OpenAI’s artificial intelligence model GPT-4 has the capability to hack websites and steal information from online databases without human help, researchers have found. That suggests individuals or organisations without hacking expertise could unleash AI agents to carry out cyber attacks.

“You literally don’t need to understand anything – you can just let the agent go hack the website by itself,” says Daniel Kang at the University of Illinois Urbana-Champaign. “We think this really reduces the expertise needed to…

Tiny, hollow wires can produce doughnut-shaped laser light that could be used to levitate small objects or transmit information.

Conventional lasers typically make beams that appear as a single, small point of light when they hit a surface. But for some novel communication technologies that use light to transfer information, it can be better to use lasers that produce hollow beams like a drinking straw, which appear as a ring of light when they hit a surface.

Such hollow laser…

Building AIs that sleep and dream can lead to better results and more reliable models, according to researchers who aim to replicate the architecture and behaviour of the human brain. But other experts say recreating the intelligence we see within ourselves may not be the most fruitful path for AI research.

Concetto Spampinato and his colleagues at the University of Catania, Italy, were looking for ways to avoid a phenomenon known as “catastrophic forgetting”, where an AI model trained to do a new task loses the ability to carry out jobs it previously aced. For instance, a model trained to identify animals could learn to spot different fish species, but then it might inadvertently lose its proficiency at recognising birds.

They developed a new method of training AI called wake-sleep consolidated learning (WSCL), which mimics the way human brains reinforce new information. People shuffle short-term memories of experiences and lessons learned throughout the day into long-term memories while sleeping. The researchers say this method of learning can be applied to any existing AI.

Models using WSCL are trained as usual on a set of data for the “awake” phase. But they are also programmed to have periods of “sleeping”, where they parse through a sample of awake data, as well as a highlight reel from previous lessons.

Take an animal identification model more recently trained on images of marine life: during a sleep period, it would be shown snapshots of fishes, but also a smattering of birds, lions and elephants from older lessons. Spampinato says this is akin to humans mulling over new and old memories while sleeping, spotting connections and patterns and integrating them into our minds. The new data teaches the AI a fresh ability, while the remainder of the old data prevents the recently acquired skill from pushing out existing ones.

Crucially, WSCL also has a period of “dreaming”, when it consumes entirely novel data made from mashing together previous concepts. For instance, the animal model might be fed abstract images showing combinations of giraffes crossed with fish, or lions crossed with elephants. Spampinato says this phase helps to merge previous paths of digital “neurons”, freeing up space for other concepts in the future. It also primes unused neurons with patterns that will help them pick up new lessons more easily.

“The point is that, as you gather new knowledge, you just combine more complex patterns,” says Spampinato. “You might combine panda, giraffe and leopard, so you’ll create this very strange mythological figure, and that will force the model to learn more complex patterns that maybe in the future can be reused.”

Spampinato tested three existing AI models using a traditional training method, followed by WSCL training. Then he and his team compared the performances using three standard benchmarks for image identification. The researchers found their newly developed technique led to a significant accuracy boost – the sleep-trained models were 2 to 12 per cent more likely to correctly identify the contents of an image. They also measured an increase in the WSCL systems’ “forward transfer”, a metric indicating how much old knowledge a model uses to learn a new task. The research indicated AI trained with the sleep method remembered old tasks better than the traditionally trained systems.

Despite these promising results, some experts say using the human brain as a blueprint is not necessarily the best way to boost AI performance. Andrew Rogoyski at the University of Surrey in the UK says AI research is still in its infancy, and designing new models is often akin to alchemy. He also warns against anthropomorphising AI architecture or rigidly mimicking the human brain in ever greater detail.

“The human brain should not be regarded as the ultimate architecture for intelligence,” he says. “It’s the result of millions of years of evolution and an unimaginably wide range of stimuli. We may develop AIs that have structures completely different from their biological designers.”

Still, Rogoyski says the sleep training method seems interesting. Instead of having the dreaming AI mimic human beings, however, he suggests another biological inspiration: dolphins, which have the ability to “sleep” with one part of the brain while another remains alert, switching as needed. After all, an AI that requires hours of sleep is not ideal for commercial applications.

Topics:

Living human cells have resonant frequencies – rates at which they will naturally oscillate if the conditions are right. This has been suspected for decades, but researchers have now measured what some of those frequencies are.

When an object vibrates, the rate at which it moves back and forth has a big effect. At many frequencies, the object will only vibrate a little or quickly diminish, but at its resonant frequencies the size of the vibration becomes large.

Javier Tamayo at the Spanish…

A supercomputer capable of simulating, at full scale, the synapses of a human brain is set to boot up in Australia next year, in the hopes of understanding how our brains process massive amounts of information while consuming relatively little power.

The machine, known as DeepSouth, is being built by the International Centre for Neuromorphic Systems (ICNS) in Sydney, Australia, in partnership with two of the world’s biggest computer technology manufacturers,…

Balls of human brain cells linked to a computer have been used to perform a very basic form of speech recognition. The hope is that such systems will use far less energy for AI tasks than silicon chips.

