Robots Talking

Imagine you are part of a massive group project to write a new book. To keep everyone’s privacy safe, no one is allowed to see each other’s notes. Instead, you all work on your own chapters and send a summary to a central editor who combines them into one masterpiece. This is essentially how Federated Learning works—a popular way to train artificial intelligence (AI) on our phones and laptops without ever looking at our private data 1, 2.

However, new research has uncovered a "silent saboteur" that could cripple this process. While we often worry about how LLMs (Large Language Models) handle our data, this study reveals that the very process of model learning itself is under a new kind of stealthy attack 3.

For years, experts believed that to "poison" an AI model—essentially feeding it bad information to make it fail—hackers had to work together. They thought attackers needed to control thousands of devices and have them communicate like a coordinated army to sneak past security 4-6. This was considered expensive, difficult to maintain, and easy to catch 5, 7.

The breakthrough in this new study is a framework called XFED. It proves that attackers don't need to talk to each other at all 3, 8. XFED is a "non-collusive" attack, meaning each hacked device acts independently, like a sleeper cell waiting to strike 3, 9, 10.

To stop hackers, current artificial intelligence systems use "Byzantine-robust" defenses. Think of these as a high-tech "vibe check" that looks at all incoming updates and throws out anything that looks like an outlier 4, 11.

XFED is brilliant because it learns how to pass this check by observing the group. Here is how it works in simple terms:

The findings are a wake-up call for the tech world. In tests across six different datasets, XFED successfully broke eight state-of-the-art defenses and outperformed six existing types of attacks 3, 20, 21. It was particularly effective at "breaking" systems that use model learning for everything from image recognition to predicting what you'll type next on a mobile keyboard 22-24.

The researchers found that XFED could degrade a model’s accuracy even when only a small fraction of devices were compromised 25, 26. This means that the security we thought we had in distributed AI training is much thinner than previously believed 3, 26.

The study doesn't just bring bad news; it offers a path forward. While many common defenses failed, one called FLTrust remained fairly consistent—though it requires a small "trusted" dataset to compare updates against, which can be its own challenge to collect fairly 20, 27, 28.

As we continue to integrate LLMs and artificial intelligence into our daily lives, this research highlights an urgent need for more robust defenses. The next time your phone gets a "smarter" update, remember that there is a silent battle happening behind the scenes to keep that model learning process safe from invisible poison 26.

Have you ever wondered why we, as humans, are so obsessed with games? From the strategic depth of Chess to the frantic tapping of Flappy Bird, we spend countless hours in digital and physical playgrounds. According to recent research, this isn't just about killing time—it’s actually a cornerstone of our General Intelligence.

Games are "structured microcosms" of the real world. When we play, we are actually practicing skills like resource management, social deduction, and physical navigation in a safe, fun environment. Now, researchers from institutions like MIT and Harvard are using this "Multiverse of Human Games" to see if artificial intelligence can finally keep up with us.

Evaluating how "smart" an AI really is has become a massive challenge. Traditional tests often focus on narrow tasks like solving a specific math problem or writing code. But being good at one thing doesn't mean a machine has the versatility of a human adult.

To bridge this gap, researchers built the AI GAMESTORE. This platform uses LLMs (Large Language Models) to automatically source and adapt popular games from the Apple App Store and Steam into standardized tests for machines. By having artificial intelligence play 100 different games—ranging from Angry Birds clones to complex puzzles—the researchers could measure its ability to learn and adapt just like a human would.

The researchers pitted seven of the world's most advanced LLMs (including frontier models like GPT-5.2 and Gemini 2.5 Pro) against 106 human players. The goal was simple: play the first two minutes of a new game and see who scores higher.

The results were a wake-up call for the tech world:

You might think a supercomputer could easily beat a human at a "casual" mobile game, but the AI GAMESTORE revealed three major "cognitive bottlenecks" where machines fail:

This research shows that while artificial intelligence is getting better at talking and coding, it still lacks the "cognitive versatility" of a typical human. The "Multiverse of Human Games" provides a way to track this progress through a "living" benchmark that can't be easily cheated or memorized.

The ultimate goal isn't just to build a better gamer. It’s to develop AI that can interact with the world as flexibly, safely, and intuitively as we do. Until then, it looks like your high score on the App Store is safe!

Decoding the Brain: How AI Models Learn to "See" Like Us

Have you ever wondered if the way an AI sees the world is anything like how you do? It's a fascinating question that researchers are constantly exploring, and new studies are bringing us closer to understanding the surprising similarities between advanced artificial intelligence models and the human brain.

A recent study delved deep into what factors actually make AI models develop representations of images that resemble those in our own brains. Far from being a simple imitation, this convergence offers insights into the universal principles of information processing that might be shared across all neural networks, both biological and artificial.

