Tech & AI Unfiltered: The Signals That Matter, Not the Noise

Grego AI, a company that didn't exist in public yesterday, just stopped a twenty-seven-point-seven million dollar blockchain exploit using an AI. This wasn't just another bug bounty; it was the largest ever awarded for an AI-discovered vulnerability, and it signals a new front in cybersecurity. In our last episode, we talked about finding the real signal in the endless scroll of tech news. Today, that signal is a breakthrough in AI reasoning that the biggest labs have been missing. This week, the story isn't about one model getting slightly smarter. It’s about a structural change in how AI operates, how it makes money, and how it's breaking things—and fixing them—at a scale that is starting to look very different.

First, let's talk about how software itself is being made. According to a new analysis from Chamath Palihapitiya, AI now generates seventy-five percent of Google’s new code. Let me say that again. Seventy-five percent. At Microsoft, it’s up to thirty percent. This isn't about chatbots helping developers. This is a fundamental shift in production. Daily code commits from Anthropic's Claude on GitHub have surpassed one hundred and thirty-four thousand. That’s not a feature, that’s a workforce. And it’s driving insane economics. Anthropic’s own revenue reportedly grew from one billion to forty-four billion dollars in just seventeen months, largely on the back of these coding agents.

The risk, of course, is that these agents are still prone to catastrophic errors. We're hearing stories of AIs deleting live production environments or wiping entire company databases. The power is scaling faster than the guardrails. Next, a quiet, tectonic shift is happening in the architecture of the models themselves. Ben Pouladian's reporting this week highlights a stunning convergence. Three of the most advanced labs—NVIDIA, Physical Intelligence, and Thinking Machines—all shipped the same new kind of AI architecture within five weeks of each other. They’re moving from turn-based models to continuous, predictive systems.

In the old model, which is what you're used to, the AI waits for you to finish your prompt. It processes, it generates, and then it stops, waiting for the next turn. As one engineer put it, "Until the model finishes generating, its perception freezes." The new architecture doesn't wait. It operates in a continuous loop, perceiving and acting at the same time. NVIDIA’s DreamZero model "dreams" of future video frames and then acts on them. Thinking Machines’ new model can process audio, video, and text simultaneously in two-hundred-millisecond micro-turns. The bottleneck is no longer the model's intelligence.

The bottleneck is the bandwidth between us and it. This new architecture is an attempt to solve that. Meanwhile, this new power is creating geopolitical divides. At the Africa Forward Summit in Nairobi this month, the conversation was about a growing split. On one side, you have frontier AI companies like Anthropic holding closed-door meetings with theologians and philosophers, trying to embed complex moral reasoning directly into their models. They are designing for normative values. On the other side, you have governments in the Global South struggling with the basics of AI governance. The Science Diplomat reported that South Africa recently had to withdraw its entire national AI policy because it was riddled with fabricated citations, likely generated by an AI.

The credibility gap is widening. UN Secretary-General António Guterres was blunt, calling for AI to be shaped by "African data, African languages, African researchers and African leadership" before this gap becomes a chasm. And in a final, quick turn of events, we're seeing market pressures create strange alliances. Despite Elon Musk's very public criticism of Anthropic's safety-focused ideology, his company, xAI, is reportedly in collaboration talks with them. Why would two ideological rivals team up? Because a common enemy—in this case, OpenAI's market dominance—makes for very strange bedfellows. It’s a reminder that for all the talk of philosophy, market dynamics are a powerful force.

So let's go back to that twenty-seven-point-seven million dollar bug bounty. Let's go back to Grego AI. Because what happened there, and what’s happening at Microsoft, is the synthesis of all these trends. It’s the payoff. For years, the story in AI security was about models that could find simple bugs. They could spot a known vulnerability pattern, or flag a suspicious line of code. It was helpful, but not transformative. Human auditors, especially in high-stakes environments like blockchain protocols, were always superior because they could understand the entire system. They could reason about how different parts of a complex codebase interact over time.

This is what just changed. Grego AI didn't just build a better model. They built a reasoning system they call Deep Invariant Analysis. According to their CTO, Gregorio Maspero, "The frontier models from the leading labs all have a big reasoning limitation... But we did. We built our own reasoning architecture that fixes that." Instead of just looking at code, their system maps the entire codebase, models multi-layer interactions between different smart contracts, and then autonomously designs and refines attack paths. It thinks like a hacker, but at machine speed and scale. It found a critical vulnerability in a major protocol that top human auditors had missed.

And it didn't just find it—it proved it could be exploited, which is why the bounty was so high. This is not a one-off event. The same week, Microsoft published a deep dive on its new security system, codenamed MDASH. It’s the exact same pattern. MDASH isn't a single model. Microsoft says it orchestrates over one hundred specialized AI agents, from frontier models to smaller, distilled ones. This system of agents discovered sixteen new, zero-day vulnerabilities in core Windows components. It now sits atop the industry-standard CyberGym benchmark with a score of eighty-eight-point-four-five percent, a full five points ahead of any competitor.

And it did so with zero false positives. Microsoft’s conclusion is the key insight for this entire week. They wrote, "The durable advantage lies in the agentic system around the model rather than any single model itself." This is the deep dive. The story is no longer the model. It's the system. It's the agent. The architectural shift we talked about—from turn-based to continuous—is what makes these agentic systems possible. An AI that has to wait for you to type a prompt cannot proactively hunt for security flaws in a live network. It can’t generate code, see the error, and correct it in a continuous loop.

The old model experiences reality in a single thread. The new models are multi-threaded by nature. They are always on, always perceiving, always acting. They are becoming persistent entities within our digital infrastructure. This explains the explosion in AI-generated code at Google. It's not that a Google engineer is sitting there typing "write me a new microservice" one hundred thousand times a day. It's that agentic systems are now integrated into the development pipeline, continuously writing, testing, and committing code as part of the workflow. This is why Anthropic's revenue is exploding. They aren't just selling a chatbot.

They are selling autonomous digital workers that can be hired to perform tasks. And this is where the risk becomes so apparent. An agent that is powerful enough to find a vulnerability missed by every human expert is also powerful enough to cause damage no human could anticipate. The stories of AIs deleting production databases aren't just funny anecdotes. They are the logical consequence of deploying autonomous agents whose goals are not perfectly aligned with their environment. The difference between finding a twenty-seven million dollar bug and causing a twenty-seven million dollar outage is a very thin line of code.

So this week, the thread that connects everything—from a stealth startup in blockchain security, to Microsoft's new platform, to the frantic pace of code generation at Google, to the very architecture of the next generation of models—is the shift away from AI as a tool you use, and toward AI as an agent that acts. We are moving out of the era of prompt-and-response. We are entering the era of continuous computation, where autonomous systems are a permanent, active layer of our reality. The conversations at the Africa Forward Summit are not a sideshow; they are a warning that our social and political structures are organized around the old, turn-based world.

They are not ready for this one. The question is no longer what we will ask the AI to do. It's what the AI is doing right now.