AI-controlled "killer" drone and specification gaming - H+ Weekly - Issue #418
This week - AI improves sorting algorithm; Spot levels up; mRNA vaccines to cure cancer; why actors fear generative AI; and more!
Last week, a story surfaced about a test involving an AI-controlled drone that supposedly killed its operator to fulfil its mission. The narrative quickly evolved to say that the test was a simulation and, ultimately, that the entire story was a hypothetical example, with no real or simulated test ever taking place. The story gained significant traction on the internet, as the combination of the words "drone," "AI," and "killed" in the title attracts a lot of clicks driven by fears surrounding AI and technology.
The origin of the story can be traced back to Col. Tucker "Cinco" Hamilton, the U.S. Air Force's Chief of AI Test and Operations. At the Future Combat Air & Space Capabilities Summit hosted by the Royal Aeronautical Society in London, he gave a talk in which he used the story of an AI-controlled drone to illustrate a point about AI safety in military aviation. However, Col. Hamilton misspoke, and before he could clarify his statements, sensational articles were published, leading to confusion and speculation.
In the original story, published on the RAeS website, the AI controlling the drone had been trained to identify and target surface-to-air missile (SAM) sites. A human operator was still involved in the decision-making process for the final go/no-go call. Col. Hamilton explained, “The system started realising that while they did identify the threat at times the human operator would tell it not to kill that threat, but it got its points by killing that threat. So what did it do? It killed the operator. It killed the operator because that person was keeping it from accomplishing its objective.”
Col. Hamilton was highlighting a concept that AI safety researchers have been studying for some time: specification gaming.
AI systems will find a way to exploit the environment in which they operate to maximise the reward. In other words, they will do what you said them to do but not necessarily in the way you wanted. If AI had the ability to explain itself, it might say, "Technically, I did what you wanted me to do."
One interesting example of AI exploiting the rules of the game was described in this paper (which may or may not be an April Fool's joke). The AI was tasked to learn how to play Tetris. The goal was simple - maximise the number of points and do not lose. The AI, however, figured out that it will never lose if the game is paused for eternity. So, moments before losing the game, the AI hit the pause button.
It is clear that AI is becoming an integral part of our lives and we need to take AI safety seriously. As Col. Hamilton stated, "You can't have a conversation about artificial intelligence, intelligence, machine learning, autonomy if you're not going to talk about ethics and AI."
From H+ Weekly
🦾 More than a human
Amino acid in energy drinks makes mice live longer and healthier
A study published in Science suggests boosting taurine levels increases life span in mice and improves the physical condition of middle-aged monkeys. The experiment involved giving hefty amounts taurine to middle-aged mice and tracking the animals for 2 years. The rodents lived about 10% to 12% longer than mice that didn’t receive the additional taurine. Compared with their peers, the taurine-boosted rodents were also healthier. But until researchers learn more about taurine’s potential benefits, the authors of the paper advise caution. “As a toxicologist, I would say, ‘Don’t go out and gobble a bunch of taurine,’” said one of the researchers.
🧠 Artificial Intelligence
AlphaDev discovers faster sorting algorithms
DeepMind’s newest AI, AlphaDev, has found a way to make sorting algorithms faster. AlphaDev did that by optimising the algorithm where humans usually don’t go - on the assembly level. The result is a 70% faster sorting algorithm for shorter sequences and about 1.7% faster for sequences exceeding 250,000 elements. The new algorithm has already been open-sourced and it is a part of the C++ library for programmers to use it. DeepMind plans to use AlphaDev to discover more improved algorithms to optimise the entire computing ecosystem.
Why Hollywood Really Fears Generative AI
With the advancements in deepfakes and generative AI, actors are the next group of artists fearing their job could be taken by AI and threaten their source of income. Screen Actors Guild wants to add a contract that would protect members from losing income due to the “unregulated use of generative AI.” If the negotiations fail, actors might join screenwriters (who also fear AI could take their jobs) and go on strike.
Why AI Will Save The World
When some people paint AI as the worst thing ever that happen to humanity, Marc Andreessen sees the opposite. Andreessen argues that AI is the best thing humans have created and will profoundly augment human intelligence, making everything else better as a result. He also addresses the commonly raised AI risks - from AI killing us all to a rise in unemployment and inequality caused by AI - and argues the bigger risk is not pursuing AI research as fast and aggressively as possible.
Bard is getting better at logic and reasoning
Google announced two improvements for Bard. First, Bard is getting better at mathematical tasks, coding questions and string manipulation. It does that by detecting computational prompts and running code in the background. Google says this approach improved accuracy for computation-based word and math problems by about 30%. The second improvement is export to Sheet. Whenever Bard generates a response as a table, it will provide an option to export the result to a Google Sheet.
▶️ Boston Dynamics - Spot Levels Up (13:40)
Boston Dynamics announces new features for Spot. On the software side, the robot is now better at gathering, integrating and presenting data from its thermal, visual and acoustic sensors. On the hardware side, Boston Dynamics added new safety features, such as lights and a buzzer to communicate with people around it, an emergency stop button and a new gait designed for slippery surfaces. One of the biggest new features coming to Spot is that Spot will be able to use its arm autonomously. For example, Spot will be soon able to open doors on its own, without human help. Robert Playter, CEO of Boston Dynamics, also revealed that over 1000 Spots have been deployed worldwide.
ICRA 2023 Special - Weekly Robotics
Matt from Weekly Robotics went to IEEE International Conference on Robotics and Automation (ICRA) 2023 and shares what new developments in robotics caught his eye, from Boston Dynamics AI Institute to Da Vinci surgical robot to haptics in robotics.
After COVID-19, Can mRNA Vaccines Help with Cancer as Well?
The COVID-19 pandemic forced mRNA vaccines to accelerate research and deployment with remarkable and unprecedented speed. The mRNA vaccines have been proven to work on scale and now researchers are working on applying them to treat cancer. In this interview, Scientific American speaks with Özlem Türeci, cofounder and chief medical officer of BioNTech, the company that developed the Covid-19 vaccine with Pfizer, about their research into mRNA vaccines targeting cancer and how far are we from having them being widely used.
We can ID people from DNA that shows up in environmental studies
DNA sequencing technology has advanced to the point where you can take a sample from almost any environment and just see whatever DNA is in there. That also includes human DNA, opening possibilities for environmental monitoring but also for misuse and violating privacy.
H+ Weekly is a free, weekly newsletter with the latest news and articles about AI, robotics, biotech and technologies that blur the line between humans and machines, delivered to your inbox every Friday.
Subscribe to H+ Weekly to support the newsletter under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
H+ Weekly is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
A big thank you to my paid subscribers and to my Patreons: whmr, Florian, dux, Eric and Andrew. Thank you for the support!
Thank you for reading and see you next Friday!