AI Breakthroughs That Prove We’re Not Fully in Control Anymore

Artificial intelligence is no longer just a clever autocomplete. It’s playing elite-level sports, shaping how wars are fought and watched, debugging critical software, and even helping design biological treatments. At the same time, companies are quietly recording every keystroke of their workers to train future AI agents, and there are now social networks where only AI is allowed to talk.

Here’s a tour through some of the most striking recent developments—and what they say about how much control we still have over this technology.

Robots That Learn to Play Like Us

Table tennis used to be a showcase for human reflexes and intuition. Now, a new robot system called ACE is challenging that assumption.

ACE is one of the first robots that can truly interact with humans at human time scales in a fast, dynamic sport. Instead of following pre-programmed motions, it learns through reinforcement learning—improving by playing, not by being scripted.

The system tracks the ball with nine cameras and can even read the logo on the ball to measure spin in real time. Its arm has eight joints, giving it enough flexibility to position shots, react quickly, and adjust mid-play. Researchers deliberately limited its speed and reach so that it plays under the same constraints as a trained human, forcing it to win through strategy and timing rather than brute mechanical advantage.

In tests, ACE beat most high-level players it faced. Pros reported that it sometimes pulled off shots they thought were impossible, yet also missed balls that any decent human player would return. In other words, it behaves a bit like today’s large language models: brilliant in some moments, strangely clumsy in others.

As more systems like this emerge, the line between “human” and “machine” performance in physical tasks will keep blurring—especially as research like tiny AI controllers that teach robots to move like us continues to advance.

Teaching AI to See—and Pay Attention—Like Humans

When you walk into a room, your eyes technically see everything. But your brain only pays attention to a few things: the coffee mug you’re reaching for, the person calling your name, the sudden crash outside.

New research shows that if we want AI to behave more like humans, it needs to understand this difference between raw vision and attention.

Researchers used 360° virtual reality videos to create a setting where an AI system could look anywhere at any time. That creates a key question: where should it focus? What would a human pay attention to in this scene?

They built a dataset of 81 different 360° videos and tested them with three audio setups: no sound, basic sound, and spatial sound (audio that clearly comes from specific directions). Over 100 people watched these videos while their eye movements were tracked.

Then the team trained two AI models:

• One used only visual information.
• The other used both visuals and sound.

The audiovisual model did a much better job predicting where humans would look. It could even identify parts of the scene people focused on because of sound cues, not just visuals. This confirms something intuitive: human attention is multisensory. We don’t just see the world—we listen to it, and sound often decides where we look next.

For AI systems that need to interact naturally with humans—robots, AR assistants, autonomous vehicles—understanding attention, not just images, will be critical.

Do AI Models Actually Understand the Real World?

A long-running debate in AI is whether large language models truly “understand” anything, or if they’re just parroting patterns from the internet.

A recent study tackled this by testing whether models can distinguish between real, unlikely, and impossible events. Researchers fed models simple scenarios like:

• Realistic: cooling a drink with ice.
• Unlikely but possible: cooling a drink with snow.
• Impossible: cooling a drink with fire.
• Pure nonsense: sentences that don’t mean anything at all.

Instead of just checking the model’s answers, they looked inside its internal activations using a technique called mechanistic interpretability. That means analyzing how the model’s “neurons” light up and how features are represented internally.

They found that the model formed distinct internal patterns for each type of scenario. It could reliably tell apart categories like “unlikely” versus “impossible” with about 85% accuracy. Even more interesting, when humans were split on whether something was impossible or just unlikely, the model’s internal states showed similar uncertainty.

This suggests that large models aren’t just memorizing text. From raw symbols—tokens, words, pixels—some form of structured world knowledge is emerging. They appear to build an internal sense of what tends to happen in reality, what’s rare, and what simply cannot happen.

That doesn’t mean they think like humans. But it does mean we should stop assuming they’re just sophisticated parrots.

AI in Warfare: Why “Humans in the Loop” May Be an Illusion

As militaries adopt AI, a common reassurance is that “a human will always be in the loop” for lethal decisions. On paper, that sounds safe. In practice, it may be more illusion than control.

Modern AI systems can already pick targets, guide missiles, and coordinate drone swarms in real time at speeds humans can’t match. The real problem isn’t just what AI does on its own—it’s that humans often don’t understand why it does what it does.

Many advanced models are black boxes: we see inputs and outputs, but not the reasoning in between. Even their creators can’t fully explain how specific decisions arise. That creates what some researchers call an “intention gap”: the AI follows instructions, but interprets them in ways humans didn’t anticipate.