The AI That Learns to See: DINOv3

The researchers in this study used a cutting-edge artificial intelligence model called DINOv3, a self-supervised vision transformer, to investigate this question. Unlike some AI models that rely on vast amounts of human-labeled data, DINOv3 learns by figuring out patterns in images on its own.

To understand what makes DINOv3 "brain-like," the researchers systematically varied three key factors during its training:

To compare the AI models' "sight" to human vision, they used advanced brain imaging techniques:

They then measured the brain-model similarity using three metrics: overall representational similarity (encoding score), topographical organization (spatial score), and temporal dynamics (temporal score).

The Surprising Factors Shaping Brain-Like AI

The study revealed several critical insights into how AI comes to "see" the world like humans:

    ◦ Early in training, the AI models quickly aligned with the early representations of our sensory cortices (the parts of the brain that process basic visual input like lines and edges).

    ◦ However, aligning with the late and prefrontal representations of the brain required considerably more training data.

    ◦ This "developmental trajectory" in the AI model mirrors the biological development of the human brain, where basic sensory processing matures earlier than complex cognitive functions.

AI Models Mirroring Brain Development

Perhaps the most profound finding connects artificial intelligence development directly to human brain biology. The brain areas that the AI models aligned with last during their training were precisely those in the human brain known for:

These are the associative cortices, which are known to mature slowly over the first two decades of life in humans. This astonishing parallel suggests that the sequential way artificial intelligence models acquire representations might spontaneously model some of the developmental trajectories of brain functions.

Broader Implications for AI and Neuroscience

This research offers a powerful framework for understanding how the human brain comes to represent its visual world by showing how machines can learn to "see" like us. It also contributes to the long-standing philosophical debate in cognitive science about "nativism versus empiricism," demonstrating how both inherent architectural potential and real-world experience interact in the development of cognition in AI.

While this study focused on vision models, the principles of how AI learns to align with brain activity could potentially extend to other complex artificial intelligence systems, including Large Language Models (LLMs), as researchers are also exploring how high-level visual representations in the human brain align with LLMs and how multimodal transformers can transfer across language and vision.

Ultimately, this convergence between AI and neuroscience promises to unlock deeper secrets about both biological intelligence and the future potential of artificial intelligence.

Decoding AI's Footprint: What Really Powers Your LLM Interactions?

Artificial intelligence is rapidly changing our world, from powerful image generators to advanced chatbots. As AI – particularly large language models (LLMs) – becomes an everyday tool for billions, a crucial question arises: what's the environmental cost of all this innovation? While much attention has historically focused on the energy-intensive process of training these massive LLMs, new research from Google sheds light on an equally important, and often underestimated, aspect: the environmental footprint of AI inference at scale, which is when these models are actually used to generate responses.

This groundbreaking study proposes a comprehensive method to measure the energy, carbon emissions, and water consumption of AI inference in a real-world production environment. And the findings are quite illuminating!

The Full Story: Beyond Just the AI Chip

One of the most significant insights from Google's research is that previous, narrower measurement approaches often dramatically underestimated the true environmental impact. Why? Because they typically focused only on the active AI accelerators. Google's "Comprehensive Approach" looks at the full stack of AI serving infrastructure, revealing a more complete picture of what contributes to a single LLM prompt's footprint.

Here are the key factors driving the environmental footprint of AI inference at scale:

Together, these four components illustrate that understanding AI's impact requires looking beyond just the core processing unit. For instance, the comprehensive approach showed a total energy consumption that was 2.4 times greater than a narrower approach.

Beyond Energy: Carbon and Water

The energy consumption outlined above then translates directly into other environmental impacts:

Surprisingly Low, Yet Critically Important

So, what's the actual footprint of a single LLM interaction? For a median Gemini Apps text prompt, Google found it consumes 0.24 Wh of energy, generates 0.03 gCO2e, and uses 0.26 mL of water.

To put that into perspective:

These figures are significantly lower than many previous public estimates, often by one or two orders of magnitude. This difference comes from Google's in-situ measurement, the efficiency of their production environment (e.g., efficient batching of prompts), and continuous optimization efforts.

The Path to an Even Greener AI

Despite these already low figures, Google's research emphasizes that significant efficiency gains are possible and ongoing across the entire AI serving stack. Over just one year, Google achieved a 33x reduction in per-prompt energy consumption and a 44x reduction in carbon footprint for the median Gemini Apps text prompt.

These dramatic improvements are driven by a combination of factors:

The Future of Responsible AI

The sheer scale of AI adoption means that even small per-prompt impacts multiply into significant overall footprints. This research highlights that a standardized, comprehensive measurement boundary for AI environmental metrics is not just good for transparency; it's essential for accurately comparing models, setting targets, and incentivizing continuous efficiency gains across the entire artificial intelligence serving stack. As AI continues to advance, a sustained focus on environmental efficiency will be crucial for a sustainable future.