OpenAI won gold in IMO this year and ICPC. DeepMind also won gold in one and silver in another. AI is beating humans in things once considered extremely difficult.
Almost all of the code written in companies born a year ago is written by AI itself. For sure, humans are still needed for system design, optimizations, business logic, infrastructure—which is way more important than code itself—but yes, 90–95% of the code in places like Anthropic is now AI-generated. We still need humans, more than enough, because we have a once-in-a-lifetime opportunity to ship amazing software.
Earlier, just having software was itself a moat. Now the real moats are network effects, brand value, access to capital, talent, and proximity to government. With a good 3–5 engineers you can still create any software as a starting point, but scaling is tough, and humans are needed more than ever.
The idea of reinforcement learning on top of transformers looks like the right path—more tokens, more FLOPs, more intelligence, and thus better results. A human in one space can explore one path. These systems, in parallel, can explore n paths, share insights, and converge toward the idea of an AI scientist assisting real-world discovery and new directions.
When I say the ceiling is broken, I mean these systems are more mature than ever. Very soon—in the next 2–5 years—we will have systems capable of doing almost all economically valuable jobs: coding, financial analysis, customer writing, art creation. The bar is raised so high that any piece of mediocrity can be replaced within a month. People are starting to appreciate this change.
When machines replaced handlooms, workers resisted—some even burned down machines. I suspect we’ll see massive job losses, especially at the entry level, since much of the routine work is being automated away. But it’s not justified to say AI can’t be better than an associate or senior—it clearly can. The reason senior roles are safer is because they make higher-level decisions and have incentives not to let AI replace them.
Some folks will try to slow down AI for obvious reasons, but the systems are cooking fast. Within the next decade, we will have systems smarter than all humans.
Countries like India haven’t even fully adopted basic software or cloud yet, so there is still a lot of space before full AI replacement happens. Language remains extremely challenging, and domain-specific data for Indian languages is so rare it barely exists. But that means everyone is equally cooked—the gap is closing fast.
My recommendation is simple: master the basics of any skill, learn AI tools end-to-end, pick a domain, and start thinking about applied research for your own problems. You won’t necessarily create a permanent moat, but you’ll get a 3–6 month edge to stay ahead of others.
Note – my view comes from being an AI engineer who also loves research. Most of what I write is shaped by what I hear in tech circles. I still firmly believe deploying AI into existing systems is tough and will require a lot of forward-looking solution engineering to make it work.Note – this is 100% written by me, he he.