ChatGPT is a product, not a model
A million people signed up in five days. The underlying model isn't new. What's new is the interface and the training to follow instructions, and that changes every software team.
OpenAI released ChatGPT last Wednesday. Greg Brockman said yesterday it passed a million users. Everyone I know has been pasting screenshots all week: it writes cover letters, explains tax forms, debugs code, writes poems about their dog, and argues confidently about things it has made up.
The model underneath isn’t a big leap. It’s from the GPT-3.5 series, fine-tuned from models that have been available through OpenAI’s API for most of this year. The techniques were described in the InstructGPT paper in March: supervised fine-tuning on examples of good answers written by people, then reinforcement learning from human feedback, where people rank several outputs and a reward model learns their preferences. The dialogue format is new. The capability mostly isn’t.
So why is this the moment everyone noticed? I think it’s because it’s a product and not a model. The API required you to know about prompts, temperature and tokens. ChatGPT is a text box. You ask something in plain language, it answers, you ask a follow-up, and it remembers what you said. It’s free. The fine-tuning makes it follow instructions and admit some limits instead of just continuing text. That combination turned a capability that developers had been playing with for two years into something anyone could try in thirty seconds.
In 2016 I wrote that conversational interfaces were at least a decade away and that when they worked, they’d come from models trained on huge amounts of text, not from hand-built dialogue trees. I got the second half right and the first half very wrong. It took six and a half years.
What I think changes for software teams:
Every product that has a text box, a search bar, a help center or a form is going to be asked whether it should be a conversation. For many, the answer will be yes, at least partly.
Coding changes faster. Copilot was the first sign last year. A general assistant that can explain an error, write a test and translate code between languages will change how junior engineers learn and how senior engineers spend time.
Evaluation becomes the hard problem. It’s easy to make a demo that impresses in five examples. It’s hard to know how often it’s wrong in production, and it’s wrong in fluent, convincing ways. The teams that do well will be the ones that build good ways to measure quality, not the ones with the best demo.
I can’t remember seeing a technology spread this quickly. When something reaches a million people in five days by being useful, not by being a game or a fad, it’s usually the start of something. My guess is that Google, Meta and everyone else with a large language model ships something like this within six months, and that “chat with a model” becomes a standard interface as quickly as touchscreens did.