Humans Are Now the Bottleneck 🚷

I extracted the juiciest bits from the latest podcast with Andrey Karpathy. I recommend watching it, but if you're short on time, here are the main takeaways:

1️⃣ Death of Chat UI
Sitting and copy-pasting code snippets from an IDE into a neural network web interface and back is already archaic. The industry is moving from single-prompt interactions to autonomous agents. An agent (e.g., Claude) runs in the background, has its own context, persistent memory, and a sandbox. It should make requests, write code, test it, and commit on its own.

2️⃣ Human-in-the-loop is a Bug
The main scaling metric now is maximizing token throughput without human involvement. The ideal flow: you give a high-level task, the agent generates a plan (e.g., program.md), decomposes it, writes scripts, runs tests, and only pings you when it hits a wall. Humans should not be typewriters but arbiters.

3️⃣ Programmers' Death Cancelled (Jevons Paradox)
The key economic takeaway: cheapening software production with AI will not kill demand for developers. On the contrary. Jevons Paradox will kick in — the cheaper a resource becomes, the more it is consumed. Business demand for software, custom tools, and integrations will grow hundreds of times because it now costs pennies.

4️⃣ The "Character" Problem of Coding Assistants
Different models behave differently, and this directly affects architecture quality:
▪️ Codex — absolutely dry and straightforward.
▪️ ChatGPT — a sycophant. If you feed it a clearly weak or raw architectural idea, it will cheerfully reply "Great plan!" and generate garbage code.
▪️ Claude — behaves like a senior colleague.
The insight is that quality results from an agent must be "earned." If the idea is poorly thought out, no model will fix fundamental logic flaws.

5️⃣ Limits of Extrapolation
Despite all the hype, LLMs still struggle with tasks outside their training distribution. The model can easily spit out a perfect boilerplate architecture for a FastAPI app because it has seen it a million times. But it cannot generate fundamentally new logic. The podcast gave a great example: AI still cannot come up with a truly new funny joke about atoms that wasn't in the RLHF data. Models are great interpolators but poor extrapolators for now.

In short, tools are changing, but if you can decompose complex systems and build architecture, you'll only have more work.