AI News and Latest Developments: Drones “Kill” Operators. DeepMind speeds up your computer.wind turbine boost

Reports that an AI drone ‘killed’ its pilot were unfounded

This month we heard from a US Air Force Colonel about an interesting AI experiment. AI-controlled drones trained to autonomously carry out bombing missions attacked human operators when instructed not to attack targets. Its programming prioritized the successful completion of its mission, and decided to forcibly remove it, believing that human intervention would get in the way.

The only problem with this story was that it was nonsense. First, as the colonel said, the test was a simulation. Second, the U.S. Air Force hastily released a statement, revealing that a colonel speaking at a British conference had “misrepresented” and that no such experiment had taken place.

new scientist When asked why people are so quick to believe AI horror stories, one expert said part of it is due to humans’ innate fascination with “horror stories you want to whisper around the campfire.” Stated.

The problem with this kind of misunderstood narrative is that it’s very compelling. This “news” was published all over the world before it was confirmed, and few publications were interested in getting the facts straight afterwards. AI poses a real danger to society in many ways, and we need an informed debate to explore and stop them, not sensationalism.

DeepMind AI speeds up algorithms that could have a global impact on computer power

AI has continued to surprise us in recent years, showing that it can spit out essays on a given topic, create photorealistic images from scratch, and create functional source code. So it’s no surprise that the news that DeepMind AI slightly improves its sorting algorithm isn’t too exciting.

But when you dig deeper, the research is interesting and certainly has real-world applications. The sorting algorithm has been run trillions of times around the world and is commonly used in software of all kinds, so coders can invoke it as needed to save them the hassle of reinventing the wheel. written to the library. The algorithms stored in these files have long been refined and tweaked by humans to be considered complete and as efficient as possible.

This month, DeepMind’s AI discovered improvements that can speed up sorting by up to 70% in the right scenarios. Improvements that could be deployed on every computer, smartphone, or anything with a computer chip could result in significant savings in energy usage and computation time. How many other commonly used algorithms can AI find efficiency gains in? Time will tell.

AI could boost the output of all wind turbines around the world

While DeepMind is looking for efficiency gains in source code, other companies are using AI to find efficiency gains within machines. Wind turbines work best when facing directly into an oncoming wind, which keeps changing directions relentlessly. Today, turbines use a variety of techniques to keep them efficient, but it looks like AI could do a little better job.

The researchers trained the AI ​​on real-world data about wind direction and found that it could come up with strategies to increase efficiency by pointing the turbines in the right direction more of the time. This required more rotation and consumed more energy, but even taking that into account he was able to squeeze 0.3% more power out of the turbine.

This number may not make much of a headline, but it is enough to add 5 terawatt-hours of electricity per year if installed in all the turbines around it, which is enough for an average Albanian or British household of 170 This amount is almost the same as the annual consumption of 10,000 households. world.

Capitalization test is a surefire way to distinguish between AI and humans

The Turing test is a well-known method for evaluating the intelligence of machines. Can a human conversing through a text interface determine if he is talking to another human or an AI? I’m fairly proficient at conducting meaningful conversations, so perhaps a new test is needed.

In recent years, 204 test suites have emerged that have been proposed as a kind of new Turing test, covering subjects such as mathematics, linguistics, and chess. However, a simpler method has just been published in a paper, which involves adding extra capital letters or words to meaningful descriptions in order to trick the AI.

Give a human a phrase like “Is CURIOSITY water, is ARCANE wetTURBULENT, or is ILLUSION drySAUNA?” And you’ll notice that only lowercase letters form logical sentences. However, researchers have shown that an AI reading the same input would be baffled. Five large language models failed the test, including OpenAI’s GPT-3 and ChatGPT, and Meta’s LLaMA.

But other experts point out that the test exists today, and AIs can be trained to understand it, so it will pass in the future. The distinction between AI and humans can become a never-ending cat-and-mouse game.

What will the future of AI look like? Google and the EU have very different ideas

Regulators and tech companies don’t seem to be heading in the same direction when it comes to AI. Some industry officials have called for research to be halted until the hazards are better understood, but most lawmakers are pushing security rules to ensure research can proceed safely, and many high-tech Companies are going full steam ahead to commercial release of AI.

EU politicians have agreed to the latest version of the AI ​​law they have been working on for years. European Commission President Ursula von der Leyen pledged to quickly introduce an AI bill when she was elected in 2019. It would require companies to disclose copyrighted content used to train generative AIs such as ChatGPT.

Meanwhile, companies like Google and Microsoft, worried about being left behind by a revolution that might rival the birth of the internet, are pushing AI into many of their products.

While technology has always outpaced the law and society struggles to minimize harm, AI is indeed advancing at an alarming rate. And the results of commercial deployment can be devastating. Google has found that even the most curated for advertising, its output can be unreliable. The potential benefits of AI are undisputed, but the key is to make sure they outweigh the harm.

topic:

• artificial intelligence/
• AI