots.
“And that future seems to be arriving rather rapidly, so it’s great the rest of you are catching up to that dream I had as a young boy reading people like Arthur C. Clarke and Isaac Asimov.”
4 key trends
Walsh said there are a number of reasons why AI is making progress today and happening at this point in history.
“AI is starting to invade our lives in some way, sometimes good and sometimes bad,” he said. “So why is it happening at this point in history? Not ten years ago and not ten years in the future.”
He said the answer lies with four exponential trends. The first exponential trend is Moore’s Law, which is officially dead because of technical issues like physical quantum limits.
“Intel has declared they are not going to double transistor count every two years going forward," he said.
“I’m not worried that’s going to hold back the field. Chip designers have been pretty lazy over the last 20 years. They have mostly been just shrinking the 806 architecture.
"There has not been so much innovation in the design of chips and we’re starting to see that with DPUs. We’re going to see a lot more interesting things in specialised hardware to do particular tasks like machine learning, which will give us more compute with the same transistor count.”
“Whilst Moore’s Law is technically dead, there’s enough innovation that’s going to happen that will give us ever-increasing compute power.”
The second exponential is the ever-escalating amount of data. “Corporations are discovering one of the most vital things they have in their business is the data they have, about their operations, about their customers.
“The data that’s available has been doubling and that’s very useful for artificial intelligence because a lot of what we do these days, particularly machine learning, is training on data,” he said.
But there’s one limitation worth pointing out about AI today, he noted, explaining machines are incredibly slow learners.
“Unlike humans, you can all learn from a single example. Machines and state-of-the-art machine learning still needs hundreds of thousands, sometimes millions of examples to learn from," he said.
"But the good news is that that data is often being collected and we’re often having datasets that we can do that number from.”
The third exponential trend is the progress being made on the algorithmic front. “In the last few years, with things like deep learning, we’ve been seeing in some cases exponential improvements in the performance of those algorithms.”
The fourth exponential trend is the amount of money flowing into the field, he noted.
“With this you can measure the amount of venture capital flowing into the field. You can see the activity, the amount of people, the number of companies and the number of startups - all of these sorts of things have been doubling again every two years or so,” he said.
“Put those four things in a pot together and that is the recipe, largely speaking, for making significant progress.”
But while progress is being made, Walsh cautioned there’s still a long way to go before the industry can build machines that can perform a broad set of tasks that humans can do.
“We can build narrow-focused tasks. We can teach them to do narrow-focused things like play Go, read x-rays, diagnose eye disease. So there’s a lot we can do, but there’s a lot still we can’t do and a number of challenges.”
Given the ongoing limitations, he said AI today can do tasks that require a ‘moment’s thought.’
“You can recognise faces with a moment’s thought and that’s what we can teach computers to do. We can teach computers to do it, but we should be very careful that when we do so, they may very well have the same biases that we have,” he said.
“So what will be get machines to do? The goal ultimately is what we call the 4D’s: the dirty, the dull, the difficult and the dangerous. All things that we don’t want humans to do. So in that sense we should be happy that we get machines to take over these jobs.”
Already, some of these jobs today involve increased amounts of automation and AI. Chatbots are a good example of a practical use of AI.
“It’s starting to invade parts of our lives - and we’re not even aware of it.”
AI is much more than just machine learning, he added. “AI is about also getting computers to make the sorts of decisions that humans make. And do that in a much more thorough, systematic and optimal way.”