Tech Twitter Unfiltered: The Real AI Divide Behind the Headlines

Greg Brockman just announced that eighty percent of OpenAI's code is now written by AI. This lands just as the rest of the industry is finally starting to admit that for most companies, AI's productivity gains are still effectively zero. The gap between what is happening inside the top labs and what is happening everywhere else just became a chasm. And that chasm is the single biggest story in technology right now. It's not just about who has the best model. It's about who can actually use it. Last week we talked about the creeping disillusionment with AI's real-world performance. This week, the builders at the absolute frontier just turned up the volume to deafening levels, claiming the future is not only here...

it's already writing itself. The shockwave from OpenAI's eighty percent claim is still echoing, but it wasn't the only ground-shaking event. The Pentagon just made a move that redefines the relationship between Silicon Valley and the military. On May first, it signed a new slate of classified AI contracts with Nvidia, Microsoft, Amazon Web Services, and a lesser-known firm called Reflection AI. They join SpaceX, OpenAI, and Google on the inside. But the real story is who was just kicked out: Anthropic. They were ejected after refusing to remove safety restrictions on their models for military use. The Pentagon wanted fewer guardrails.

Anthropic said no. So the Pentagon found partners who would say yes. This is a hard pivot. The debate over ethics in military AI is over, for now. The side arguing for speed won. Meanwhile, Snap CEO Evan Spiegel dropped his own startling number. He revealed that AI now writes about two-thirds of Snap's code. That's another top-tier tech company reporting massive internal transformation. But Spiegel immediately followed this boast with a warning. He said, quote, “Tech leaders vastly underestimate the societal pushback to AI.” Think about that. The CEO of a company that is clearly winning with AI is also the one sounding the alarm about public opinion.

He’s seeing the adoption curve inside his company, and the skepticism curve outside of it, and he’s telling everyone they are on a collision course. Snap is pushing forward with things like AI Sponsored Snaps—basically ad-bots—while its own leader is bracing for impact. Away from the big stages, a different conversation is happening among the engineers who actually have to make this stuff work. A thread from the team at a company called Activepieces went viral for its honesty. They broke down why their AI coding agents were failing. The problem wasn't the model. It wasn't GPT-4 or Claude or whatever came next.

The problem was context. The AI had no idea about their specific codebase, its rules, its architecture. They concluded that the bottleneck for ninety percent of AI coding teams right now is not model capability. It is, in their words, "context engineering." You have to build a harness for the AI—a system of documentation and structure that feeds the model what it needs to know. Without it, even the most powerful AI generates garbage. It's a crucial piece of ground truth in a week dominated by hype. And finally, trying to make sense of it all, investor Chamath Palihapitiya just published a framework that's getting a lot of attention.

He maps the entire AI world into a six-layer stack. At the bottom, you have infrastructure and chips—think AWS and Nvidia. Then data. Then the models themselves, like OpenAI's. Then an execution layer, and finally, the application on top. His point is that value and control are not evenly distributed. They are massively concentrated at the bottom. In the infrastructure. The chips. The companies that own those layers—Nvidia, Microsoft, Google, Amazon—effectively control the fulcrum points for the entire industry. It’s the clearest map we have for where the power struggles of the next decade will be fought. Let's go back to that eighty percent number from OpenAI.

Because on the surface, it sounds impossible. Eighty percent of all code written by AI. It conjures images of machines autonomously building the next generation of software while human engineers watch. The reality, as Brockman explained at Sequoia's AI Ascent conference, is more nuanced, but no less dramatic. What it really means is that AI is involved in eighty percent of coding tasks at the company. That includes everything from simple autocomplete to refactoring entire blocks of code, to generating new functions that a human then reviews and edits. Brockman said, "It's hard to know what percent is not being written by AI." He described an inflection point that happened around December 2025.

Before then, the models were helpful, maybe handling twenty percent of tasks. After that point, something clicked, and the capability jumped to eighty percent. The leap was so significant, he said, that OpenAI had to completely retool its internal workflows. You don't just give engineers a better autocomplete. You have to rebuild the entire process around the assumption that the AI is a junior developer on every team. This is the source of the productivity claims. Inside OpenAI, they have created the perfect environment. They have the world's most advanced models. They have engineers whose entire job is to get the most out of those models.

