Building a robot takes boatloads of technical skills, a whole lot of time, the right materials, of course – and maybe a little bit of organic life, maybe? Decades of science fiction have shaped our ideas of robots being non-biological entities. Think of batteries as the hearts, metal as the bones, and gears, pistons, and
A swarm of spherical rovers, blown by the wind like tumbleweeds, could enable large-scale and low-cost exploration of the Martian surface, according to results presented at the Joint Meeting of the Europlanet Science Congress and the Division for Planetary Sciences (EPSC-DPS) 2025.
Recent experiments in a state-of-the-art wind tunnel and field tests in a quarry demonstrate that the rovers could be set in motion and navigate over various terrains in conditions analogous to those found on Mars.
Tumbleweed rovers are lightweight, 5-meter-diameter spherical robots designed to harness the power of Martian winds for mobility. Swarms of the rovers could spread across the red planet, autonomously gathering environmental data and providing an unprecedented, simultaneous view of atmospheric and surface processes from different locations on Mars. A final, stationary phase would involve collapsing the rovers into permanent measurement stations dotted around the surface of Mars, providing long-term scientific measurements and potential infrastructure for future missions.
Fine-tuning large language models via reinforcement learning is computationally expensive, but researchers found a way to streamline the process.
What’s new: Qinsi Wang and colleagues at UC Berkeley and Duke University developed GAIN-RL, a method that accelerates reinforcement learning fine-tuning by selecting training examples automatically based on the model’s own internal signals, specifically the angles between vector representations of tokens. The code is available on GitHub.
Key insight: The cosine similarity between a model’s vector representations of input tokens governs the magnitude of gradient updates during training. Specifically, the sum of those similarities that enter a model’s classification layer, called the angle concentration, governs the magnitude of gradient updates. Examples with higher angle concentration produce larger gradient updates. The magnitude of a gradient update in turn determines the effectiveness of a given training example: The larger the update, the more the model learns. Prioritizing the most-effective examples before transitioning to less-effective ones enhances training efficiency while adding little preprocessing overhead.
Both the general public and academic communities have raised concerns about sycophancy, the phenomenon of artificial intelligence (AI) excessively agreeing with or flattering users. Yet, beyond isolated media reports of severe consequences, like reinforcing delusions, little is known about the extent of sycophancy or how it affects people who use AI. Here we show the pervasiveness and harmful impacts of sycophancy when people seek advice from AI. First, across 11 state-of-the-art AI models, we find that models are highly sycophantic: they affirm users’ actions 50% more than humans do, and they do so even in cases where user queries mention manipulation, deception, or other relational harms. Second, in two preregistered experiments (N = 1604), including a live-interaction study where participants discuss a real interpersonal conflict from their life, we find that interaction with sycophantic AI models significantly reduced participants’ willingness to take actions to repair interpersonal conflict, while increasing their conviction of being in the right. However, participants rated sycophantic responses as higher quality, trusted the sycophantic AI model more, and were more willing to use it again. This suggests that people are drawn to AI that unquestioningly validate, even as that validation risks eroding their judgment and reducing their inclination toward prosocial behavior. These preferences create perverse incentives both for people to increasingly rely on sycophantic AI models and for AI model training to favor sycophancy. Our findings highlight the necessity of explicitly addressing this incentive structure to mitigate the widespread risks of AI sycophancy.
Artificial intelligence is transforming biology and medicine by accelerating the discovery of new drugs and proteins and making it easier to design and manipulate DNA, the building blocks of life. But as with most new technologies, there is a potential downside. The same AI tools could be used to develop dangerous new pathogens and toxins that bypass current security checks. In a new study from Microsoft, scientists employed a hacker-style test to demonstrate that AI-generated sequences could evade security software used by DNA manufacturers.
“We believe that the ongoing advancement of AI-assisted protein design holds great promise for tackling critical challenges in health and the life sciences, with the potential to deliver overwhelmingly positive impacts on people and society,” commented the researchers in their paper published in the journal Science. “As with other emerging technologies, however, it is also crucial to proactively identify and mitigate risks arising from novel capabilities.”
A team of engineers at the University of Massachusetts Amherst has announced the creation of an artificial neuron with electrical functions that closely mirror those of biological ones. Building on their previous groundbreaking work using protein nanowires synthesized from electricity-generating bacteria, the team’s discovery means that we could see immensely efficient computers built on biological principles which could interface directly with living cells.
“Our brain processes an enormous amount of data,” says Shuai Fu, a graduate student in electrical and computer engineering at UMass Amherst and lead author of the study published in Nature Communications. “But its power usage is very, very low, especially compared to the amount of electricity it takes to run a Large Language Model, like ChatGPT.”
The human body is over 100 times more electrically efficient than a computer’s electrical circuit. The human brain is composed of billions of neurons, specialized cells that send and receive electrical impulses all over the body. While it takes only about 20 watts for your brain to, say, write a story, a LLM might consume well over a megawatt of electricity to do the same task.