GPT Software is Getting Really Good
Exploring the latest advances in GPT reasoning models and their impressive capabilities in code generation and problem-solving with a virus simulation demo.
My jaw is on the floor. GPT software is getting really good. Perhaps six months ago I would have considered it interesting, certainly a tool worth understanding. I’ve been talking with many startups and founders using GPT and have done my own apps and experiments since GPT2, but the latest reasoning models are something else.
The idea behind a reasoning model is that it starts out creating a chain of thought. Its prompt instructs it to ask a number of questions, and create a somewhat sophisticated thought process and plan. It will reason through the plan, question assumptions and only then start to construct the answer. This in combination with generally improved models with particular attention to code completion is what has me baffled.
I want to make a simple js simulation of a virus. I want 1000 pixels on screen with different vectors. All green except for one that is red. When pixels intersect with the red pixel they turn red five seconds later. I want to chart the progression of the contagion in a little chart on screen.
With just a few back and forths I got the following demo: