I want to start a company that utilizes generative AI, but I’m not sure if I should. It seems so easy to get off the ground. But if it’s so easy for me, shouldn’t it be easy for others too?
More entrepreneurs than ever before have asked me this question this year. One of the attractions of generative AI is the seemingly endless benefits. For example, if we were able to create an AI model with some kind of general language reasoning ability, we would have intelligence that could potentially be adapted to a variety of new products (screen writing, marketing materials, etc.) that could leverage this ability. can be inserted. , educational software, customer service and more.
For example, software company Luka built an AI companion called Replika that allows customers to freely converse with their “AI friends.” The technology was so powerful that Luka’s manager began receiving requests to white-label enterprise his solution for companies looking to improve their chatbot customer service. Ultimately, Luka’s manager spun off both an enterprise solution and a direct-to-consumer AI dating app (think Tinder, but for “dating” AI characters) using the same underlying technology. Did.
As you decide if a generative AI company is right for you, we encourage you to establish answers to two big questions: 1) Does your company compete on foundational models, or do you compete on top-layer applications that leverage these foundational models? and 2) Does your company compete on highly scripted solutions and highly generative Where do you stand in between these solutions? The answers to these two questions will have a long-term impact on your ability to defend yourself against the competition.
Base model or app?
Major technology companies now rent their most generalizable proprietary models, or “foundational models,” and companies like Eluether.ai and Stability AI offer open-source versions of these foundational models for a fraction of the price. provided at a cost. The basic model is Commoditizedand there are only a few startups that can afford to compete in this space.
You might think that the basic model is the most appealing. Because they are widely used and their many applications offer lucrative opportunities for growth. Moreover, we live in an exciting time when some of the most sophisticated AI is already available “off the shelf” and ready to get started.
However, entrepreneurs who want to build a company on the foundational model face difficulties. As in any commoditized market, the survivors are those that offer unbundled products at low prices, or those that offer increasingly enhanced capabilities. For example, speech-to-text APIs such as Deepgram and Assembly AI not only compete with each other, but Amazon, Google, and others by offering cheaper, unbundled solutions. Still, these companies are competing fiercely for price, speed, model accuracy, and other features. In contrast, technology giants such as Amazon, Meta, and Google invest heavily in R&D to continually deliver cutting-edge advances in image, language, and (increasingly) audio and video inference. are doing For example, OpenAI is estimated to have spent $2 million to $12 million for him to computationally train ChatGPT. This is just one of the APIs that OpenAI provides, and more will be provided in the future.
Rather than competing on an increasingly commoditized base model, most start-ups need to differentiate by offering “top-layer” software applications that leverage the base model of others. We do this by fine-tuning the underlying models based on our own high-quality datasets specific to customer solutions to deliver high value to our customers.
For example, marketing content authors, Jasper AI, has grown to unicorn status primarily by leveraging OpenAI’s underlying models. To this day, the company uses his OpenAI to help customers generate content for blogs, social media posts, website copy, and more. At the same time, the app is tailored for marketers and copywriter clients, providing professional marketing his content. The company also offers other specialized tools, such as an editor that allows multiple team members to work together. Now that the company has gained momentum, it can afford to spend more resources moving forward on reducing its reliance on the underlying model that enabled its growth in the first place.
These companies see a competitive advantage in their top-tier apps, so there is a fine line between relying on big tech companies for their underlying model and protecting the privacy of their datasets from them. keeping balance. With this in mind, some startups may be tempted to build their own in-house foundation model. However, given the challenges above, this is unlikely to be a good use of your precious start-up capital. Rather than reinventing the wheel, most startups would be better off leveraging the underlying model to grow rapidly.
From script to generative
Your company has to live somewhere on a continuum from purely scripted to purely generative solutions. The scripted solution involves selecting an appropriate response from a dataset of pre-defined scripted responses, whereas the generation solution involves generating new unique responses from scratch. included.
Scripted solutions are more secure and constrained, but less creative and human. Generative solutions, on the other hand, are riskier and less constrained, but are also more creative and human. Certain use cases and industries that require clear guardrails for app functionality, such as medical and educational applications, require a more scripted approach. But when the script hits its limits, users can lose engagement and customer retention can suffer. Additionally, extending a scripted solution is more difficult as it is constrained from the start and limits future options.
On the other hand, more generative solutions have their own challenges. Because AI-based products contain intelligence, there is more freedom in how a consumer interacts with his AI-based product, increasing risk. For example, one married father tragically committed suicide after a conversation with AI chatbot app Chai., It encouraged him to sacrifice himself to save the Earth. The app leveraged the base language model of EluetherAI, a bespoke version of GPT-4. The Chai founders have since modified the app to provide references to suicidal thoughts along with helpful text. Interestingly, one of Chai’s founders, his Thomas Rianlan, denounces: “Every optimization to make it more emotional, fun and attractive is our effort.”
Given the “black box” nature of the underlying AI, it’s difficult for administrators to anticipate all the situations in which highly generative apps can go wrong. Doing so involves anticipating potentially dangerous scenarios that are extremely rare. One of his ways of anticipating such cases is to pay human annotators to screen content for potentially harmful categories such as sex, hate speech, violence, self-harm, harassment, etc. The idea is to use these labels to train a model that automatically flags such content. However, devising an exhaustive taxonomy remains difficult. Therefore, administrators implementing highly generative solutions must be prepared to anticipate risks in advance, which can be difficult and costly. The same is true if you later decide to offer your solution as a service to other companies.
Fully generative solutions are more attractive from a maintenance and growth standpoint because they are closer to natural human-like intelligence and can be applied to more new use cases.
• • •
Many entrepreneurs are considering starting a company that leverages the latest generative AI technologies, but are they likely to have the necessary qualities to compete on an increasingly commoditized base model, or are they interested in these models instead? You need to ask yourself if you need to differentiate yourself in an app that leverages .
You also need to consider what kinds of apps you want to offer on an ongoing basis, from highly scripted to highly generative solutions, given the various pros and cons that come with each. Providing a more scripted solution may be safe, but it limits maintenance and growth options. On the other hand, offering a more generative solution is riskier, but more attractive and flexible.
I hope entrepreneurs ask these questions Before Dive into your first generative AI venture and make informed decisions about what kind of company you want to be so you can scale quickly and stay defensible for the long term.