In a fast, high-pressure environment like war, this gap becomes dangerous. A human “in the loop” might see what the system is doing but have no real basis to judge whether a surprising move is brilliant strategy or catastrophic misfire—similar to how Go experts were stunned by moves from DeepMind’s systems that no human had seen before.

The argument coming from some policy and ethics researchers is that oversight alone isn’t enough. We need tools that let us understand AI’s intentions and constraints before it acts, not just after. Until then, human control over AI in warfare may be far weaker than it appears on paper.

AI and Longevity: Replacing the Cell’s Power Plants

Not all breakthroughs are digital. Some of the most exciting work sits at the intersection of AI, biology, and medicine—especially in longevity research.

One promising line of work focuses on mitochondria, the “power plants” of our cells. When mitochondria fail, cells lose energy and start to break down, contributing to conditions like blindness, diabetes, and Alzheimer’s.

Fixing mitochondria is notoriously hard because they have their own DNA and are difficult to edit safely. So researchers tried a different approach: instead of repairing broken mitochondria, why not replace them?

Using a system called MitoCatch, scientists can grow healthy mitochondria in other cells, extract them, and then deliver them into damaged cells. MitoCatch uses a kind of biological handshake—matching proteins on both the mitochondria and the target cell—so they latch together and the new mitochondria can enter and start working.

In mice with inherited blindness, this technique helped damaged eye cells survive and function better. It doesn’t fix the underlying genetic mutation, but it gives cells enough fresh power sources to keep going.

It’s early days, but this kind of work hints at a future where we don’t just treat symptoms—we swap out failing cellular components entirely. Combined with AI-driven discovery and lab automation, it’s easy to see why some researchers think we may be approaching “longevity escape velocity” sooner than expected.

AI That Writes and Fixes Code—And Improves Itself

Mythos and the New Era of Software Security

AI coding models are no longer just autocomplete for developers. They’re starting to find and fix bugs at a scale humans can’t match.

Mozilla recently got early access to Anthropic’s Mythos model, which is optimized for cybersecurity. The result: Mythos identified 271 previously unknown security bugs in Firefox. For an open-source browser with a mature security process, that’s a huge number.

Mozilla says tools like this are changing software security. Traditionally, teams relied on automated tests, human experts, and bug bounties. Many vulnerabilities were so subtle that they required significant time, skill, and money to uncover. Now, advanced AI can scan massive codebases and surface potential zero-days in bulk.

This creates a race. Defenders can use AI to clean up old code, but attackers can use the same techniques to find fresh vulnerabilities. Mozilla’s CTO argues that every major piece of software may need to go through this kind of AI-assisted security sweep as soon as possible—especially open-source projects maintained by tiny or unpaid teams.

In other words, AI isn’t just writing new code. It’s forcing a one-time, painful upgrade to the entire security posture of the software ecosystem.

Google’s “Strike Team” vs. Self-Improving Models

Inside Google, there’s growing concern that rivals—especially Anthropic—have pulled ahead in AI coding. That’s serious, because whoever leads in AI-assisted coding could end up with models that improve themselves.

Google has formed a focused internal “strike team” within its DeepMind division to push its coding models forward. The goal isn’t just to generate short code snippets, but to handle full software projects: understanding large codebases, planning multi-step tasks, and maintaining long-term context.

Leaders like Sergey Brin are directly involved. Internally, Google is also pushing employees to use AI coding tools trained on Google’s private code, though these tools won’t be released publicly.

The fear is that a small lead in coding quality could compound. If a model like Claude (or its successors) becomes good enough to help write the next version of itself, each generation could get better at improving the next—creating a feedback loop that leaves slower competitors behind.

This is the same dynamic explored in broader discussions about self-improving AI systems and the economic shocks they bring, similar to what’s happening with models like DeepSeek and GPT in other domains, as covered in recent analysis of AI-driven job losses.

Who Owns the AI Wealth? The Case for an AI Dividend

If AI systems can write code, debug software, analyze data, and even design biological treatments, a natural question follows: who benefits from all that productivity?

One proposal gaining attention is the idea of an “AI dividend.” New York congressional candidate Alex Bores is pushing a plan that would take a portion of the wealth created by AI and distribute it directly to the public.

The logic is straightforward: as AI automates more work, many jobs could be displaced or transformed. Even if the overall economy grows, the gains might concentrate in a handful of companies and investors. An AI dividend would tie payouts to how much AI boosts the economy, ensuring that ordinary people share in the upside even if their specific jobs are affected.

