TLDR 1: Dataset is no longer a hard requirementEven an extremely small dataset can produce valuable results .TLDR 2: It has very broad use casesYou can never trust interns to be 100% right, but they can provide useful value with proper supervision, guidance, and planning.Classic unit economics model for SaaS and online services will be a challenge due to the extremely high running cost of most LLM
However, the size and quality of the dataset were still one of the predominant factors in training an AI.Chart showing traditional AI model needing decent dataset before being useful, with diminishing returns in AI model quality with increased datasets For example, at uilicious.com , we used our limited dataset to train an AI model . We found that it was spewing garbage half the time, leading us to set the model aside and grow the company without AI while building up our dataset.
In many cases: they are really good at learning new specialized knowledge when given the datasets in an appropriate format Chart showing how building on LLM, moved the curve, in allowing interesting AI prototypes to be made quicker, with less training data than traditional models. Data was still, in a way, king. It may have moved the curve to be more accessible, but it still required a team to build up large datasets and tune the model.
It has unlocked the possibility of creating usable AI with extremely small datasets - something that most startups have access to or can create by hand. This is a fundamental shift in how we think about AI training.
Alternatively, if you can get chatGPT to act in a way that you find useful and of value to your startup, you can build it as a dedicated AI service. Ultimately still, this limit on accuracy and reliability is only a critical issue in sensitive industries . For most cases, it is merely a distraction in finding use cases. Once the AI crosses the “good enough” threshold .
And so on and so forth - where at no point in the process should the human be removed from the loop in supervising and iterating with the intern.A one-day intern can be quickly built using prompt engineering; anything you manage to get chatGPT to do fits into this category. The downside is there is a practical limit of what you can fit into the training data. The plus side is you can experiment and set this up in seconds, and it's very easy to test with chatGPT.