Startups must add AI value beyond ChatGPT integration

The AI hype train is going full swing. At this point, it is hard to name an industry not affected by this disruptive technology. Startups feel these hype waves like no one else as the inventors’ demands rise and competition grows.

It’s becoming increasingly challenging for startups to raise investment without some AI element in their product. So far, ChatGPT integration — relatively easy, affordable, and fast — has been sufficient to keep up with the race and capture a market share. But it looks like this will no longer be the case.

Soon enough, making a statement will take more than just plugging in an open source model.

Venture capitalists are already stressing the importance of adding value for startups, not just using a ChatGPT API.

Okay, but how do we provide this added value?

As CEO of a product development business, I’ve been consulting clients on AI integration solutions and bringing more value to their apps. I’ll explain and share tips on delivering additional value to an AI-driven product by fine-tuning open foundational models.

Why simply adopting a foundational model is no longer an advantage

With the dynamic rise of AI, winning an investment is challenging for startups void of AI involvement. This has led to the influx of “ChatGPT wrappers” — apps parasitizing on the foundation model and bringing null value regarding technological novelty or user flow. VCs are already overwhelmed with the inflow of “ChatGPT for X” startups, labeling them as solutions unlikely to survive in a year or two.

Yet, neglecting to take advantage of the AI race may be a strategizing mistake. To get success, you need to be a step ahead of the game, offering a value-added technology that augments foundation models and provides competitive advantages. Fine-tuning an AI model using collected or synthetic data can deliver a competitive advantage to your startup.

How to deliver added value to your AI-based startup via fine-tuning

When seeking how to involve AI in your product, there are a few ways to go. Here are the main ones:

1. Start with Chat GPT. Start by building a prototype with ChatGPT, and then develop specialized models for optimization. This approach involves refining the broad-spectrum functionality of foundational models to adapt them to more specific and targeted tasks.
2. Fine-tuning. Create a specialized model custom-tailored for specific tasks and gradually adapt ChatGPT over time to these specialized roles — actually fine-tuning. While simply integrating bulky foundational models like ChatGPT can be too expensive and resource-demanding, fine-tuning can provide more flexibility, better performance, and adaptation to business needs while being much more affordable. These qualities make it a perfect fit for early-stage startups seeking shorter time-to-market and full decision-making autonomy: no need for external approval or negotiations with third parties.

What is fine-tuning?

Fine-tuning is a process in machine learning where a pretrained model is trained (typically on a smaller dataset) to refine or adapt its knowledge to a specific task. The process of fine-tuning usually involves three steps:

1. Finding the proper training data.
2. Preparing a script.
3. Processing this data.

You can innovate, iterate, and deploy solutions at an accelerated pace. This speed and flexibility are essential for the competitive startup market, so it can be a significant advantage in capturing market share or meeting a critical market need before competitors.

Why data is the key

Data is the key to successful fine-tuning. That said, the type of data you need in every case depends on your product’s purpose. Startups with access to proprietary data have a solid advantage in fine-tuning.

Proprietary data is the user information businesses collect for gaining insights, making informed decisions, and differentiating themselves from competitors. So, to make sure your business is valuable in the dynamic AI market, prioritize proprietary data collection, fine-tuning the open source model, and ensuring output accuracy.

Flexibility of fine-tuning

It is way too easy to collect data for classification purposes than it is for detection tasks. For example, you might have access to 10,000 images for detection purposes and 1,000,000 for classification. As an alternative, you could train the detection model on the 10,000 images and track the difference in performance and accuracy.

The difference might be striking, as a detector trained on as few as 10,000 images is bound to learn less effectively. In that case, when there is no opportunity to gain more data, you can utilize a network trained on facial-recognition data. It still would perform better than no data training at all, as tangential data is still better than no data. For example, if you have a model initially tailored for detecting cars, it would take much less effort to specifically retrain it to detect heavy-load vehicles.

How does fine-tuning optimize resources?

Let’s say you want to build a product with an AI detection feature. You have two ways to go:

• Integrate a closed-model API designed for detection purposes.
• Take a network initially trained for image classification tasks and fine-tune it to perform people-detection tasks.

At first glance, going with the first approach is faster and easier. Yet, there are some hidden stones. Usually, open source models or closed models’ APIs can be too expensive, robust, and big, requiring more computational power, which can be specifically challenging for startups. In cases like these, opting for a basic AI model created for classifying images and then fine-tuning it for detection purposes can be more efficient. This process would involve two phases:

• The training phase, where the network is taught the specifics of detection.
• The prediction phase, where the network uses this knowledge to identify and detect objects in new images.

For the training phase, the model requires a dataset of images annotated with the coordinates of bounding boxes. These annotations enable the model to learn the spatial grid where objects are located.

During the prediction phase, the adapted model processes a new image to identify and provide coordinates for these bounding boxes, allowing for the detection of objects within the image. These detected boxes can then be manually drawn or visualized as needed.

For example, a Ukraine-based AI avatar generation DYVO needed an AI network to generate myriads of custom avatars for one person. So they took an AI open model aimed at image generation and fine-tuned it to create realistic images of people by training it with actual people’s pics.

Drawing up

To advance in the highly competitive market, AI-based startups should provide added value as a compelling advantage. Fine-tuning is an affordable and flexible way for startups to build a specialized AI model adapted to their business needs.