How Someone Had Said That They Were In “Awe” After An LLM Had Built A React App To Add Two Numbers In 2020

In July 2020, founder Sharif Shameem posted a video showing GPT-3 building a functioning React app from a plain English description, a moment widely cited as an early sign of the AI coding wave. Six years later, agentic coding models like Anthropic's Claude Opus 4.6 can complete tasks that take humans 14.5 hours with 50% success, and OpenAI's GPT-5.3-Codex was instrumental in building itself, with most code at OpenAI now written by AI agents.

Six years can feel like several lifetimes in the AI era. In July 2020, a founder named Sharif Shameem posted a video on Twitter that briefly became the talk of the tech world. He had built a tool called debuild.co, and in the demo, he typed a plain English sentence into a text box: a button that says “Add $3” and a button that says “Withdraw $5,” then show me my balance. A few seconds later, GPT-3 had written a working React component, complete with a constructor, state management, and two buttons that actually changed a number on the screen. “I just built a functioning React app by describing what I wanted to GPT-3. I’m still in awe,” Shameem wrote. At the time, this was genuinely stunning. GPT-3 had only been available to a small circle of developers for a few weeks, and the idea that a language model trained to predict the next word in a sentence could also produce syntactically correct, functionally accurate JSX was not something most engineers expected. The demo spread quickly, and along with a handful of similar clips that same week — English descriptions turning into Figma layouts, SQL queries, and regex patterns — it became one of the moments people point to when they talk about the beginning of the current AI wave. Debuild itself never grew into a lasting company. Shameem later moved on to found Lexica, a diffusion model and image search startup, before joining OpenAI’s models team, where he currently works. Watching that video today feels almost quaint, and that gap is worth sitting with for a second, because the distance between a chatbot that could produce a two-button balance tracker and where coding models stand now is enormous. From toy demos to actual software engineers The clearest sign of how far things have moved is what today’s models are being asked to build, and what they are managing to build without much hand-holding. Where GPT-3 needed a carefully worded prompt and produced a single component, current agentic coding models are given open-ended tasks and left to work through them largely unsupervised, reading error logs, installing dependencies, and iterating on failures much like a human engineer would. Until recently, Anthropic’s Claude Opus 4.6 had held the longest task-completion time horizon https://officechai.com/ai/10-best-agentic-coding-and-terminal-use-models-march-2026/ of any model tracked by METR, with a fifty percent success rate on tasks that would take a human roughly fourteen and a half hours to complete, meaning it can stay coherent on a single problem for the better part of a working day. The scale of what these systems can produce has shifted just as much as their endurance. In one demonstration, sixteen instances of Claude Opus 4.6 working in parallel wrote a C compiler in Rust https://officechai.com/ai/10-best-agentic-coding-and-terminal-use-models-march-2026/ from scratch, capable of compiling the Linux kernel, a task that sits several tiers above generating a button that adds to a counter. OpenAI has said its own coding model, GPT-5.3-Codex, was the first version of Codex that was instrumental in building itself, with early checkpoints used internally to debug their own training runs and diagnose evaluation failures. OpenAI executives have also described a shift already underway inside the company, where the majority of code https://officechai.com/ai/more-and-more-code-at-openai-is-being-written-by-ai-agents-openai-exec/ being written is increasingly authored by agents operating in their own environments rather than developers typing line by line. Winning the competitions humans built to test each other Coding progress has been mirrored almost exactly in mathematics, and the two fields have effectively become one story. Within the last year, AI models from OpenAI and Google DeepMind have each delivered gold medal performances https://officechai.com/ai/google-deepmind-officially-wins-a-gold-medal-at-the-international-math-olympiad/ at the International Mathematics Olympiad, a competition that has run since 1959 and is regarded as the hardest math contest a pre-university student can enter. Google later went on to win gold at the International Collegiate Programming Contest https://officechai.com/ai/google-deep-think-gold-medal-performance-a-international-collegiate-programming-contest/ as well, a competition built specifically to separate the strongest programmers on the planet from everyone else. OpenAI has since claimed a similar result at the International Informatics Olympiad https://officechai.com/ai/openai-delivers-gold-medal-performance-at-international-informatics-olympiad-for-programming/ , placing sixth among human competitors and first among any AI system entered. These wins are not confined to competition problems with known answers, either. Models have started turning up results on open mathematical questions that had gone unsolved for decades, with researchers at multiple labs treating frontier models less like novelty demos and more like genuine collaborators https://officechai.com/ai/gpt-5-6-sol-helps-prove-that-non-sofic-groups-exist-in-big-math-breakthrough/ on unsolved problems, an idea that would have sounded like science fiction back when Shameem’s video was making the rounds. What actually changed None of this happened because language models suddenly understood code the way a human does. It happened because reasoning was pushed into the models directly, with reinforcement learning teaching systems to check their own work, backtrack, and try again before producing a final answer, rather than simply predicting the most statistically likely token. The AIME math benchmark is a useful marker of how quickly this compounded: GPT-4 scored close to zero on it, and within about a year its successors were clearing ninety percent. Coding followed a similar curve, with OpenAI’s o3 reaching an Elo rating on Codeforces that placed it above all but roughly 174 human programmers. There is a real human cost tucked inside these numbers too, and it shows up whenever the story stops being about benchmarks and starts being about people. A professional mathematician described watching AI take Olympiad gold as a gut punch https://officechai.com/ai/ai-winning-international-math-olympiad-gold-feels-like-a-gut-punch-says-professional-mathematician/ to something he had built his entire identity around, and it is not hard to imagine software engineers feeling some version of the same thing as agentic coding tools take on more of the work that used to define the job. Shameem’s old tweet still holds up as a genuinely important artifact, not because the app itself was impressive by today’s standards, but because it captured the exact moment a lot of people stopped treating language models as text generators and started treating them as something closer to a general reasoning tool. Six years on, models are writing compilers that can build the Linux kernel and winning gold medals at competitions designed to filter out everyone except the best human minds in the world. The balance tracker in that video went from -5 to 0 with a single click of a button. It is worth remembering that the field it came from has moved by several orders of magnitude more than that in the time since.