How do you build an AI business from scratch?

January 26, 2023


How do you improve a model like GPT-3 if you are a small business owner?

For one thing, it's extremely complex/or impossible to do it at scale.

Instead, my main argument here is that, by simply verticalizing (creating a ChatGPT for every niche) this incredible tool, we can make it much more factual and grounded, of course, also more limited.

In fact, one of the main issues of this current technology (large language models) is that they work (for the first time) extremely well on generalized-tasks.

And if you were to restrain them too much, you would get. As a result, a much more constrained ChatGPT.

So, as a way to learn, experiment, and limit the drawback of a technology/product like ChatGPT, we created, an AI business assistant, which has been fine-tuned on a crafted and curated Knowledge Base from FourWeekMBA.

This, fine-tuning process, helps verticalizing GPT-3, while trying to limit its ability to allucinate, or give inaccurate answers.


Well, let me break down the proces:

- First thing is to make sure the AI model can be fine-tuned, on a set of highly curated data/content (Knowledge Base).

- Then, you need to make it sure that the prompt, which is "the natural language coding interface" - if you wish - keeps the model grounded to the context it was given, while still being able to give a wide range of answers.

- A last piece of the puzze is the reinforcement learning process, where you enable users to provide a feedback into each answers, and based on that prioritize again the learning and fine-tuning of the model.

At least, as entrepreneurs, building tools on top of OpenAI, this is what we have control over, unless we build a foundational model ourselves (which might now might be a multi-million dollars endeavour).

Thus, for now, is a limited version of the above, and if it gains traction I'll be investing more resources to further fine-tune, and leverage on reinforcement learning to make this a much better vertical tool for business people.

In addition, for now, to keep things simple, this has been trained on the FourWeekMBA's knowledge base.

Yet imagine, how, in the future this might get enriched from other knowledge bases of publishers that might want to opt-in.

The interesting part? It is a traffic generator, in short, the tool speeds up the process of giving you the short, answer, to very specific business questions, and it sends you back the the article where this answer can be found.

Thus, that is a win-win-win (a win for the user, a win for the publisher, and a win for the tool provider).

Feel free to play, and beyond the feedback which you might be able to live when you get an answer, feel free to ping me back and let me know what you think!

Keep in mind this is a first iteration, the interface is still quite simple and we're planning to further fine-tune it.

But for now, you can start playing with it!