Will AI ever counter human stupidity?
Artificial intelligence is quickly bringing exciting smart tools ever closer, but there are questions over our ability to properly harness it — and dubious or even illegal uses
AI never gets bored or tired, and can be better than humans for some tasks. That opens up amazing new horizons for us, but also worries because humanity is deeply flawed.
There’s plenty of good AI stuff already: medical researchers use it to spot early onset of diseases, and authorities can have AI chew through spatial information from satellites to predict flooding. Traffic management where cars, cyclists and pedestrians are directed towards non-congested routes will probably be with us before improved public transport as it’s cheaper to implement and of course, “smart”.
Local heroes Cacophony believe AI could even build better rodent traps that recognise and are selective in what they catch and use sound lures to attract pests. Fingers crossed that Cacophony’s open source effort works and takes off, because New Zealand is overrun by pests. If AI enabled traps can’t deal with the pest problem humans have created, it’s hard to imagine what will.
There are lots of good applications for AI, to build tools that can truly help solve very difficult problems. After many decades of theorising, we’re finally at the stage where AI can think like people; almost, just a little, enough to create tireless tools to interact with us and our environment.
It is that ability of computers to reach directly into our lives that has people worried, and no doubt one reason why Google assembled a defensive showcase of AI for good in Tokyo earlier this month.
At the Solve with AI event, Google research lead Jeff Dean asserted the company’s ethical stance when working with the emerging technology. Google’s powerful AI tools are open source available to anyone however, and can be used for dubious and even illegal purposes.
Furthermore, AI systems are created and trained by biased people and can make mistakes which can be difficult to detect.
Even with the clever and good AI tools displayed, you are left thinking that these tech solutions only nibble away at the edges of much larger and more serious problems — and maybe even hide them.
Sure, it’s satisfying to see implementations such as policing of rainforests in the Amazon using discarded Android phones that are solar powered and listen out for illegal logging.
Thanks to the devices, local tribes are able to catch illegal loggers and burn their trucks to dissuade them from returning to fell trees. Will it scale up and put paid to the global business of large-scale illegal logging, an ecological disaster that’s threatening life on Earth though?
Indian cotton farmers able to
recognise bollworm infestations of crops by taking pictures of bugs stuck to traps, with AI analysing the images using an app. Cultivating a single kilogram of cotton requires 22,500 litres of water. India is running out of water.
And what was Google thinking, presenting in Japan of all places, how to track humpback whales with AI that can accurately identify their sounds from up to 20 nautical miles away?
Perhaps when well-trained machine learning models that have been running for a while co-operate with one another and independently come to conclusions and insights, we’ll have a SuperAI that is able to consider human stupidity and counter it.
Such a SuperAI would require substantial improvements for computer hardware however. Currently, Google’s third-generation tensor processing unit (TPU) board provides 420 tera floating point operations per second of maths crunching prowess.
Compared to just 10 years ago, that’s a huge number. The TPU boards can be combined into larger systems providing hundreds of peta-FLOPS which enables today’s limited AI uses a large amount of power in the process.
Estimates I’ve seen suggest an AI system being able to do what the brain does with a power consumption of 15 to 20 Watt, would require tens of megawatts. From a sustainability point of view, it would be stupid to build a power-guzzling AI like that.
Chip makers such as Intel are working on low-power processors such as the recent Pohoiki Beach system, with 64 Loihi chips for cognitive computing research, but they are a long way from being able to match the brain’s capability.
How far? Pohoiki Beach provides eight million neurons, and Intel hopes to have 100 million neuron systems ready by the end of the year.
The human brain has somewhere in the region of 100 billion neurons. We have a few years left before the neuromorphic engineers best us, in other words.
It is that ability of computers to reach directly into our lives that has people worried.