Korva is a scroll of the internet’s best ideas.
Amjad's Law stopped being true
Learning to program
Amjad Masad learned a bit of programming at 7 or 8, growing up in Jordan in the 1990s, and what annoyed him was the minutiae: the tedious stuff that should be automated. Years later, as a founding engineer at Codecademy, he heard stories like a former fitness instructor who built a fitness app and made $10,000 a month. Coding still suited a certain personality, one with a high tolerance for pain, and it didn't attract the general population.
Starting Replit
So he started Replit to blur learning and building into one step. He had seen GPT-2 in 2019 and figured machines that understand human language could be turned toward code, and in 2020 he wrote a blog post saying programming was on the verge of being solved, with people needing to learn less and less of it.
Coding without the code
In 2025, with Opus and the newer Claude models, he began building things without looking at the code, before Karpathy had coined vibe coding. Saying on TBPN that people no longer need to code was painful after 15 years spent making coding easier. Engineers on Twitter and Hacker News attacked him and warned that people would build insecure apps. His answer was that any problem AI could create, AI could solve, and a security agent could handle it. By mid-2026, engineers at his company only look at the code when it's about to get deployed.
How did programming get easier over time?
Early programmers wrote machine code or assembly, instructions tied to one specific processor. Fortran arrived in 1957 and let scientists write formulas in something close to algebra, and a compiler translated them. BASIC followed in 1964, created at Dartmouth so that students outside computer science could program. Python, which Guido van Rossum released in 1991, pushed readability further by making code look closer to plain English.
Why did GPT-2 get so much attention in 2019?
OpenAI announced GPT-2 in February 2019 and at first withheld the full model, citing worries about misuse. It had 1.5 billion parameters and was trained on WebText, a collection of text from about eight million web pages. Given a short prompt, it could continue with several paragraphs of fluent prose. OpenAI released the full 1.5 billion parameter model in November 2019, after staged releases of smaller versions.
Where did the term vibe coding come from?
The AI researcher Andrej Karpathy used the phrase in a February 2025 post on X. He described a way of programming where you give in to the vibes, accept the AI's code, and mostly stop reading the changes. The term spread quickly through developer communities. Collins Dictionary later named vibe coding its word of the year for 2025.
More from Casey Newton
