Tech Twitter Unfiltered: The Real Stories Shaping AI’s Fiercest Rivalries

Elon Musk’s xAI secretly trained its Grok models on data from its rival, Anthropic. This wasn't an accident — it was a months-long operation, continuing even after their official access was revoked in January. Last week, we talked about AI going on the offensive in state-sponsored cyber warfare. Today, the battle lines are being drawn between the AI labs themselves. The Information reports that xAI engineers engaged in what's called "model distillation." They fed prompts to Anthropic's Claude, took the high-quality outputs, and used them to teach their own, smaller model.

When Anthropic cut off their official access, they didn't stop. They just got sneakier. They used personal accounts. They routed requests through intermediary services like Blackbox AI. This wasn't a casual lookup. This was an industrial-scale operation. Anthropic itself identified over twenty-four thousand fraudulent accounts making sixteen million requests as part of these attacks. Musk’s defense? He calls it "industry standard practice." This, from the man currently suing OpenAI for its own lack of transparency. The hypocrisy is stunning.

The fight for AI supremacy is getting dirty. And while the giants are stealing from each other, the ground beneath them is shifting. Google Research just dropped a bombshell called TurboQuant. It’s an algorithm that compresses AI memory usage by six times. Without retraining the model. The result? Up to eight times speedups on NVIDIA's top-of-the-line H100 GPUs. A key-value cache that used to eat thirty-one gigabytes of memory now takes up just four. This is a software fix for a hardware problem, and it changes the economics of AI inference overnight.

Cloudflare CEO Matthew Prince is calling it Google’s “DeepSeek moment,” a breakthrough that could rewrite the rules of who can afford to run powerful AI. But all this AI activity — the training, the stealing, the inference — it's having a consequence. A huge one. The very fabric of the internet is changing. The Ladybird browser project just took a radical step. On June seventh, they banned all public code contributions. All of them. Why? Because the project was being flooded by AI-generated code. Cheap, low-quality, sometimes malicious patches that were overwhelming the human maintainers.

The trust model of open source is breaking. And it’s not just code. It’s the whole web. New data from Cloudflare shows that bots now account for 57.4 percent of ALL web traffic. And the biggest chunk of that — 50.6 percent of the entire internet — is just AI training crawlers. The primary audience for the internet is no longer human. It's machines. Machines that don't click ads, don't subscribe, and don't pay publishers for the content they are consuming at a planetary scale. This isn't a future problem. It's happening right now.

Let's unpack this. Because these three stories aren't separate. They are one story. The story of an industry so consumed with itself, it’s starting to eat its own tail. And in the process, it's devouring the open, collaborative systems that built it. First, go back to xAI and Anthropic. The technical term is distillation. The real term is theft. But the most important part isn't the act itself. It's the context. While xAI was siphoning data from Claude, Elon Musk's other company, SpaceX, was renting 220,000 GPUs to Anthropic for that same Claude model's training.

It gets worse. Google, Anthropic's biggest backer, is paying SpaceX 920 million dollars a month for AI compute. So you have Musk's rocket company providing the picks and shovels for a gold rush, while his AI company tries to steal the gold from the miners using those very tools. Musk’s claim that this is "industry standard" isn't just a defense. It's a mission statement. It declares a new reality where the only rule is that there are no rules. Where outputs are fair game, terms of service are suggestions, and the goal isn't just to compete, but to consume your competitors' work as fuel for your own.

This isn't collaboration. It’s a form of technological cannibalism. The models are learning from other models, in a closed loop that is increasingly detached from original human data and, apparently, human ethics. This is the high-stakes, zero-sum game being played in the C-suites. Now, layer on the second development. Google's TurboQuant. On the surface, this looks like a purely positive development. Making AI cheaper and faster. A six-fold reduction in memory is a massive leap. It means you can run bigger, more powerful models on the same hardware.

It means a startup could potentially run an inference task that used to require a whole server rack on a single GPU. This is democratizing. This is progress. Right? Not so fast. Remember the Jevons Paradox. It's an economic theory from the 1860s that says when you make a resource more efficient to use, you don't use less of it. You use WAY more. You find new applications. You drive demand through the roof. When coal engines became more efficient, Britain burned more coal, not less. A Morgan Stanley analyst put it perfectly: "TurboQuant leads to more intense computing rather than dimming demand." This isn't a threat to NVIDIA.

It's rocket fuel. Every company that was holding back on a new AI feature because the inference cost was too high just got a green light. Every developer with a crazy idea just saw their budget multiply. TurboQuant doesn't solve the AI energy and resource problem. It supercharges it. It makes the race to build and deploy AI even more frantic. It pours gasoline on the fire that's already raging at the top of the industry. The fire that leads to companies like xAI deciding that "industry standard" means taking whatever you can get your hands on.

Efficiency gains aren't leading to a more sustainable, thoughtful AI ecosystem. They are enabling a bigger, faster, hungrier one. And that brings us to the collateral damage. The Ladybird browser and the death of open trust. For decades, open-source software ran on a simple, powerful idea: many hands make light work. A global community of volunteers, contributing code, finding bugs, building public goods together. It was based on trust. The assumption was that most contributors were acting in good faith. Generative AI just shattered that assumption.

It made the cost of creating plausible-looking but ultimately garbage code drop to zero. Projects like Ladybird and curl—a foundational piece of internet infrastructure—are now drowning in it. They're getting fake bug reports from bounty hunters using AI to generate noise. They're getting code submissions that look fine on the surface but are inefficient, insecure, or just plain wrong. The cost of reviewing these cheap contributions is higher than the value they provide. So they're doing the only logical thing. They're shutting the door.

Andreas Kling, the founder of Ladybird, said it best: "Only developers directly responsible for the code’s real-world consequences should be the ones introducing it." Think about what that means. It's the end of the drive-by contribution. The end of the global volunteer army. It’s a retreat into a high-trust, closed-circle model of development. It’s a necessary survival tactic. But it’s also a tragedy. The very model that built the modern internet is now too vulnerable to exist on the modern internet. This is the direct consequence of the first two stories.

The relentless, scaled-up AI development, fueled by stolen data and efficiency hacks, is creating a digital environment that is fundamentally hostile to human-scale collaboration. The bots aren't just traffic. They are a denial-of-service attack on human systems of trust. When 57 percent of your users are machines, you don't have a human-centric platform anymore. You have a machine-feed. Publishers are realizing this. Google and Microsoft just rushed out opt-out toggles for their AI features, a desperate attempt to placate creators who are watching their work get scraped, summarized, and served up without credit or compensation.

So what does this week mean? It means the AI industry has entered a new phase. A frantic, inwardly-focused, and self-devouring cycle. The competition is no longer just about building the best model. It's about consuming the competition's output, optimizing resource consumption not for sustainability but for greater intensity, and in the process, overwhelming the digital commons we all rely on. The walls are going up. The open source community is closing its doors to strangers. Publishers are trying to fend off the scrapers.

The internet's promise was one of connection, of open access and shared creation. That was the sales pitch. But the underlying infrastructure is now being repurposed. It’s becoming a substrate for a machine-to-machine economy that is largely invisible, and totally unaccountable, to us. This week’s events set up a fundamental conflict. The conflict between the open, human-centric web and the closed, ruthlessly efficient world of AI development. And right now, the machines are winning. The internet was built for humans.

The machines just served us an eviction notice.