AI roundup this month: Drone ‘kills’ operator; DeepMind accelerates

Reports of an AI drone “killing” its operator have led to nothing

This month we heard about a fascinating AI experiment by a US Air Force colonel. An AI-controlled drone trained to autonomously perform bombing missions had turned to its human operator when told not to attack targets; his scheduling prioritized completing missions successfully, so he saw human intervention as an obstacle in his way and decided to forcibly take him out.

The only problem with the story was that it didn’t make sense. First, as the colonel said, the test was a simulation. Secondly, a US Air Force statement was hastily issued clarifying that the Colonel, speaking at a conference in the UK, had “misspoke” and that no such tests had been carried out.

New scientist asked why people are so quick to believe AI horror stories, with one expert saying it was partly due to our innate attraction to “horror stories we like to whisper around the fire.

The problem with this kind of misunderstood story is that it’s so compelling. The “news” was published around the world before any facts could be verified, and few of those publications had an interest in setting the record straight afterwards. AI poses a real danger to society in many ways and we need informed debate to explore and prevent them, not sensationalism.

DeepMind AI accelerates algorithm which might have global impact on computer power

Artificial intelligence has brought surprise upon surprise in recent years, proving capable of spitting out an essay on a given topic, creating photorealistic images from scratch, and even writing functional source code. So you’d be forgiven for not being too excited about news of a DeepMind AI slightly improving a sorting algorithm.

But dig deeper and the work is interesting and has solid real-world applications. Sorting algorithms are executed trillions of times around the world and are so commonly used in all kinds of software that they are written in libraries that programmers can call as and when needed to avoid having to reinvent the wheel. These archived algorithms had been perfected and optimized by humans for so long that they were considered complete and as efficient as possible.

This month, DeepMinds AI saw an improvement that can speed up sorting by up to 70%, in the right scenario. Any improvement that can be implemented on every computer, smartphone, or anything with a computer chip can bring huge savings in power consumption and computation time. In how many most commonly used algorithms can AI find efficiency gains? Time will tell.

Artificial intelligence could boost the output of all the world’s wind turbines

While DeepMind looks for efficiency gains in source code, others are using artificial intelligence to find them in machines. Wind turbines work best when they face the oncoming wind directly, but the breeze stubbornly keeps changing direction. Turbines currently use a variety of techniques to maintain efficiency, but it appears that artificial intelligence may be able to do a slightly better job.

The researchers trained an AI on real-world data on wind direction and found that it could come up with an efficiency-boosting strategy by keeping the turbine facing in the right direction most of the time. This resulted in more spin, which used more energy, but even taking that into account they were able to squeeze 0.3% more power out of the turbines.

This figure may not make a big headline, but it is enough to boost electricity generation by 5 terawatt hours per year, about the same amount consumed annually by Albania, or 1.7 million average UK homes, when distributed to every turbine around the world.

The caps test is a surefire way to sort AIs from humans

The Turing test is a famous way to evaluate the intelligence of a machine: can a human being conversing through a text interface tell if he is talking to another human or an artificial intelligence? Well, large language models like ChatGPT are now quite adept at holding lifelike conversations, so perhaps we need a retest.

In recent years, we’ve seen a suite of 204 tests proposed as a kind of new Turing Test, covering subjects like math, linguistics, and chess. But a much simpler method has just been featured in a paper, in which capital letters and superfluous words are added to sensible statements in an attempt to derail the AI.

Give a human being a phrase like CURIOSITY waterARCANE wetTURBULENT or ILLUSION drySAUNA? and they are likely to notice that lowercase letters alone form a logical sentence. But an AI reading the same input would be confused, the researchers showed. Five large language models, including OpenAIs GPT-3 and ChatGPT and Metas LLaMA, failed the test.

But other experts point out that the test exists now, AI can be trained to understand it, and so it will pass in the future. Distinguishing AI from humans could become an endless cat-and-mouse game.

What is the future of AI? Google and the EU have very different ideas

Regulators and tech companies don’t seem to be moving in the same direction on AI. While some industry players have called for the research to halt until the dangers are better understood, most lawmakers are pushing safeguards rules to ensure it can progress safely, and many tech companies are racing full speed ahead to release commercial AI.

EU politicians agreed on an updated version of its AI law, which took years to work on: European Commission President Ursula von der Leyen promised to urgently introduce AI legislation when she was elected in 2019. Laws will now require companies to disclose any copyrighted content used to train generative AI like ChatGPT.

On the other hand, companies like Google and Microsoft are continuing to implement artificial intelligence in many of their products, worried about being left behind in a revolution that could rival the birth of the Internet.

While technology has always outpaced legislation, leaving society scrambling to ensure harm is kept to a minimum, AI is indeed moving forward at an astonishing pace. And the results of its commercial launch could be catastrophic: Google has discovered that its output can be unreliable even when selected for advertising. The potential benefits of AI are undisputed, but the trick will be to make sure they outweigh the harms.

Themes:

• artificial intelligence/
• TO THE