In recent years, zero-shot generation has taken the world by storm. The ability to generate a complete asset, whether that be a document, image, or 3D model by using a single prompt has amazed people across the globe and opened the door to endless new generations of products.
Many machine learning models and products have been released during this short period and the rate of growth is not slowing down. In fact, research continues to expand while new commercial enterprises are formed daily.
Zero-shot generation has been quite successful in making incredible progress on several key metrics, including efficiency, ease-of-use, and affordability. In combination, these advantages position zero-shot generation as a highly appealing option for all kinds of people.
Nonetheless, the evolution carries on, and many have encountered the limitations of zero-shot generation. Despite its inevitable role in future products, zero-shot generation has several inherent problems to be solved. The central problem is all about control, with accuracy, embodying the vision, and even privacy being important concerns as well.
This book focuses on addressing these problems directly through what is called “human-in-the-loop” (HITL) workflows. While HITL has only found a place in a few products thus far, many insiders would agree that it is the next wave of artificial intelligence.
HITL workflows are so powerful and compelling because they give humans varying degrees of control over the output. With the right abstractions, users can be as refined or as generalized as they need for a given task.
It’s worth exploring how zero-shot generation and HITL differ in the marketplace. It’s quickly intuitive that both methods will have a place in the future. While the jury may still be out on all of the potential uses for either approach, there are some clear signals for startups and researchers to consider.
Naturally, professionals are generally going to want more control and accuracy in their work. It’s not as common that a professional can accept a result that is essentially random. While there is definitely room for overlap, it is self-evident that professionals will prefer HITL workflows generally.
On the other hand, beginners in a given field are going to be much less interested in creating the details themselves. They usually want a close-enough result that they can use quickly. It can even be fun to enter a single prompt and get a complete output.
The good news for startups and entrepreneurs is simple: artificial intelligence will find a place in every industry and every category of the market. While it may sometimes be helpful to differentiate between the beginner-friendly and professional approaches, there is no lack of opportunity across verticals.
There is another, perhaps unexpected, factor when examining zero-shot generation and HITL. This is the dichotomy of local machine learning vs cloud machine learning. As businesses, it can be tempting to jump to profitable conclusions, but it’s often better to adapt to the future than it is to try and profit immediately. Consider the following details.
Zero-shot generation generally requires massive machine learning models to create an output, alongside heavy-duty graphics hardware. The inference times are slow, and oftentimes the results are still not very great.
HITL models rarely carry these same issues. This is because HITL models are generally more targeted and narrow than the general-purpose models. This means that it can be run on a user’s local hardware, rapidly, and on modest hardware - even CPU.
These advantages for HITL lead to other marketable features which will increase sales and reinforce your customer-oriented brand. The most commonly mentioned in privacy. Privacy is inherent in models that run on local hardware as the data is not sent across the network and through some opaque pipeline. Another very important aspect is the unlimited usage capability in local HITL models. There are no limits and no extra costs associated.
As a startup or entrepreneur, you might wonder how you can make money with this approach. It’s actually not complicated. Users are happy to pay for substantive features such as real-time collaboration or cloud storage. As your product catches on, new opportunities will come for partnerships and large enterprise sales.