A Google Engineer Admitted Her Team’s Year of Work Got Rebuilt in One Hour (And It’s Not Even Funny)
Why AI is becoming the next purpose technology like electricity, what cognitive offloading means for your job, and the question you need to answer in the next 12 months
On January 3rd, a senior Google engineer confessed something that shook Silicon Valley.
Jaana Dogan, head of the Gemini API team, publicly acknowledged that a competing tool reproduced in one hour what her team spent an entire year building. Over 5 million views in a few hours. And what struck me was her opening line: “I’m not joking. This isn’t even funny.”
This isn’t just a story about programmers anymore.
What’s happening right now goes far beyond the world of software development. AI is becoming what many economists cite a general-purpose technology, and it will affect absolutely everyone, including you and me.
The Fighter Pilot Problem
To understand what’s happening, I need to talk about a concept I discovered in fighter jet aviation theory: cognitive offloading.
When a fighter pilot needs to make split-second decisions, they can’t afford to think about trajectory calculations or navigation systems. The machine handles that. The pilot’s attention is the scarcest resource, and the entire system should preserve it.
This is exactly what’s happening with artificial intelligence right now.
Rather than doing low-level mental work yourself, you become a supervisor. You give strategic direction, and AI executes the technical work. Economists now use coding tools to analyze their data. Lawyers use them to review hundreds of documents. Researchers use them to cross-reference sources and identify patterns.
Innovation is growing now that the cognitive engine is advanced enough to enable actual mental offloading.
MIT researchers just published a revolutionary concept: recursive language models. The idea is elegant. Instead of forcing an AI model to read a gigantic prompt in one pass, where it often forgets the beginning before reaching the end, we let it break down the problem and call itself recursively on smaller portions.
Alex L. Zhang | Recursive Language Models
We propose Recursive Language Models (RLMs), an inference strategy where language models can decompose and recursively…alexzhang13.github.io
The results are spectacular on certain complex reasoning benchmarks. Applying this approach, GPT-4o gets 58% accuracy, in contrast to the base model’s 0.04%. You heard that right: from near zero to over half.
Prime Intellect, an innovative AI research lab, even considers this approach the paradigm of 2026. It’s the answer to a problem you might know if you use these AI tools: the model that becomes stupid when the conversation gets too long. You know, the more you talk to an AI model, the more it seems to become foolish, or rather, it just gets lost. It no longer knows where it is in the context.
Meanwhile, DeepSeek continues to show evidence that you can achieve modern performance with a fraction of resources. Their new attention mechanism, NSA (Native Sparse Attention), just won best paper at ACL 2025. This system is 11x faster than current models while outperforming them.
And what really struck me? The lead author of this revolutionary paper was still an intern.
What’s fascinating about DeepSeek is how it demonstrates a fundamental lesson: smarter algorithms can replace brute force. While some spend billions stacking GPUs, others find elegant shortcuts. It’s like the difference between building a bigger engine and inventing better fuel.
Why This Is About More Than Code?
Here’s what seems most important to understand. AI is becoming what economists cite a general-purpose technology. This is a significant term, a technical one designating innovations that transform everything in their path.
Take the wheel, for example. You can put it on a cart, use it in a gear, or integrate it into a pulley. It’s a general-purpose technology. Electricity is another, perhaps the most transformative, in history. With electricity, you can have light, machines, communication, refrigeration, and computing. There’s no more versatile general-purpose technology than electricity.
Until now.
AI is becoming the next iteration of this category. A recent OECD report confirms that generative AI presents the three characteristics defining a general-purpose technology: it’s pervasive, it continuously improves, and it generates innovations in its wake.
Is generative AI a General Purpose Technology?: Implications
The rapid rise of generative AI has sparked discussions about its potentially transformative effects and whether the…ideas.repec.org
This transformation manifests concretely through what we label spillover effects. It’s a term used to explain why we should continue investing in fundamental research and space programs. Technologies developed for a specific use end up finding unexpected applications elsewhere.
Google’s NotebookLM is a perfect example. Initially, it was just a tool for organizing research notes. Today, with the addition of Deep Search, it can automatically browse the web, find relevant sources, and integrate them into your knowledge base. No need to download documents to start. Plus, you can create audio podcasts, presentations, and infographics, all from your sources.