“This is just proof-of-concept to show we can do the job,” says Feng Guo at Indiana University Bloomington. “We do have a long way to go.”

Brain organoids are lumps of nerve cells that form when stem cells are grown in certain conditions. “They are like mini-brains,” says Guo.

It takes two or three months to grow the organoids, which are a few millimetres wide and consist of as many as 100 million nerve cells, he says. Human brains contain around 100 billion nerve cells.

The organoids are then placed on top of a microelectrode array, which is used both to send electrical signals to the organoid and to detect when nerve cells fire in response. The team calls its system “Brainoware”.

New Scientist reported in March that Guo’s team had used this system to try to solve equations known as a Hénon map.

For the speech recognition task, the organoids had to learn to recognise the voice of one individual from a set of 240 audio clips of eight people pronouncing Japanese vowel sounds. The clips were sent to the organoids as sequences of signals arranged in spatial patterns.

The organoids’ initial responses had an accuracy of around 30 to 40 per cent, says Guo. After training sessions over two days, their accuracy rose to 70 to 80 per cent.

“We call this adaptive learning,” he says. If the organoids were exposed to a drug that stopped new connections forming between nerve cells, there was no improvement.

The training simply involved repeating the audio clips, and no form of feedback was provided to tell the organoids if they were right or wrong, says Guo. This is what is known in AI research as unsupervised learning.

There are two big challenges with conventional AI, says Guo. One is its high energy consumption. The other is the inherent limitations of silicon chips, such as their separation of information and processing.

Guo’s team is one of several groups exploring whether biocomputing using living nerve cells can help overcome these challenges. For instance, a company called Cortical Labs in Australia has been teaching brain cells how to play Pong, New Scientist revealed in 2021.

Titouan Parcollet at the University of Cambridge, who works on conventional speech recognition, doesn’t rule out a role for biocomputing in the long run.

“However, it might also be a mistake to think that we need something like the brain to achieve what deep learning is currently doing,” says Parcollet. “Current deep-learning models are actually much better than any brain on specific and targeted tasks.”

Guo and his team’s task is so simplified that it is only identifies who is speaking, not what the speech is, he says. “The results aren’t really promising from the speech recognition perspective.”

Even if the performance of Brainoware can be improved, another major issue with it is that the organoids can only be maintained for one or two months, says Guo. His team is working on extending this.

“If we want to harness the computation power of organoids for AI computing, we really need to address those limitations,” he says.

Topics:

Ukrainian attack drones equipped with artificial intelligence are now finding and attacking targets without human assistance, New Scientist has learned, in what would be the first confirmed use of autonomous weapons or “killer robots”. While the drones are designed to target vehicles such as tanks, rather than infantry, it is almost certain that the resulting explosions are killing Russian soldiers without a direct command from a human operator, although no casualties have been confirmed.

The drone is a quadcopter called …

An artificial intelligence has consistently beaten champion drone pilots in races for the first time, achieving lap times no human was able to match. The technology could be used to speed up drones carrying out everyday tasks.

The sport of drone racing involves humans piloting small quadcopters around a course at speeds of more than 100 kilometres per hour, with the vehicles subject to g-forces of up to 5 g. The drones must …

The alarm bells have been rung. In May, computer scientist Geoffrey Hinton, known as the “godfather of AI”, quit his role at Google to warn of the “existential threat” posed by artificial intelligence. The Center for AI Safety followed up with an open letter, signed by Hinton and hundreds of others, warning that advanced AI could destroy humanity. “Mitigating the risk of extinction from AI should be a global priority,” read the statement.

This sudden surge of concern seems to have been motivated by the rapid advance of AI-powered chatbots like ChatGPT and the race to build more powerful systems. The fear is that the tech industry is recklessly accelerating the escalation of AI’s capabilities. All of which sounds scary.

But the warnings are also suspiciously vague. When you scrutinise the scenarios typically put forward for precisely how AI could wipe out humans, it is hard to avoid the conclusion that such fears aren’t well-founded. Many experts are instead warning that fretting over long-term doomsday scenarios is a distraction from the immediate risks posed by existing AIs.

Generally, the people who talk about existential risks reckon that we are on a trajectory towards artificial general intelligence (AGI), roughly defined as machines that can out-think humans. They predict that people will invest advanced AIs with more autonomy, giving them access to vital infrastructure, such as the power grid or financial markets, or even putting them at the forefront of warfare – at which point they could go rogue or otherwise resist our attempts to control them.

But it remains to be seen if AIs will ever reach the kind of super-intelligence …

Drones being evaluated by the US military could soon be equipped with an artificial intelligence that is claimed to be better than humans at identifying targets, although the classified nature of the work makes it difficult to verify this claim.

Stephen Bornstein of Athena AI, the Australian company behind the system, says the AI will assist human drone operators, who can lose concentration after hours spent looking at streaming video.

“The AI will do a lot of the heavy lifting for them,” says …