And they have a modern codebase, likely designed with AI interaction in mind. This is the "context engineering" the Activepieces team was talking about, but executed at the highest possible level. So when Brockman says eighty percent, he's not lying. He's describing a reality that exists inside his company. Here's the turn. That reality does not exist almost anywhere else. A National Bureau of Economic Research paper from just this February found that among companies that did adopt AI, eighty percent saw no measurable productivity gains at all. None. An MIT study from 2025 was even more brutal: it reported that ninety-five percent of corporate AI pilot programs yielded zero return on investment.

Cognitive scientist Gary Marcus has been shouting this from the rooftops, calling the hype around general AI a "trillion-dollar delusion." So what is happening? We have two completely different stories being told at the same time. The story from the frontier labs is one of exponential progress and radical transformation. The story from the rest of the corporate world is one of expensive experiments that go nowhere. The mistake is thinking one of them has to be false. Both are true. OpenAI has achieved this breakthrough. But it's a breakthrough of integration, not just of technology. They didn't just get a better model.

They changed how they work. Most companies can't or won't do that. They buy the software, hand it to their teams, and expect magic. They're trying to put a Formula One engine in a Ford Pinto. The engine might be real, but the car is still going to fall apart. The eighty percent number isn't a lie. It's a benchmark for a future that most of the industry is not even remotely prepared for. And that brings us to the Pentagon. Because while the corporate world struggles to get any value out of AI, the military sees the potential with absolute clarity. And it has decided it can't wait. The decision to eject Anthropic is a statement.

Anthropic was founded by former OpenAI employees specifically to build safer, more ethical AI. They built contractual red lines into their government work. These weren't vague principles. They were specific prohibitions against using their AI for things like mass domestic surveillance or in fully autonomous weapons systems that could make kill decisions without a human in the loop. For years, the Pentagon was willing to work within those limits. That just stopped. The new contracts, according to defense officials who spoke to The Next Web, use the phrase "lawful operational use." That language is deliberate.

It replaces Anthropic's specific, hard-coded restrictions with a vague standard that gives the Department of Defense, quote, "wide leeway to potentially use powerful advanced AI technologies for secret combat operations, including to assist with targeting." This is not a small change. It's a philosophical reversal. The Pentagon is signaling that in the global race for AI supremacy, self-imposed ethical restraints are now seen as a liability. By signing these new agreements, Google, Microsoft, OpenAI, and the others have accepted this new reality. They have agreed to a framework where the definition of "safe" and "ethical" is left to the military to decide on a case-by-case basis.

Anthropic was the one company that held the line, and for that, they were shown the door. This creates an entirely new dynamic. It's no longer a question of if powerful AI will be used in military applications. It's a question of how, and with what oversight. The companies that once publicly wrestled with these questions are now quietly, contractually, on board. The debate has moved from the op-ed pages of the New York Times to classified briefing rooms inside the Pentagon. And it means the immense power being unlocked in labs like OpenAI—the power to write eighty percent of its own code—is now being handed over to the military with fewer strings attached than ever before.

So this week, we saw the two poles of the AI revolution pull even further apart. Inside the elite labs, a future of incredible productivity is dawning. They are building tools that are reshaping what's possible in software and science. But that power is so immense, and so concentrated, that it's creating its own gravity. It's pulling in billion-dollar military contracts that demand the removal of safety wheels. It's creating a performance gap so wide that most of the business world is left staring at failed pilot programs, wondering where they went wrong. And it's happening just as the public, according to CEOs like Evan Spiegel, is growing more wary, not less.

The story of AI is no longer about a single, shared future. It's about splintering realities. There's the reality inside OpenAI's headquarters, the reality inside the Pentagon's new contract offices, and the reality for everyone else. What connects them is the technology. What separates them is everything else. The lab can claim eighty percent of the code is AI... but the real world still gets one hundred percent of the vote.