Korva is a scroll of the internet’s best ideas.
Renamed ChatGPT hours before launch
GPT-3's first use
When GPT-3 launched, the only commercial use that really worked was copywriting. Hiring a marketing firm for a landing page cost about 20 dollars, and the same job cost 20 cents when the model at OpenAI wrote it.
Chatting in the playground
Developers were also chatting with the model in a testing interface called the playground. That took effort, because the model had not been tuned for conversation, so users had to supply a few examples of what chatting means before it would do it. People liked it anyway, and a lesson from Y Combinator, to follow what users are already doing, pointed OpenAI toward building a good chatbot.
A cautious launch
When GPT-4 was finished and in use inside the company, it was clearly a big deal. Hard questions came up: whether it would create fake news, say really offensive things, or get OpenAI in trouble. So the chat interface went out first with the weaker GPT-3.5, and a GPT-4 product was planned for a few months later.
The chat interface was not meant to be a product, and nobody expected a huge hit. The hope was that it would help the world catch up and realize something was going on. The model turned out to be over the threshold where people said it was awesome, and even without much utility yet, it was an incredible moment for people to feel AI progress with something they enjoyed using.
Why did early GPT models need examples to chat?
GPT-3 was trained to predict the next word in a text, not to answer a person. Given a bare question, it might continue with more questions, because that is what lists of questions often look like online. Showing it a few sample exchanges first gave it a pattern to continue, a technique called few-shot prompting. OpenAI's 2020 GPT-3 paper was titled Language Models are Few-Shot Learners.
How did feedback training make models conversational?
In 2022 OpenAI published InstructGPT, which fine-tuned GPT-3 using human feedback. People wrote example responses and ranked competing outputs, and the model was adjusted to favor the answers people preferred. In OpenAI's evaluation, labelers preferred outputs from a 1.3 billion parameter InstructGPT model over those from the 175 billion parameter GPT-3. OpenAI described ChatGPT as trained with similar methods.
Why do labs release weaker models first?
Staged release has precedent at OpenAI. In February 2019 it described GPT-2 as potentially too risky to release in full and published a smaller version first. The gap gave researchers time to study how it might be misused before larger versions followed. The full 1.5 billion parameter GPT-2 model was released in November 2019.
More from Invest Like the Best with Patrick O'Shaughnessy
