2024: The K Got Sharper — Sella Rafaeli

Last week OpenAI announced o3, a new reasoning model. On ARC-AGI - a benchmark specifically designed to be hard for AI and easy for humans, where previous models barely registered - it scored roughly 76% to 88%, depending on how much compute it was allowed. On competitive programming, its reported rating puts it among the top few hundred human competitors in the world.

A year ago, at the start of 2024, I wrote that you should bet on AI and bet against the hype. Looking back over the year, I think that was right - but I underestimated the first half.

AI exceeded my expectations

In one year, we went from GPT-4 being the clear best model to a whole field of excellent models from several companies, with prices dropping dramatically. Coding assistants went from autocomplete to editing entire codebases. A quarter of Google’s new code is written by AI. AI pioneers won two Nobel prizes. And now reasoning models are solving problems that were considered years away.

I’ve been programming for twenty-four years. I’ve never seen anything improve this fast. I’m more excited about building software than I’ve been since I was a teenager.

The hype also failed as expected

At the same time, almost every specific hype prediction from January was wrong:

The AI gadgets flopped.
The “AI software engineer” did not replace software engineers. It became a tool for them.
Rushed AI features embarrassed some very big companies.
Most “AI startups” funded this year still haven’t found a business model, and the $600 billion question is still open.

Both things were true all year: the technology was better than the optimists said, and the products were worse than the marketers said. That’s not a paradox. It’s what a real technological shift looks like.

The K got sharper

The theme I kept returning to this year was the K-shaped developer market. After watching it all year, I’m more convinced than ever.

The top arm went up. Experienced developers who embraced AI had their best year ever. They built more, faster, often alone or in tiny teams, and were rewarded for it. In conversations with senior developers all year, I heard the same thing again and again: “I’m doing the work of a team now.” Companies that understood this started paying for judgment and outcomes instead of headcount and hours.

The bottom arm went down. Entry-level hiring stayed painfully low. Developers whose work was mostly implementing well-specified tasks felt the ground shift under them. And the comfortable middle - the “pretty good at a framework” career - got noticeably smaller.

I don’t think that’s going to reverse. If anything, models like o3 suggest the bottom of the K will be pushed further down, and the leverage at the top will keep growing.

What I think 2025 brings

Agents that actually work, for well-defined tasks. Not autonomous engineers - but reliable assistants that can take a task, do it, and hand back a reviewable result.
Review as the core skill. The best developers will be measured by the quality of what they approve, not the quantity of what they write.
Smaller teams, bigger output. The most impressive products of 2025 will be built by surprisingly small teams of very senior people.
A premium on trust. As AI-generated everything floods the world, people and companies with a track record of good judgment will be worth more.
The junior question gets serious. At some point the industry will realize it needs to grow the next generation of seniors, and the companies that figure out how will win.

My advice, one more time

If you’re a developer: get on the top arm of the K, deliberately. Use the tools every day. Own outcomes. Learn to review. Find your people.

If you hire developers: senior talent with AI is the best deal in the history of software. And train some juniors, because someone has to.

And to everyone reading: may 2025 be a calmer, safer year. We could all use one.

The future belongs to people with judgment. AI just made that obvious.