GPT-3: what a founder should actually think about it
In summer 2020, OpenAI released GPT-3 - the largest language model at the time. What it actually means for business right now.
In late June 2020, OpenAI opened beta access to GPT-3. Within a few weeks it was the only topic in the professional community. The demos were striking: the model writes code, generates text, answers questions, mimics writing styles. One well-known developer wrote that for several days he could not tell whether he was reading a person or the model.
The reaction is typical of any major technology event: some people declare the end of an era, others say "nothing new here". For a founder or manager, neither position is useful.
I will try to say what seems accurate from a practical standpoint.
What GPT-3 actually does well
The model handles tasks that require transforming text into text according to a pattern. Rephrase, extend, classify, answer a question given context, write a draft from a structure.
That is not trivial. Previously each such task required a separately trained model, labelled data, and a team capable of working with it. Now much of this works with a few examples in the prompt.
This changes the economics of prototyping. Testing an idea, building a first draft of a feature - both become cheaper and faster.
Where the boundary is
GPT-3 does not reason in a strict sense. It is statistically plausible - which is fundamentally different from being correct. The model confidently produces wrong facts. It does not know when it does not know.
For tasks where an error is expensive - legal documents, medical recommendations, financial calculations - this is not a tool for autonomous operation. A human in the loop is mandatory.
Another boundary: the model does not know your business. It was trained on publicly available text. It does not have your data, your domain knowledge, your internal terminology, your rules.
What this means for product decisions
If your product or process contains tasks that are currently done by hand and look like "take text, transform by a rule, return the result" - it is worth checking what GPT-3 can do with those tasks.
That does not mean "replace everything". It means "find a few specific bottlenecks and test a hypothesis".
A good sign that a task is suitable: you can explain the rule in words, you have examples of correct outputs, and you are prepared to check the model's output before using it.
Three questions before starting
- Do we have a task where both input and output are text, and the transformation rule can be described with examples?
- Who will check the model's output before it reaches a customer or enters a process?
- What happens if the model makes a mistake - how critical is that?
If you have answers to the first two, and the third does not worry you, that is a solid starting point for an experiment.
GPT-3 is not magic and not a threat. It is a tool with clear strengths and clear weaknesses. Treat it like a tool.