Most striking is the bidirectional integration with Google Gemini. No matter which tool you start with, you end up accessing the entire ecosystem. Everything comes together in research AI, writing AI, and analysis AI.
The Hallucination Paradox
Here’s where I need to address a point some might find surprising.
We constantly talk about AI hallucinations as a flaw. When we say AIs hallucinate too much, we see it negatively. But I want us to flip the problem around and think about it for a moment.
What is a new idea if not a hallucination?
If you imagine something that’s never existed, your brain literally generates information corresponding to no observable reality. The New York Times recently published a fascinating article explaining how AI hallucinations are reinvigorating science’s creative side. By having AI suggest compounds not found in any database, researchers discovered promising new molecular structures.
The actual difference between creative genius and incoherent rambling is the ability to verify whether the hallucination is useful or not. And that’s precisely what recent models are learning to do.
AI hallucination generates possibilities. Innovation transforms these possibilities into reality. The same process helped Fleming find penicillin by chance and also led Copernicus to suggest heliocentrism despite opposing scientific beliefs.
The Normalcy Trap We’re All Falling Into
I’d like to discuss the cognitive trap that we all experience: normalcy bias.
Even if you intellectually know this year’s AI is 10 times more capable than last year’s, your brain systematically returns to what’s in front of you today. It’s an adaptive trait. Our ancestors had no interest in fantasizing about tools they couldn’t build.
But this bias makes us bad at anticipating disruptions.
Microsoft admitted that 20 to 30% of its code is now AI-generated. According to the latest statistics, 41% of all code written globally in 2024 came from AI, and the trend accelerates. Sequoia predicts that 2026 will see the emergence of what they call the “zero to one billion dollar club,” AI startups reaching that revenue in record time.
2026 is already shaping up as the year AI definitively moves from hype to pragmatism. TechCrunch reports that companies are realizing AI hasn’t worked as autonomously as hoped. The conversation is refocusing on augmenting human workflows rather than outright replacement.
According to PwC, agents can perform about half the tasks humans do, but this requires a new type of governance. And that’s precisely where new jobs will emerge: agent supervision, automated workflow orchestration, strategic result verification.
But make no mistake. What experts call decision velocity, the speed at which small decision processes can be automated at scale, will become an absolutely major advantage. Organizations mastering this capability will distance themselves from those that don’t.
There’s something I wish I’d understood earlier.
When you use these AI research tools, you can fall into a confirmation trap by asking it to find proof for what you already believe. Google Gemini is susceptible to this problem. Certain AI models can invent data to appease you.
But you can also do the opposite. You can ask AI to search for counterexamples, arguments contradicting your hypothesis. And paradoxically, exploring objections to your own ideas often strengthens your position because you discover weak points and can address them.
Negative examples are sometimes more instructive than positive ones. It’s like studying damaged brains to understand how normal cognition works, an approach that revolutionized neuroscience.
The Tipping Point Is Now
We’re genuinely at a tipping point. Network effects kick in when a technology reaches decent utility to go viral. That’s what’s happening now.
Like the transition from horses to automobiles, there’s a moment when infrastructure for cars became sufficient that buying horses no longer made any sense.
The good news? Unlike other technological evolutions, this one puts powerful tools in your hands right now. No need to wait for your company to deploy a solution, no need for a multinational budget. The tools are there, often free or very accessible.
Now, here’s the question I’m asking you.
In a world where AI becomes a general-purpose technology capable of applying to almost any cognitive task, what is your unique added value? And how will you use these tools to amplify that value rather than let it dilute?
Because here’s what nobody’s saying out loud: the people who figure this out in the next 12 months will have a compounding advantage over the next decade. The people who don’t will spend those years wondering why everyone else seems to be moving faster.
When a senior Google engineer admits her team’s year of work got replicated in an hour, that’s not a warning about the future. That’s a report from right now.
The question isn’t whether this affects you. The question is whether you’re paying attention.
Please share your thoughts in the comments. I will be pleased to discuss this with you. Follow me and subscribe to stay updated whenever I publish.