Google Says AI Isn't Building Itself Yet—Even Though It Writes 75% of the Code

The 75 Percent Paradox

Google recently dropped a stat that should have sent shockwaves through every development studio in the world: more than 75% of the company's new code is now generated by AI. On paper, that sounds like the beginning of the singularity. If the machines are writing three-quarters of the software, we should be living in a world of lightning-fast innovation and shrinking product cycles.

But then Google DeepMind released its first internal audit of AI-driven research, and the reality check hit hard. Despite all that synthetic code, they found no measurable speedup in AI progress. The machines are typing faster, but they aren't thinking faster. As a founder, this is the most important signal we've seen in the last year. It confirms what many of us suspected: we are confusing activity with progress.

The Illusion of Velocity

We’ve been told that AI is a force multiplier. In some ways, it is. If you need a boilerplate React component or a Python script to scrape a public database, an LLM will give it to you in seconds. That saves you from a trip to Stack Overflow, sure. But Google’s internal audit suggests that these efficiency gains are being swallowed by the complexity that comes with them.

When 75% of your codebase is generated by a model, the human job shifts from creation to verification. This is a subtle but dangerous pivot. Writing code is actually the easy part of building software; understanding the logic and ensuring it doesn't break the rest of the system is where the real work happens. If you spend ten minutes generating a thousand lines of code and then three hours debugging why it fails in production, you haven't actually moved faster. You've just traded one type of labor for another.

Why Research Stalled

The DeepMind report is particularly damning because it looks at the high end of the spectrum—research. This is where the next breakthroughs in architecture and efficiency are supposed to happen. If AI were truly "building itself," we would see a recursive loop where the AI discovers a new optimization, implements it, and then uses that better version to find the next one.

That isn't happening. The report indicates that while AI can assist in the grunt work of data labeling or running simulations, it still lacks the intuition to navigate the "unsolved" problems. It can reorganize what we already know, but it isn't manifesting new knowledge. For builders, this means the competitive moat isn't going to be who has the best AI coding assistant. The moat is still the human ability to identify a problem that hasn't been solved yet.

The High Cost of Cheap Code

There is a hidden cost to this 75% figure that Google isn't talking about: technical debt. When a human writes code, there is (ideally) a mental map of why certain trade-offs were made. When an AI generates code, it’s a statistical prediction of what the next token should be. It doesn't have a "why."

By flooding repositories with AI-generated scripts, we are effectively creating a massive pile of legacy code that no single human fully understands. If Google, with the brightest minds in the world and the best hardware, can't find a way to turn this volume of code into measurable progress, what makes a small startup think they can? We are building on sand, and the tide of maintenance is coming in.

The Founder's Perspective: Focus on the 25%

If you're building in the crypto or AI space right now, the temptation is to automate everything. You want to ship daily. You want to use every API and agentic tool available to minimize your headcount. But the DeepMind audit suggests you should be doing the opposite. You should be obsessing over the 25% of code that isn't generated by AI.

The value is in the logic that the LLM can't guess. It’s in the novel consensus mechanism, the unique privacy layer, or the specific way you integrate a decentralized database. If your entire product can be written by an AI, then your product is a commodity, and a trillion-dollar incumbent will eventually swallow you whole. Real innovation is still a slow, painful, human-centric process.

The Myth of the Self-Building Machine

We need to stop waiting for the AI to take the wheel. The narrative that AI is building itself is a marketing play to drive stock prices and VC valuations. The internal data from the world's leading AI lab tells a different story. It tells a story of a tool that is excellent at mimicking the past but currently incapable of inventing the future.

Google is writing more code than ever, yet they are finding that the distance between "working code" and "scientific breakthrough" remains unchanged. The machine can copy the syntax, but it can't duplicate the spark of discovery. As builders, we should use these tools to clear the brush, but we shouldn't expect them to build the house for us.

Takeaway for Builders

Don't confuse a high volume of output with high-value progress. Google’s data proves that even if you automate 75% of your workflow, you might still be standing still. Focus on the hard problems that require human judgment, because that’s the only place where true value is being created. AI is a faster typewriter, not a better author. If you want to win, you have to be the one writing the story, not just generating the text.

Read the original at Decrypt →