On facial recognition tech in Ray-Bans:
Jon Burmeister writes: What possibly could go wrong? “Oh! See that pretty girl! Who is she? Where does she live?”
“Ow! That cop busted my head. Who is he? Where does he live?”
Now I expect the first example to be shrugged off, but the second one will get the establishment’s attention.
Nick Thurn writes: Didn’t Google try something like this about 15 years ago — Google Glass, I think — sunk without a trace [Editor’s note: Google Glass was launched in 2013 and pulled from the market in 2015]. Users were christened “Glassholes”.
There’s nothing special about this tech, especially when it can phone home via 5G — we’re already wandering round uniquely identified and tracked by our smartphones — even with all data-gathering turned off and while running a VPN.
I expect China, which is ramping up a pervasive “social credit” system, will equip their police with this type of tech at some point.
Frank Dee writes: Orwell’s surveillance in 1984 looks old-fashioned. The authorities have us at their fingertips. No more revolutions, no more protests, no more activism. We could all wear masks, but that in itself would cause suspicion. But there is one factor…
The amount of information that Meta, Instagram, etc have on us is only as good as the amount of information that we give them. A choice is coming; do you want social media, or do you want privacy?
On using nuclear to power AI:
Roberto writes: Can we please, please, please differentiate between the AIs?
There is AI that does important, laborious work, like looking for tumours or predicting earthquakes, that works hand in hand with human experts to better their outputs. Due to the relatively small sizes of their data sets they don’t burn huge amounts of energy, plus they can be ethical despite companies like I-MED dropping the ball
And then there’s Generative AI – a glorified predictive text algorithm that consumes tonnes of energy while stealing the work of millions of humans, so as to put these same humans out of work by producing increasingly incorrect and/or biased outputs, while making a handful of toxic, billionaire tech bros richer and more powerful.
Andrew Holliday writes: Note that Berg’s view rests on this assumption: “To the best of our knowledge, the current generation of AI follows a simple scaling law: the more data and the more powerful the computers processing that data, the better the AI.” We know that assumption is wrong. AI would be better if the data it accessed was moderated and filtered for accuracy (quality data) rather than quantity (it’s quantity that’s producing the distorted and biased BS examples of AI).
To be fair to AI, people are exactly the same. Reading peer-reviewed journals and quality text books and analysis results in a much better informed person than someone falling down the rabbit-hole of “doing their own research” — even though theoretically that is limitless.
AI is having the same problem – more data isn’t the solution, it’s the problem. More garbage in, more garbage out. So the power needs aren’t really a known quantity yet, or even a vaguely viable guesstimate – because the process itself is what needs a lot more work, not a lot more data. That way lies madness. Or, as we currently label it, alternative facts.
Rod M writes: Missing from this “debate” is the amount of potable water used by both nuclear power stations and AI data centres. My understanding is nuclear power plants use more than coal stations, which use significant amounts of fresh water.
According to Forbes: “Tech giants have significantly increased their water needs for cooling data centers due to the escalating demand for online services and generative AI products. AI server cooling consumes significant water, with data centers using cooling towers and air mechanisms to dissipate heat, causing up to 9 liters of water to evaporate per kWh of energy used.”
Bryn S writes: Yes, high quality neural network “AI” systems require a lot of processing to train the model. Lots of processing involves lots of cost, both in hardware and the supporting infrastructure, including electricity. (A side note: once your model is trained, using it to make predictions, answer questions, etc requires trivial amounts of processing.)
I suggest that we can reframe the requirement here: to develop new AI models you want lots of cheap power.
The CSIRO GenCost report indicates nuclear power is more expensive than firmed renewables. So, not cheaper. Dutton’s shadow of a plan involves a tiny fraction (5% from memory) of Australia’s power needs being supplied by nuclear power. So, not lots of power.
Perhaps we can become a low-cost place to train AI models using renewable power? For the lowest cost, run the training on more processors when energy prices are cheapest (middle of the day powered by rooftop solar) and slow it down at peak demand times.