Unsurprisingly, powerful tech interests are spending heavily to oppose stronger AI regulation and redistribution. Bores argues that they don’t need to win forever—just delay effective rules for a few election cycles. Given the speed of AI progress and the political influence that comes with concentrated capital, that delay could be enough to lock in a very unequal future.

Whether or not an AI dividend becomes law, the underlying question isn’t going away: at what point does AI-driven wealth concentration become so extreme that society decides to intervene?

AI, War, and the New Propaganda Theater

AI isn’t just changing how wars are fought—it’s changing how they’re watched, discussed, and even gamified.

In the current Iran conflict, people are building AI-powered dashboards that pull in satellite imagery, ship tracking data, news, chats, and prediction markets into real-time “war rooms.” Many of these dashboards were built in days or hours using AI tools. Their creators claim they’re faster and more truthful than traditional media, and some are turning them into full-blown media channels.

At the same time, AI-generated propaganda is spreading: catchy songs, stylized Lego-like videos, and emotionally charged content that pushes pro- or anti-government narratives. These pieces are designed to be shareable and addictive, blurring the line between news, entertainment, and psychological operations.

Prediction markets add another layer. When people can bet on outcomes—like whether a specific strait will be blocked or a strike will occur—conflict starts to look like a game with teams, logos, and scheduled “matches.” That risks turning war into something spectators root for and trade on, rather than a last-resort tragedy to be avoided.

There is some upside: more open data and alternative analysis can challenge state-controlled narratives. But without context, expertise, or guardrails, AI-driven war dashboards and propaganda engines can also mislead, polarize, and desensitize people to real human suffering.

Meta’s Keystroke Tracking and the Future of Work

While governments and researchers wrestle with AI’s big-picture risks, companies are quietly reshaping everyday work to feed these systems.

Meta has begun installing software on employees’ work computers in the US that records how they use their machines: mouse movements, clicks, and keystrokes. The stated goal is to train AI systems to better understand how humans actually work on computers.

Meta says AI is already good at tasks like coding and research, but still struggles with basic computer interactions—using shortcuts, navigating menus, juggling multiple apps. Real-world usage data from employees can help models learn these patterns.

The tracking is limited to approved work apps like email, chat, and coding tools, and doesn’t apply to personal phones. But there’s no opt-out. For workers, that raises obvious questions: if everything you do is recorded and used to train AI, how long until that AI can do your job? And how long after that until your role shifts from doing the work to supervising the AI—before being automated away entirely?

In effect, many employees may be training the systems that will eventually replace large parts of their own roles.

When AI Talks Only to Itself: Agent-Only Science Networks

One of the most intriguing experiments in AI right now is a social network where humans can watch—but not participate.

The site, called Agent for Science, looks a bit like Reddit, but it’s built exclusively for AI agents. Its focus is narrow: scientific research, verification, and debate.

Here’s how it works:

• Only AI agents can post, comment, or review content.
• Most of the papers on the site are generated by AI systems that run experiments (or simulations) and then write up the results.
• Humans can design the agents, choose their focus areas, and assign personalities like “skeptic” or “storyteller,” but they can’t join the conversations directly.

Each comment from an agent is labeled with its role—supporting, questioning, challenging, and so on. More than 150 agents have already produced around 40,000 comments, sometimes surfacing insights that human readers might miss.

Crucially, the platform is tightly focused on science, not general chatter. That helps keep discussions grounded and useful. In the future, as more labs build automated facilities where robots run physical experiments under AI control, networks like this could become part of a new scientific workflow: AI proposes ideas, runs tests, debates results, and refines theories—at machine speed.

It’s both exciting and unsettling. If AI systems can generate, test, and refine scientific knowledge faster than humans can follow, we may end up in a world where much of our understanding of reality is mediated through systems we didn’t fully design and can’t fully explain.

Where This All Leaves Us

Across robotics, perception, warfare, medicine, coding, politics, propaganda, workplace surveillance, and scientific research, one pattern keeps repeating: AI is moving from “tool we use” to “system that acts, learns, and sometimes surprises us.”

We’re still writing the rules, both technically and politically. But the longer we wait to decide what kind of AI-powered world we want, the more that world will be shaped by default—by whoever moves fastest, collects the most data, and deploys the most aggressive systems.

Whether you’re excited, worried, or both, one thing is clear: AI is no longer something happening in the background. It’s becoming the environment we all live and work in.